基于VMD-Self Attention-LSTM的短期电力负荷预测

Short Term Power Load Forecasting Based on VMD Self Attention-LSTM

摘要: 针对传统神经网络对长期时序特征提取能力较低，在电力负荷预测中精确度不高，提出一种VMD-Self Attention-LSTM混合预测方法，首先通过变分模态分解将负荷序列分解多个模态分量，以降低负荷序列的非平稳性和强非线性；其次，利用注意力机制模块改进长短期记忆网络，提取高维重构权重特征，获得负荷序列长期依赖关系；最后，将变分模态分解和改进长短期记忆网络结合用于电力负荷预测；此外，选择ANN和VMD-LSTM模型为对照组，针对混合模型的精度进行对比，经验证，VMD-Self Attention-LSTM混合模型的均方误差较其他方法有所下降，在对验证集的未来七日预测验证中，其平均绝对误差较ANN、VMD-LSTM模型分别下降3.49%、1.25%，该模型的预测性能更优异。

Abstract: The traditional neural network has low ability to extract long-term time series features and low ac-curacy in power load forecasting. Therefore, a VMD-Self Attention-LSTM hybrid forecasting method is proposed. Firstly, the load series is decomposed into multiple modal components through varia-tional modal decomposition to reduce the non-stationary and strong nonlinearity of the load series; Secondly, the attention mechanism module is used to improve the short-term and short-term memory network, extract the high-dimensional reconstruction weight feature, and obtain the long-term dependence of load series; Finally, the combination of variational mode decomposition and improved short-term and short-term memory network is applied to power load forecasting; In addition, the ANN and VMD-LSTM methods were used as the control group to compare the accuracy of the mixed model. It was verified that the mean square error of the VMD-Self Attention-LSTM mixed model was lower than that of other methods. In the next seven days’ prediction verification of the validation set, its average absolute error was 3.49% and 1.25% lower than that of the ANN and VMD-LSTM models, respectively, and the prediction performance of the model was better.