Transparent Federated Load Forecasting at the Secondary-Substation Level for DSO Flexibility

To participate in local flexibility markets (LFMs) or activate flexible connection agreements (FCAs), distribution system operators (DSOs) need a day-ahead load baseline at the secondary substation that is accurate, transparent enough to be reproduced by third parties, and respectful of the privacy of the smart-meter data underneath. Mainstream forecasters are black boxes, and pooling per-customer data centrally raises sovereignty and GDPR concerns. We present a transparent, interpretable, edge-deployable secondary-substation (SS) load forecaster trained by federated learning across substations—only model parameters are exchanged, raw consumption data never leaves a node—feeding a downstream congestion-flexibility application. The forecaster reuses Wangari Global’s causal-augmented, deterministic-first method [1] and specialises it to the DSO SS level. On real, public UK Power Networks LV-feeder smart-meter data (8 feeders rolled up to 6 secondary substations, half-hourly), the interpretable model reaches 29.8% day-ahead MAPE at the substation-total level with calibrated 90% intervals (empirical coverage 0.906), beating a seasonal-naive baseline (35.5%). Federation cuts aggregate MAPE from 33.8% (local-only) to 23.9%—within 0.3 pp of a centralised, data-pooled bound (23.5%)—and lifts a thin-history substation from 111% to 50% MAPE (a +61 pp cold-start gain), all with parameters-only sharing. The model is genuinely edge-deployable (∼0.6 KiB artifact, sub-millisecond inference, within a Raspberry Pi 4 budget). We report honestly that absolute MAPE remains above the 5% operational target because these public-proxy substations aggregate only 1–2 LV feeders (≈ tens of homes); the contribution is the right granularity, transparency, privacy-preserving federation, and a congestion-flexibility schedule, with <5% the in-project objective on the far-more-aggregated transformer-level pilot data. The evaluation harness is released openly for reproducibility.