Traffic flow prediction is a critical area of research within Intelligent Transportation Systems (ITS), aiding in traffic planning, control, and more. Real-world traffic flow exhibits inherent non-stationarity, presenting a significant challenge for deep forecasting models. Current non-stationary traffic flow prediction models often decompose the sequence into multiple frequency components, process them in groups, and then recombine to get the output. This approach can lead to a loss of information from the original sequence. In this paper, we propose a novel framework for non-stationary traffic flow prediction called Trend-Variation Separated Learning Framework (TVSLF). TVSLF addresses these limitations by decomposing the raw traffic flow input into trend and variation terms. These terms are processed separately through NET1 and NET2, respectively, before being combined to generate the final output. TVSLF effectively mitigates two key challenges in traffic flow forecasting: underfitting the detailed variations of true values in multi-horizon predictions and lagging in single-horizon predictions. Our method's effectiveness and efficiency are validated against existing methods using four real-world datasets.