Mechanism-dependent antagonism of auxiliary information in substation-level load disaggregation for distribution network planning

Open-source energy system models disaggregate zonal electricity demand to substations, and modellers routinely layer auxiliary information onto that allocation, assuming more information cannot hurt. The pipeline gives no warning when it does. We fix the Voronoi assignment and test 22 configurations crossing uniform, Grid Point Model (GPM), and graph neural network (GNN) weighting with post-correction, prior-loss, and feature-fusion integration on 1891 British, 143 Australian, and 13 German substations, with Germany descriptive. At unit correction intensity (β=1), Nighttime Light and Proximity cut RMSE by 37–43% on the tested uniform and GPM bases, raise the British GNN value by 21%, and lower the Australian one by 30%. Absolute accuracy is study-case- and estimand-dependent: pooled predictive R2 is positive for selected British configurations, whereas region-mean predictive R2 is negative for all 22 configurations and for the pipeline-matched grid reference in both main study cases. Under geometric re-normalization and log-MSE, correction helps exactly when factor–residual alignment exceeds half the factor’s relative dispersion. In the implemented pipeline it classifies original-scale RMSE direction in 391/435 configurations (89.9%, against a 74.7% majority-class base rate), rising to 97.0% outside the displacement band and 67.3% inside. ... mehrWe derive the mapping from grid-point correction to the criterion’s substation factor. The condition characterizes error direction, not magnitude, and evaluating factor–residual alignment and residual dispersion requires metered substation demand; it informs integration-mechanism trade-offs, not absolute accuracy or downstream planning decisions.