Electric Power Load Forecasting Method Based on Contrastive Learning and Attention Mechanism

Currently, as the proportion of new energy sources continues to increase, a large amount of uncertainty and randomness leads to a decline in the flexibility of the power grid companies' adjustment capabilities. Improving the accuracy of load forecasting can help power companies optimize generation planning, reduce the need for reserve capacity and lower operational costs. We propose a load forecasting method based on contrastive learning and attention mechanisms that can reduce load forecasting errors. Firstly, we proposes a contrastive learning method that constructs a pre-task to pretrain the Transformer encoder, which enables it to extract similarity features. Thereby it can reduce the redundancy of data and enhancing the correlation of samples. Secondly, we introduce a BVA(Bi-view attention module) that can perform weighted processing on feature vectors, making the information aggregated into the feature vectors more effective. Numerical case studies are conducted on real data sets, our proposed method can effectively reduce the error in load forecasting.