For the improvement of the effect of traffic multi-step forecasts using short-term data of loop in Sydney coordinated adaptive traffic system(SCATS), on the basis of data preprocessing, traffic data and time points at time t over a sampling period of n intervals were included in the traffic state feature vector, Euclidean distance was used to measure the closeness between current traffic state and historical traffic state, the number of nearest neighbors corresponding to the minimum error of travel multi-step forecasts was selected, and the weights of k-nearest neighbors were identified by normalizing the reciprocal of the distance between traffic states, a new bi-level method of multi-step forecasting using k Nearest Neighbor(k-NN) algorithm was designed, including a multi-step forecasting method and a predictable steps on-line estimation method. The validity of the proposed method was tested with data measured from a megacity. The results indicate that the proposed method can further improve the effect of short-term traffic multi-step forecasts.