Abstract: The weather factors that affect the output of photovoltaic power generation systems have great volatility and discontinuities. Therefore, it is necessary to create a suitable prediction model to accurately predict the characteristics of photovoltaic output to ensure the efficient operation of the power generation network. This paper selects the appropriate historical photovoltaic power generation data through the maximal information coefficient, uses it as one of the features to reconstruct the input data, and attention mechanism is introduced on the Encoder-Decoder model constructed by LSTM neurons to obtain an attention-based Encoder-Decoder photovoltaic power generation prediction model. The analysis of actual photovoltaic power plant examples verifies the accuracy and applicability of the proposed model in predicting photovoltaic power generation.

Key words: photovoltaic power generation, maximal information coefficient, LSTM neural network, Encoder-Decoder framework, attention mechanism