AI Summary & Key Takeaways
AI InsightWithin these counterfactual scenarios, most of the reduction therefore come from the shared congestion-triggered routing architecture rather than from the forecast-risk terms.
📌 3 Key Takeaways
Abstract
Short-term congestion at public electric-vehicle charging stations depends on both near-future demand and how vehicles are distributed across the network. We forecast cumulative station-level charging initiations and use the forecasts in congestion-triggered station assignment. The analysis reuses 441077 publicly released charging transactions from 13 stations in Jiaxing, China, aggregated at 15-minute resolution. Daily and weekly seasonal baselines are compared with Poisson, Tweedie, hurdle, and quantile LightGBM models for cumulative 15-, 30-, and 60-minute demand. Because physical arrivals and queueing times are not observed, recorded charging starts served as simulation-entry events in the counterfactual analysis. The forecasts are embedded in a counterfactual capacity-and-routing simulation with transaction-level service durations, inbound reservations for redirected vehicles, partial recommendation acceptance, an eight-minute switching-time penalty, queue abandonment, demand growth, and charger outages. All routing policies are evaluated on immutable scenarios with common random numbers over 21 test days stratified by weekday and charging-initiation volume. Relative to the stronger daily seasonal baseline, the Tweedie model reduces mean absolute error by 17.6-19.8% and root-mean-square error by 21.7-26.8%. Under normal demand and full capacity, the routing policies remain inactive and introduce no unnecessary switching. Under the evaluated stress scenarios and full recommendation acceptance, dynamic assignment reduces simulated mean total delay by 75.0-78.7% and simulated abandonment by approximately 95-100%, while rerouting fewer than 10% of vehicles. Risk-aware routing has the numerically lowest aggregate simulated delay, but under the combined demand-and-outage scenario it differs from current-load routing by only 0.001 minutes per vehicle, and the pairs 95% confidence interval included zero. Within these counterfactual scenarios, most of the reduction therefore come from the shared congestion-triggered routing architecture rather than from the forecast-risk terms.
Özet
Short-term congestion at public electric-vehicle charging stations depends on both near-future demand and how vehicles are distributed across the network. We forecast cumulative station-level charging initiations and use the forecasts in congestion-triggered station assignment. The analysis reuses 441077 publicly released charging transactions from 13 stations in Jiaxing, China, aggregated at 15-minute resolution. Daily and weekly seasonal baselines are compared with Poisson, Tweedie, hurdle, and quantile LightGBM models for cumulative 15-, 30-, and 60-minute demand. Because physical arrivals and queueing times are not observed, recorded charging starts served as simulation-entry events in the counterfactual analysis. The forecasts are embedded in a counterfactual capacity-and-routing simulation with transaction-level service durations, inbound reservations for redirected vehicles, partial recommendation acceptance, an eight-minute switching-time penalty, queue abandonment, demand growth, and charger outages. All routing policies are evaluated on immutable scenarios with common random numbers over 21 test days stratified by weekday and charging-initiation volume. Relative to the stronger daily seasonal baseline, the Tweedie model reduces mean absolute error by 17.6-19.8% and root-mean-square error by 21.7-26.8%. Under normal demand and full capacity, the routing policies remain inactive and introduce no unnecessary switching. Under the evaluated stress scenarios and full recommendation acceptance, dynamic assignment reduces simulated mean total delay by 75.0-78.7% and simulated abandonment by approximately 95-100%, while rerouting fewer than 10% of vehicles. Risk-aware routing has the numerically lowest aggregate simulated delay, but under the combined demand-and-outage scenario it differs from current-load routing by only 0.001 minutes per vehicle, and the pairs 95% confidence interval included zero. Within these counterfactual scenarios, most of the reduction therefore come from the shared congestion-triggered routing architecture rather than from the forecast-risk terms.