A weighted pattern recognition algorithm for short-term traffic flow forecasting

The k-nearest neighbor (k-NN) nonparametric regression is a classic model for single point short-term traffic flow forecasting. The traffic flows of the same clock time of the days are viewed as neighbors to each other, and the neighbors with the most similar values are regarded as nearest neighbors and are used for the prediction. In this method, only the information of the neighbors is considered. However, it is observed that the “trends” in the traffic flows are useful for the prediction. Taking a sequence of consecutive time periods and viewing the a sequence of “increasing”, “equal” or “decreasing” of the traffic flows of two consecutive periods as a pattern, it is observed that the patterns can be used for prediction, despite the patterns are not from the same clock time period of the days. Based on this observation, a pattern recognition algorithm is proposed. Moreover, empirically, we find that the patterns from different clock time of the days can have different contributions to the prediction. For example, if both to predict the traffic flow in the morning, the pattern from the morning can lead to better prediction than same patterns from afternoon or evening. In one sentence, we argue that both the pattern and the clock time of the pattern contain useful information for the prediction and we propose the weighted pattern recognition algorithm (WPRA). We give different weights to the same patterns of different clock time for the prediction. In this way, we take both virtues of the k-NN method and the PRA method. We use the root mean square error (RMSE) between the actual traffic flows and the predicted traffic flows as the measurement. By applying the results to actual data and the simulated data, about 20% improvement compare with the PRA is obtained.