1:45 PM - 3:15 PM
[MGI26-P03] Parameter Estimation of Local Particle Filter by Bayesian Optimization
Keywords:Data assimilation, Machine learning, Non-Gaussian, Parameter estimation
Existing data assimilation methods such as the 4D-Var and Ensemble Kalman Filter (EnKF) assume that the forecast error follows a Gaussian distribution. However, around convective clouds and low-pressure systems that bring torrential rainfall, the forecast error tends to be non-Gaussian due to strong nonlinearities, and therefore, the optimal analysis values is not always obtained.
On the other hand, the particle filter (PF) that does not assume a Gaussian distribution may provide optimal analysis values. However, PF requires a large amount of computational resources for the convenience of resampling. The local particle filter (LPF) is one of the methods to operate PF with limited computational resources, and it addresses this problem by increasing the number of apparent particles (ensembles) through localization.
However, LPF has high parameter sensitivity, and if the localization and inflation are not properly tuned, the accuracy of the analysis will be greatly degraded. In this study, we constructed a data assimilation system consisting of the Lorenz-96 model (L96) and LPF, then investigated how the accuracy of the analysis changes by estimating its parameters through Bayesian optimization.
The results showed that when inflation is optimized offline, the estimation converged to the optimal value by manual tuning, and the accuracy of the analysis improved. We will present the newest results up to the time of the meeting.
On the other hand, the particle filter (PF) that does not assume a Gaussian distribution may provide optimal analysis values. However, PF requires a large amount of computational resources for the convenience of resampling. The local particle filter (LPF) is one of the methods to operate PF with limited computational resources, and it addresses this problem by increasing the number of apparent particles (ensembles) through localization.
However, LPF has high parameter sensitivity, and if the localization and inflation are not properly tuned, the accuracy of the analysis will be greatly degraded. In this study, we constructed a data assimilation system consisting of the Lorenz-96 model (L96) and LPF, then investigated how the accuracy of the analysis changes by estimating its parameters through Bayesian optimization.
The results showed that when inflation is optimized offline, the estimation converged to the optimal value by manual tuning, and the accuracy of the analysis improved. We will present the newest results up to the time of the meeting.