Short-term PV output prediction based on CNN-BiLSTM-attention and Kendall-DBSCAN feature extraction

In order to address the limitations in prediction accuracy caused by the inherent volatility and uncertainty of photovoltaic power generation, this study developed a short-term photovoltaic power output prediction model that integrates meteorological feature selection and weather clustering. The model specifically utilizes the Kendall-density-based spatial clustering of applications with noise (DBSCAN)-convolutional neural network (CNN)-bidirectional long short-term memory (BiLSTM)-attention algorithm. First, the Kendall coefficient was utilized to quantify the similarity between meteorological factors and photovoltaic output. Meteorological factors exhibiting higher similarity were selected as input features for the prediction model. The DBSCAN clustering algorithm was then employed to categorize historical operational data into four typical operating conditions: sunny, cloudy, overcast, and adverse weather. Second, an evaluation model based on feature similarity and mutual information entropy is constructed to calculate the similarity between the target day and each cluster, and the optimal historical similarity day dataset is selected. Finally, a CNN-BiLSTM-attention composite neural network is used for photovoltaic power output prediction. The findings demonstrate that the CNN-BiLSTM-attention neural network employing Kendall-DBSCAN feature extraction attains mean absolute errors of 1.2842, 1.2553, 1.6503, and 1.2486 for the four weather types, respectively. In comparison with alternative models, the root mean square error is reduced by 8.6%–58.4%, thereby demonstrating excellent predictive performance.