StabilizeNet: a novel framework for alleviating non-stationarity in time series

Time series prediction is widely used in many fields in modern life, and its importance has become increasingly prominent. Non-stationarity is one of the main issues affecting the accuracy of time series predictions. Due to the statistical characteristics of time series data changing over time, it is difficult to stably apply the patterns learned from historical data to future predictions, thus increasing the difficulty and uncertainty of predictions. In order to solve this problem, a novel framework called StabilizeNet was proposed，which was designed to reduce the non-stationarity of time series data. The framework consisted of three parts: centralization and scaling transformation, linear transformation, and reverse transformation. By introducing a learnable normalized linear transformation matrix, it optimized data information retention and enhanced the model’s ability to capture time series dynamics. Compared with advanced time series prediction models such as Informer, SCINet, Pyraformer, FEDformer, and Crossformer, StabilizeNet demonstrated effectiveness and superiority on multiple datasets. This framework provides a new preprocessing method for time series prediction, which helps to improve the prediction performance of time series predictive models.