Open Access

ARTICLE

Distributed Photovoltaic Power Prediction Technology Based on Spatio-Temporal Graph Neural Networks

Dayan Sun¹, Xiao Cao^2,*, Zhifeng Liang¹, Junrong Xia², Yuqi Wang³

1 Power Dispatching and Control Center, State Grid Corporation of China, Beijing, 100031, China
2 New Energy Research Institute, China Electric Power Research Institute Co., Ltd., Nanjing, 210003, China
3 Grid Technology Center, State Grid Shandong Electric Power Company Electric Power Scientific Research Institute, Jinan, 250002, China

* Corresponding Author: Xiao Cao. Email:

(This article belongs to the Special Issue: AI-Driven Innovations in Sustainable Energy Systems: Advances in Optimization, Storage, and Conversion)

Energy Engineering 2025, 122(8), 3329-3346. https://doi.org/10.32604/ee.2025.066341

Received 05 April 2025; Accepted 11 June 2025; Issue published 24 July 2025

Abstract

Photovoltaic (PV) power generation is undergoing significant growth and serves as a key driver of the global energy transition. However, its intermittent nature, which fluctuates with weather conditions, has raised concerns about grid stability. Accurate PV power prediction has been demonstrated as crucial for power system operation and scheduling, enabling power slope control, fluctuation mitigation, grid stability enhancement, and reliable data support for secure grid operation. However, existing prediction models primarily target centralized PV plants, largely neglecting the spatiotemporal coupling dynamics and output uncertainties inherent to distributed PV systems. This study proposes a novel Spatio-Temporal Graph Neural Network (STGNN) architecture for distributed PV power generation prediction, designed to enhance distributed photovoltaic (PV) power generation forecasting accuracy and support regional grid scheduling. This approach models each PV power plant as a node in an undirected graph, with edges representing correlations between plants to capture spatial dependencies. The model comprises multiple Sparse Attention-based Adaptive Spatio-Temporal (SAAST) blocks. The SAAST blocks include sparse temporal attention, sparse spatial attention, an adaptive Graph Convolutional Network (GCN), and a temporal convolution network (TCN). These components eliminate weak temporal and spatial correlations, better represent dynamic spatial dependencies, and further enhance prediction accuracy. Finally, multi-dimensional comparative experiments between the STGNN and other models on the DKASC PV dataset demonstrate its superior performance in terms of accuracy and goodness-of-fit for distributed PV power generation prediction.

---

Keywords

---