An Explainable KAN--Shapley Framework for Many-to-Many Electric Vehicle Charging Dispatch in Intelligent Transportation Systems

Mobile Charging Services (MCS) can alleviate the shortage of fixed charging infrastructure, but real-world dispatching in dense cities must jointly handle time-varying traffic and grid-price signals, hard feasibility constraints such as SOC safety, service time windows, and service regions, and the need for auditable decisions. Existing optimization-based schedulers rely on hand-crafted costs and provide limited decision transparency, whereas end-to-end learning-based dispatchers are often opaque and may violate constraints at inference time. This paper proposes an Explainable KAN–Shapley Dispatch Network (XKSDN) for many-to-many dispatch between EV charging request clusters and multiple Mobile Charging Vehicles (MCVs). XKSDN first abstracts massive requests via spatiotemporal–energy clustering, and then learns a structured utility function with a Kolmogorov–Arnold Network (KAN), whose edge-wise one-dimensional spline functions expose nonlinear effects such as distance decay and SOC-urgency transitions. To turn learned utilities into executable schedules, a constraint-aware matching layer integrates feasibility masking (Big-$M$), virtual nodes for partial matching, and node replication for limited MCV capacity, and solves the resulting assignment with an improved Hungarian algorithm. For accountability, XKSDN computes local Shapley-value attributions within clusters, quantifying the marginal contributions of distance, SOC, congestion, and price to each dispatch command. Experiments on New York City trip data with SUMO simulations show that XKSDN improves service success rate and response delay while reducing operational cost against representative optimization and deep learning baselines, and provides consistent, auditable explanations for safety-critical deployments.