How Remote Sensing Can Be of Value to the Energy Sector
Date:
Abstract: Recent work has demonstrated how one can use daytime satellite imagery to assist the prediction of individual residential building consumption levels upon connection. Using six years of longitudinal data of electricity consumption for a large cross-section of grid customers from Kenya, we apply convolutional neural networks (CNNs) to daytime satellite images to predict expected levels of residential electricity consumption for individual customers throughout the country.
Achieving universal electricity access through cost-effective use of resources benefits from accurate estimates of expected electricity consumption of the anticipated consumers. While not the only criteria, demand estimates are a crucial input to electrification planning. Such estimates can be inferred from income or recent utility experience in the area, or scarce survey data, or through indirect but widely available sources such as nighttime lights and satellite images. Leveraging Convolutional Neural Networks (CNNs), this work presents a novel data-driven approach trained on a sample of labeled geo-referenced utility bills to predict demand for individual buildings using daytime satellite imagery. The training dataset consists of 0.01% of Kenya’s residential electricity bills and predictions are made for the entire population using high-resolution satellite imagery of the country. This work shows that richer predictions are obtained with satellite images compared to other widely available options.