Addressing the limitations of traditional models in capturing long-term dependencies and generalization ability in meteorological sequences, we propose a novel air temperature forecasting model based on sparse attention and adaptive seasonal and trend decomposition using loess (ATFSAS) in this paper. ATFSAS employs an encoder-decoder architecture, and integrates a sparse attention mechanism to effectively capture long-term dependencies in meteorological observation data. An information distillation method is introduced to reduce redundancy during the encoding process. The model refines the periodic and trend components in the forecast signals by combining a multi-layer decoder with an adaptive timing decomposition unit, achieving precise air temperature forecast. Based on the climate dataset in Jena, Germany, 24 h refined air temperature forecast is performed by ATFSAS, abtaining a mean absolute error of 1.7108℃. Compared to the LSTM model, ATFSAS demonstrates superior performance in medium-short term daily average air temperature and multi-region single-day average air temperature forecasts based on the China ground climate daily dataset, and their mean absolute errors get improved by 35.56% and 23.66%, respectively.