PSGformer: A novel multivariate net load forecasting model for the smart grid

• Published in: Journal of computational science Vol. 78; p. 102288
• Main Authors: Zhang, Qingyong, Zhou, Shiyang, Xu, Bingrong, Shen, Zhanwei, Chang, Wanfeng
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.06.2024
• Subjects: Graph neural network; Power load forecasting; Smart grid; Transformer; Transformer; Power load forecasting; Smart grid; Graph neural network
• ISSN: 1877-7503; 1877-7511
• DOI: 10.1016/j.jocs.2024.102288

Smart grid intelligently transforms modern power system through the introduction of various advanced techniques, promoting the growth of distributed renewable energy sources and improving the efficiency of load management. The construction of smart grid recognizes the importance of net load forecasting, which represents the difference between the demanded load and the renewable energy generation, and has a significant impact on power system monitoring and control. Recently, some Transformer-based models have shown excellent performances in load forecasting. However, most of them perform weakly in face of net load forecasting due to the lack of emphasis on the specific trend variations and the inter-feature dependencies in cyclical temporal patterns. To address the two issues, we present the PSGformer that consists of multiscale series decomposition, seasonal module and trend module, which can focus on unique properties that affect the accuracy of net load forecasting by processing different components obtained from the series decomposition. The seasonal module seamlessly introduces graph attention network into the Transformer to compensate for the capability of capturing local inter-feature dependencies. The trend module adopts a new segment-wise iterative approach to capture trend patterns. Compared to the six baseline models, PSGformer has higher forecasting accuracy and stability on four load datasets from different countries. •The model considers the impact of trend-seasonal variations and relevant factors affecting the net load on prediction accuracy.•Graph attention network is introduced into Transformer to capture the inter-feature dependencies in the seasonal component.•A segmented iterative model is proposed to process the trend component.