Daily Parking Demand Interval Forecasting with Similar-Day Matching Diagnostics

Daily parking demand varies with calendar patterns, weather, and recent demand states, creating substantial uncertainty. Reliable prediction intervals support parking-resource scheduling. Existing similar-day methods mainly use retrieved neighbors for point forecasting, but seldom exploit retrieval-based uncertainty information that reflects the reliability of historical analogs and can inform interval construction. This study proposes a framework for daily parking-demand interval forecasting with similar-day matching diagnostics. Similar-day retrieval and kernel-weighted aggregation generate point forecasts and seven diagnostics of matching quality, neighbor-demand consistency, recent demand variation, and historical sample support. A residual conditional quantile model combines these diagnostics with conventional covariates to construct asymmetric, day-specific initial intervals. Rolling conformalized quantile regression (CQR) calibrates recent coverage, while adaptive conformal inference (ACI) updates the miscoverage rate from observed coverage errors. Rolling experiments on three real parking facilities show that, with both methods achieving empirical coverage close to 80%, the proposed method reduces macro-average relative interval width and Winkler score by 6.78% and 10.22%, respectively, compared with Conformal PID (proportional–integral–derivative) control, the best external baseline by macro-average Winkler score. The framework provides day-specific uncertainty estimates, with potential applications in next-day parking-resource planning.