As of today, energy poverty remains a significant global challenge. Over 600 million people still live without access to electricity, with most of this population concentrated in Sub-Saharan Africa and parts of South Asia.

Fortunately, there are solutions in place to address this challenge, particularly in developing nations. Traditionally, energy planners and governments would extend grid infrastructure to reach communities without electricity. However, this can be costly depending on the distance from the existing grid. For governments in the Global South with limited budgets, finding more cost-effective solutions is absolutely necessary.

Grid extension, mini-grids, and Solar Home Systems

For example, solar mini-grids can be a feasible option for communities that are too far from existing grid infrastructure but still have enough demand for a shared electricity system. In other cases, for communities with very small or dispersed populations, Solar Home Systems can be more practical and cost-effective because they provide electricity directly to individual households.

SEforALL and integrated energy planning

International organizations such as the United Nations, the World Bank, and Sustainable Energy for All have played an important role in helping improve electricity access through initiatives, funding, research, and partnerships. Sustainable Energy for All, also known as SEforALL, works with governments and partners to support Sustainable Development Goal 7: affordable, reliable, sustainable, and modern energy for all.

This summer, I was fortunate to have the opportunity to work with SEforALL’s Integrated Energy Planning team. Their work focuses on least-cost electrification solutions. In other words, they help identify the most cost-effective and feasible energy solutions for countries, especially in the Global South.

Understanding OnSSET

One of the tools used in this area is called OnSSET, which stands for the Open Source Spatial Electrification Tool. OnSSET is a mapping and modelling tool that energy planners can use to calculate cost-effective ways to bring electricity to communities in a particular country.

OnSSET can produce many different scenarios for each country of interest. Let’s consider Mozambique as an example. Each scenario provides a data-backed pathway for how the country could work toward its electrification goals by a certain year, such as 2030. The results of each scenario depend on different assumptions, such as the cost of solar PV modules, electricity demand levels, carbon price, and grid connection limits.

In essence, this information is extremely helpful for planners and policymakers because it helps them understand different pathways to improve energy access in their respective countries.

What I worked on

There is no doubt that OnSSET is a valuable tool. However, understanding and accessing the data from the scenarios quickly is still a challenge. The scenario results are often found in long and complicated CSV files. Policymakers and planners may not have the time to manually go through these files to find the insights they need.

This is where my project came in.

Building ChatOnSSET

For the summer, I worked on building ChatOnSSET, an AI application that reads OnSSET scenario CSV files, makes calculations in Python, and explains the results in simple language. A user can ask an energy-planning question, and the application provides a response based on the CSV file it has access to.

A key part of the project was making sure that Python performs the calculations, while the AI explains the results. This was important because I did not want the AI to guess numbers or create unsupported answers. The goal was to make the responses traceable to the actual scenario data.

Testing the prototype with Mozambique

Most of the summer involved testing the accuracy of the application and ensuring that the chatbot provided the correct responses by comparing the outputs to the CSV files. I focused my project on Mozambique. The OnSSET/GEP scenarios for each country are publicly available online through energydata.info. For the prototype, I tested across three scenarios for Mozambique.

The logic was that if we could prove the application works accurately for one country and several scenarios, then the same structure could eventually be scaled to include more countries and more scenarios.

So far, the application has been successful in providing accurate responses to specific energy-planning questions. One example question an energy planner might ask is:

What is the national investment required by technology by 2030?

The application can then calculate the investment required for different technologies, such as grid extension, mini-grids, and Solar Home Systems.

I have attached a demo below of what the application looks like so far.

My experience working with a team and building an application from scratch

Overall, this has been an incredibly rewarding experience for me. This was my first time working in a professional capacity, and receiving feedback and guidance from the Integrated Energy Planning team was valuable in many ways.

My mindset this summer was to absorb as much information as possible and ask as many questions as possible. Along with that, I took the initiative to learn about frameworks like OnSSET and understand the terminology that energy planners are familiar with.

Lessons from the development process

Regarding the development process, I realized that taking it step by step was the best approach. I constantly tested the application after every major update and made sure the chatbot was answering questions without errors. Explaining these updates to the IEP team and doing live demos was also important in helping me make progress.

Sometimes it felt like one step forward and two steps back. Even though that can be discouraging, I learned that this is a normal part of building applications that people may actually use.

Communication Skills

I also improved my communication skills through live demos. Explaining technical concepts and walking through my code taught me how important it is to make things simple. A technical tool is only useful if people can understand what it does and trust how it works.

Final reflection

All in all, this project taught me how useful maps and geographic information systems can be. For context, I am majoring in geospatial data science at university, and working on an impactful project rooted in maps, spatial data, and energy access has inspired me to become more technical in this field.

Executive Summary

Recent global energy market shocks, intensified by geopolitical conflicts in Ukraine and the Persian Gulf, have shifted the paradigm of the clean energy transition from merely a climate mandate to a fiscal necessity. For emerging economies, particularly across the Global South, rising demand paired with intense commodity price volatility has made prolonged reliance on foreign fossil fuel exporters economically untenable. According to the framework outlined in the BloombergNEF New Energy Outlook 2026, accelerating the deployment of cheap and local renewable generation is a necessity. Under the Economic Transition Scenario, domestic clean technology serves as a vital macroeconomic shield that protects national balance sheets, ensures sovereign energy security, and retains critical capital within the domestic ecosystem.

Macroeconomic Import Liability

One of the countries that has been significantly impacted by commodity price volatility is Vietnam. As shown in the chart below, Vietnam maintained a negative net energy import status from 1995 to 2005. This is attributed to the fact that Vietnam’s domestic production of primary energy commodities, like crude oil and coal, outpaced internal demand, acting as a net financial contributor to the state. However, since 2006, Vietnam has shifted structurally to become an energy-import-reliant nation. In 2020, Vietnam reached its peak liability: 46% of the country’s total energy use came from net imports. During this time frame, Vietnam rapidly scaled into a global manufacturing and electronics supply chain hub, creating an exponential demand for power that quickly outmatched domestic energy production. According to the BloombergNEF New Energy Outlook 2026, this extreme reliance on volatile foreign commodity markets forced emerging economies to spend between 3% to 6% of their annual GDP purely to cover energy import liabilities.

Capital Mobilization & The System Flexibility Challenge

Clean electricity has undoubtedly become insulated from fossil fuel price shocks, driven by global technology supply overcapacity and steep cost deflation curves. Under the BNEF Economic Transition Scenario, solar is projected to become the world’s single largest source of electricity generation by 2032. Currently, the primary barrier to clean power integration has shifted from technology cost to system flexibility. Specifically, grid infrastructure capacity, transmission upgrades, and localized permitting timelines. Furthermore, a massive scaling of capital mobilization is required to achieve this macroeconomic shield and build more renewable energy project pipelines. To successfully integrate this variable renewable generation, global battery storage deployment must scale a staggering 17-fold by 2035, jumping from a baseline of 223 GW to 3.8 TW.

Sources

Datasets I used to generate Figure 1:

• World Bank Open Data
• World Bank Open Data

BloombergNEF New Energy Outlook 2026 Executive Summary:

https://assets.bbhub.io/professional/sites/44/NEO_2026_Executive_Summary.pdf

The Kerosene Tax

The West Nile Region in Northern Uganda is facing a major challenge: a lack of electricity access. This is a microcosm of a broader and systemic problem in Sub-Saharan Africa.

Picture a family of refugees who just escaped the ongoing conflict in South Sudan and settled in the West Nile. They are completely disconnected from the grid. As a result, they must depend on kerosene to perform their daily activities. Kerosene, which is extremely harmful to the environment, requires families in the West Nile region to spend up to 25% of their income to burn kerosene for only a few hours of toxic, dim light.

However, the solution for this already exists and is popular throughout countries addressing energy poverty challenges: off-grid solar systems. At first, the greatest challenge solar companies faced in implementing an off-grid solution was to make their products affordable to their customers. Today, solar has largely become affordable, even to the poorest demographics, through various payment schemes.

In the West Nile region, it is Trust, not price, that is impeding solar adoption amongst its population. The off-grid solar market in Northern Uganda was flooded with cheap, low-quality solar lamps. Many farmers saw that these lamps broke within a week. This led farmers to believe that cheap solar was a scam.

Why Charity Backfires

A common suggestion to improve electricity access in the poorest communities is to simply give solar products free of charge. Although this idea seems intuitive, it has continuously been proven that it does not address the root problem.

