We present a multiscale sequential data assimilation (M-SDA) system and apply it to short-term traffic flow prediction. Assimilation models in traditional sequential data assimilation (T-SDA) systems, which are usually constructed using historical measurements, are always disturbed by local noises. Simultaneously, the accuracy of assimilation results is also affected. To reduce the effects of these noises on assimilation models and the accuracy of results, an M-SDA system combining a T-SDA system and noise separation methods is constructed. This paper comprises four main parts: (1) a T-SDA system for short-term traffic flow prediction and multiscale noise separation methods are briefly discussed, and an example of denoised measurements with separated multiscale noises is given; (2) an M-SDA system for short-term traffic flow prediction is established; (3) the impacts of different noise separation scales on the accuracy of assimilation results are analyzed; and (4) applications of the M-SDA system to short-term traffic flow prediction are presented and compared with those of a T-SDA system. Experimental results were acquired from traffic flow measurements collected from a sub-area of a highway near Liverpool and Manchester, UK. The gap between the true and predicted values was evaluated by the root mean square error (RMSE) and mean absolute percent error (MAPE). By comparison with the prediction results from the T-SDA system, it was experimentally shown that the M-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.