Multivariate Time Series Forecasting of Integrated Energy Systems Based on Fast Fourier Transform Fully Connected Space-Time Graph

In the current context of energy conservation, the issue of energy emissions has become a focal point in resource management. The manageable prediction of integrated energy systems plays a crucial role in adjusting the energy input structure in a reverse manner, leading to a substantial reduction in energy system waste and unnecessary expenses. Moreover, precise energy load forecasting can greatly enhance energy conservation efforts, facilitating environmental protection and promoting efficient resource utilization. The purpose of multi time series prediction usually involves learning time step features of multi-dimensional time series, followed by forecasting a time series over a period. However, traditional methods of extracting time series do not yield the same segmentation effect as a sentence, making it challenging to capture semantic information. Therefore, extracting local information is crucial for analyzing their relationships. A method has been proposed to aggregate time steps into subsequences called patch to enhance locality and extract semantic information that cannot be obtained at the point level, then put the semantic sub-series into the structure of a Fast Fourier Transform Graph Neural Network to learn the features in space and time dimensions. The proposed method demonstrates superior prediction accuracy compared to CNN, MLP, LSTM, RNN, GNN, StemGNN, MTGNN with great results: a mean absolute percentage error (MAPE) of 0.251%, an explained variance score of 99.256%, and an R2-score of 99.138%.