Exploiting Fourier Transform for Multi-scale Electric Load Forecasting

Niangxi Zhuang, Xiaobing Yang, Peilin Yang, Chaohui Liang, Lulu Sun, Choujun Zhan

Abstract: Accurate electric load forecasting is one important issue in many applications for power grid systems. However, traditional forecasting models have limitations in capturing nonlinear dynamic properties across different time scales. Hence, the uncertainty and complexity of power grid systems makes accurate prediction of power loads a challenging task. To overcome the bottleneck, we propose a deep learning model called Multi-Scale Fourier Net (MSFNet), which can effectively capture dynamic properties of different scales of power load time series by employing multi-scale analysis through Fourier Transform. Unlike conventional methods, the MSFNet uses FFT to identify primary periodic features as time scales for multi-scale analysis, while also utilizing Gated Recurrent Units (GRU) and the channel attention mechanism to capture long-term dependencies. The PJM electricity market dataset in the United States is adopted to evaluate the performance of the MSFNet. Experimental results indicate that MSFNet outperforms benchmark time series forecasting models in various evaluation metrics, with particularly notable performance in the mean squared error (MSE) and root mean squared error (RMSE) metrics.