Predictive Constraint-Verified EV Routing: A Machine-Learning-Guided Weighted Partial MaxSAT Framework for On-Demand Charging-Stop Selection on Long-Distance Indian Routes

Range anxiety among electric vehicle (EV) drivers stems less from battery capacity itself than from the difficulty of trusting a dashboard range estimate and choosing, quickly, among charging options of unequal speed, price and compatibility. The authors' earlier work encoded this choice as a Boolean Satisfiability problem solved once, offline, using a fixed energy-consumption rate on short intra-city routes. This paper extends that formulation to Predictive Constraint-Verified EV Routing (PC-EVR), a system that (i) predicts a vehicle's achievable range from live telemetry using a trained XGBoost regressor rather than a fixed rate, and (ii) re-solves the next charging decision at every real stop as a Weighted Partial MaxSAT instance, using the RC2 core-guided solver, in which a 20% state-of-charge floor is a non-negotiable constraint that filters candidates outright rather than a factor the solver merely weighs in. The range predictor is selected after a controlled, identical-split comparison against Linear Regression, Random Forest and a neural network, and its selection is validated by a dedicated ablation study rather than an offline error metric alone. PC-EVR is benchmarked against two purpose-built greedy baselines (nearest-station, cheapest-station) under identical safety and traffic assumptions, across five real long-distance Indian trips and India's national government charging-station registry of 39,641 stations. The proposed system completes every trip without a single safety-constraint violation while reducing average trip time by approximately 8.7 hours against the nearest-station baseline and 21.6 hours against the cheapest-station baseline. The ablation study further shows that routing quality is insensitive to small prediction-accuracy differences among tree- and network-based models but degrades measurably once accuracy falls below a threshold, a finding with direct implications for range-model selection in future systems.