Time series, characterized by its intricate components, poses a unique challenge across analytical applications. While decomposition models excel at learning the representation of components, they face limitations in noise sensitivity and feature extraction. In response, we introduce a novel Seasonal Trend Encoder (STEncoder), a robust decomposition for time series forecasting, which utilizes channel-independent attention mechanisms and a temporal convolution network to learn representations of the series. Meanwhile, the attention mechanism often struggles with the inherent noise present in time series. We specifically propose Adaptive Wavelet Block (AWB) that harnesses the discrete wavelet transform to attenuate the model’s noise sensitivity and to capture non-linear dependencies that are obscured by noise via adaptive thresholding. Furthermore, the block manifests a robust capability when applied to a variety of datasets. Extensive experiments conducted across nine real-world datasets have demonstrated the superior performance of the STEncoder over eleven baseline methods. It particularly excels in datasets with pronounced seasonal characteristics.