NUMERICALLY OPTIMIZED EMPIRICAL MODELING OF HIGHLY DYNAMIC, SPATIALLY EXPANSIVE, AND BEHAVIORALLY HETEROGENEOUS HYDROLOGIC SYSTEMS- PART 2

摘要

Modeling highly dynamic hydrologic systems on a spatially expansive scale is challenging because behaviors vary discontinuously both spatially and temporally. Building representative empirical models requires sufficient data to capture diverse causes and effects. Measured variables may be either categorical or dynamic (time series or signals), and relations between predictor and response variables are often nonlinear. Previous efforts described a modeling approach and two applications that use a sequence of numerically optimized data mining algorithms. "Time series clustering" can optimally segment a large collection of signals into "classes" that are dynamically similar. Artificial neural networks are a multivariate, nonlinear curve fitting technique that can optimally model each class's behavior. The first application was a model that auto-regressively generated spatially continuous aquifer water level predictions from a finite set of water level signals and the spatial coordinates of their monitoring sites. The second application predicted hourly stream temperatures in western Oregon from climate signals and categorical stream habitat and basin attributes. This paper describes an application in Wisconsin that, like the Oregon model, predicts stream temperatures from climate signals and categorical site attributes. However, it is conceptually different from the earlier applications, which used concurrently measured signals. Most of the Wisconsin stream temperatures were measured at different times over 13 years. This required a new time series clustering algorithm that would still segment the signals according to their dynamic.