<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Christian Van Engers on Medium]]></title>
        <description><![CDATA[Stories by Christian Van Engers on Medium]]></description>
        <link>https://medium.com/@christian.van.engers?source=rss-bd1a30573dfa------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*dmbNkD5D-u45r44go_cf0g.png</url>
            <title>Stories by Christian Van Engers on Medium</title>
            <link>https://medium.com/@christian.van.engers?source=rss-bd1a30573dfa------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 16:14:38 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@christian.van.engers/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Improving customer support by predicting solar panel issues using Machine Learning]]></title>
            <link>https://medium.com/@christian.van.engers/improving-customer-support-by-predicting-solar-panel-issues-using-machine-learning-fac3fbc28f86?source=rss-bd1a30573dfa------2</link>
            <guid isPermaLink="false">https://medium.com/p/fac3fbc28f86</guid>
            <dc:creator><![CDATA[Christian Van Engers]]></dc:creator>
            <pubDate>Thu, 23 Jul 2020 04:08:34 GMT</pubDate>
            <atom:updated>2020-07-23T07:06:22.461Z</atom:updated>
            <content:encoded><![CDATA[<h3>Introduction</h3><p>The World Health Organization estimates that in 2018, 50% of the population in Sub-Saharan Africa lacked access to electricity[1]. To address the ongoing need for electricity access and the challenges surrounding climate change, renewable energy sources are emerging as an important power source, with a reported 5-fold increase between 2012–2015 for wind and solar in this region[2]. <a href="https://www.fenixintl.com/">Fenix International</a> currently helps supply over 800,000 households with energy through their lease-to-own solar home systems.</p><p>These solar home systems consist of a smart-battery, solar panel and a series of accessories such as lights and a phone charger. The hardware running the battery also contains a GSM module that can connect to the mobile phone network. The module facilitates lease-to-own payments and allows remote customer support. The ability to perform remote service is essential when dealing with a customer base that may face financial and logistical challenges to visiting one of Fenix’s Service Centres.</p><p>I spent 6 months volunteering for Fenix as a Data Fellow. I had used Python for data analysis during my time in academia and was excited to try my hand at a data science application in the real world. One of my projects involved building machine learning (ML) models to predict issues for devices brought to a service centre, using data from GSM-enabled systems from the weeks leading up to the visit. My goal was to support improving Call Centre service through better remote diagnosis, and to expedite diagnosis when a customer comes into a Service Centre.</p><h3>The data</h3><p>The GSM data I used came in aggregate form, such as the amount of energy the battery received over the course of a day. Since different customers use their devices differently (and due to different weather and geography), the data has inherent noise. This means that to be able to develop a ML model that outputs reliable predictions, data from a large number of devices is needed.</p><p>First, I mapped out what issues customers were having. The most common issue had just been fixed with a firmware update (which devices downloaded remotely via the GSM module!). The most viable alternative was to look at issues related to solar panels (~ 4000 customer support tickets relating to this issue).</p><h3>Results</h3><p>For each panel that had a customer support ticket, the GSM data collected on the days preceding that ticket was retrieved. The panels were then classified according to their diagnosis, being either <em>faulty</em> or <em>working</em>. Next, the data was split into two groups. One group was used to train the Machine Learning model, and the other (consisting of data the model had never seen!) was used to test the model’s predictive performance.</p><p>The results of that test are shown in the figure below. The y-axis shows the number of devices in the test group. The x-axis shows the so-called “prediction threshold”. The prediction threshold is the estimated chance that the model produces a correct prediction for a particular outcome — in this case whether the panel is faulty and needs to be swapped out with a new one <em>(</em><strong>P</strong><em>faulty)</em>. Only two outcomes are predicted: whether a panel is working or faulty. This means that the chance the model correctly predicts that the panel is working is: <strong>P</strong><em>working</em> = (1 - <strong>P</strong><em>faulty</em>). As such, at a prediction threshold for faulty panels of <strong>P</strong><em>faulty</em> = 0.3, the prediction threshold for working panels is <strong>P</strong><em>working</em> = 0.7.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*FDMeqxDvND6hsxhy5nK0mg.png" /></figure><p>The central band in the figure, consisting of the brown and dark blue bar sections, show the correct predictions. Faulty panels that the model classified as faulty and should be swapped are shown in brown. Correctly classified working panels are shown in blue. The outer red and light blue bar sections respectively show the proportion that would have been falsely swapped, and the proportion that would have been falsely classified as working but should have been swapped instead.</p><p>At the default threshold of 0.5, the model gets it right about 70% of the time. In practice this means that if a Service Centre Associate (SCA) were to blindly follow the model, they would end up saving a lot of time: Testing a panel is done by placing the panel out in the sun and letting it charge a battery the SCA knows is working, a process taking hours. However, the SCA would also falsely swap almost 20% of all panels. And, perhaps even worse, would send almost 10% of customers home with a broken panel!</p><p>One of Fenix’s core values is ensuring a good customer experience. This means minimizing the number of panels that should have been swapped (light blue). At the same time, it is important to avoid false swaps (in red), since they constitute capital cost. The top left and bottom right of the graph show some potential uses where negative outcomes are minimized. For example:</p><blockquote>(1) If the model predicts the likelihood of faulty as 0.2 or less, recommend returning the panel to the customer without a test (and possibly with some education on proper panel use). This saves the Service Centre Associate 11% of panel tests with an error rate of 9% (1% of total).</blockquote><blockquote>(2) On the other hand, if the model predicts the likelihood of a faulty panel is 0.8 or greater, recommend a panel swap without a test. This would save 18% of tests — albeit with a false swap rate of 18% (3% of total)</blockquote><p>The potential costs and benefits of such an implementation need to be carefully considered. Until we have more data to improve the model, its best use may be to identify broken panels when the Service Centre is very busy: this avoids customers not being able get help — which would force them to return at a later date or spend the night in town. However, it is exciting to see that even with little data we already get predictions that could have an impact.</p><p>Over time data will accumulate and the model will improve, becoming a viable tool to assist or expedite diagnosis at the Service Centre. More work is needed, but like bringing electricity to Sub-Saharan Africa using renewable energy sources, it is a great start.</p><p>— — — — — — — — — <strong><br></strong>[1] https://data.worldbank.org/ — Access to electricity (% of population) — Sub-Saharan Africa<br>[2] https://data.worldbank.org/ — Electricity production from renewable sources, excluding hydroelectric (% of total) — Sub-Saharan Africa</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fac3fbc28f86" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>