When NGOs realized that trust was broken, they tried to give solar lamps to local agents on credit to sell door-to-door. In 2011, GIZ, a German development agency, partnered with a local company called Matrix Electronics in the Arua District to run a pilot. Before they started this pilot, they understood that most people did not have the capital to buy solar lamps to start a business. They used the consignment model, which involved freelance technicians handling boxes of solar lamps. The goal was to sell as many of these solar lamps to locals, and each freelance technician would earn a small commission for every sale. Matrix Electronics would retain most of the revenue.

For a few months, the pilot seemed to be working well. However, by the end of the 11-month pilot, GIZ noted that 46 solar lamps were unaccounted for. The technicians hadn’t returned the unsold lamps, nor had they brought back the cash for the ones they did sell. From a corporate perspective in Germany, it looked like blatant theft.

There is much to this case study, though. In economics, there is a concept called a “liquidity buffer.” In other words, this is emergency cash for people living in extreme poverty. It was plausible to believe that the freelance technicians sold the solar lamps at a discount and used that cash to help their families. For example, if a technician needed to pay for their child’s Malaria treatment, they would not hesitate to make quick cash to pay for their child’s well-being. Ultimately, Matrix Electronics would not receive any of the revenue.

This case study highlights a crucial insight: energy access challenges in the West Nile are a structural failure, not a moral one.

In summary, charities think they are distributing technology, but to a desperate family, they are distributing emergency currency. You cannot fix the energy gap by handing out unsecured credit. Instead, you have to fix the accountability.

Neighbourhood Bank

If unsecured credit and charity are the wrong tools for the job, what is the right one? It’s a neighbourhood piggy bank.

While communities in the West Nile region might lack electrical grids, they are not completely disconnected from each other. They are connected through Village Savings and Loan Associations. Think of a VSLA as a hyper-local, peer-run credit union. In these groups, neighbours meet weekly to pool their spare change and lend it to one another.

In contrast to the failed GIZ pilot, these groups boast an incredible 99% repayment rate. The reason for this success is peer accountability. You might default on an NGO, but you are much less likely to default on your neighbour who is sitting across from you.

Navo Solar leverages this existing trust network to distribute solar products safely and effectively. To ensure the system remains secure and transparent, Navo Solar divides the distribution process into three distinct pillars.

First, there is the Community-Based Trainer. They act as the teacher. Their role is to run the meetings, explain how the solar technology works, and endorse the product. Crucially, they do not hold inventory and are strictly banned from selling the products. This separation ensures they remain a neutral, trusted voice rather than a biased salesperson.

Second, there is the Commercial Agent. This is a member recruited from within the VSLA who handles the actual hardware and sales. Unlike the freelance technicians in the GIZ pilot, this agent must purchase the inventory with their own money. By requiring upfront purchase, the agent understands the significant responsibility his role comes with. This structural change ensures the solar lamps do not become an emergency liquidity buffer for the agent.

Third, there is the VSLA Internal Loan Fund. If a family wants to purchase a solar lamp to escape the kerosene tax but lacks the upfront cash, they do not need to rely on unreliable mobile money or external credit checks. Instead, they borrow the money directly from their group’s collective pool.

Through these three pillars, Navo Solar solves the core issue. By separating the teacher from the salesman and anchoring the financing in peer accountability, they are fixing the distribution channel so that affordable, high-quality solar can reach the families who need it most.

The Ladder of Trust

To regain the trust of people who were previously scammed, you must conduct a low-risk pilot. It is not feasible to convince people to purchase an expensive $120 multi-room solar system without established trust.

Therefore, it is reasonable to start by selling a $15 USD Pico Solar kit to a family. It provides basic lighting and a radio. Once it survives for a considerable amount of time (proving it is not a counterfeit), the family’s trust is restored.

Next, the family uses a VSLA to upgrade to a $50 solar system. This system powers three lights and includes phone charging capabilities.

Eventually, after families are satisfied with their products, they can purchase a $120 Solar Home System. This is a full infrastructure kit that is capable of powering entire homes and small businesses.

To hold each other accountable, VSLA members will meet once a week to make their repayments. Considering the loan is internal, the repayment feels much closer to a contribution to their neighbourhood rather than a burden. By allowing families to pay off a Pico kit in small weekly installments, even the largest home systems eventually become an attainable reality rather than an impossible luxury.

The Engine of Sustainability

Beyond restoring trust, a sustainable energy project must answer one simple question: Who pays for it? Many solar initiatives in Africa have collapsed because they were treated as permanent charities rather than businesses. When the initial grant money from an international donor runs out, the supply chain breaks, the agents stop working, and communities are left without electricity.

Navo Solar is designed as a self-sustaining commercial channel that relies on local economics rather than continuous donations. The local agent earns a real commission, which motivates them to continue selling and providing after-sales service. To optimize this, the model follows evidence from a 2025 Living Goods study in Uganda, which proved that a 50:50 split between a flat stipend and performance-based commission is optimal for keeping workers active and productive.

Furthermore, Navo Solar utilizes a local advantage to beat the counterfeit market: the Ugandan government’s 2024 UECCC price subsidy program. This program provides a 60% discount on certified, high-quality solar technology. By leveraging this subsidy, Navo Solar can offer verified products at prices that are actually lower than the dangerous, low-quality fakes found on the street. This ensures that the business is not only bankable, targeting a projected revenue of $1.07 million over three years, but that it is built to last.

Powering the Last Mile

The world has spent billions of dollars developing incredible solar technology, but that technology is useless if it cannot cross the last mile into rural homes. As evidenced, the barrier to energy access in the West Nile is a lack of trust between technology and users, not affordability.

Think back to the family of refugees from South Sudan mentioned earlier. Their goal is not just to have reliable electricity; it is to have a safe home, a way to charge a phone to speak with relatives, and the ability to save the 25% of their income previously lost to kerosene.

Navo Solar proves that by fixing the distribution channel instead of just the solar panel, we can solve the energy crisis through the power of community.

References

[1] GIZ (Deutsche Gesellschaft für Internationale Zusammenarbeit). (2011). Solar Lamps Field Test Uganda: Field Test Report. Energypedia. energypedia.info/images/7/72/GIZ_Solar_Lamps_field_Report_Uganda_Webversion.pdf

[2] Elven, M., et al. (2025). Behavioral savings sessions increase the pursuit of solar products among refugees in Uganda. npj Climate Action. Nature Publishing Group. DOI: 10.1038/s44168–025–00212-x nature.com/articles/s44168–025–00212-x

[3] CARE International. (2025). Village Savings and Loan Associations Training Manual: A Step-by-Step Guide for Community-Based Trainers. Revised December 2025. CARE USA. care.org/wp-content/uploads/2026/01/CARE-VSLA-Manual.pdf

[4] CARE International. (2018). Village Agent Model Learning Document. CARE Canada. care.ca/wp-content/uploads/2018/12/Village-Agent-Model-Learning-Document.pdf

[5] CARE International. (2024). Collective Investment Manual. Revised December 2024. CARE USA. care.org/wp-content/uploads/2025/10/Collective-Investment-Manual_ENG.pdf

[6] Living Goods Uganda / Frontiers in Health Services. (2025). Effects of financial compensation structures on community health worker performance, motivation, and retention: evidence from a multi-arm quasi-experiment in Uganda. Frontiers in Health Services. DOI: 10.3389/frhs.2025.1687782 frontiersin.org/journals/health-services/articles/10.3389/frhs.2025.1687782/full

[7] Solar Sister / United Nations Framework Convention on Climate Change (UNFCCC). Solar Sister: Uganda, Rwanda, South Sudan. UNFCCC Momentum for Change: Lighthouse Activities. unfccc.int/climate-action/momentum-for-change/lighthouse-activities/solar-sister

[8] Uganda Energy Credit Capitalisation Company (UECCC). (2024). UECCC’s Price Subsidy Program for Clean Energy Technologies Officially Launched. UECCC Press Release, December 5, 2024. ueccc.or.ug/uecccs-price-subsidy-program-for-clean-energy-technologies-officially-launched

[9] Uganda Energy Credit Capitalisation Company (UECCC). (2024). UECCC Partners with Energy Service Companies to Extend Price Subsidies for Clean Energy Technologies. ueccc.or.ug/ueccc-partners-with-energy-service-companies-to-extend-price-subsidies-for-clean-energy-technologies

[10] GOGLA / Lighting Global. (2020). Off-Grid Solar Market Trends Report 2020. GOGLA. gogla.org

[11] ENGIE Energy Access Uganda. (2022). ENGIE Energy Access Uganda: Refugee Investment Network Profile. Refugee Investment Network. refugeeinvestments.org/opportunities/engie-energy-access-uganda

[12] TaroWorks / Global Distributors Collective (GDC). (2020). Last Mile Distribution: How Software Can Serve This Sector. TaroWorks Blog. taroworks.org/last-mile-distribution

[13] NURI (Northern Uganda Resilience Initiative). VSLA Training Manual: Northern Uganda. NURI Programme Documentation. Documents CBT recruitment in Arua via ARUDIFA and Lugbara language requirements for Arua/Yumbe.

