Short-term photovoltaic power prediction based on CEEMDAN-PE and BiLSTM neural network

The volatility and uncertainty associated with photovoltaic (PV) energy production impose considerable challenges to the reliable operation of power grid systems. In order to address this challenge, it is necessary to obtain accurate forecasts of the output. In this paper, a hybrid model is proposed, which incorporates complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN), permutation entropy (PE) and bidirectional long short-term memory (BiLSTM) networks. Firstly, CEEMDAN is utilized to decompose PV power series into multiple intrinsic mode functions (IMFs) to reduce non-stationary and volatility impacts on prediction. Then PE is used to reconstruct the decomposed IMFs into new simplified sequences. This approach reduces computation complexity while effectively retaining fluctuation characteristics of original signals. Secondly, the minimum meteorological factors that have a great impact on PV power are identified through Pearson correlation analysis. Subsequently, a BiLSTM model is built to predict each reconstructed new sequence, final results are obtained by superimposing the reconstructed sequences, which exploits their bidirectional spatiotemporal correlations. Finally, model performance is evaluated with four evaluation metrics, outlier tests and Friedman tests. Results demonstrate that under different weather conditions, the CEEMDAN-PE-BiLSTM hybrid model exhibits higher accuracy, better generality, and stronger robustness compared to other similar models.

© 2017 Elsevier Inc. All rights reserved.