Ultra-short-term Power Prediction of Photovoltaic based on CNN-AM-BiLSTM

The power of photovoltaic power generation is affected by a variety of factors such as weather, radiation and temperature, exhibit a high degree of randomness and uncertainty. To address this problem, this paper proposes a ultra-short-term photovoltaic power prediction model based on Convolutional Neural Networks (CNN) and attention mechanism and bidirectional long short-term memory (Bi-LSTM). Firstly, Pearson's correlation coefficient was used to filter out the meteorological factors which have significant influence on the historical power. Subsequently, in order to capture the spatial features of the input data, a convolutional neural network is utilized for feature extraction and an attention mechanism is introduced to strengthen the model's ability to discriminate key features. Meanwhile, the Bi-LSTM network is utilized to capture the temporal dependencies in the time series data, and finally the PV power prediction value is obtained. Finally, the validation is based on data from actual PV power plants, and the results show that the proposed Convolutional Neural Networks and attention mechanism (CA) and Bi-LSTM Ultra-short-term PV power prediction model outperforms the single model in terms of prediction accuracy, reflecting the effectiveness and superiority of the hybrid model.