[14] SolarAid / SunnyMoney. Ashden Award Case Study. Ashden. ashden.org/awards/winners/solaraid

Supplemented by: Turner Kirk Trust on SolarAid. turner-kirk.org/impact/solaraid-the-sunnymoney-brains-unit

[15] West Nile Innovation Hub. West Nile Region: Population and Agricultural Dependence Data. westnileinnovationhub.org/region

[16] IPC Uganda. (2025). IPC Acute Food Insecurity and Acute Malnutrition Report: Uganda, April 2025 — February 2026. Integrated Food Security Phase Classification. ipcinfo.org

[17] VeraSol. Product Database: Sun King and d.light Certifications. data.verasol.org/products/sek

[18] Sun-Connect. Uganda Solar Distributor Directory. sun-connect.org/distributors/uganda-2

[19] AFARD (Agency for Accelerated Regional Development). Project Documentation: Arua and West Nile Operations. afard.net/our-projects

Introduction

The bottleneck for advancing the energy transition in the Global South is not a technology gap or a lack of sunlight. It is credit risk.

Off-grid solar companies in countries like Kenya have led by example by accelerating energy access by using a Pay-As-You-Go financial model. It makes electricity affordable through small, frequent, and manageable payments.

Although this model has had tremendous success, there are some major challenges that have yet to be overcome. For example, in many Global South Nations, users of PAYG solar systems usually work in agriculture.

This means that the customer’s incomes are almost entirely tied to harvest production. If a severe drought devastates a farmer’s land, in all likelihood, they will default on their solar payments. Given the worsening climate crisis, droughts and other severe weather phenomena have become more common and have made it more difficult to predict. As a result, capital allocators and PAYG solar companies are more hesitant to invest in Global South countries that are vulnerable to climate change.

To put it simply, if we can’t map and monitor the agricultural risk, we can’t finance the energy transition, thus slowing progress.

I wanted to see if we could solve this capital allocation problem from space.

In this project, I built a machine learning model to answer a specific question: Can we use historical rainfall and vegetation health (NDVI) to predict financial stress for off-grid communities before it happens?

The Target: De-risking 2 Million People

Before predicting credit risk, we needed to know where the risk actually mattered.

In a previous geospatial analysis, I built a suitability model through ArcGIS Proto identify the most viable markets for mini-grid deployment in Kenya. Instead of just looking for the sunniest regions, I filtered the country for the “Sweet Spot” zones that balanced three conflicting variables:

1. High Solar Yield: GHI > 2000 kWh/m²
2. Infrastructure Gap: Validated using VIIRS Nighttime Lights and >10km from the national grid.
3. Market Scale: High-density population clusters.

This analysis highlighted 5 priority investment zones representing over 2 million people (such as the massive 864k-person cluster near Tsavo). These regions are ecologically perfect for solar but are currently underserved by the main grid.

These clusters represent a massive opportunity for PAYG solar companies. However, these areas are located in semi-arid regions that are vulnerable to climate shocks. Before companies deploy solar home systems in these areas, they must understand the relationship between climate and credit risk.

This is where my machine learning model comes in.

The Tech Stack

To predict credit risk for 2 million people, we need to process decades of satellite imagery. To perform my analysis, I relied on a cloud-native geospatial stack:

• Google Earth Engine: GEE allows us to query and process planetary-scale satellite data directly on Google’s servers.
• Python (geemap & pandas): used the geemap library to interface with GEE via Python, eventually extracting the spatial data into structured DataFrames.
• Scikit-Learn: The machine learning library I used to train the Random Forest algorithm. I chose Random Forest because it is highly parallelizable and handles noisy, high-dimensional environmental data very well.

The Data: Measuring Risk from Space

To mathematically model credit risk, we need to represent two things: the climate catalyst and the economic impact.

1. The Catalyst (Precipitation): I sourced historical rainfall data using CHIRPS (Climate Hazards Group InfraRed Precipitation with Station data). CHIRPS is ideal for this type of analysis because it combines high-resolution satellite infrared imagery with ground-based measurements, providing a reliable historical baseline of drought conditions.
2. The Economic Impact (Vegetation Health): To measure how the land reacted to the lack of rain, I used NDVI (Normalized Difference Vegetation Index). NDVI acts as a direct proxy for agricultural yield. If the NDVI drops over a prolonged period, crops are failing. When crops fail, the farmer’s income disappears, and the probability of a PAYG solar default skyrockets.

Defining the Default Zone - The 0.3 Threshold

Before feeding this data into a Random Forest algorithm, I needed to visualize the historical data. I plotted the NDVI for all 5 priority clusters over a three-year period (2021 to 2023), specifically looking at how these regions responded to the severe droughts in East Africa.

But a line chart of vegetation health isn’t actionable for a financial institution. To make it useful, I established a Critical Stress Threshold at an NDVI of 0.3.

In remote sensing, an NDVI below 0.3 generally indicates sparse vegetation, severe drought stress, or barren land. For an off-grid solar company, this line represents the “Default Zone.” If a cluster’s vegetation health dips below this line and stays there, the local agricultural economy is effectively paralyzed.

Analyzing the Data

There are several useful insights that can be drawn from the chart above. Clusters like Embu (blue line) remained highly resilient, consistently hovering well above the 0.3 threshold even during dry seasons. These are low-risk markets. If we look at the Tsavo cluster, we can see the NDVI values took a massive plunge in 2022 due to a severe drought. If we look at Mombasa, the values virtually stayed below 0.3 for the entire 3 years. This is because Mombasa is an urban city, which makes its land unsuitable for farming.

The Ingestion Pipeline: Pixels to Pandas

A Random Forest model cannot ingest raw satellite images; the data must be spatially aggregated and temporally structured.

Instead of downloading thousands of heavy images locally, I utilized Google Earth Engine’s reduceRegions function. This calculated the average NDVI and total rainfall for my specific geographic boundaries directly in the cloud.

To bridge the gap between Google’s cloud servers and my local machine learning environment, I used the geemap Python library. A single command (geemap.ee_to_df()) converted the cloud-based outputs into local Pandas DataFrames.

However, this created a new engineering challenge: Temporal Alignment. Satellites do not take pictures every day. The CHIRPS rainfall data is recorded in 5-day intervals (pentads), while the optical NDVI data depends on satellite flyovers and cloud cover. To train an ML model, the “Rain” and “Vegetation” data must be perfectly aligned on the same timeline.

To solve this, I used a pandas merge_asof function with a 5-day tolerance. This matched the closest rainfall reading to the exact date the satellite photo was taken, effectively synchronizing the climate catalyst with the economic impact.

With a clean, synchronized master dataset, it was time to build the engine.

The Engine: Feature Engineering & The Baseline Model

I focused the machine learning model specifically on the Tsavo Cluster, as it historically exhibited the most severe climate volatility.

Before training the algorithm, I had to translate how nature works into variables a machine could understand. Plants do not react to rain instantly; there is a lag effect. Therefore, feeding the model “today’s rain” is useless. Instead, I engineered three specific features for the algorithm to learn from:

1. Past_30d_Rain_Avg: A rolling 6-period mean of the CHIRPS pentad data.
2. Prev_NDVI: The vegetation health from the previous time step.
3. Month: To account for predictable, seasonal agricultural cycles.

I chose a Random Forest Regressor to process these features. To establish a baseline capability, I ran Experiment A (The “Lab Test”). I took all the historical data and performed a standard random split (80% training, 20% testing).

The results were strong. The Random Forest successfully interpolated the data, achieving a high positive R² score. It accurately mapped the relationship between the 30-day rainfall averages and the resulting vegetation health.

The Temporal Split

In the real world, you cannot predict the future using a random split of time. You must train a model entirely on the past and test it exclusively on the future.

