Using satellite imagery and deep learning to predict energy demand and measure sales performance

Abstract

The traditional approach to predicting energy demand uses time series forecasting models with historical ground truth data. However, this data is not always publicly available at the geographical level required. This paper offers a new method to predict energy demand using freely available data that has been increasingly applied to predict social and business phenomena, satellite imagery. We collected monthly nighttime satellite imagery covering the entire Brazilian extension between 2012 and 2019. After extracting the luminosity per Brazilian state, we developed deep learning, time series, CNN model to predict energy demand based on the nighttime satellite imagery. The model achieved an MSE of 3% compared with ground truth data collected from the Brazilian government's Energy Research Office. Based on energy demand data, we propose an easy-to-implement method to measure total market demand (i.e., the overall sales performance in an industry) and measure the sales performance of one electrical materials producer. We labeled it as the delta (Δ) sales performance index. Our contributions are robust and flexible in predicting energy demand and measuring sales performance at any desired geographical level.