This study aims to develop a physically consistent, dual-attention bidirectional Long Short-Term Memory (BiLSTM) framework for accurate hourly forecasting of PM2.5 concentrations in New Taipei City, Taiwan, and to determine whether multi-scale atmospheric signal conditioning and attention-based temporal modelling can improve forecasting performance and generalisation across heterogeneous coastal-mountain atmospheric regimes. Hourly PM2.5, co-located gaseous pollutant, and meteorological observations from five monitoring stations, were pre-processed through wavelet-based denoising, Box–Cox variance stabilisation, and robust scaling. A residual dual-attention BiLSTM incorporating adaptive feature gating, dual global pooling, and a softplus output layer was trained using Huber loss and the Adam optimiser, and benchmarked against a baseline BiLSTM and traditional statistical models (VAR and ARIMA). Forecasts were combined through an inverse-error-weighted ensemble, and performance was assessed using MSE, RMSE, MAE, MAPE, R², and explained variance, complemented by an ablation analysis at each station. The proposed model achieved R² between 0.80 and 0.85 across all five stations, consistently outperforming the baseline BiLSTM and the ARIMA/VAR benchmarks, which produced near-zero or negative explanatory power. Improvements associated with individual architectural components. The model shows attenuation for extreme concentrations (>50–60 µg/m³), residual heteroscedasticity persists at high pollution levels, and the results are based on a single metropolitan area over the study period, so wider spatial and temporal validation would strengthen the generalisability of the conclusions. The framework provides a computationally efficient, physically constrained tool for hourly PM2.5 forecasting that can support real-time public health advisories and emission-management decisions in coastal-mountain metropolitan areas. The study introduces a station-adaptive, physically consistent dual-attention BiLSTM that combines signal conditioning, adaptive gating, and error-weighted statistical ensembling without requiring predefined spatial graph structures, offering new insight into the conditions under which architectural complexity benefits PM2.5 forecasting.