Hybrid Deep-Learning Approach with Geoffrey E. Havers-Based Loss Function and Evaluation Metric for Multilocation Traffic-Flow Forecasting

Abstract

Traffic forecasting can enhance the efficiency of traffic control strategies such as routing decisions, variable speed limits, and ramp metering, resulting in a decrease in congestion, pollutants, and expenses, and an improvement in journey time predictability. Traffic forecasting, however, remains challenging because of the complex, heterogeneous, and cyclic nature of traffic data. To address this complexity, this research employs a multi-input hybrid deep self-attention network (MIHDSAN) for multilocation forecasting. The model inputs are selected using correlation analysis. New tunable loss and evaluation metrics formulations are proposed based on the traffic-modeling Geoffrey E. Havers (GEH) statistic. The proposed method was validated on two independent real-world traffic datasets from Stockton and Oakland, California. The weekly periodicity was the more relevant periodic input feature compared with daily variations; however, the daily variation was also significant for the Stockton dataset. The inclusion of weekly traffic periodicity (>95% correlated) improved the performance of the model by 3%. Adding daily periodicity was only beneficial for the Stockton dataset (91% correlated). The proposed GEH metric and its standard acceptance criterion offer both quantitative and qualitative means of evaluating the forecasts produced. The GEH loss function was consistent and outperformed current industry-standard methodologies of mean absolute error (MAE) in 80% and mean squared error (MSE) in 94% of cases. Therefore, this research presents evidence to suggest that the proposed GEH loss and evaluation functions validated in this paper become a standard criterion for traffic forecasting.