Ozone is a major air pollutant in the Yangtze River Delta (YRD) region, posing significant public health risks due to elevated concentrations. This study introduces the Site-Informer model, specifically designed for long-term forecasting of ozone concentrations. Building upon the Informer framework, the model incorporates ProbSparse self-attention and introduces spatial correlation-based site aggregation, enabling precise forecasts at both hourly (12, 24, 48, 96, 168) and daily (1, 2, 3, 7, 14) resolutions. Experimental results demonstrate that the Site-Informer model outperforms existing models, including Informer, Transformer, and GRU, across various time-series lengths. The Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) values are reported within the ranges of 0.381–0.562 μg/m3 and 0.508–0.718 μg/m3, respectively. Additionally, a transfer learning approach is incorporated, resulting in the TL-Site-Informer model, which significantly improves daily prediction accuracy, reducing MAE and RMSE to 0.375–0.484 μg/m3 and 0.474–0.596 μg/m3, respectively. The findings indicate a reduction in prediction errors of 2% to 6% compared to existing models across different time series lengths. These results validate the model’s efficacy for long-term trend analysis and air quality monitoring, establishing a benchmark for future forecasting research.