TJ4.4 Improving Forecasts of Extreme Values by Machine Learning Models Using Occam's Razor

Tuesday, 9 January 2018: 11:30 AM
Room 7 (ACC) (Austin, Texas)
William W. Hsieh, Univ. of British Columbia, Vancouver, BC, Canada

Nonlinear regression models (e.g. artificial neural networks, extreme learning machines, etc.) can give very poor predictions when given predictor values which lie beyond the range used in model training, as extrapolation by nonlinear models is notoriously unreliable. Hence current nonlinear machine learning models tend not to perform well in extreme value forecasts. A new approach based on Occam's Razor is proposed, where simple linear extrapolation is used instead of unreliable nonlinear extrapolation. The linear extrapolation is built on the nonlinear regression model. This approach has been tested separately on forecasting streamflow, precipitation amount and suspended sediment concentration, resulting in improved forecasts for all three.
- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner