Short-term photovoltaic power forecasting with feature extraction and attention mechanisms

The uncertainty of weather conditions has always been a major challenge limiting the performance of photovoltaic (PV) power prediction. Enhancing the accuracy and stability of PV power prediction is crucial for optimizing grid operation. In response to this challenge, this study introduces a hybrid model that integrates an attention mechanism with the Convolutional Neural Network (CNN) and Bidirectional Long Short-Term Memory Network (BiLSTM). This model aims to mitigate the adverse impact of weather variability on the accuracy of PV power prediction by effectively extracting key features from multidimensional time series data. Additionally, through ablation experiments, this research further assesses the contribution of each model component to performance. Validated with actual data collected from a 1MW PV power station in China, our proposed model demonstrates significant performance advantages compared to eight advanced prediction models. Ablation study results reveal that removing the CNN component led to a 58.2% increase in MAE and a 53.9% increase in RMSE, while the removal of the attention mechanism resulted in an 83.8% increase in maximum error. These findings underscore the substantial enhancement in prediction accuracy achieved through the integration of CNN and BiLSTM, and the introduction of the attention mechanism significantly boosts the model's prediction stability.