NEURAL NETWORK METHODS FOR TRAFFIC INTENSITY FORECASTING

Abstract

This paper is dedicated to solving the problem of short-term traffic flow intensity forecasting using modern artificial neural network architectures, including SimpleRNN, LSTM, GRU, and 1D-CNN. The dataset for the research consists of one-dimensional time series containing counts of vehicles at specific intersections at various points in time. The accuracy of time series forecasting is crucial for tasks involving adaptive traffic light control, traffic mode planning, and road infrastructure load assessment. Standard time series forecasting methods are typically based on statistical models, which may be insufficient for describing non-linear traffic dynamics and daily and weekly seasonality. A specific problem is the proper consideration of cyclical time features (hour of the day, day of the week), where standard encoding creates discontinuities at cycle boundaries. The proposed approach involves data preprocessing, including outlier removal and trigonometric cyclical feature encoding. That allows neural networks to correctly interpret daily and weekly seasonality while maintaining the continuity of the time cycle. During the study, computational experiments across various forecasting horizons (1–12 hours) were conducted using real-world data, enabling an evaluation of the proposed approach's effectiveness in short-term forecasting. The results of the comparative analysis demonstrate a reduction in the Mean Absolute Percentage Error (MAPE) compared to traditional methods. It was established that the 1D-CNN model achieves the shortest training time, while LSTM and GRU achieve the highest forecasting accuracy. The obtained results can be used to develop adaptive traffic-light control systems within the Smart City concept