Cross-Domain Reformer-XL Model for Battery-Swapping Demand Forecasting in Electric Vehicles

Battery-swapping networks for electric vehicles need short-term, station-level demand forecasts that are still reliable after transferring a model from a mature service area to a new city, where historical reservations are sparse and station behavior changes rapidly. This paper introduces a cross-domain Reformer-XL model that integrates locality-sensitive hashing attention, segment-level recurrence, domain-conditioned normalization, and a station-context gate for multi-horizon battery-swapping demand forecasting. The model is tested on a constructed engineering dataset that contains 18.7 million swapping orders from 214 stations across four urban areas, and meteorological, calendar, traffic, tariff and station-inventory covariates have been aligned at 15-minute intervals. Compared with LSTM, temporal convolution, standard Transformer, Informer and non-transfer Reformer baselines, the proposed model reduces the average MAE to 10.96 swaps per station-hour and lowers the weighted RMSE by 11.7% on the target-domain test set. With only 14 days of target data in cold-start transfer, the model still reached 92.3% of the full-data accuracy and increased peak-hour recall from 0.781 to 0.846. Ablation experiments show that recurrent memory provides a relatively large improvement for a long horizon, and domain-conditioned normalization mainly stabilizes the weekend and rainfall scenarios. As shown in the above results, sparse recurrent attention can support scalable and transferable demand forecasting for battery-swapping infrastructure planning and real-time station operation.