Adaptive long-term traffic state estimation with evolving spiking neural networks

Due to the nature of traffic itself, most traffic forecasting models reported in literature aim at producing short-term predictions, yet their performance degrades when the prediction horizon is increased. The scarce long-term estimation strategies currently found in the literature are commonly based on the detection and assignment to patterns, but their performance decays when unexpected events provoke non predictable changes, or if the allocation to a traffic pattern is inaccurate. This work introduces a method to obtain long-term pattern forecasts and adapt them to real-time circumstances. To this end, a long-term estimation scheme based on the automated discovery of patterns is proposed and integrated with an on-line change detection and adaptation mechanism. The framework takes advantage of the architecture of evolving Spiking Neural Networks (eSNN) to perform adaptations without retraining the model, allowing the whole system to work autonomously in an on-line fashion. Its performance is assessed over a real scenario with 5 min data of a 6-month span of traffic in the center of Madrid, Spain. Significant accuracy gains are obtained when applying the proposed on-line adaptation mechanism on days with special, non-predictable events that degrade the quality of their long-term traffic forecasts.