<p>The integration of AI models with environmental fields continues to encounter challenges related to the adaptability between model algorithms and specific application scenarios. Ground-level ozone (GLO) is a critical air pollutant characterized by substantial fluctuations and its multifactorial variability complicates accurate predictions when using conventional modeling methods. This presents new challenges for environmental control and quality management. In this study, two transformer frameworks were proposed: ST-Transformer (based on Wavelet Transform(WT)) and DE-Transformer (based on Non-Stationary Attention), to effectively capture the short-term non-stationary characteristics resulting from periodic trends in ozone data. Precursor species concentration, ozone levels, and meteorological data were selected and processed in these two models. The WT decomposed the original time series data into several sub-sequences while eliminating redundant information before integrating it into the traditional Transformer model. This resulted in a 53% improvement in predictive performance for the ST-Transformer. On the other hand, the DE-Transformer employed Non-Stationary Attention to segment time series data into several sequence blocks based on mean and variance values, and utilized a multi-layer perceptron as a projector to reintegrate fluctuation data. Due to its increased sensitivity to factors influencing concentration peaks and fluctuations, the DE-Transformer exhibited superior predictive performance compared to the ST-Transformer. Consequently, the DE-Transformer can be effectively applied in analyzing high-volatility environmental data scenarios, thereby effectively improving the accuracy of environmental quality predictions.</p>