Residual Calibration for Electricity-Load Prediction Intervals

This study examines whether updating the residual-calibration distribution improves future electricity-load prediction intervals when the point forecaster is held fixed. Thirty consumers were selected deterministically across training-period variability ranks. Models were trained through September 2013, initialized using October–December 2013 residuals, and evaluated sequentially over all 8,760 hours of 2014. Static split conformal, rolling 28-day, exponentially weighted, fixed-reference adaptive conformal inference (ACI), and rolling-reference ACI policies shared identical point forecasts. Across 262,800 customer-hour forecasts per policy, pooled marginal coverage was 0.8923 for static calibration and 0.8987–0.9007 for the updating policies. Pooled coverage above each customer’s training-period 80th demand percentile increased from 0.8068 to 0.8475–0.8544, while equal-customer peak means increased from 0.7646 to 0.8033–0.8118. Mean interval width increased from 61.27 to 65.36–65.97 kWh, but interval scores improved. Updating residual calibration therefore improved empirical reliability at a modest sharpness cost, although peak coverage remained below the nominal 0.90 target. The findings are descriptive evidence from 30 selected series and do not establish conditional or population-wide coverage.