Photovoltaic Power Prediction Based On CEEMDAN Decomposition With Multi-Objective Optimization LSTM

Fossil energy is gradually unable to meet human needs due to its non-renewability an d environmental pollution problems.In contrast,solar energy as a clean,non-polluting,safe and economical renewable energy source,has been widely used.However,under the influence of complex changes in the natural environment,the output power of photovoltaic power generation often shows the characteristics of randomness and volatility,and when it is integrated into the grid on a large scale,it is easy to produce impact on the grid and threaten the stable operation of the grid.Therefore,accurate prediction of PV power generation is of great significance to the large-scale development of PV industry and the safe and stable operation of the power grid.In this paper,after an in-depth analysis of the principle of PV power generation,the influencin g factors of PV power,the advantages,disadvantages and characteristics of different prediction models,a PV power prediction method based on CEEMDAN decomposition and multi-objective arithmetic optimization algorithm(MAOA)is proposed to improve the prediction accuracy and prediction stability of PV power by combining the adaptive noise-complete ensemble empirical modal decomposition(CEEMDAN),long short-term memory network(LSTM)and multi-objective arithmetic optimization algorithm(MAOA)in a complementary way.Firstly,CEEMDAN is used to decompose the historical PV sequences to eliminate noise and extract the main features of the original data,and to measure the complexity of each component subsequence obtained from the decomposition using the alignment entropy,and the sequences with similar alignment entropy values are reconstructed to obtain new subsequences with different frequencies to reduce the complexity of the subsequent prediction calculation.Secondly,to address the problem that the cmanual selection of LSTM hyperparameters is too limited,the root mean square error(RMSE)and standard deviation error(SDE)between prediction results and actual power are used as two fitness functions for optimization,and the archival-based multi-objective arithmetic optimization algorithm(MAO A)is used to find the number of hidden layer neurons and learning rate of LSTM in order to improve the prediction performance of the model.Finally,to verify the generalizability of the hybrid model in this paper,a subset of data from different months is constructed for the ablation study and compared with three different hybrid models(CEEMDAN-PE-MOPSO-LSTM,CEEMDAN-PE-MOWOA-LSTM,and EEMD-PE-MAOA-LSTM),and the prediction results show that the hybrid model in this paper outperforms other prediction models in all evaluation indexes in different months and can achieve both accuracy and stability of prediction.