I restructured the data using a Temporal Split, forcing the model to train on data before 2022, and testing it blindly on the 2022–2023 Horn of Africa drought.

When confronted with the 2023 data, the model failed.

The Engineering Lesson: The Limits of Random Forest

Why did a model with engineered features and a strong baseline score fail so catastrophically?

The failure was a fundamental mathematical limitation of the Random Forest architecture. Decision trees work by finding the average of historical outcomes. Because of this, Random Forests cannot mathematically extrapolate beyond the bounds of their training data. The algorithm had never seen a drought this severe in its pre-2022 training set. Therefore, it lacked the mathematical capacity to predict an NDVI value lower than the historical minimum it was trained on.

This was a massive engineering lesson. Machine learning is excellent at interpolating historical patterns, but it is fundamentally flawed when predicting unprecedented climate shocks.

Conclusion

This project proved that while Geospatial AI is an incredible tool for monitoring risk, using standard machine learning to predict the future of a rapidly changing climate is a flawed strategy.

Here is the GitHub with the code: https://github.com/kevint2007/solarstressindex/blob/main/Kenya.ipynb

Author Bio

I’m Kevin Theivendran, an 18-year-old computer science and geography student interested in leveraging geospatial data and policy to advance the Sustainable Development Goals.

To reach me, contact me via:

LinkedIn: https://www.linkedin.com/in/kevin-theivendran-0ab344233/

Email: kevintheivendran9@gmail.com

Introduction

Sri Lanka is a small island nation in South Asia, home to about 22 million people. Rice is the backbone of its farming sector, forming the staple diet and main source of income for millions of rural families. Most rice growers are smallholders and cultivate family plots that rely on seasonal rainfall and age-old irrigation systems.

Sri Lanka’s Ancient Irrigation Method: The Cascading Tank System

For thousands of years, Sri Lankan farmers have relied on the cascading tank system to irrigate their fields. The aim of this system is simple: to prevent rainwater from monsoons from being discharged into rivers quickly. Instead, it utilizes topography to collect rainwater. As shown in the diagram, small reservoirs collect rainwater at high elevation. These reservoirs are connected to middle reservoirs and, by extension, the big reservoirs at the lowest elevation. Farmers rely on water from the Big Tank to irrigate their rice paddy.

Water Scarcity and Climate Unpredictability

Rice farming in Sri Lanka operates within a dual seasonal system that has structured rural life for centuries. The Maha season runs from October through February, driven by the Northeast monsoon and the Inter-monsoon rainfall. The Yala season spans April through September, fed by the First Inter-monsoon and the Southwest monsoon rains. These two seasons are not equal: Maha is the country’s major growing season, responsible for the bulk of annual rice production, while Yala serves as a supplementary harvest, particularly important in irrigated areas.​

Case Study: The 2016–2017 Drought

The 2016–2017 drought marked a turning point in Sri Lanka’s agricultural vulnerability. Consecutive monsoon failures, both the 2016 Northeast monsoon and the Southwest monsoon, left irrigation tanks critically depleted. Total paddy production in 2017 fell nearly 40% to just 2.7 million tonnes, down from 4.4 million tonnes in 2016.​

The Maha season of 2016–2017 saw a 49% collapse in output, dropping to 1.478 million tonnes. The subsequent Yala season fared little better, with production falling 48% to 789,000 metric tonnes. Beyond the harvest shortfall, the impact on food security was immediate and severe: rice prices spiked to all-time highs, forcing 900,000 people into food insecurity. Households that traditionally grew their own food were forced to purchase from markets at inflated prices, leading many to reduce consumption and shift to cheaper, less nutritious alternatives.​

Case Study: The 2024 Eastern Crisis

The 2024 season revealed an even more perverse vulnerability: the simultaneous occurrence of destructive floods and water shortage within a single year. In the eastern provinces, particularly around the Senanayake Samudra reservoir on the Gal Oya river, monsoon intensification in May–June 2024 brought unprecedented rainfall exceeding 200 millimeters in a single event. Flash floods, river overflows, and coordinated dam releases devastated recently planted fields.

Just months later, the same region faced severe water scarcity. Water levels in the Senanayake Samudra reservoir, normally high during Maha, reached dangerously low levels by mid-season, forcing the Irrigation Department to release stored water into saturated lands. Farmers who had just replanted their Maha fields after flood damage watched as their new crops were destroyed again. Over 235,000 people were affected by these floods.

This oscillation, from flood to drought to flood again within a single year, has become common.

Climate Change Intensifying the Pattern

Research from the Rice Research and Development Institute emphasizes that climate-driven droughts, elevated temperatures, and floods are now the defining feature of rice cultivation in Sri Lanka. Leading researchers note that unpredictable rainfall patterns, shifting rain intensities, and changing monsoon timing are making it increasingly difficult for farmers to plan cultivation cycles.

Water Waste and Inefficient Usage

While agriculture in Sri Lanka uses about 90% of the nation’s freshwater, only around 20% of farmable land receives water through modern, reliable irrigation year-round. This means most farmers continue to rely on unpredictable rainfall. Historic tanks and canals lose significant amounts of water through evaporation and leakage. On the other hand, diesel-powered pumps are expensive to operate and maintain as the country deals with a fuel shortage.

For the minority of Sri Lankan farmers with mechanical irrigation, they mostly rely on diesel-powered water pumps. These pumps cost 15,000–25,000 Sri Lankan Rupees monthly for fuel alone, which is a significant expense for the country’s smallholders.

Farmer Persona and the Case for Solar Irrigation

Most Sri Lankan rice farmers are smallholders operating under tight financial constraints. Although detailed public data on their montly incomes or diesel expenditures is limited, available case studies can provide a realistic profile.

Typical Smallholder Paddy-Farmer Profile

• Land size: 0.5–2 hectares, typical of dry-zone smallholders
• Livelihood: Reliant on seasonal rice cultivation (Maha and Yala), with yields sensitive to rainfall variability
• Irrigation method: Rainfed or diesel-powered pumps, with costs and reliability varying by monsoon performance
• Financial constraints: Limited ability to invest upfront in modern irrigation technologies

The Productivity Potential: What Solar Irrigation Unlocks

Supplementing Irrigation for Existing Systems
While roughly 70–80% of Sri Lanka’s paddy land is irrigated through tanks, reservoirs, and canal schemes, access to water is not always reliable. Seasonal variability, tank inefficiencies, and occasional droughts can leave farmers without sufficient water for their crops. Solar-powered pumps provide a climate-resilient solution by supplementing existing irrigation infrastructure: they can lift water from tanks, canals, or shallow wells to fields that otherwise depend on variable tank levels or costly diesel pumps. By replacing diesel, solar pumps reduce operational costs by LKR 18,000–25,000 per month and enable supplemental irrigation, making double cropping (two harvests per year) feasible. Integrated pay-as-you-go functionality and IoT monitoring ensure continuous, efficient water delivery while protecting the pump investment. In this way, solar pumps enhance productivity, reduce risk, and improve economic outcomes for irrigated smallholder farmers.

Research shows Sri Lankan farmers are roughly 71% technically efficient; with improved technology like solar pumps and better farming techniques, average yields could grow by nearly 30%. Trials have shown profit increases of up to 400% and yield doubling when solar is paired with improved water management. Solar irrigation enables double-cropping (two rice harvests per year) without the financial and environmental burden of diesel.

Solar Irrigation as an Economically Friendly Solution

Solar-powered pumps present a favorable financial path. Over five years, total costs, including purchase and maintenance, are lower than those of diesel systems once fuel cost savings are included. Investment in solar irrigation allows farmers to pay monthly, avoiding heavy upfront costs and lowering ongoing expenses. Leases and pay-as-you-go arrangements make the technology accessible immediately, even for low-income farmers. On average, a typical farmer can save 1.8–2.5 million Rupees over ten years by switching from diesel to solar irrigation.

Potential Collaboration with Ennos Sunlight Pumps

Ennos Sunlight Pumps are engineered with a strong emphasis on durability, reliability, and precision, which makes them well-suited for smallholder farming contexts such as Northern Sri Lanka. Designed in Switzerland and manufactured through partnerships in South Africa and India, Ennos pumps are built to perform in challenging rural environments with minimal maintenance requirements.

