부산대학교 도서관

Efficient and accurate short-term electric load forecasting plays a significant role in energy conservation and reducing carbon emissions. Recurrent neural networks (RNN) and their derived deep learning models have continuously improved the accuracy of short-term load predictions. However, traditional deep learning models, constrained by the one-dimensional structure of time series data, struggle to capture the relationships within and between periods. And when performing load forecasting tasks, these models tend to establish temporal relationships in the time dimension while overlooking the relationships between different feature variable dimensions. In order to address both, this paper proposes a Crossformer-based TimesNet-LSTM method for short-term electric load forecasting. The proposed approach takes historical load data as input and leverages the unique structure of TimesNet to convert the one-dimensional time series into a two-dimensional space for information extraction. The Crossformer model with double attention mechanisms is then employed to capture the relationships between sequences, time, and feature variables in different dimensions. Finally, the LSTM computes the output results. Experimental calculations on publicly available datasets from Australia and the United States demonstrate the superior performance of the proposed model compared to traditional single models and other hybrid models in short-term forecasting of multidimensional electricity load data. The Mean Absolute Percentage Error (MAPE) achieved on the Australian dataset is 0.52%, while on the U.S. dataset it is 0.53%. These outstanding results highlight the universality and robustness of the model. The proposed Crossformer-based TimesNet-LSTM method not only overcomes the limitations of traditional one-dimensional deep learning models but also enhances the accuracy of short-term electric load forecasting. Its application has significant implications for energy saving and carbon emission reduction.