Short-term load forecasting based on hybrid strategy using warm-start gradient tree boosting

Abstract

A deep-learning-based hybrid strategy for short-term load forecasting is presented. The strategy proposes a novel tree-based ensemble method warm-start gradient tree boosting (WGTB). Current strategies either ensemble submodels of a single type, which fail to take advantage of the statistical strengths of different inference models, or simply sum the outputs from completely different inference models, which does not maximize the potential of the ensemble. Inspired by the bias-variance trade-off, WGTB is proposed and tailored to the great disparity among different inference models on accuracy, volatility, and linearity. The complete strategy integrates four different inference models of different capacities. WGTB, then, ensembles their outputs by a warm-start and a hybrid of bagging and boosting, which lowers bias and variance concurrently. It is validated on two real datasets from the State Grid Corporation of China of hourly resolution. The result demonstrates the effectiveness of the proposed strategy that hybridizes the statistical strengths of both low-bias and low-variance inference models.