Electricity Forecasting in Kenyan Off-grid Microgrid: Forecasting Accuracy Versus Multi-Year Load Growth

Low-cost microgrids have a pivotal role in extending universal access to electricity. However, for simplicity, their energy management system usually relies on simple forecasting-free techniques like load following or cycle charging, which are sub-optimal. The availability of accurate electricity forecasting plays a key role in achieving cost effectiveness in advanced predictive techniques. However, the accuracy of forecasts can affect performances, especially considering the extreme load growth that microgrids experience. For these reasons, this study investigates the effectiveness of machine-learning techniques in rolling-horizon short-term load forecasting for off-grid rural regions in developing countries, considering using real multiyear data. Several models, including shallow neural networks, linear regression and basic techniques are evaluated using high-resolution data collected over 4 years for a Kenyan microgrid in Faza island. Trained on day-before/day-after pairs in a rolling-horizon approach along the years, these models predict 24-hour demand based on historical data, with performance compared to basic forecasting methods. Results indicate the ability of machine-learning models to better perform over simple baseline predictions, suggesting their potential for real-time energy dispatch optimization. This research contributes to reducing costs and advancing sustainable energy practices, which are essential to achieve universal electricity access in developing regions.