Nuclear power has significant potential in many countries as a reliable power source with a small carbon footprint. To realise this potential, safety (and also the public perception of safety) is of the utmost importance, and both existing and new design nuclear power plants strive to improve safety, maintain availability and reduce the cost of operation and maintenance. Moreover, plant life extensions and upgrades push the demand for new tools for diagnosing faults in nuclear power plants. Current approaches for diagnosis and prognosis, which rely heavily on operator judgement on the basis of online monitoring of key variables, are not always reliable. This project brought together three UK Universities (Leeds Beckett, Portsmouth, and Liverpool) and the Bhabha Atomic Research Centre in India to directly address current challenges in developing online monitoring tools for nuclear power plants.
To perform diagnosis tasks, the project made use of artificial intelligence tools, where mathematical algorithms that emulate biological intelligence are used to solve difficult modelling, decision making and classification problems. This involved optimising the number of inputs to the models, finding the minimum data requirement for accurate prediction of possible untoward events, designing experiments to maximise the information content of the data, training, validating and analysing the models. We used neural network models to diagnose loss of coolant accidents and pinpoint their severity. The accuracy and robustness of various neural network models was studied and analysed. The prediction of possible radioactive release for various accident scenarios, and, in order to facilitate emergency preparedness, the post release phase, was modelled to predict the dispersion patterns for the scenarios under consideration.