A Numerical Study of Accounting for Structural Error and Uncertainty in Environmental Models
Zhulu Lin1, Bruce M. Beck2, Gang Shen3
1Department of Agricultural and Biosystems Engineering, North Dakota State University, Fargo, ND, United States; 2Warnell School of Forestry and Natural Resources, University of Georgia, Athens, GA, United States; 3Department of Statistics, North Dakota State University, Fargo, ND, United States

The significance of model structure error and uncertainty (MSEU) is rarely adequately recognized. What is more, it is not a matter of little consequence to the formation of policy for environmental protection and ameliorating the prospective effects of climate change. The real challenge in dealing with structural error and uncertainty in a model lies in at least two aspects: (1) how to identify this source of uncertainty in models, including unambiguously differentiating it from other sources; (2) how to account for the model error and uncertainty in making forecasts. Many conceptual and computational methods have been proposed to tackle this challenge. However, crucial to the approach proposed herein is the treatment of the model's parameters as stochastic processes, themselves modeled as Generalized Random Walks. A known biological system, with nonlinear dynamics, is specified as the prototypical case study for assessing and comparing the performance of our proposed approach with three other, more familiar approaches to accounting for MSEU in model-generated predictions: model fitting error; the expansion of parametric uncertainty; and Bayesian model averaging. Our predictive test case is constructed around future conditions in which the sequence of input disturbances of the system's behavior is in general terms identical to the past observed pattern of input disturbances used for prior model calibration, but in specific terms significantly different. The test reveals that while our novel approach and that of Bayesian model averaging perform well, the performance of the approaches using model fitting error and the expansion of parametric uncertainty are clearly inferior.

Keywords: Model structural error and uncertainty; Model averaging; Recursive estimation; Stochastic process

Biography: Dr. Zhulu Lin is currently an assistant professor in Department of Agricultural and Biosystems Engineering in North Dakota State University at Fargo, USA. He received his Masters degree in Statistics and his Ph.D. in Soil and Water Resources Management both from University of Georgia. He is interested in applying both mechanistic and statistical modeling approaches to environmental systems analysis.