Abstract
This paper presents a hybrid approach for predicting pollutant dispersion in urban street canyons, taking into account noise barriers. The methodology combines detailed CFD modeling and a surrogate model based on the BiLSTM neural network architecture with an attention mechanism. Configurations with barrier heights of 0.1H, 0.2H, and 0.3H were studied. CFD calculations revealed a nonlinear effect of barrier height on aerodynamics and the formation of pollutant accumulation zones, with the most complex non-stationary behavior observed at a height of 0.2H. The surrogate model successfully predicts concentration evolution for both the barrier-free and 0.1H barrier cases, demonstrating an average absolute percentage error of less than 15%. For a 0.2H barrier, accuracy decreases in zones of intense turbulence due to the highly non-stationary nature of the process. This approach significantly reduces computational costs while maintaining physical accuracy, which is promising for decision support systems in urban ecology. The model provides accelerated forecasting compared to CFD calculations by 7–8 orders of magnitude, so the inference time showed 1–5 ms, while one CFD simulation takes about 54 hours on the CPU.