Session: 08-01: New Facility Planning/ Environmental Management (EM)/ Health & Safety

Paper Number: 107995

107995 - Regularization Parameter Identification for Nonlinear Inverse Modeling of Atmospheric Emissions 

Inverse modeling technique, relying on both environmental observations and atmospheric transport simulations, has been widely developed to estimate the release rate of airborne contaminants, such as chemicals, volcanic ash, greenhouse gases and radionuclides. The behaviors of inverse modeling would be constrained by the spatiotemporal coverage of the observations. However, the environmental observations are typically spatiotemporally sparse because of the poor network of measurement sites, the lengthy minor interval of sampling, and the frequent changes in the transport direction of pollutants. The temporal absence of observations is one of the signs, which may cause release missing in the estimate as observed in previous studies based on the Fukushima accident, endangering corresponding consequence assessments.

The subjective approach to supplement the lost information because of temporal absence is combining various forms of observations, meteorological monitoring and event information, based on expert judgments. This approach has stringent requirements for the type and quality of the supplied data and necessitates manual repetitive comparison. The objective approach for compensating for lost information is introducing additional a priori information or features about releases through regularization. Existing methods are designed to suppress the ill-posed nature of inverse modeling and, therefore, remove the artificial oscillations of release rate, rather than recovering the missing releases. For instance, assuming the release rate follows a Gaussian distribution to penalize excessive release rate, follows semi-Gaussian or lognormal to pose a positive constraint. Additionally, the assumptions of smoothness and sparsity have been used; the former blurs sharp releases while the latter reduces the amount of temporal releases. Adding an a priori release rate may enhance the inverse modeling but the construction is challenging.

Targeting the incomplete observations, we provide a total variation (TV) regularized inverse modeling to adaptively compensates for information loss in an objective way without an a priori release rate or release process assumptions. This method poses a piecewise-constant feature to the estimated release rate, which is controlled by the regularization parameter. However, the nonlinear property of the TV term challenges a fast automatic selection method for the parameter, which may cause over-weighted release features and confuse the real release. In this study, we applied four methods to choose the optimal parameter for both the nonlinear Tikhonov method and the TV regularized inverse modeling, including the discrepancy principle, generalized cross-validation, L curve method, and a method based on minimizing the cost function with a series of parameter values. The performance of different selection methods was investigated to estimate the Cs-137 release rate in the Fukushima accident with the hourly suspended particulate matter (SPM) observations of Cs-137. Quantitative and Qualitative evaluations were performed in terms of the temporal estimated release rate and resulting air concentration simulations.

Presenting Author: Xinwen Dong Institute of Nuclear and New Energy Technology, Tsinghua University

Presenting Author Biography: Xinwen Dong, Ph.D. student, Institute of Nuclear and New Energy Technology (INET) of Tsinghua University in Beijing, China. Supervised by associate professor Sheng, Fang. Recent work focuses on air dispersion modeling, and source inversion towards spatiotemporal incomplete observations.