For AgriLanka’s pilot phase, the 0.5 HP Ennos Sunlight Pump is the most appropriate configuration. This capacity aligns well with the needs of smallholder farmers cultivating small plots, home gardens, and fragmented fields typical of post-conflict and drought-prone regions in Northern Sri Lanka. The 0.5 HP system is sufficient for lifting water from shallow wells, canals, or storage tanks and distributing it efficiently through basic surface or hose-based irrigation methods.

A key advantage of Ennos pumps is their integrated 4G IoT technology, which enables remote monitoring and pay-as-you-go functionality. This allows farmers to access solar irrigation without high upfront costs, paying only for the water they pump through flexible, credit-based financing. The system also provides real-time performance data, fault detection, and theft alerts, reducing operational risk and improving asset security during the pilot.

Ennos pumps are designed to operate maintenance-free for over ten years, significantly lowering lifecycle costs and minimizing downtime for farmers. This long-term reliability is particularly valuable in remote villages where access to repair services is limited.

By deploying a 0.5 HP Ennos Sunlight Pump through a pay-as-you-harvest or pay-as-you-pump model, AgriLanka can test an affordable, scalable solar irrigation solution that directly addresses water access, income stability, and climate resilience for smallholder farmers in Northern Sri Lanka.

Deploying Solar Irrigation in Sri Lanka: Geographic and Social Context

Sri Lanka receives solar radiation levels of 5kWh/m²/day which is ideal for year-round pump operation. Government targets call for 80% renewable energy by 2040, making solar irrigation a strategic focus. Neighboring India and Bangladesh have successful models, giving Sri Lanka valuable blueprints.

Institutions such as IWMI and UNDP actively support modernization, offering technical training, funding, and project support. Awareness of climate risk is rising among farmers, and cooperative structures in regions like the Northeast are ready for innovation.

References

ennos. (2020, May 15). ennos — portable solar water pump. Ennos. https://ennos.ch/

Mutlu Özdoğan, Wang, S., Ghose, D., Fraga, E., Fernandes, A., & Varela, G. (2025). Field-Scale Rice Area and Yield Mapping in Sri Lanka with Optical Remote Sensing and Limited Training Data. Remote Sensing, 17(17), 3065–3065. https://doi.org/10.3390/rs17173065

Philips, R. (2025, January 12). Sri Lanka’s Rice Crisis: Scarcity In Spite Of Self-Sufficiency — Colombo Telegraph. Colombo Telegraph. https://www.colombotelegraph.com/index.php/sri-lankas-rice-crisis-scarcity-in-spite-of-self-sufficiency/

Rashmitha Diwyanjalee. (2025, February 3). Sri Lanka’s Rice Fields in Crisis: How Climate Change is Reshaping the Nation’s Staple Crop — Climate Fact Checks. Climate Fact Checks. https://climatefactchecks.org/sri-lankas-rice-fields-in-crisis-how-climate-change-is-reshaping-the-nations-staple-crop/

Samarasinghe, B. K. D. J. R., Zhu, Y., Abeynayake, N. R., Zeng, X., Mathavan, B., Wanninayake, R. W. W. M. P. K., Rasheed, S., & Bah, A. S. (2025). Climate change and rice production in Sri Lanka: short-run vs. long-run symmetric and asymmetric effects. Frontiers in Sustainable Food Systems, 9(9). https://doi.org/10.3389/fsufs.2025.1592542

Singh, S. K. (2007). THIS REPORT CONTAINS ASSESSMENTS OF COMMODITY AND TRADE ISSUES MADE BY USDA STAFF AND NOT NECESSARILY STATEMENTS OF OFFICIAL U.S. GOVERNMENT POLICY. https://apps.fas.usda.gov/newgainapi/api/Report/DownloadReportByFileName?fileName=Grain%20and%20Feed%20Annual_New%20Delhi_Sri%20Lanka_CE2025-0004.pdf

Executive Summary

HABGuardAI is an early-warning system designed to predict harmful algal blooms before they disrupt desalination operations. Around the world, desalination plants provide freshwater for hundreds of millions of people in coastal and arid regions, but growing bloom frequency is creating costly, unpredictable shutdowns. In the Middle East alone, home to over 300 desalination plants producing more than 40 million cubic meters of water daily, HABs have already caused multi-million-dollar losses and emergency water shortages.

Blending satellite ocean data with machine learning, the HABGuardAI predicts bloom risk up to two weeks in advance, enabling operators to act in time before contamination strikes. With the system, desalination facilities across the globe avoid shutdowns, protect membranes, and maintain continuous freshwater supply, strengthening global water security against one of the fastest-growing threats in the ocean.

Introduction

Desalination plants have emerged in the last few decades as a viable option to increase freshwater access worldwide. Despite its lucrative potential, there are significant barriers that prevent desalination from becoming a truly sustainable, scalable, and global solution. High costs and demand of electricity, excessive carbon emissions, and regular maintenance pose as glaring pain points. In the midst of all these challenges, an overlooked threat to many desalination plants, especially in the Middle East, is harmful algae blooms.

When the Ocean Turns Against Desalination

In countries surrounding the Arabian Gulf and Red Sea, desalination is a necessity. More than 90% of the UAE’s fresh water comes from seawater desalination, as does approximately 70% of Saudi Arabia’s. Throughout the entire Gulf region, more than 300 plants together produce over 40 million cubic meters of water daily, supplying drinking water to tens of millions of people.

Harmful algal blooms happen when microscopic algae, under conditions like high temperatures, a calm sea, and nutrient-rich runoff, multiply very rapidly. These events are gradually becoming more frequent. Satellite records and regional marine studies recorded that the frequency of HABs in the Arabian Gulf has increased by about 20% over the past decade, which has been linked to climate change and coastal pollution.

When a bloom drifts toward a plant’s intake, it can clog filtration systems, foul reverse osmosis membranes, and release biotoxins into feedwater. The result is an immediate operational halt to prevent contamination.

The red tide event of 2008–2009 along the coast of the UAE gave a serious signal of the vulnerability of desalination infrastructure: more than 70 plants were driven offline, compelling the need for emergency imports of water and losses in excess of $200,000 per plant per day. Similar incidents have been documented in Oman, where HABs caused multi-day disruptions, and in Saudi Arabia, where jellyfish and algal events affected Red Sea desalination units near Jeddah.

Across the globe, HABs cause annual shutdowns or capacity losses in 10% to 15% of large SWRO plants, roughly 150 in number. A single bloom may persist for five to ten weeks, and most plants lose an average of four to fifteen days of production per event.

The High Cost of Being Unprepared

Every unplanned shutdown cascades into financial and environmental losses:

• $100,000–$300,000 per day in halted production.
• $500,000–$1 million to replace fouled membranes.
• Millions in emergency water sourcing and energy waste.

Despite these risks, most plants rely on reactive measures: visual satellite images, manual sampling, or delayed environmental reports. By the time the water looks bad, it’s already too late.

The ocean’s unpredictability remains desalination’s most expensive blind spot.

The Status Quo of Tracking HABs

Most desalination plants today appear to rely on a predominantly reactive, low-resolution monitoring system despite the fact that shutdowns due to HABs come at very high economic and operational costs.

Monitoring currently is done through three major categories, including remote sensing, manual sampling, and public environmental reporting. Each offers a partial view, and taken together, they compose a fragmented defense system that struggles to deliver timely and actionable intelligence.

1. Satellite Remote Sensing

Most desalination utilities rely on Chl-a detection from satellites to infer the presence of blooms. NASA MODIS-Aqua, Copernicus Sentinel-3 OLCI, and NOAA VIIRS are some programs providing imagery at spatial resolutions between 300 m and 1 km and revisit intervals of 1–2 days.

These sensors, while effective for detecting large-scale bloom formations offshore, often have insufficient temporal and spatial resolution to monitor the narrow, nearshore intakes of most desalination plants. Data gaps as high as 40–60% can result from cloud cover, aerosol interference, and sun-glint reflection during these periods.

Even when data is available, algorithmic misclassification remains a major problem. Empirical chlorophyll algorithms designed under open ocean conditions usually overestimate or underestimate the algal biomass in optically complex coastal waters dominated by suspended sediments and colored dissolved organic matter. Several studies conducted in the Arabian Gulf and Red Sea document retrieval errors for Chl-a exceeding 30 to 50%, hence providing unreliable data for operational forecasting.

2. Manual Sampling and Laboratory Analysis

In-situ water sampling is carried out on a weekly or biweekly basis by most desalination operators, near intake sites. Samples are filtered, preserved, and analyzed via fluorometry, microscopy, or HPLC to determine algal concentration and species composition.

