Neural network assessment of thermal stresses in the pipelines of a VVER reactor plant

Abstract

Background

Transient processes occurring at operating VVER reactors contribute to the overall damage to equipment and pipelines. The damage is determined by stresses calculated using indirect sensor data at control points, yet with assumptions and a large margin. Thus, the development of more advanced methods for determining stresses is of great interest.

Aim

To use a neural network for determining stresses at arbitrary points of equipment and pipelines based on readings from external thermocouples and pressure in the coolant circuit.

Materials and methods

The present study applies a neural network approach to solve the inverse problem of thermoelasticity. The developed neural network establishes a relationship between measured and predicted values. The stress range is selected as the criterion for calculation accuracy.

Results

We determined the stress values based on the readings of external surface thermocouples at control points for a number of pipelines in the primary circuit of the VVER reactor plant and branch pipe connection to the main circulation pipeline. The difference between predicted values of stress ranges and those obtained in solving the direct problem falls within 10%. The main approaches to increasing the stability of the resulting solution with insufficient quality of input data are considered.

Conclusion

Neural networks of simple configuration can be used in the monitoring system, since they quickly and quite accurately calculate the stresses in the pipelines of the VVER reactor plant. The proposed approach has great potential for further development and application at NPPs. The work for determining the optimal hyperparameters of the neural network should be carried out to improve its predictive ability.