ROBUST FORECASTING OF TRAFFIC FLOW INTENSITY

Abstract

Accurate and timely information about current and forecast parameters of traffic flows is an important condition for the functioning of intelligent transport systems. The use of high-quality data will make it possible to solve the problem of adaptive traffic flow management more effectively, reduce travel time, improve the accuracy of travel route planning, and generally improve the efficiency of using transport infrastructure.
 This article presents studies of the impact of traffic light objects on the registration of traffic flow parameters by transport detectors. It is established that the location of the control zone of transport detectors within the area of the formation of the queue of vehicles leads to artificial noise of the collected data. The use of robust forecast of transport flow parameters in solving problems of adaptive traffic flow management is justified.
 The solution of the problem of robust traffic intensity prediction using the LightGBM model based on the gradient boosting algorithm is shown. The feature space of the model included the traffic intensity log, the date and time of the data received, the weekend features, and the lane group ID. The robustness of the forecast is ensured by using the smoothed time series model error function using the LOWESS method, when applying the original, non-smoothed time series to the model input.
 The model trained on real data has properties of robustness of the traffic intensity forecast, provided filtering of stochastic fluctuations and outliers of the measured ones. At the same time, it distinguishes the features of the daily profiles for working and weekend days in all groups of bands, the forecast delay was observed only if there was a significant deviation of the observed intensity from the daily trend.