The invention discloses a dynamic STKNN model-based short-time traffic prediction method. The method comprises the steps of: a, feature representation: averaging traffic conditions of each road segment in history days, and carrying out feature representation on each road segment by utilizing an average speed vector; b, automatic traffic mode recognition: automatically recognizing similar traffic modes of road networks through an AP cluster algorithm; c, automatic time interval division: aiming at the automatically recognized traffic modes, automatically dividing time intervals through a K-Means algorithm; and d, self-adaptive STKNN model construction: aiming at each time interval of each traffic mode obtained in the step 2 and 3, respectively constructing an STKNN model. The method is capable of directly helping the traffic management department to generate reasonable and efficient strategies to ease traffic congestion, thereby realizing redistribution of road network traffic flows andhelping the public to realize correct path planning.