Zagirska I. Modeling and forecasting of radionuclide transfer from soil into plants using dynamic Bayesian network

The study is aimed at estimating and forecasting the transfer coefficient of radionuclide from soil to agricultural plants based on the real data collected in the areas affected by the Chornobyl disaster. The model was developed in the form of a dynamic Bayesian network, which is an element of novelty, since the use of this tool for radio-ecological modeling was not previously carried out. Also the Bayesian networks exhibit the following advantages: the possibility for using of discrete and continuous variables in the frames of the same model; the model dimensionality can be very high (hundreds of variables); today there exists highly developed alternative techniques for model structure construction and for inference performing; usually the models reflect actual causal relationships. Besides, statistical data model can include expert estimates for variables values, parameter restrictions etc. Today adaptive approaches are very popular in application to model building because of their high flexibility to possible changes in data and modeled system dynamics. To construct the model in the form of DBN we proposed special adaptive computation scheme that supposes the use of two optimization criteria directed towards maximization of total probability for the network constructed. Thus, the scheme allows performing structural and parametric adaptation with new data that is coming from the measurement system. The best model of the candidates estimated is selected on the final step of its application. The problem considered in this study is of a high priority, since the human body internal exposure is mainly caused by the presence of contaminated plants on the lower level of the food chain, and mathematical modeling of the processes is still not common in general. The factors affecting the radionuclide transfer coefficient were analyzed, and the dependencies for transfer level change were identified, depending on the humidity, acidity, soil type, depth of the root system, the content of K + and 2Ca +. The dynamic approach hired allows tracking the changes of plant contamination within the period over 80 months with a time step equal to 1 month. The junction tree algorithm was used for inference as the network consists both of continuous and discrete nodes. The results obtained demonstrate high accuracy in accordance with general criteria: the standard deviation does not exceed the value of about , mean absolute percentage error does not exceed 5,5% for all measurements, the error variance is close to zero, that justifies the use of dynamic Bayesian networks a good alternative to solve this problem. Also the possibility of this approach usage while solving problems of the same class in general was considered. The model allows creating long-term scenarios to identify the possible way to agriculture development over the areas affected by the Chornobyl disaster and similar anthropogenic disasters. On the basis of data processing techniques, models constructed and the set of criteria used we developed decision support system allowing to substantially decrease the time required for constructing the best possible model and to generate alternative decisions. Further extension of functionality of the system developed is easily possible by using new techniques of intellectual data analysis and data-mining approach together with modern statistical data processing techniques. Keywords: radionuclide contamination of soil, mathematical modeling, dynamic Bayesian network, probabilistic inference.