HYBRID TIME-SERIES FORECASTING MODELS FOR TRAFFIC FLOW PREDICTION

Abstract

Traffic flow forecast is critical in today’s transportation system since it is necessary to construct a traffic plan in order to determine a travel route. The goal of this research is to use time-series forecasting models to estimate future traffic in order to reduce traffic congestion on roadways. Minimising prediction error is the most difficult task in traffic prediction. In order to anticipate future traffic flow, the system also requires real-time data from vehicles and roadways. A hybrid autoregressive integrated moving average with multilayer perceptron (ARIMA-MLP) model and a hybrid autoregressive integrated moving average with recurrent neural network (ARIMA-RNN) model are proposed in this paper to address these difficulties. The transportation data are used from the UK Highways data-set. The time-series data are preprocessed using a random walk model. The forecasting models autoregressive integrated moving average (ARIMA), recurrent neural network (RNN), and multilayer perceptron (MLP) are trained and tested. In the proposed hybrid ARIMA-MLP and ARI-MA-RNN models, the residuals from the ARIMA model are used to train the MLP and RNN models. Then the efficacy of the hybrid system is assessed using the metrics MAE, MSE, RMSE and R2 (peak hour forecast-0.936763, non-peak hour forecast-0.87638 on ARIMA-MLP model and peak hour forecast-0.9416466, non-peak hour fore-cast-0.931917 on ARIMA-RNN model).