Station-Aware Patch-Gated Temporal Transformer for Multihorizon Forecasting in Smart City Applications

Multihorizon time series forecasting (TSF) tasks are a crucial component of smart city systems, enabling decision-makers to take proactive measures in traffic, energy, transportation, and environmental management. Although existing deep learning models are effective in capturing hierarchical temporal dependencies, they often assume static inter-station dependencies, rely on single temporal operators that fail to capture both short-term transients and long-term cycles, and use autoregressive decoding that amplifies early errors via exposure bias. To tackle these limitations, we introduce a Station-Aware Patch-Gated Temporal Transformer (SAPGTT), an end-to-end framework for robust multi-horizon forecasting. Specifically, we propose a station-aware spatial mixer that applies self-attention across stations at each time step, dynamically learning time-varying dependencies without relying on fixed dependency assumptions. Then, we introduce a hybrid temporal encoder that incorporates bidirectional long-short-term memory (Bi-LSTMs) for short-to-medium dynamics, along with temporal self-attention to capture long-range seasonal patterns. In addition, a patchwise gated convolutional forecaster segments sequences into motifs and employs gating to suppress noise, producing direct multi-horizon outputs that eliminate error accumulation from iterative rollouts. Extensive experiments on a benchmark urban air quality fore-casting task demonstrate that SAPGTT consistently outperforms state-of-the-art deep learning baselines, achieving up to a 17.0% reduction in MAE and 18.8% in RMSE, delivering accurate, stable, and horizon-consistent forecasts.