While accurate, this process introduces a 24–72 hour time lag between sampling and result dissemination which is far too long when bloom conditions can shift hourly with wind and current. In the Gulf region, where bloom transport speeds average 10–20 km/day, a 2-day delay can mean the difference between early intervention and unplanned shutdown.

In addition, data are usually only collected at fixed-point stations, typically 1–2 per plant, by most operators, which spatially under-represents the coastal hydrodynamics. Without a distributed sensor network or continuous flow cytometry, blooms moving laterally or below the surface layer often go undetected until visible discoloration occurs.

3. Government Bulletins and Regional Early Warning Systems

Several national and regional agencies publish HAB bulletins. Examples of such bulletins include but are not limited to NOAA’s Harmful Algal Bloom Operational Forecast System for the Gulf of Mexico and ROPME bulletins for the Arabian Gulf.

However, these are coarse in scale, 10–25 km grids, and normally updated once every 2–3 days, mostly for ecological monitoring rather than operational decision support of desalination plants. The data dissemination pipelines are manual, with reports sent via PDFs or email summaries, not API-linked dashboards.

In effect, most plants receive alerts only after surface discoloration has already been confirmed which leaves operators with hours to respond.

4. The Technical Gap

This system-level fragility arises from three major data challenges:

Spatial undersampling: Conventional satellite imagery cannot resolve sub-kilometer gradients or subsurface chlorophyll maxima near intakes. Intakes are generally found 100–500 m from the shore at a depth of 5–20 m.

Temporal latency: Remote and field data streams are asynchronous, updated at intervals far longer than bloom transport rates.

Predictive limitations: current systems are diagnostic, meaning they describe what has happened, not prognostic-they predict what will happen. Also, there are no machine-learning-based probabilistic forecasts tailored for the desalination intake conditions.

5. Consequences

The result is a reactive operational model: shutdown decisions are made only after turbidity spikes, filters clog, or water turns visibly discolored. Plants often take 4–15 days to recover from HAB-induced fouling, at membrane replacement costs of $500,000-$1 million per incident. In effect, the world’s desalination infrastructure operates blind to one of its fastest-growing environmental threats, with detection tools lagging reality by days or even weeks.

Introducing HABGuard

HABGuard is an early-warning system that predicts harmful algal blooms 7–14 days before they reach desalination plant intakes.

By combining satellite ocean data and machine learning, HABGuard transforms the way desalination operators monitor risk. It turns reaction into prediction, giving utilities the foresight to act before contamination strikes.

How HABGuard Works

1. Collects satellite data including Chlorophyll-a (bloom biomass), sea surface temperature (SST), wind speed, and ocean currents from NASA MODIS, Copernicus Sentinel-3, and NOAA archives.
2. Trains on historical patterns of harmful algal blooms in regions like the Arabian Gulf, Gulf of Mexico, and Mediterranean.
3. Predicts bloom risk near desalination intakes using AI models such as Random Forests and ConvLSTMs
4. Generates actionable alerts with simple recommendations:
   ⚠ WATCH: Bloom likely near Fujairah intake (6 days).
   Recommendation: Switch to offshore intake and increase pretreatment flow.

The system integrates seamlessly with plant dashboards or SCADA systems. offering a software-only solution with instant operational impact.

The Features of the ML Model

HABGuard identifies early bloom “fingerprints” invisible to the human eye:

• Chlorophyll-a spikes that signal algal growth.
• Seawater surface temperature anomalies showing warm, stagnant waters.
• Low-wind periods when algae accumulate near shore.
• Drift vectors from current models that forecast bloom movement.

Training data from past events, including the UAE 2008 red tide and Red Sea 2019 blooms, provide ground truth labels for supervised learning.

In initial replay tests, HABGuard correctly flagged 98% of high-risk events at least a week before they reached desalination intakes. That’s enough time for operators to switch intakes, recalibrate filters, or prepare alternative sources.

Cost-Savings from HabGuard

HABGuardAI does not need new tanks, pipes or chemical dosing skids. It only needs data.

Plants that use the forecast avoid days when no water is produced. Each lost day costs about USD 100,000–300,000. Early warning also means the plant stops or slows intake before the cells destroy the reverse osmosis sheets. Membrane replacements drop by roughly 30–50 percent. Emergency power for extra pumps and trucks for last-minute chemicals are also called for less often, which lowers fuel and labor bills.

References

Anderson, D. M., Boerlage, S., Dixon, M., & Linders, J. (2017). Harmful algal blooms (HABs) and desalination: A guide to impacts, monitoring and management (IOC Manuals and Guides, Vol. 78). Intergovernmental Oceanographic Commission of UNESCO. https://repository.oceanbestpractices.org/handle/11329/646

Caron, D. A., Dennehy, P., & Krasner, S. W. (2010). Harmful algae and their potential impacts on desalination operations off Southern California. Water Research, 44(2), 385–416. https://dornsife.usc.edu/caron/wp-content/uploads/sites/263/2023/11/2010_Caron_etal_WR.pdf

Gobler, C. J., Doherty, O. M., Hattenrath-Lehmann, T. K., Griffith, A. W., & Seneca, S. (2017). Ocean warming since 1982 has expanded the niche of toxic algal blooms in the North Atlantic and North Pacific oceans. Proceedings of the National Academy of Sciences, 114(19), 4975–4980. https://doi.org/10.1073/pnas.1619575114

Wells, M. L., Trainer, V. L., Smayda, T. J., Karlson, B. S., Trick, C. G., Kudela, R. M., Ishikawa, A., Bernard, S., Wulff, A., Anderson, D. M., & Cochlan, W. P. (2015). Harmful algal blooms and climate change: Learning from the past and present to forecast the future. Harmful Algae, 49, 68–93. https://doi.org/10.1016/j.hal.2015.07.009

Jafari, M., Vrouwenvelder, J. S., Atik, C., & Cornelissen, E. R. (2021). Cost of fouling in full-scale reverse osmosis and nanofiltration installations in the Netherlands. Desalination, 500, 114865. https://doi.org/10.1016/j.desal.2020.114865

Energy Recovery Inc. (2023). The economics of downtime. https://energyrecovery.com/wp-content/uploads/2023/11/White-PaperEconomics-of-Downtime.pdf

Middle East Institute. (2025, January 6). The Gulf’s water crisis: Why cooperation is crucial and complicated. https://www.mei.edu/publications/gulfs-water-crisis-why-cooperation-crucial-and-complicated

NASA Goddard Space Flight Center. (2015, July 27). MODIS chlorophyll-a product documentation. https://modis.gsfc.nasa.gov/data/dataprod/chlor_a.php

Tilstone, G. H., Moore, G. F., Shutler, J. D., Smyth, T. J., Miller, P. I., & Groom, S. B. (2021). Performance of Sentinel-3 OLCI, MODIS-Aqua, and VIIRS ocean colour products for satellite chlorophyll-a algorithms. Remote Sensing of Environment, 262, 112502. https://doi.org/10.1016/j.rse.2021.112502

Kim, J., Kim, H., Kim, S., Lee, S., & Kim, S. (2024). Recent advances in algal bloom detection and prediction using machine learning. Science of the Total Environment, 916, 168984. https://doi.org/10.1016/j.scitotenv.2024.168984

Zhou, G., Xu, K., Yang, X., & Yu, C. (2024). A spatiotemporal attention-augmented ConvLSTM model for ocean remote sensing reflectance prediction. International Journal of Applied Earth Observation and Geoinformation, 126, 103434. https://doi.org/10.1016/j.jag.2023.103434

Veolia. (2014, February 28). Fujairah 2 reverse osmosis desalination plant, United Arab Emirates. https://www.africa.veolia.com/en/our-clients/fujairah-2-reverse-osmosis-desalination-plant-united-arab-emirates

Khaleej Times. (2015, April 4). Red tide forces temporary shutdown of salination plant. https://www.khaleejtimes.com/uae/red-tide-forces-temporary-shutdown-of-salination-plant?amp=1

I just returned from the Villars Institute in Villars-Sur-Ollon, Switzerland, after a week-long, intensive program on ecopreneurship. If you don’t know what ecopreneurship is, it’s entrepreneurship with the aim of addressing problems in the climate space. At Villars, I had the opportunity of meeting 24 other young changemakers working in the climate space. Each student came in with a project in mind and was eager to take the next steps. Some were in the ideation phase, and others had already raised venture capital! Regardless, all of the people there were bold, motivated, and ambitious. My project, AgriLanka, a venture that strives to make solar irrigation pumps affordable to Sri Lankan farmers through microfinancing, was still in the early stages. Although it was in the early stages, I had received invaluable mentorship from Sofi Khwaja and Sofia de Meyer. These two outstanding ecopreneurs were our program directors. On top of that, the Villars Institute invited industry experts who had expertise in a host of disciplines, including economics, public speaking, and systems thinking. The program, which ran from August 3rd to 9th, struck the perfect balance between intense learning and reflection. Each day from the 4th to the 8th was dedicated to a specific theme, giving us time to dive deep into complex topics while still being surrounded by the breathtaking beauty of Villars-sur-Ollon in the Swiss Alps.

