Power Prediction of Distributed Photovoltaic Plants Based on Dynamic Feature Extraction and CNN-BiLSTM

To address the weaknesses of the conventional PV power prediction model that the prediction accuracy decreases during weather fluctuation, this paper introduces a distributed power prediction model for PV plants based on dynamic feature extraction and CNN-BiLSTM. Firstly, the K-means is applied for secondary clustering to mine the PV power fluctuation trend of abrupt weather with two hours as the time scale, and the clustering results show that the fluctuating weather presents richer trend changes in short-term sequences. Secondly, PV power fluctuation trend features are extracted by introducing moving average (SMA) and Bollinger band indicator (BOLL). Then, CNN network is used to extract nonlinear local features among multidimensional input data, as well as BiLSTM network is used for bi-directional feature extraction of time-series data and output the prediction results of PV power. The final experimental findings reveal that the model introduced in this paper surpasses other prediction models in terms of accuracy.