基于LSTM组合模型的短期电力负荷预测

Short-Term Power Load Forecasting Based on LSTM Combination Model

针对电力负荷时序难以进行精度预测的难题,提出了一种基于自适应白噪声的完备集合经验模态分解(CEEMDAN)的TCN-LSTM短期电力负荷组合预测方法.首先使用CEEMDAN分解方法将原始负荷序列进行分解,该方法与集合经验模态分解(EEMD)方法相比,能使序列分解更加完备,且具有更小的重构误差.然后为了降低非平稳序列对预测精度的影响,通过平稳性检验将分解后的序列按照平稳性质分类,将非平稳序列合并后输入LSTM网络预测,平稳序列则计算排列熵后重组成高排列熵的平稳序列和低排列熵的平稳序列,再分别输入到LSTM网络和TCN网络中进行预测,最后对预测结果进行叠加得到最终的预测结果.实证结果表明:通过按照平稳性分类和计算排列熵的方式来对CEEMDAN分解后的序列进行重新组合的方法,不仅提高了模型的运算效率,同时比其他预测方法具有更高的预测精度.

Considering the difficulty of accurate prediction of power load time series,this research proposes a TCN-LSTM short-term power load combination forecasting method based on complete ensemble empirical mode decomposition with adaptive noise(CEEMDAN).First,the original load sequence is decomposed using CEEMDAN decomposition method.Compared with ensemble empirical mode decomposition(EEMD)method,this method can make the sequence decomposition more complete and has less reconstruction error.Then,in or-der to reduce the influence of non-stationary sequences on the prediction accuracy,the decomposed sequences are classified according to their stationary properties through stationarity test,and the non-stationary sequences are combined and input into the LSTM network for prediction.The stationary sequences are then reconstructed into stationary sequences with high and low permutation entropy by calculating the permutation entropy.Then they are input into LSTM network and TCN network respectively for prediction.Finally the prediction results are superim-posed to get the final prediction results.The empirical results show that the method of recombining the decom-posed sequences of CEEMDAN based on stationality classification and calculation of permutation entropy not only improves the efficiency of the model,but also has higher prediction accuracy than other prediction methods.