Time series prediction involves static and dynamic features. Extraneous input information hampers model performance, and the statistical attributes of time series change over time, affecting distribution. Targeted processing of input data for feature and distribution dynamics is vital. This article introduces AdaDynaTrans, an attention-based optimization model. It dynamically combines features from static covariate extraction and temporal variable evolution modules, learning time relationships at various scales. Using self-designed dynamic residual and feature selection units, it suppresses irrelevant information.AdaDynaTrans tackles time covariate shift through temporal segmentation and segmented training mechanisms. Additionally, a relative position-optimized transformer is employed to capture local dependencies within temporal data, thereby achieving exceptional performance within real-world scenarios. Through comprehensive evaluations on both public datasets and industrial scenarios, the considerable efficacy of AdaDynaTrans is demonstrated. Ablation experiments are conducted to analyze the effectiveness of each constituent element, and the interpretability of the model is showcased using a case study involving Mooney viscosity prediction in rubber production. This research contributes substantively to the field by presenting a model that not only achieves high performance but also offers insights into the temporal dynamics of complex industrial processes.