I’ve put together a full reflection, including the biggest lessons I took away, in a YouTube video below.

Our mission: Build the financial and technological infrastructure to power Africa’s smallholder farmers into prosperity.

On Farming In Malawi

In Malawi, over 80% of the population is employed in agriculture. Despite this, only 2% of the population working in agriculture has access to modern irrigation tools. This means that Malawian farmers have to rely on rain and traditional practices to cultivate their crops. Due to the growing climate change concerns, relying on these methods is unsustainable, as droughts have become more frequent. Therefore, farmers remain in a cycle of subsistence. [1]

One simple yet powerful tool can change this: solar-powered irrigation pumps

With solar-powered irrigation pumps, Malawian farmers can double their harvest production, lengthen harvest seasons, and increase their output by 3x. Not only can we improve the lives of Malawian farmers, but we are also accelerating the transition to clean energy.

However, these pumps cost around $750 to $1000 upfront, which is a number far too expensive for the average Malawian smallholder farmer.

That is where AgriMalawi comes in.

The biggest problem with selling solar-powered irrigation pumps is that that are very expensive for farmers in Sub-Saharan Africa. The market value for solar-powered irrigation pumps is anywhere from $800 USD to $1000 USD. For context, the average smallholder farmer in Malawi makes $1840 USD worth of crops. Only 13% of that income is used for profit. In other words, farmers only have $239 of disposable income.

Building a Financial Model that Works for Farmers

Existing solutions for financing solar-powered irrigation pumps include the pay-as-you-go solution. This model has historically had success in Africa but low-repayment rates remain a huge concern. Experts at the International Renewable Energy Agency suggest that payment structures should align with harvest seasons in Malawi.

At AgriMalawi, we developed a new model that farmers could trust: co-operatives and pay-as-you-harvest.

Here is how it works:

• We partner with agricultural co-ops (10 farmers per group).
• We provide the co-op with a solar pump at no upfront cost.
• The farmers pay back the pump in affordable installments, tied to harvest cycles, which will now occur twice a year, thanks to the pump
• After 2 years, they own the pump outright.

Each farmer pays ~$48 per harvest — around 10% of their increased seasonal income. This structure ensures affordability while keeping repayment aligned with income flow.

Our Technology

Payments are collected via a hybrid platform: an offline solution that involves a simple USSD text interface for low-tech phones and a mobile-friendly dashboard for co-op leaders with smartphones. Payments are sent directly to AgriMalawi’s platform and logged transparently.

Our Strategy

AgriMalawi is building a financially and environmentally sustainable platform that empowers farmers. Our ultimate goal in Malawi is to reach the 3.1 million smallholder farmers.

Year 1: Pilot

• Deploy 5 solar pumps through 5 co-operatives (50–100 farmers).
• Monitor repayments, crop yields, and household income changes
• Establish local partnerships with district agriculture offices, the Government of Malawi, and NGOS.
• Build the first version of the hybrid payments platform (USSD + dashboard

Year 2: Expand

• Scale to 500 pumps across central and southern Malawi (reaching 10,000 farmers).
• Integrate remote sensing + crop analytics to offer tailored repayment support.
• Launch AgriMalawi Digital: an improved mobile dashboard for farmers and co-op leaders.

Year 3–4: National Scale

• Operate in every district across Malawi with 10,000 solar pumps deployed (100,000 farmers).
• Partner with the Malawian Ministry of Agriculture to integrate AgriMalawi into national irrigation goals.
• Bundle crop insurance and savings tools into the AgriMalawi platform.
• Use farmer repayment data to generate Malawi’s first solar-credit risk profiles.

Year 5 and Beyond

Once AgriMalawi has reached every district in Malawi, our goal is to diversify our product inventory and expand beyond Malawi’s borders. In addition to solar-powered irrigation pumps, we plan to offer other affordable, climate-smart agricultural tools, such as solar-powered cold storage units to reduce post-harvest losses and low-cost soil testing kits. Our Pay-as-you-Harvest model is adaptable to smallholder farmers in other sub-Saharan African countries and even in Asia. Some of the countries that have tremendous potential to benefit from our model include Zimbabwe, Mozambique, and Tanzania.

Our Revenue Model

AgriMalawi earns revenue by:

• Taking a small margin on repayment collections through interest
• Offering value-added services like insurance, savings, and crop analytics.
• Partnering with governments and donors to subsidize high-impact infrastructure in the early stages.

As our repayment model proves itself, we’ll reduce reliance on grants and move toward self-sustaining operations.

Our Commitment to Equity

Women produce over 70% of food in Malawi, yet they face significant barriers to finance and land. AgriMalawi is committed to gender inclusion. We will prioritize co-ops with strong female participation, designing accessible interfaces, and partnering with organizations that empower female farmers.

The Vision

At AgriMalawi, we are empowering a generation of farmers with the tools, technology, and trust they need to prosper. By building the right financial infrastructure for Africa’s most essential workforce, we’re helping millions escape subsistence, build wealth, and accelerate and propel social and economic development in Africa.

Source

https://www.irena.org/-/media/Files/IRENA/Agency/Publication/2025/Apr/IRENA_POL_DRE_agriculture_Malawi_2025.pdf

The big talk in the news at these times is climate change and how we can limit its effects. With the rapid advancements in artificial intelligence, we can create climate models that can forecast the impact of climate change. As a result, this has been crucial in influencing government policy to combat climate change effectively.

In this project, I used AI to predict the Global Temperatures in 100 years. You might have heard before that global temperatures are slowly increasing due to climate change. Although the changes in these temperatures are small, they are substantial. Even a small change in global temperature can cause glaciers to melt. For example, the effects are drastic considering that many animals rely on glaciers for their survival.

The Linear Regression Model

It is important to understand the objective of this project: to show how much global temperatures will increase in 100 years and ultimately why we must act quickly.

What is a Linear Regression Model?

A Linear Regression is a mathematical approach to approximate data with a linear equation (y=mX+b). This equation can be used to predict the output value of any input. In other words, this is a form of machine learning.

Building the Model

Importing Libraries

I built this model on Google Colab. First, I had to import the Libraries.

import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import seaborn as sns
import plotly.offline as py
py.init_notebook_mode(connected=True)
import plotly.graph_objs as go
import plotly.tools as tls
%matplotlib inline

Then, I imported my dataset. I searched through Kaggle and found a collection of datasets of Earth surface temperature data. For this project, I used a dataset of average surface and land temperature every year from 1750 to 2015. This is perfect for a linear regression model. The input variable will be the year and the output will be temperature.

Here is the code that reads, displays, and describes the dataset I imported:

data= pd.read_csv('/content/GlobalTemperatures.csv')
data.head()
data.shape

Visualizing the Data

This code will output a line graph with all the data points. You will notice that I had to clean up the data. The dataset had many categories with times that included the month and day. I only need the year as my input value and the combined temperature of the land and ocean.

#Visualizing the upward trend of global temperatures
data = data[['dt', 'LandAndOceanAverageTemperature']]
data.dropna(inplace=True)

# Convert 'dt' to datetime and extract the year
data['dt'] = pd.to_datetime(data['dt'])
data['year'] = data['dt'].dt.year

# Group by year and calculate mean temperature
data = data.groupby('year')['LandAndOceanAverageTemperature'].mean().reset_index()

# Plot the trend
plt.figure(figsize=(16, 6))
ax = sns.lineplot(
    x=data['year'],
    y=data['LandAndOceanAverageTemperature']
)
ax.set_title('Average Global Temperature Movement')
ax.set_ylabel('Average Global Temperature')
ax.set_xlabel('Year (1750 - 2015)')
plt.show()

Here is the graph:

Do you notice a trend? That’s right, temperatures have increased steadily over time.

