Development of a Photovoltaic power generation power prediction model based on BiLSTM-AM and Bayesian optimization

Accurate prediction of photovoltaic power generation is essential for improving grid absorption capacity and optimizing energy management. However, existing prediction methods face challenges such as sudden weather changes, insufficient exploration of historical data correlations, and inadequate hyperparameter optimization, leading to limited prediction accuracy. To address these challenges, this paper proposes a photovoltaic power prediction model that integrates bidirectional long short-term memory networks (BiLSTM), attention mechanisms, and Bayesian optimization. The model uses BiLSTM to identify the spatio-temporal correlations between historical power data and multi-dimensional meteorological data. The attention mechanism is integrated to dynamically weight the influence of critical meteorological events, thereby further enhancing the model's robustness. Additionally, Bayesian optimization is employed to perform adaptive hyperparameter search, further improving the model's predictive capabilities. The experimental results show that the proposed integrated model outperforms the traditional LSTM model, the independent bidirectional LSTM model, and the bidirectional LSTM-AM model, with the MAE and MAPE of the predictions reduced to 0.43 and 5.2% respectively. Furthermore, the model's performance under normal, sandstorms, and snowy weather conditions was validated. The validation results indicated that the model exhibits strong adaptability to various weather conditions.