Abstract: Time series predication has become an undeniable technology in various domains of modern life. The non-stationarity of time series data is a significant challenge, leading to changes in their statistical properties and giving rise to distribution shift issues, which make accurate time series predictions more challenging. To address these challenges, a simple yet effective method called Reversible Scaling Instance Normalization (RaIN) is proposed. The RaIN method combines general normalization and optional inverse normalization techniques, and has structural symmetry. The method's symmetric structure is designed to remove and restore non-stationary information within time series data, resulting in a substantial improvement in prediction performance. To validate the effectiveness of RaIN in addressing distribution shift problems, extensive quantitative and qualitative analyses are implemented.