SFLGNN: Power Load Forecasting Based on Spectral-Frequency Learning and Graph Neural Network

With the continuous development of modern power systems, accurate power load forecasting has gradually become a key issue in energy management. Compared to deep learning models, existing statistical-based prediction models are not well suited to learn how to represent power load data and mine the rich information in the data. In addition, power load data needs to consider both intra-series temporal correlation and inter-series correlation. Recently, attempts have been made to capture both correlations simultaneously, but most of the studies capture only temporal correlations in the time domain and use predefined a priori as inter-sequence relationships. In this paper, SFLGNN (Spectral-Frequency Learning Based Graph Neural Network) is designed for power load forecasting. SFLGNN effectively solves the problem by capturing both variable correlation and inter-sequence correlation using graph neural networks, and extracts important features through CP decomposition with multi-attention mechanism. In addition, SFLGNN enhances representation learning through Spectral-Frequency Learning. We performed comparative experiments, ablation experiments, and hyper-parameter sensitivity experiments on our collected dataset. The experiments demonstrate the effectiveness of our proposed model in the power load forecasting task.

I. Introduction

Accurate power load forecasting plays an integral role in the optimal planning and efficient operation of smart grids. It is an essential tool for improving the overall economy and stability of modern power systems while minimizing the associated waste and costs. With the increasing complexity of the residential electricity situation and the integration of some emerging industries such as new energy vehicles with the power grid, the power system is facing increasing volatility and complexity, which in turn poses a greater challenge to accurate power load forecasting.