Short-Term Power Prediction Method for Photovoltaic Power Plants Based on ECA-MTGNN Integration

To address the issue of low accuracy in short-term power prediction for photovoltaic (PV) power plants, this paper proposes a power prediction method that integrates an Efficient Channel Attention Mechanism and a Multivariate Time Series Graph Neural Network (ECA-MTGNN). Firstly, to tackle challenges such as data missing and anomalies, data cleaning and imputation were performed on historical weather data and power data collected from a PV power plant located in northern China. Then, the ECA-MTGNN method was utilized to yield short-term power prediction for this plant. Finally, the prediction performance of the proposed method was evaluated using the prediction accuracy assessment criteria specific to PV power plants in the northern China region and the general evaluation metrics for time series prediction. Experimental results demonstrated that, compared to the MTGNN model, the ECA-MTGNN model achieved improvements in prediction accuracy of 8.82%, 6.27%, 2.89%, and 14.88% over the four days involved, respectively. In comparison with commonly used models for time series prediction, such as LSTM and CNN, the ECA-MTGNN model demonstrated improved prediction accuracy under conditions of gentle power variations, and more significant enhancements under conditions of sharp power fluctuations. Moreover, it exhibited a superior ability to fit actual power curves.