Air pollution poses a global challenge, carrying serious risks to both human health and the environment. Fine particulate matter (PM2.5) is a key factor contributing to respiratory and cardiovascular diseases. Therefore, accurate PM2.5 prediction is essential for analyzing pollution trends, safeguarding public health, guiding environmental planning, and shaping policy decisions. In this study, we conducted extensive experiments to enhance PM2.5 prediction and developed a robust prediction system. Our system incorporates multiple components, including AirBox and EPA data preprocessing, data fusion, feature engineering, feature selection, and the proposed prediction model, DCRNN-GS. Designed for iterative multi-step forecasting, our model utilizes data from the previous 24 h to predict PM2.5 levels for the next 24 h. Experimental results demonstrate that our system surpasses state-of-the-art methods in PM2.5 prediction performance.