Road Traffic Forecasting using Stacking Ensembles of Echo State Networks

Road traffic forecasting is arguably one of the practical applications related to Intelligent Transportation Systems where Machine Learning models have impacted most significantly in recent years. The advent of increasingly sophisticated supervised learning methods to capture and generalize complex patterns from data has unchained a flurry of research analyzing the performance of different models when learning from real data collected in road networks of very diverse nature. Nonetheless, the community has paid little attention to the use of reservoir computing models for traffic prediction. This field comprises several different modeling approaches ranging from liquid state machines to echo state networks, all sharing in common recurrence and randomness between neural processing units. This paper builds upon this research niche by exploring how ensembles of Echo State Networks can yield improved traffic forecasts when compared to other machine learning models. Specifically, we propose a regression model composed by a stacking ensemble of reservoir computing learners. As evinced by simulation results obtained with real data from Madrid (Spain), the synergistic combination of stacking ensembles and reservoir computing allows our proposed model to outperform other machine learning models considered in our benchmark.