Short-term prediction of traffic parameters is very important for proactive transportation management. The past two decades have witnessed many forecasting models being developed, yet none of them could consistently outperform the others. Layered predicting approaches take a 'divide and conquer' strategy by classifying traffic flow series into simpler regimes so that regime-specific models can be trained to handle simplified traffic states. However, traffic regimes are inherently drifting for real world traffic systems. In this paper, a layered real-time neural network ensemble learning approach is developed to address the nonlinearity and non-stationarity of traffic data. A flexible dynamic weighted ensemble forecasting system that combines adjustable numbers of neural network predictors according to their real-time prediction performance is implemented. Experimental results demonstrate that the proposed approach can effectively deal with the changing traffic regimes and produce satisfactory online short-term traffic forecasts.