An adaptive-margin support vector regression for short-term traffic flow forecast

The day-to-day volatility of traffic series provides valuable information for accurately tracking the complex characteristics of short-term traffic such as stochastic noise and nonlinearity. Recently, support vector regression (SVR) has been applied for short-term traffic forecasting. However, standard SVR adopts a global and fixed ε-margin, which not only fails to tolerate the day-to-day traffic variation, but also requires a blind and time-consuming searching procedure to obtain a suitable value for ε. In this work, on the ground of stochastic modeling of day-to-day traffic variation, we propose an adaptive SVR short-term traffic forecasting model. The time-varying deviation of the day-to-day traffic variation, described in a bilevel formula, is integrated into SVR as heuristic information to construct an adaptive ε-margin, in which both local and normalized factors are considered. Comparative experiments using field traffic data indicate that the proposed model consistently outperforms the standard SVR with an improved computational efficiency.