Abstract
To develop a hybrid methodology combining computational fluid dynamics (CFD) and machine learning (ML) to predict pollutant concentrations from road transport in an urban canyon. The study aims to analyze the impact of noise barriers of varying heights on pollutant dispersion. Based on three-dimensional RANS CFD models verified against experimental data, a dataset of time series of concentrations at nine control points for four scenarios was generated. A bidirectional recurrent neural network with long short-term memory (BiLSTM) with an attention mechanism was used to construct a surrogate model. The model quality was assessed using the MAE, MAPE, and R² metrics. The model provides high forecast accuracy (MAPE < 7%) in scenarios without a barrier and with a 0.1H barrier. It was found that a barrier height of 0.2H creates the most difficult turbulent regimes to predict, leading to a significant increase in error (MAPE up to 17.56%, R² up to -4.5). Increasing the barrier height to 0.3H partially restores forecast accuracy due to the stabilizing effect on the flow. The developed hybrid CFD-ML model is effective for rapid air quality forecasting. The results are significant for environmental planning and urban development optimization, highlighting the need to consider the nonlinear aerodynamic effects of medium-height barriers and the feasibility of using complex ML architectures to describe them.