Data Assimilation Based on Sequential Monte Carlo Methods for Dynamic Data Driven Simulation

Simulation models are widely used for studying and predicting dynamic behaviors of complex systems. Inaccurate simulation results are often inevitable due to imperfect model and inaccurate inputs. With the advances of sensor technology, it is possible to collect large amount of real time observation data from real systems during simulations. This gives rise to a new paradigm of Dynamic Data Driven Simulation (DDDS) where a simulation system dynamically assimilates real time observation data into a running model to improve simulation results. Data assimilation for DDDS is a challenging task because sophisticated simulation models often have: 1) nonlinear non-Gaussian behavior 2) non-analytical expressions of involved probability density functions 3) high dimensional state space 4) high computation cost. Due to these properties, most existing data assimilation methods fail to effectively support data assimilation for DDDS in one way or another. This work develops algorithms and software to perform data assimilation for dynamic data driven simulation through non-parametric statistic inference based on sequential Monte Carlo (SMC) methods (also called particle filters). A bootstrap particle filter based data assimilation framework is firstly developed, where the proposal distribution is constructed from simulation models and statistical cores of noises. The bootstrap particle filter-based framework is relatively easy to implement. However, it is ineffective when the uncertainty of simulation models is much larger than the observation model (i.e. peaked likelihood) or when rare events happen. To improve the effectiveness of data assimilation, a new data assimilation framework, named as the SenSim framework, is then proposed, which has a more advanced proposal distribution that uses knowledge from both simulation models and sensor readings. Both the bootstrap particle filterbased framework and the SenSim framework are applied and evaluated in two case studies: wildfire spread simulation, and lane-based traffic simulation. Experimental results demonstrate the effectiveness of the proposed data assimilation methods. A software package is also created to encapsulate the different components of SMC methods for supporting data assimilation of general simulation models. INDEX WORDS: Data assimilation, Dynamic data driven simulation, Sequential Monte Carlo, Particle filter, Spatial-temporal systems, Statistical inference DATA ASSIMILATION BASED ON SEQUENTIAL MONTE CARLO METHODS FOR DYNAMIC DATA DRIVEN SIMULATION