Forecasting river runoff through Support Vector Machines

Abstract

How “wet” or “dry” a year is predicted to be has many impacts. Public utilities need to determine what percentage of their electric energy generation will be hydro power. Good water years enable the utilities to use more hydro power and, consequently, save oil. Conversely, in a dry year, the utilities must depend more on steam generation and therefore use more oil, coal, and atomic fuel. Agricultural interests use the information to determine crop planting patterns, ground water pumping needs, and irrigation schedules. Operators of flood control projects determine how much water can safely be stored in a reservoir while reserving space for predicted inflows. Municipalities use the information to evaluate their water supply and determine whether (in a dry year) water rationing may be needed. Currently a combination of linear regression equations and human judgment is used for producing these forecasts. In this paper, we describe a Support Vector Machine based method for river runoff forecasting. Our method uses Smola/Scholkopf's Sequential Minimal Optimization algorithm for training a Support Vector Machine with a RBF kernel. The experimental results on predicting the full natural flow of the American River at the Folsom Dam measurement station in California indicates that our method outperforms the current forecasting practices.