Improved Long Short-Term Memory Network Based Short Term Load Forecasting

The power load sequence is a kind of complex nonlinear time series. As an improved RNN, the long short-term memory (LSTM) network has been applied to short term load forecasting because it can capture the temporal dynamic of nonlinear time series. In order to further improve the precision of load forecasting results, we establish the CEEMD-AE-LSTM load forecasting model. Firstly, complementary ensemble empirical mode decomposition (CEEMD) is used to decompose the load sequence into several subsequences. Secondly, the approximate entropy (AE) of each subsequence is calculated separately, and the subsequences with similar entropy values are merged into new sequences. Finally, we establish the LSTM networks based on the new sequences, and the output result of each LSTM network is superposed as the final load forecasting result. The experimental results show that the CEEMD-AE-LSTM load forecasting model outperform single RNN and LSTM network by comparing the relative error and mean absolute percentage error. Therefore, the CEEMD-AE-LSTM load forecasting model effectively forecast the short term load and upgrade the precision of short term load forecasting.