The Machine Learning

I imported the Sci-kit Learn Library which does the machine learning.

from sklearn.model_selection import train_test_split
from sklearn.linear_model import LinearRegression

Now, for a machine learning model, I must split the data into two categories: training and testing. I set 20% of the data to be used for testing. The other 80% will be used to train the model. The code:

X = data[['year']]  # Feature (Year)
y = data['LandAndOceanAverageTemperature']  # Target (Temperature)

# Splitting the dataset into the Training set and Test set
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=0)

regressor = LinearRegression()
regressor.fit(X_train,y_train)
y_pred = regressor.predict(X_test)

plt.figure(figsize=(10,6))
plt.scatter(X_train,y_train,color ='green', label ="Training Set")
plt.plot(X_train, regressor.predict(X_train), color='blue', label='Linear regression line')
plt.title('Average Global Temperature (Training Set)')
plt.xlabel('Year')
plt.ylabel('Temperature (°C)')
plt.legend()
plt.show()

Here is the graph for the testing set. This is the graph that matters for predicting the temperature.

The Prediction

On google colab, you can use any year you want and predict the global temperature. I used 2125 which is 100 years from now.

year = 2125
predicted_temp = regressor.predict(np.array([[year]]))
print(f"The predicted temperature for {year} is {predicted_temp[0]:.2f} °C.")

The Result: The predicted temperature for 2125 is 16.24 °C.

The Impact

With this project, I got to visualize how the earth’s temperature is increasing over the last 300 years. Furthermore, as the earth gets warmer, we must acknowledge the devastating consequences it will have on our ecosystems.

Author Bio

I’m Kevin Theivendran, a 17-year-old researching clean energy solutions in developing nations with the help of financial technology and artificial intelligence.

To reach me, contact me via:

LinkedIn: https://www.linkedin.com/in/kevin-theivendran-0ab344233/

Email: kevintheivendran9@gmail.com

See you next time!

I t is not a stretch to say that renewable energy has taken the world by storm. Today, we frequently hear about governments implementing policies that support the adoption of renewable energy. Recently, the Global AI Summit took place in Paris, France. Notably, Canadian Prime Minister Justin Trudeau announced that it was a G7 priority for artificial intelligence to be powered by clean energy. Trudeau’s statement highlights an important realization: society cannot progress in AI development at the expense of climate change.

The relationship between clean energy and artificial intelligence works multi-fold. We require clean energy to power data centers and we can utilize artificial intelligence to make clean electricity generation more efficient. That is exactly what I will be talking about in this article: the relationship between wind power and artificial intelligence.

The Challenge: Why Does Wind Power Need AI?

Wind energy, like many other forms of renewable energy, is intermittent. In other words, there are only certain periods throughout the day that can be optimized for wind energy usage. Unfortunately, this is one of the reasons why fossil fuel energy sources are still heavily relied upon to this day.

Wind speeds fluctuate due to many factors including natural wind patterns, geographical location, and season. When the wind speed is low, adequate electricity is not produced. On the other hand, when wind speed is high, the wind turbine operates close to or at its full potential. In fact, the relationship between wind speed and potential output can be represented by the equation p = w³. Given the fact that wind speed is variable, artificial intelligence is needed to maximize efficiency and output. Deep Learning algorithms, specifically neural networks, can help.

How Neural Networks Optimize Wind Energy

Before I explain that, what are neural networks? Neural networks are an advanced AI model modelled after the human brain. It is made up of interconnected layers which are supposed to represent neurons. These neurons process data and learn patterns like our brains. In this case, neural networks can predict wind power output, optimize energy distribution with the power grid, and eliminate uncertainty with wind electricity.

1. Input Layer — Data Collection

The neural network receives different data categories such as weather data, turbine data, and grid demand. Weather data includes past wind speeds, wind direction, temperature, and air pressure. Turbine data encompasses rotor angles, blade pitch angles, and power output. Lastly, grid demand data shows how much electricity is needed at certain times of the day. This data is crucial for forecasting future wind electricity generation.

2. Hidden Layer

The hidden layer is where the mathematical work is done to come up with an output. During this process, the hidden layer analyzes the input data, identifies relationships and correlations with the data, and works on improving prediction accuracy. One should be familiar with the concepts of weights. In essence, certain factors will have a different “weight” on the outcome. For example, if wind speed plays a huge role in forecasting, then it will have a greater impact on the prediction. Backpropagation is responsible for fine-tuning the neural network over time by adjusting the weight factors. The cool thing about backpropagation is that it will account for new data added.

3. Output Layer

The output layer is where we can see our predictions. For instance, the neural network can predict how much wind electricity will be generated in a certain period. Depending on the kind of neural network, there can be a different number of outputs.

Google DeepMind: Using AI for Renewable Energy

Google DeepMind has been one of the greatest developments in AI in the last decade. DeepMind has worked to solve many problems in various disciplines such as biology, physics, and healthcare. Even then, DeepMind has applied artificial intelligence to optimize wind energy production.

In 2018, DeepMind worked with wind farms across the Central United States by applying machine learning algorithms to predict wind power output. Specifically, with the help of neural networks, the farms could predict their wind power output 36 hours ahead. Because of this, the value of wind energy has increased. But why?

When electricity is more predictable, the value of it goes up. Power grids rely on various sources of electricity to power the homes of millions of citizens. If wind farms can communicate with the power grid about how much electricity they can produce, it’s a game changer. Furthermore, the power grid can plan by preparing backup sources if there isn’t enough electricity produced from wind. From the perspective of wind farms, they can make more money as the value of wind energy increases.

All in all, DeepMind’s AI-driven wind energy forecasting can make the power grid cleaner, more efficient, and more reliable.

What’s The Future of Wind Energy?

Wind energy has surplus potential to become a dominant source of electricity, ultimately combating climate change and limiting fossil fuel reliance.

In 2023, the global wind industry had increased its power generation by 50% compared to the previous year. More and more wind turbines are being built across the world.

Increased Performance

Wind turbines are becoming larger, more efficient and more powerful. For example, there is a lot of buzz surrounding offshore wind farms. Interestingly, the strongest winds are found near the ocean and sea. The objective of these offshore wind farms is to capture these winds over the sea.

Energy Storage

Another important aspect of clean energy adoption is energy storage. Given that clean energy sources are intermittent, we must conserve the energy we get. Excess wind energy generated must be stored in battery storage systems or converted to green hydrogen. The most popular batteries right now are lithium-ion, sodium-ion and solid-state. Right now, scientists are figuring out ways to improve battery storage systems, specifically the material of the battery.

Artificial Intelligence

Of course, as demonstrated in the article, artificial intelligence will continue to play a critical role in optimizing wind energy. For instance, AI-powered predictive maintenance will help detect potential failures in wind turbines before they happen, reducing downtime and maintenance costs. Additionally, AI-driven smart grid integration will enable energy distribution without human intervention. This ensures that wind power is efficiently utilized and stored when supply exceeds demand. Lastly, as machine learning models become more advanced, they will improve wind power forecasting even further, making wind energy an even more reliable and valuable resource for the global energy grid.

TL;DR

Artificial intelligence, particularly neural networks, is transforming wind energy by improving power output predictions, optimizing energy distribution, enhancing grid reliability, and increasing its value. Google DeepMind has already demonstrated this by enabling wind farms to predict energy production 36 hours in advance. Moreover, AI will continue to boost wind energy efficiency through better forecasting, smart grid integration, and predictive maintenance, helping make wind power a more reliable and dominant clean energy source.

Thank you for reading this article and stay tuned for more content on the intersection of AI and climate change.

Author Bio

I’m Kevin Theivendran, a 17-year-old researching clean energy solutions in developing nations with the help of financial technology and artificial intelligence.

To reach me, contact me via:

LinkedIn: https://www.linkedin.com/in/kevin-theivendran-0ab344233/

Email: kevintheivendran9@gmail.com

See you next time!

Sources

Elkin, Carl, and Sims Witherspoon. “Machine Learning Can Boost the Value of Wind Energy.” Google DeepMind, 26 Feb. 2019, deepmind.google/discover/blog/machine-learning-can-boost-the-value-of-wind-energy/? Accessed 17 Feb. 2025.

Mishra, Rishabh. “How Wind Energy Is Reshaping the Future of Global Politics.” World Economic Forum, 14 May 2024, www.weforum.org/stories/2024/05/how-wind-energy-is-reshaping-the-future-of-global-power-and-politics/.