Photovoltaic power prediction based on multi-head attention Bi-LSTM

To address the challenge of predicting photovoltaic (PV) power generation caused by its inherent randomness and uncertainty, a multi-head attention-based bidirectional long short-term memory (Bi-LSTM) model was proposed. Meteorological and PV system operation data highly correlated with power generation were extracted through Pearson correlation analysis, and data preprocessing was performed by combining outlier handling with standardization. Outlier handling reduces the impact of extreme data and mitigates gradient oscillation during training, while standardization ensures consistency in feature scales to avoid imbalanced weight learning. Different weights were assigned to critical temporal features by the multi-head attention mechanism to enhance the ability to capture PV power trends. Forward and backward dependencies in the time series were processed by the Bi-LSTM to improve adaptability to complex sequential data. Experimental results show that in the application to a PV plant in Xinjiang, the proposed model achieves significantly higher prediction accuracy than the traditional LSTM network, featuring lower errors and good promising application prospects.