基于VMD-LSTM-Attention模型的短期负荷预测研究

Research on short⁃term load forecasting based on VMD⁃LSTM⁃Attention model

针对电力负荷数据非线性强以及影响因素多等问题,提出一种基于变分模态分解(VMD)和注意力机制的长短期记忆(LSTM)网络的组合预测方法。首先通过VMD对历史负荷数据进行频率分解,分解成不同频率的分量,并在考虑不同影响因素(如温度、湿度等)的条件下,通过LSTM网络对不同模态分别进行预测,再通过自注意力机制进一步挖掘数据内部的相关性,最后将预测的不同模态分量重构相加作为预测结果。实验采用中国某地区实际用电负荷数据,经过与不同基准模型的对比和分析,验证了提出的模型提高了预测精度,具有一定的应用潜力。

In view of the strong nonlinearity of power load data and too many influencing factors,a long short⁃term memory(LSTM)network combination forecasting method based on variational mode decomposition(VMD)and attention mechanism is proposed.The VMD is used to decompose the historical load data into components with different frequencies.Under the condition of considering different influencing factors(temperature,humidity,etc.),the LSTM network is used to predict different modes respectively.And then,the self⁃attention mechanism is used to further explore the internal correlation of data.The predicted components of different modes are reconstructed and added as the prediction results.The actual power load data of a certain region in China is used in the experiment.After comparison with models with different benchmarks,it shows that the proposed model can improve the prediction accuracy and has a certain application potential.