PREDICTIVE ANALYSIS OF TRAFFIC VOLUME USING ML

Abstract

Road traffic in growing cities has increased far faster than the capacity of the road network, and congestion now causes long travel delays, higher fuel consumption, and increased air pollution. Traffic authorities can manage congestion more effectively if they know in advance how many vehicles are likely to use a road at a given time. This paper presents a predictive analysis system for traffic volume that uses machine learning to forecast hourly vehicle counts from historical sensor data, calendar information, and weather conditions. The data used by the system comes from roadside vehicle detection equipment such as inductive loop detectors, infrared counters, and camera-based counters, which record the number of vehicles passing a point in each time interval. These counts are combined with calendar attributes including hour of the day, day of the week, month, and public holidays, and with weather attributes including temperature, rainfall, snowfall, cloud cover, and a general weather description. The raw sensor records are checked for gaps, duplicate timestamps, and faulty readings caused by detector malfunction, and are then aggregated into a clean hourly time series. Feature engineering plays a central role in the system. Cyclical encoding is applied to hour and month so that the model understands that eleven at night is close to midnight, and lag features and rolling averages capture the recent behaviour of the traffic stream. Several regression models are trained and compared, including linear regression, a decision tree regressor, k-nearest neighbours, random forest, gradient boosting, and extreme gradient boosting. The models are evaluated using mean absolute error, root mean square error, and the coefficient of determination on a chronologically separated test period. The best performing model is used to produce forecasts for the next several hours and for the following day. The results are displayed on a web dashboard that shows the predicted volume curve alongside actual counts, highlights expected peak periods, and classifies each hour into light, moderate, or heavy traffic. Feature importance analysis shows the relative influence of time of day, weekday, holidays, and weather, which helps traffic planners understand the causes of variation rather than only the forecast values. The system is intended to support signal timing adjustment, deployment of traffic police, scheduling of road maintenance, and advance information to commuters. Because it relies on commonly available sensor counts and public weather data, it can be applied to other road segments with modest effort. Future work includes real-time streaming of detector data, forecasting across multiple connected junctions, and the use of sequence models for longer forecast horizons.