基于多模态模型的分布式光伏电站多步功率预测方法

Abstract We developed a multimodal photovoltaic (PV) power forecasting method that integrates visual and physical information. It addresses the shortcomings of traditional forecasting methods in handling changes in the state of PV panels . Time-series images obtained by monitoring equipment were used to observe changes in PV states over time. The Pearson correlation coefficient was used to assess the relationship between meteorological and environmental factors and PV power. A three-stage hybrid data reconstruction method was proposed to solve the problems of missing data, high noise, and synchronization of timestamps in the PV system . A convolutional feature extraction network was used to analyze the impact of PV panel obstruction on power generation efficiency . A novel learnable weight cross-attention feature fusion was incorporated to overcome the limitation of a single data fusion strategy to capture complex correlations. The experimental results demonstrated that the proposed method outperformed seven other methods for single-step forecasting, achieving the lowest mean absolute error of 1.53. The mean square error improvement of the model for multi-step forecasting was, on average, 45.40% higher than that of the baseline methods.