A dual-attention mechanism LSTM model for power output forecasting of high-penetration photovoltaic systems

Abstract

Photovoltaic (PV) power generation in high-penetration renewable energy systems exhibited pronounced fluctuations and substantial uncertainty. To improve PV power forecasting accuracy, this study proposed a hybrid forecasting model, termed DA-LSTM, which integrated a dual-attention (DA) mechanism with a long short-term memory (LSTM) network. The model leveraged LSTM temporal encoding to capture key meteorological features from multidimensional input data. In addition, attention mechanisms were introduced at both the feature and temporal levels to extract task-relevant representations and identify historical time steps most similar to the current prediction target, thereby enabling the modelling of non-linear temporal dependencies. To evaluate model performance, a series of comparative experiments was conducted against multiple benchmark models. The results showed the following. First, compared with the benchmark models, DA-LSTM achieved a root mean square error (RMSE) of 795.36 kW, a mean absolute error (MAE) of 580.15 kW, and an R² value of 0.9658. Second, ablation experiments demonstrated that removing the feature attention module increased the RMSE by 4.8%, which confirmed the effectiveness and necessity of the dual attention mechanism. Third, the proposed model exhibited superior robustness and adaptability under conditions characterised by abrupt weather variations and high-penetration power fluctuations. Overall, the experimental results demonstrated that incorporating dual attention into the LSTM framework significantly improved the forecasting accuracy of PV power generation. The proposed approach also provided a practical technical solution for enhancing prediction performance and operational stability in high-penetration renewable energy grids.