Seasonality-Aware Dual-Memory Pricing for Electric Vehicle Charging Stations under Non-Stationary Demand

Abstract Electric-vehicle (EV) charging prices are selected from partial price-response feedback while demand combines recurrent calendar patterns with structural changes. The central learning challenge is to retain reusable seasonal information while adapting to recent market conditions. We develop and evaluate seasonality-aware dual-memory pricing (SA-DMP), which combines long- and short-horizon seasonal demand forecasts through context-specific weights updated from previously observed prediction losses. In six controlled 100-day demand environments with 20 paired seeds per scenario, SA-DMP achieves lower mean dynamic pseudo-regret than an architecture-matched fixed-weight dual-memory policy in all six environments. Its macro-average oracle-normalized reduction is 0.441 percentage points (95% seed-cluster bootstrap interval: 0.315–0.564 percentage points). A separate exploratory, post-development calibration study using disjoint test seeds yields a 0.463-percentage-point macro-average reduction relative to a validation-selected fixed blend. An eight-connector operational simulation complements the learning comparison by characterizing multi-hour occupancy, abandonment, and time-of-use procurement in the evaluation of contribution margin. At the same time, absolute contribution margin remains below a validation-selected time-of-use tariff in the tested operational environments, delineating the deployment boundary of the adaptive gate. The findings identify the incremental value of loss-adaptive seasonal memory weighting for non-stationary EV charging-price learning under bandit feedback.