A novel short-term load forecasting approach for data-poor areas based on K-MIFS-XGBoost and transfer-learning

Sufficient historical power load data is crucial for establishing accurate load forecasting model. However, in newly developed areas or areas with relatively backward power metering infrastructure, historical data are often very limited. With a focus on this problem, a transfer forecasting framework for short-term loads is proposed in this paper. To improve the fitting ability of the forecasting model, a modified K-means method combined with mutual information feature selection (MIFS) algorithm is proposed to extract the hidden features of different seasonal loads. Then with corresponding load data (source area), a Bayesian hyperparameter optimization approach is proposed, to establish the optimal extreme gradient boosting (XGBoost) load forecasting model. To solve the problem of limited historical load data (target area), we transfer the source area model to the target area by fine-tuning model parameters and weighting holiday samples. For demonstration, two case studies, with power load datasets in two regions of China and GEFcom2012 datasets, are performed. The results of comparative experiments show that the forecasting error can be reduced by a large margin with the knowledge learned from other areas.