Boosted Multi-Task Learning for Inter-District Collaborative Load Forecasting

Abstract

This paper proposes a boosted multi-task learning framework for inter-district collaborative load forecasting. The proposed framework involves two subsequent stages: in the first stage, districts would collaborate under a seamlessly-integrated federated learning scheme to capture the global load pattern; in the second stage, districts would withdraw and perform local training to capture the local load patterns. The probabilistic Gradient-Boosted Regression Tree (GBRT) is applied as the bottom-level machine learning algorithm, which would allow for an easy and intuitive embodiment of the generalized multi-task learning framework. We further propose two candidate district withdrawal mechanisms to connect the two stages: the simultaneous withdrawal, which prioritizes prediction accuracy, and the dynamic withdrawal, which prioritizes training efficiency and district incentivization. The follow-up performance analyses and the case study on 11 districts of the Zhuhai city confirm the superiority of the proposed framework and district withdrawal mechanisms.