Resilience-Oriented Daily Peak Load Forecasting on Utility Open Data Using a Dual-Stream Seasonal ARIMA and Kolmogorov–Arnold Network

Short-term load forecasting (STLF) is a core input to unit commitment and operating-reserve scheduling. However, average error metrics can mask tail failures during extreme events (e.g., typhoons and heatwaves) that drive the highest operational risk. This paper proposes a reproducible, resilience-oriented STLF framework for utility open data that decomposes demand dynamics into (i) a seasonal ARIMA stream for periodic memory and (ii) a Kolmogorov-Arnold Network (KAN) stream that learns an interpretable nonlinear correction of the ARIMA residual. Using Taipower open datasets (2023–2024) and stress-testing on Typhoon Gaemi (2024), the proposed ARIMA–KAN achieves 1.65% day-ahead MAPE and 2.35% week-ahead MAPE, while reducing worst-case typhoon MAPE from 12.4% (SARIMA) to 3.6%, shortening recovery time from 4.0 to 0.5 days, and decreasing degradation area from 25.5 to 2.8 %-days. In addition to accuracy and event-window resilience metrics, we provide interpretability via KAN spline functions and SHAP analysis to support operational trust and post-event diagnosis. The resulting pipeline is lightweight, auditable, and directly reusable for other utilities that publish load/reserve and meteorological open data.