A Novel Multi-task Single-Step Traffic Congestion Forecasting Framework for Large-Scale Road Networks

Forecasting traffic congestion in large-scale transportation networks is a challenging problem of great importance in intelligent transportation systems. Most previous studies performed traffic congestion forecasting using single-task learning (STL), where machine learning algorithms were utilized for training, testing, and making forecasts for a single road segment. A key limitation of the STL is reduced accuracy and increased computational costs, as they cannot exploit the commonalities across multiple road segments. This paper proposes a novel multi-task learning (MTL) framework for forecasting single-step traffic congestion in a large-scale road network. The proposed framework involves the following two steps: (i) cluster the road segments in a network using a clustering algorithm, (ii) build a multi-task single-step traffic congestion forecasting model by simultaneously providing the normalized training data of all the road segments in a cluster, and (iii) carryout single-step forecasting of traffic congestion on all of the road segments in a cluster simultaneously. Experimental results on two different real world datasets, 1-h traffic congestion and 5-min traffic congestion, demonstrated that the proposed framework developed using the LSTM has not only improved the forecasting accuracy, but was also executed faster with higher energy efficiency as compared with the existing STL-based machine learning algorithms. In particular, we have observed that the proposed framework built an efficient model by consuming 15% less time and 50% less energy than the state-of-the-art approach.