Location-Aware Curtailment Restores Railway Voltage Compliance

Power limits on AC railway feeding sections are often applied as blunt section-wide caps: every train in the protected area is throttled when the electrical constraint binds. That is safe, but it can waste timetable capacity when the actual voltage problem is local to a small set of trains. This paper tests a narrower alternative for a degraded 25 kV corridor: identify which train is causing the far-field voltage constraint to bind, cap that train at the relevant time, and leave the others alone.

Core Contribution

The paper’s main claim is not that aggregate demand can be reduced. In the reported case, the aggregate one-minute RMS peak is unchanged at 8.98 MW under every policy. The contribution is a location-aware demand-shaping rule that treats voltage feasibility as a spatial constraint rather than a whole-section power target. On the representative GB corridor used in the study, that distinction matters: the static section cap restores compliance by adding 210 seconds of total journey-time penalty, while the online controller restores the same solver-validated feasibility with 7 seconds by advising one high-speed train.

The result is useful because constrained rail electrification is often managed under degraded feeding, where a single-ended supply can make remote voltage the binding condition. If the binding train can be identified online, operators get a way to recover electrical headroom without paying the delay cost of throttling traffic that is not part of the constraint.

Technical Approach

The controller evaluates candidate controls as one cap level per train. To keep the search near real time, the authors precompute and cache single-train runs: power and position over time for each train and cap level. A candidate plan is then scored by assembling the aggregate demand on a shared time base, evaluating an analytic voltage estimate at the instants with the largest power-by-distance loading, and computing three outputs: infeasible seconds, summed journey-time penalty, and peak one-minute RMS demand.

The paper is careful about the role of the estimate. It forms the online control decision, but it is not the evidence used to claim feasibility. Every committed plan for the uncontrolled case, static cap, offline genetic algorithm, and online controller is rescored with a full multi-conductor AC power-flow solver at one-second resolution. The controller also screens against $U_{min} + \Delta$ with $\Delta = 1.5$ kV because the analytic estimate is optimistic near the binding instants.

Figure 4 is the central evidence: the uncontrolled corridor falls below the 17.5 kV compliance line, while the static cap, offline GA, and online controller all stay above it under the full solver.

Results and Analysis

The baseline comparison is unusually direct. The uncontrolled case has 57 infeasible seconds and reaches a minimum section voltage of 16.35 kV. The incumbent static limit eliminates infeasibility and raises the minimum voltage to 19.26 kV, but at a cost of 210 seconds of aggregate journey time because it caps multiple trains in the section. The offline GA also eliminates infeasibility, reaches 19.01 kV, and costs 7 seconds, but needs 64.9 seconds of computation. The online controller returns the same committed plan as the GA: 0 infeasible seconds, 19.01 kV minimum voltage, 7 seconds of delay, and 1.45 seconds of runtime.

That is the paper’s strongest quantitative point. The online method matches the offline optimum on this case while cutting the incumbent delay cost by 30×. The caveat is that the offline GA is a comparator for this scenario, not a proof that the online policy will always equal the optimum. The authors acknowledge that the controller is sequential and non-anticipative: it commits a train’s cap at band entry using information available then, so a later-binding train cannot force a previous cap to be revised.

Figure 5 explains why the result is not visible in the aggregate peak. The peak one-minute RMS demand is 8.98 MW for all policies; the active constraint is the voltage drop at a far-field location, not the maximum whole-line load.

The estimate-versus-solver table is also important. For the uncontrolled plan, the estimate gives 16.57 kV while the solver gives 16.35 kV, so it is only 0.2 kV optimistic at the binding configuration. For capped plans, the estimate is about 1.5 kV higher than the solver: static limit 20.71 kV estimated versus 19.26 kV solved, and GA/controller 20.55 kV estimated versus 19.01 kV solved. The margin absorbs that optimism in this case, but it is a design assumption that would need route-specific validation before deployment.

The evidence supports the narrow claim: on one solver-selected, restorable outage-feeding scenario, selective curtailment restores compliance with far less delay than a static cap. It does not yet establish performance across traffic levels, binding locations, or a real deployment with measured train electrical states.