Intelligent Forecasting of Energy Consumption using Temporal Fusion Transformer model

Abstract

The increasing demand for electrical energy is a major problem of carbon dioxide emissions. As well, the inefficient use of electricity is also concerning. Smart grids can make electricity use more efficient by integrating other technologies into the electrical system. Forecasting is one of those technologies which can improve electricity consumption efficiency. Precise forecasting can balance the demand and supply of electrical generation, and with the growing use of renewable energy sources such as solar and wind, more accurate forecasting is necessary. Our objective was to find a forecasting model that can best fit demand forecasting. Transformer is the name of the model that we applied in the forecasting task. The N-BEATS and N-HiTS models were used to compare with Transformer. The result is shown in mean absolute percentage error (MAPE). The Transformer model had the lowest MAPE (4.5980%) compared to the N-BEATS (5.0266%) and N-HiTS (7.9865%) models, indicating that it provides a more accurate prediction. The model's hyperparameters were set to the same values so that their results could be compared properly.