| Accurate traffic forecasting is of great interest for commercial, security, and energy ef- ficiency applications. In this work we focus on forecasting the movement of people in a building and vehicles on a roadway. Traditional forecasting methods use statistical models, learned from historical data. However, these methods fail during the presence of anomalies. In such cases, the forecast can deviate significantly from historical averages. We have de- veloped a method to recognize the occurrence of an anomaly, and adjust the forecast to take this into account. Our approach is to first train a background model to forecast the data. In order to do this, we developed a new ensemble predictor that takes outputs from a set of models, thus achieving a more accurate result than any individual model alone. We then look for intervals where large deviations occur, between the forecasted data and the actual data. Clustering and modeling these yields a number of potential anomaly models. We then train a classifier to recognize the anomalies as they begin to occur. We then intro- duce a novel method to combine the anomaly-based forecasts using a Bayesian approach, to produce a more accurate ensemble forecaster. We demonstrate the efficacy of the approach on actual building sensor datasets, as well as a vehicle traffic dataset, and show that the new approach incorporating anomaly detection achieves significantly better accuracy than the ensemble forecaster without anomaly detection. |
