The increasing urbanization has intensified air pollution exposure, seriously threatening public health. This study examines how various machine learning and deep learning models can enhance the optimization of traffic changes within an urban planning framework, with a particular emphasis on air quality. Using local pollutant data from stations located in Novi Sad (Serbia), the models estimate emission changes from major transport infrastructure modifications. Two complementary approaches are used: (1) multi-output regression aimed at predicting air pollutant concentrations, and (2) multi-output classification focused on predicting the air quality index (AQI). For daily predictions, the Long Short-Term Memory (LSTM) model achieved the best R2 = 0.5, while the Balanced Random Forest classifier reached 0.775 balanced accuracy for AQI classification. Finally, the most accurate models were applied to two prospective city network scenarios projected - (1) construction of two river bridges rerouting heavy traffic, and (2) creation of an extended pedestrian zone downtown - to quantify the expected reduction in pollutant concentrations and estimate possible changes in AQI categories. Depending on the pollutant, the results indicate a relative daily decrease between \(0.7\%\) and \(2.9\%\) on average for both scenarios. Although the improvement might seem small, this reduction is obtained for 13 street segments within a 100-meter radius around air stations, and a potential reduction for the whole city could be more substantial. Additionally, to turn forecasts into policy-ready evidence for traffic system decisions, we paired predictive models with explainable artificial intelligence (XAI) methods. Applying XAI methods reveals that meteorological factors, including relative humidity, total precipitation, temperature, and wind speed, primarily account for the predictions of developed models. Overall, the XAI methods highlight the combined influence of temporal, spatial, and meteorological factors on AQI.