A hybrid short-term load forecasting model based on variational mode decomposition and long short-term memory networks considering relevant factors with Bayesian optimization algorithm

• Published in: Applied energy Vol. 237; pp. 103 - 116
• Main Authors: He, Feifei, Zhou, Jianzhong, Feng, Zhong-kai, Liu, Guangbiao, Yang, Yuqi
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.03.2019
• Subjects: algorithms; Bayesian optimization algorithm; Bayesian theory; China; energy; Long short-term memory network; lymphocytes; prediction; processing technology; regression analysis; Relevant factors; Short-term load forecasting; temperature; Variational mode decomposition; China; Long short-term memory network; Variational mode decomposition; Relevant factors; Short-term load forecasting; Bayesian optimization algorithm
• ISSN: 0306-2619; 1872-9118
• DOI: 10.1016/j.apenergy.2019.01.055

•A hybrid load forecasting model with hyper-parameters optimization is proposed.•Nonlinear mapping is introduced to map the relevant factors.•Bayesian Optimization Algorithm (BOA) is used in hyperparameter optimization.•Proposed model periodically moves the data window which has high practicability.•Proposed method is validated by seven contrast methods in Keras python framework. Short-term load forecasting plays an essential role in the safe and stable operation of power systems and has always been a vital research issue of energy management. In this research, a hybrid short-load forecasting method with Variational Mode Decomposition (VMD) and Long Short-Term Memory (LSTM) networks considering relevant factors which optimized by the Bayesian Optimization Algorithm (BOA) is studied. This method firstly decomposition with VMD which is a non-recursive signal processing technology that can decompose a signal into a discrete number of modes, then, consider the relevant factors and extend to the sequence according to the coefficient of association. Specifically, for the day type and higher or lower temperature, the nonlinear mapping is used and optimized by the BOA. Finally, the subsequences are predicted by LSTM which is a special Recurrent Neural Network with memory cells and reconstructed. To validate the performance of the proposed method, two categories of contrast methods including individual methods and decomposition-based methods are demonstrated in this study. The individual methods which without decomposition processes including LSTM, Support Vector Regression, Multi-Layered Perceptron Regressor, Linear Regression, and Random Forest Regressor, and the decomposition based methods including Empirical Mode Decomposition-Long Short-Term Memory, and Ensemble Empirical Mode Decomposition-Long Short-Term Memory. The simulation results, which developed in four periods of Hubei Province, China, show that the prediction accuracy of the proposed model is significantly improved compared with the contrast methods.