Multiple Load Forecasting at Power Substations with Temporal Fusion Transformer

Abstract

Energy demand forecasting plays a crucial role in the efficient management of the electricity system, seeking to optimise resources and minimise waste. This study investigates the predictive potential of the Temporal Fusion Transformer (TFT), an innovative approach that allows multiple time series to be dealt with simultaneously, guaranteeing the interpretability of the results. Using historical data from a New Zealand power company, this research evaluates the effectiveness of the TFT in predicting multiple power substation time series, with a predictive horizon of 24 hours into the future. The results indicate significant accuracy of the proposed model, with mean absolute errors of less than 2%.