Interpretable load forecasting via representation learning of geo-distributed meteorological factors

Meteorological factors (MF) are crucial in day-ahead system-level load forecasting as they significantly influence electricity consumption behaviors. Numerous studies have incorporated MF into load forecasting models to achieve higher accuracy. Selecting MF from a representative location or using averaged MF as inputs is a common practice. However, significant differences in MF across various locations within a region pose challenges in selecting the appropriate MF. This paper proposes a representation learning framework to extract geo-distributed MF while considering their spatial relationships. Additionally, it employs the Shapley value in a graph-based model to reveal connections between MF collected from different locations and loads. To reduce the computational complexity of calculating the Shapley value, an acceleration method based on Monte Carlo sampling and weighted linear regression is adopted. Experiments on two real-world datasets demonstrate that the proposed method improves day-ahead forecasting accuracy, particularly in extreme scenarios like the “accumulation temperature effect” in summer and “sudden temperature change” in winter. We also find a significant correlation between the importance of MF in different locations and the corresponding area’s GDP and mainstay industry. Practical recommendations are offered to operators for weather stations construction and feature selection, especially in extreme weather events. • Proposes a graph-based framework to consider the spatial relationships of weather features. • Provides an accelerated Shapley value-based graph-based model interpretation algorithm. • Experiments demonstrate improved forecasting accuracy, particularly in extreme scenarios. • Reveal correlations between geo-distributed MF importance and economic, such as GDP.