Benchmarking Model Complexity for Short-Term Surrogate Forecasting of High-Resolution Urban WRF Outputs During an Extreme Heatwave in Chongqing

Abstract

High-resolution urban Weather Research and Forecasting (WRF) simulations resolve interactions among urban form, near-surface meteorology, atmospheric dynamics, and building-energy processes, but their computational cost limits repeated scenario analysis. This study compares eight pointwise temporal surrogates (Persistence, previous-day same-hour, Ridge, Random Forest, XGBoost, multilayer perceptron (MLP), gated recurrent unit (GRU), and long short-term memory (LSTM)) for recursive 24 h prediction of T2, Q2, W10, and WRF-BEM diagnostic air-conditioning power density (AC) from a 120 h WRF-BEP/BEM-LCZ heatwave simulation over Chongqing. Model development used three expanding-window folds, fold-specific scaling, fixed monitoring cells, and five fixed seeds for stochastic families; the h97–h120 full-domain test remained untouched until all modeling choices had been finalized. Station comparisons provided a limited check on the physical plausibility of the WRF meteorological fields; AC was not validated against metered energy use. Model performance depended strongly on the target variable. Random Forest performed best for T2 (1.081 ± 0.002 °C), LSTM for Q2 (0.969 ± 0.063 g/kg), and MLP for AC (0.505 ± 0.187 W/m2), whereas deterministic Ridge was best for W10 (1.095 m/s). W10 had weaker temporal memory (median lag-1/lag-24 ACF 0.581/0.211). Ridge outperformed the Previous-day baseline for 75.0% of grid cells, although errors increased sharply in the top wind-speed decile and during rapid changes. For this single-event, single-city benchmark, recurrent models did not improve W10 accuracy over Ridge. Model choice depended on the target and the balance between accuracy, spatial fidelity, and computational cost; transfer to other events and meteorological regimes remains untested.