In recent years, sustained efforts to mitigate air pollution have significantly improved overall air quality. However, infrequent pollution events pose new challenges for accurate forecasting. Concurrently, the growing availability of monitored datasets offers opportunities for data-driven deep learning approaches in this field. Most deep learning models rely solely on time series methods, neglecting the influence of meteorological factors. Although some studies incorporate these factors with complicated structures, they often fail to capture the intrinsic relationships between meteorological conditions and pollution outcomes, especially for week-ahead predictions. This study presents an adaptive method for modeling the impact of evolving meteorological conditions on pollution outcomes. Inspired by the annual periodicity of meteorological patterns and their consistent impact on pollution, we align historical pollution events with the current window, adaptively shifting them in response to the corresponding meteorological patterns. Specifically, we represent meteorological variations as changes in the coefficients of orthogonal polynomials and employ a gating mechanism to dynamically transfer the influence of historical meteorological data to the current window, thereby indicating potential pollution process under current conditions. Experiment Results demonstrate that our method can enhance a linear model to outperform existing pure time series approaches on week-ahead prediction, emphasizing the importance of considering meteorological conditions and pollution results in air quality forecasting. Moreover, our findings align with expert domain knowledge and achieve further performance improvements when incorporating forecasted meteorological data.