Short-Term Electricity Demand Forecasting for Kathmandu Valley, Nepal

Abstract

Accurate electricity demand forecasting for a short horizon is very relevant aspect for managing day-to-day operation control, scheduling, and planning. The deterministic variables such as type of days, and weather variables such as temperature are the major factors that affect the forecasting accuracy. Since the automation systems are continuously increased and implemented by smart meters and internet of things, static models computations are replacing accordingly by dynamic real time robust forecasting models. Therefore, time series, regression, machine learning, and deep learning models are designed and implemented on the electricity demand dataset of Kathmandu Valley, Nepal. Accuracy improvement is also considered during model design. The result shows that the deep learning model, long short term memory (LSTM) performs outstanding performance in-terms of mean absolute percentage error (MAPE) value 1.56%, and root mean square error (RMSE) value 3.12 MW. While analyzing the regression coefficients, electricity demand during Dashain shows the lowest variation while Tihar (Dipawali/Laxmi Puja) shows the highest (peak) demand variation.