To address the challenge of difficulty in selecting hyperparameters for gated recurrent units (GRU), a combined model (CNN-GRU-SSA) was proposed, integrating convolutional neural networks (CNN), gated recurrent units, and the sparrow search algorithm (SSA). Firstly, CNN for feature extraction from the multidimensional input dataset was utilized. Subsequently, the features extracted by CNN into the GRU model are inputted. Lastly, the hyperparameters of the GRU model using the SSA algorithm were optimized and were applied to the model to predict PM_2.5 concentrations. Chengdu in the western region and Hangzhou in the eastern region was selected as research areas. Hourly data of air pollutants, meteorological factors, boundary layer height (BLH), and precipitable water vapor (PWV) from December 1, 2021, to February 13, 2022, were utilized for modeling. The aim was to predict the variation in PM_2.5 concentrations in both cities from February 14 to February 28, 2022. Experimental results indicate that the CNN-GRU-SSA model demonstrates a noticeable improvement in prediction accuracy compared to other models, with Chengdu exhibiting the closest proximity between predicted and actual values.