A transformer-based time series forecasting model with an efficient data preprocessing scheme

Abstract

Time series forecasting with cyclicality is key to the development of green energy, particularly wind energy, due to its high volatility. Accurate forecasting allows for optimal use of energy storage systems and balancing of power grids. In this article, the authors have developed a model for forecasting time series in wind energy through the combined use of Fourier transform and an adapted transformer architecture to solve the time series forecasting problem. The use of Fourier transform provided the ability to detect and account for hidden periodicities that may not be obvious in simple time series analysis, and allowed for the separation of random fluctuations from significant cyclical components, contributing to more accurate data analysis. The use of transformer architecture made it possible to effectively account for both short-term fluctuations and long-term trends in wind patterns, creating more accurate and reliable forecasts of wind energy production. The results show that the model outperforms methods such as transformers, long short term memory (LSTM), LSTM with Fourier transform, and DeepAR in forecast accuracy, taking into account seasonal, weather, and daily cycles of wind data.