Short-term Load Forecasting Based on Temporal Convolutional Network and Transfer Learning

Short-term load forecasting is crucial for optimizing energy supply and demand balance, which improves the efficiency of power system operation. In regions where historical load data are scarce, traditional machine learning prediction methods face great challenges. To address the short-term load forecasting issue in the context of sparse data, we propose a hybrid prediction model based on temporal convolutional network (TCN) and transfer learning. Firstly, a source domain selection method combining Mahalanobis distance and maximum mean discrepancy is introduced to measure the transferability of the source domain. This method aims to identify a source domain that is highly similar to the target domain in terms of data distribution and feature differences. Secondly, the snow ablation optimizer (SAO) is utilized to optimize the hyperparameters of the TCN model, thereby establishing an SAO-TCN transfer model to enhance overall prediction performance. Finally, the proposed hybrid model is evaluated and validated using real-world datasets. Experimental results demonstrate that in comparison to traditional machine learning methods, the hybrid model is capable of expanding the range of source domain selection and enhancing its accuracy as well, and it reduces the root mean square error by at least 15%.