基于注意力机制的CNN-BiLSTM日前电价预测

CNN-BiLSTM Day-ahead Electricity Price Prediction Based on Attention Mechanism

日前电价预测结果的准确性对于存在多元化竞争格局的电力市场具有重要意义,因此提出一种基于注意力机制的CNN-BiLSTM日前电价预测模型。该模型考虑诸多因素对日前电价预测的影响,采用皮尔逊系数进行相关性分析,得到各因素对日前电价预测的影响;利用卷积神经网络提取历史电价序列中的特征;将提取的特征信息输入双向长短期记忆网络,充分挖掘特性的变化规律进行训练;然后引入注意力机制来突出重要信息的影响并赋予权重;最后在全连接层通过激活函数加权求和计算出最终预测值。通过实例验证了该模型的准确性,其中RMSE、MAPE、MAE分别减少了33.07%、28.39%、27.08%。

The accuracy of the day-ahead electricity price prediction results is of great importance to the electricity market with diversified competition situation.In this paper,we propose a CNN-BiLSTM day-ahead electricity price prediction model based on the attention mechanism.The model considers the influence of many factors on day-ahead price prediction,adopts Pearson coefficient for correlation analysis to obtain the influence of each factor,uses convolutional neural network to extract the features in the sequence of historical electricity price,inputs the extracted feature information into bi-directional long-and short-term memory network,and fully utilizes the change rule of the features for training.The attention mechanism is then introduced to highlight the influence of important information and assigns the weights,and the final prediction value is calculated by weighted sum of activation functions in the fully connected layer.The accuracy of the proposed model is verified by examples,in which the RMSE,MAPE and MAE are reduced by 33.07%,28.39%and 27.08%,respectively.