基于改进迁移学习和多尺度CNN-BiLSTM-Attention的短期电力负荷预测方法

Short-term power load forecasting method based on improved hierarchical transfer learning and multi-scale CNN-BiLSTM-Attention

针对目标域负荷数据样本不足导致模型训练不充分从而预测精度不高的问题,提出一种改进的分层级迁移学习策略结合多尺度CNN-BiLSTM-Attention模型的短期电力负荷预测方法。设计串并行相叠加的多尺度CNN作为特征提取器,把提取到的特征作为输入传递到两层Bi LSTM结构进行进一步的学习,引入注意力机制调节捕获的信息向量权重。按照基础模型的结构划分层级,将源域数据按照拟合优度值的高低输入至模型进行分层级的迁移学习训练,保留每一层最优的训练权重,之后使用目标域数据对模型进行微调得到最终的预测模型。经实验证明,所提出的多尺度CNN-BiLSTM-Attention模型能有效提高负荷预测精度,在负荷数据样本不足时,改进的分层迁移学习策略相比于直接迁移学习能有效降低预测误差。以目标域6个月数据为例,MAPE指标降低13.31%,MAE指标降低15.16%,RMSE指标降低14.37%。

Insufficient power load data samples in the target domain result in inadequate model training and low prediction accuracy. Thus an improved hierarchical transfer learning strategy combined with a multi-scale CNN-BiLSTM-Attention model is proposed for short-term power load forecasting. A multi-scale CNN superimposed, linked and in parallel is designed as a feature extractor, and the features are then passed as input to two BiLSTM structures for further learning. Then an attention mechanism is introduced to adjust the weight of the captured information vector. This paper divides the layers according to the structure of the basic model, and inputs the source data into the model according to the level of the goodness-of-fit value to perform hierarchical transfer learning training. It then retains the optimal training weight of each layer, and uses the target domain data to carry out model training and obtain the final predictive model after fine-tuning. Experiments show that the proposed multi-scale CNN-BiLSTM-Attention model can effectively improve the accuracy of load prediction. When the load data samples are insufficient, the improved hierarchical transfer learning strategy can effectively reduce the prediction error compared with traditional transfer learning. Taking six months of data in the target domain as an example, compared with traditional transfer learning, the MAPE is reduced by 13.31%, MAE is reduced by 15.16%, and RMSE is reduced by 14.37%.