Load Forecasting in Data-scarce Areas Based on Multi-source Transfer Learning and XGBoost

Accurate electric load forecasting, especially with small sample sets, is a critical topic in the field of power system. In this paper, we propose an attentive multi-source transfer learning framework based on extreme gradient boosting (XGBoost) for load forecasting in data-scarce scenarios. In the proposed framework, a new quantifiable framework is developed by exploiting the feature-output mapping relationship constructed by XGBoost. To effectively assess the transferability across domains, a similarity metric is introduced grounded in SHAP value distribution. Models of multiple sources that exhibit high similarity to the target domain are selected as the transfer models, and attentively integrated using the optimal combining coefficients, aiming to enhance overall transfer performance. Experiments conducted on real-world data sets illustrate that, compared with another popular transfer learning method, the proposed framework not only effectively identifies the potential source tasks that can produce positive transfer, but also achieves improved forecasting accuracy for residential electric load with limited data.