Accurate air-quality forecasting is increasingly central to urban environmental governance, public-health protection, and data-driven smart-city management. This manuscript presents a polished and publication-ready account of a hybrid particle swarm optimisation (PSO), long short-term memory (LSTM), and recurrent neural network (RNN) framework for forecasting urban air quality using the University of Utah Air Pollution Monitoring Network dataset for Salt Lake City, Utah. The empirical setting is based on 25 pollution sensors, hourly aggregation, and a study window spanning 2019-07-26 to 2021-05-14. The modelling pipeline integrates data pre-processing, missing-value handling, outlier treatment, recurrent sequence learning, and PSO-based hyperparameter optimisation, followed by a curiosity-based motivation mechanism that strengthens the hybrid recurrent design. In the reported experiments, the proposed model achieves a mean absolute error of 0.0082 and an R2 score of 0.1227 in the principal benchmark comparison, while 10-fold cross-validation yields an average RMSE of 0.013, MAE of 0.010, MAPE of 3.8%, and R2 of 0.94. Additional analysis shows that hyperparameter tuning materially improves predictive performance, and ablation results indicate that the full integration of LSTM, RNN, PSO, and curiosity-based motivation produces the strongest results. Framed for the urban development and smart-city literature, the study demonstrates that hybrid recurrent forecasting can support municipal decision-making in air-quality surveillance, exposure mitigation, traffic-responsive environmental control, and sustainable service planning.