基于深度长短时记忆网络的区域级超短期负荷预测方法

超短期负荷预测为实时电力市场运行提供重要依据,预测准确度的提升对于揭示负荷变化的不确定性以及日前预测偏差具有重要意义。基于电力系统中含有的丰富大数据资源,提出了一种针对区域级负荷的深度长短时记忆网络超短期预测方法,该方法包括输入数据的预处理、深度长短时记忆(long short-termmemory,LSTM)网络的构建以及模型的训练和超参数的寻找等步骤。其中采用随机搜索的方法寻找最优超参数,并在该超参数下选择泛化能力最优的模型,与前沿机器学习预测算法进行对比。实验结果证实,深度LSTM网络可以取得更好的预测效果,适合于离线训练实时预测。此外,通过对隐藏层激活向量的可视化展示和相关关系定量计算,首次直观展示了深度LSTM算法对负荷数据中含有的抽象特征提取情况,证实了深度LSTM具有对输入负荷数据特征学习以及长短期相关性挖掘的能力。

A Very Short-term Load Forecasting Method Based on Deep LSTM RNN at Zone Level

Very short-term load forecasting provides basis for real-time electricity market. Improvement of its forecasting accuracy is important for revealing the uncertainties of the load and the deviation of day-ahead load forecasting results. Based on big data resources in power system, a deep long short-term memory (LSTM) load forecasting method at small area level is proposed, including preprocessing of input data, construction of deep LSTM network, and search of hyper-parameters using random search method. Under such hyper-parameters, forecasting performance comparison between the proposed method and machine learning methods is conducted. Results prove that the proposed method can achieve better forecasting accuracy, suitable for offline training and online predicting. Also, through visualization of hidden layer activation units and calculation of correlation, ability of feature learning and correlation mining of deep LSTM network is proved.