Short-term traffic flow prediction plays an important role in intelligent transportation systems (ITSs). Sequential data assimilation (SDA) is very effective in the short-term traffic flow prediction of expressways because of its real-time reflections of local fluctuations of fast-changing traffic flow values in the time and space domains. Assimilation models in a traditional SDA (T-SDA) system are usually constructed using historical measurements. However, historical data are always disturbed by local noises, greatly affecting the accuracy of constructed assimilation models and predicted results. To deal with the problem, we propose to adopt the extreme-point symmetric mode decomposition (ESMD) method to conduct historical data denoising for improving the assimilation model performance in the SDA system. First, the original historical measurement signals are decomposed into a series of simple signals called intrinsic mode functions (IMFs) by ESMD to further analyze and seek useful information and local stochastic noises. Second, the denoised historical traffic data are used to construct an assimilation model, and the denoised SDA (D-SDA) system for short-term traffic flow prediction is established. Third, the applications of the D-SDA system for short-term traffic flow prediction are presented and compared with those of the T-SDA system. Experimental results showed that compared with the T-SDA system, the D-SDA system can successfully reduce the effects of noises in historical measurements on assimilation model construction and improve the accuracy of short-term traffic flow prediction results.