Abstract:

In order to improve the accuracy of photovoltaic output power prediction, a model that combines complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) algorithm with improved long short-term memory (LSTM) neural network is proposed to predict short-term photovoltaic power. Firstly, the photovoltaic power sequence is decomposed by CEEMDAN algorithm and the subsequence components are obtained. Then, improved LSTM neural network is used to predict each subsequence component. Particle swarm optimization (PSO) algorithm is used to optimize the number of hidden layer neurons, learning rate and training times of LSTM neural network. The attention mechanism is used to optimize the probability distribution in the process of training LSTM neural network. Finally, the final predicted value is obtained by adding the predicted results of each subsequence component. Example analysis shows that prediction evaluation indexes of the proposed model MAE, RMSE and R²are all the best by comparison with other algorithms, which can verify the superiority of the proposed model.

Key words: photovoltaic output prediction, complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN), long short-term memory (LSTM) neural network, particle swarm optimization (PSO), attention mechanism