Mechanism Neural Network Model For Predicting Radioactive Liquid Waste Pollution Under Accident Condition

Precise and rapid prediction of the dispersion process of radioactive liquid effluent is significant in water security management of inland nuclear power station in accident conditions. Previous studies mainly focus on coastal nuclear power station or normal operation condition of inland nuclear power station and very few researches concern this issue of inland station in accident conditions.Currently, mechanism models that describe the dispersion of radionuclide are usually based on hydrodynamic theory, which is time-consuming for iteration and has rigorous requirements of boundary conditions. Therefore, these models have poor ability in rapid prediction after accident. Meanwhile, data-driven models, which have the ability to output prediction results quickly, are black-box models. They need huge samples and cannot reflect the physical characteristics of study objects. To overcome above defects, a new kind of neural network model based on mechanism is proposed. With the proposed model, two chaotic system are taken as examples to test the effect of model prection and its factors. On the other hand, an inland nuclear power station, which will be established in Xianning, Hubei Province, China recently and its receiving water body, Fushui reservoir, were selected as study objects. The hypothetical nuclear power accident refered to the accident occurred in Fukushima, Japan,2011. The main research contents are as follows:(1) Training Samples requirement of Mechanism neural network modelThe mechanism model of radioactive liquid effluent dispersion and its calculation method were systematically discussed. Then, combined with topography of reservoir and hydrologic data, the simulation of flow field in typical year was calculated, which took inland nuclear power in Xianning and its receiving water body Fushui reservoir as study objects. A hybrid calculation method which linked steady and unsteady flow together was proposed and was applicable in conditions with limited observations. Based on the fluid field simulation result, transport of nuclides with different half-life periods in Xianning nuclear power station after accident refered to the Fukushima nuclear power accident was simulated, and the difference of influence range and duration of effect of these two nuclides were compared. Furthermore, the effects of leakage points caused by hydrogeological conditions on nuclide transport were also studied. Results indicated that the concentrations of nuclides with short half life period were 1-2 orders of magnitude lower than the nuclides with long half life period at the site of the dam. More exactly, the influence duration of short half-life period nuclides was about 50 days, while for long half-life period nuclides it was larger than 100 days. Furthermore, we found local eddy could delay nuclides elimination, and the pollutant will have some effect upstream and on tributaries. Based on simulation, effects of radioactive liquid effluent on Fushui reservoir, aquatic organism, and human health are evaluated. It can be found that the receiving water would be polluted severely. More exactly, radioactive substance concentrations were about 2550 and 25 times larger than the safety criterions in partial area and fish respectively. Analysis about radioactive substance enrichment in our human bodies indicated that eating fish was more harmful than drinking water in polluted regions.(2) Establishment and test of mechanism neural network model based on Chen and Mackey-Glass chaotic systemsIn order to make the neural network has the ability to reflect research objective’s mechanism and improve generation performance, a new network model is proposed. Mathematic function of mechanism is regard as priori knowledge, added into network model as penalty term. For more precision, we replaced gradient descent algorithm with IPSO algorithm to improve the iterate speed and the global search ability. Chen chaotic system and Mackey-Glass chaotic system are taken as calculation examples, results indicate that priori knowledge can significantly improve generalization performance of neural network model. For the three compenents of Chen system, MAPE of new neural network model reduce 21.58 、 21.2 、 1.247 than the general BP model; for Mackey-Glass system, MAPE reduces about 53%. The topological structure of network, number of training samples and learning accuracy are the factors affecting the ability of priori knowledge. When number of training samples is small and learning accuracy is low, priori knowledge performs better in improving generalization performance.(3) Establishment and test of mechanism neural network model for prediction of radioactive contaminantA physical-based priori feed forward network model was established in this paper to reduce the iteration time and overcome the inflexibility of mechanism model. We emphatically analyzed the possibility of adding component transport equation which reflects the physical mechanism of research objects to artificial neural networks (ANNs) as the priori knowledge in form of penalty function, which might increase the prediction accuracy and generalization performance of ANNs. Priori neural network model is also used in the case to predict concentration of radionuclide after nuclear accident. Results show that priori knowledge indeed has the ability to increase the prediction accuracy of ANNs. Due to reflection of physical characteristics, mechanism priori knowledge has a stronger binding force and better modeling accuracy than monotonous priori knowledge which was only generated by samples. In addition, in the prediction of the changes of nuclide concentrations in long time series, mechanism priori knowledge could also increase prediction accuracy. Relative error of the whole prediction period reduce from 243%to less than 50%. However, the accuracy would gradually decrease with time because of the error transfer and amplification with inverse normalization. Considering prediction result precision and the requirement at the first stage of accident, this model is suitable for real accident decision policy.With the research topic of the dispersion of radioactive liquid effluent of inland nuclear power station in accident conditions, a useful method for water security management by applying both theoretical research and case study was proposed. Meanwhile, the research fronts of this area in future were pointed out based on the shortage in this paper.