Application of hybrid model based on empirical mode decomposition, novel recurrent neural networks and the ARIMA to wind speed prediction

Ming-De Liu,Lin Ding,Yu-Long Bai
DOI: https://doi.org/10.1016/j.enconman.2021.113917
IF: 10.4
2021-04-01
Energy Conversion and Management
Abstract:<p>Wind speed is the key factor of wind power generation. With the increase of the proportion of wind power generation in total power generation, the accurate prediction of wind speeds plays an important role in the stable operations of power grids. However, the strong randomness of wind speeds makes it difficult to accurately predict wind speeds. Thus, a wind speed prediction model combining empirical mode decomposition (EMD) with some novel recurrent neural networks (RNN) and the autoregressive integrated moving average (ARIMA) is proposed to solve the problem. The selected RNNs are long short-term memory network (LSTM) and the gated recurrent unit (GRU) network. In this model, EMD is used to decompose the wind speed sequence to reduce the complexity and non-stationary of the series. The entropy of the samples of the sub-sequences after decomposition is calculated. Consequently, LSTM is applied to predict the high frequency sub-sequences with large entropy while the ARIMA is employed to predict the remaining low frequency sub-sequences and one residual. Finally, the prediction results of each sub series are combined to obtain the final prediction results. To verify the accuracy and stability of the model, four wind speed data sets form Inner Mongolia, China, are used to test the proposed methods. Five models are established in four practical cases and their performances are compared with the performances of the proposed model. The results in this paper show the following: (1) the EMD method can improve the wind speed prediction performance when it is combined with LSTM; (2) after decomposition, LSTM is suitable for predicting high complexity subsequences and the ARIMA is suitable for effectively predicting low complexity subsequences based on the different sample entropies; and (3) the root mean squared errors (RMSEs) of the hybrid model on the four wind speed data sets are 0.4163, 0.2085, 0.1613, and 0.2790, respectively, which are basically lower than those of the five models compared. Therefore, it is feasible to apply the hybrid model to wind speed prediction.</p>
energy & fuels,thermodynamics,mechanics
What problem does this paper attempt to address?