Spatio - Temporal Traffic Forecasting under Weather Perturbations in Tashkent: TFT vs CNN - LSTM with Google Maps and OWM
Abstract
Urban congestion forecasting in data-scarce cities remains challenging due to heterogeneous road dynamics and exogenous perturbations such as weather. We study multi-horizon forecasting of the Travel Time Index (TTI) on major corridors of Tashkent, Uzbekistan, under varying meteorological conditions. We construct a multimodal dataset by combining Google Maps travel times with OpenWeatherMap (OWM) forecasts and calendar features. We compare a Temporal Fusion Transformer (TFT) - designed for interpretable multi-horizon forecasting with known-future covariates-against a strong CNN-LSTM baseline augmented with spatial neighborhood inputs. Across 5, 15, and 60-minute horizons, TFT achieves lower MAE and sMAPE and provides calibrated quantiles, while CNN-LSTM remains competitive for short horizons on stable weather days. Ablation studies quantify the utility of weather and calendar signals, showing gains particularly during precipitation and heat events. We release the problem formulation, evaluation protocol, and reproducibility checklist to facilitate future research on Central Asian urban traffic.