Explainable spatiotemporal multi-task learning for electric vehicle charging demand prediction

Abstract

This paper introduces an explainable multi-task learning framework designed to accurately predict zonal-level, multi-dimensional charging demand characteristics for electric vehicles (EVs), including the occupancy of charging piles, charging volumes, and charging durations. The proposed framework is structured into two interconnected phases. In the prediction phase, the study develops a temporal GraphSAGE model adept at capturing spatiotemporal nuances. This model is seamlessly integrated within the multi-task learning framework, which encompasses multiple prediction tasks – occupancy, volume, and duration – and promotes sharing of data representations across related tasks to enhance domain knowledge transfer. During the interpretation phase, the ”mask-compute-analyze” technique is employed to assess the significance of model components by nullifying corresponding inputs and evaluating their performance impacts using Shapley values. Building upon this, the approach incorporates small-world network theory, significantly reducing the computational complexities associated with the interpretability of spatial inputs across large transportation networks. Additionally, the framework adopts a dual analysis strategy, conducting both extra and intra-analysis, to comprehensively investigate extensive network effects as well as localized phenomena. The proposed method is validated through a realistic case study in Shenzhen, China, using real-world data from charging stations. We demonstrate that the multi-task learning framework not only improves the MAPE of occupancy prediction by 25.87% but also enhances the performance of volume prediction by 8.15% and duration prediction by 26.10%, compared to learning each task individually. In terms of interpretability, our analysis reveals that feature interactions during the model training process significantly boost predictive accuracy in the multi-task learning framework, while during the implementation phase, the prediction performance primarily depends on feature data directly related to the specific task. Additionally, we find that the absence of data from surrounding nodes had a negligible impact on individual nodes, attesting to the superiority and resilience of the proposed prediction framework.