Use of Attention-Based Neural Networks to Short-Term Load Forecasting in the Republic of Panama

Abstract

One pillar of our society is the use of electricity as an engine of development, Short-Term Load Forecasting (STLF) contributes to the resilience and security of electrical supply, by predicting the amount of electricity that should be generated in the near future. Humanity is currently moving from an energy mix based on fossil fuels to a sustainable energy mix, a green one. One challenge of this shift is to forecast, as accurately as possible, the amount of energy load at any moment. This study compares STLF performed by state-of-the-art Neural Network and SARIMA model. First, demand is predicted with SARIMA model and then with a neural network with attention, in this occasion, the Temporal Fusion Transformer (TFT), next, both techniques are compared. The results show that SARIMA is suitable for STLF, with average performance metric values of MAPE and RMSE, of 0.064 and 101.4 MWh, respectively; when use TFT, prediction accuracy increases with a MAPE of 0.044, and RMSE of 69.2 MWh. This research is presented as a review of the state-of-the-technique and thus establishes a baseline that can be used to forecast National Energy Load in the Republic of Panama.