Traffic Prediction Based on the GRU Neural Network

Abstract

Traffic prediction is an important and challenging task in the last decade. In this project, we trained a GRU neural network to predict traffic at four junctions using a normalization and differencing transform to achieve a stationary time series. Given the varying trends and seasonality observed at each junction, we employed unique strategies to make each stationary, and evaluated model performance using the root mean squared error. Additionally, we plotted predicted values against original test data. Data analysis reveals that the number of vehicles at Junction 1 is increasing at a faster rate than at Junctions 2 and 3, while sparse data precludes conclusions regarding Junction 4. Moreover, Junction 1 exhibits stronger weekly and hourly seasonality, while the other junctions display a more linear trend. The results of this study can effectively inform vehicle flow rate prediction. However, GRU models may suffer from slow convergence rates and low learning efficiency, leading to prolonged training times and underfitting. Despite these limitations, the use of GRU models remains the most accurate, achievable, and straightforward method for calculating traffic predictions.