A number of researches have been carried out so far on estimation of the Departure from Nuclear Boiling Ratio (DNBR) in a nuclear reactor. However, they were mainly static approaches and therefore, have not considered temporal changes due to the dynamics of a reactor and correlation of major parameters in the reactor.

In this paper, a Continually Running Fully Recurrent (CRFR) network is developed to predict and capture the temporal variations of the DNB ratio, paired with a reactor core simulator which simulates the neutronic and thermal hydraulic dynamics of a 40MW pressurised heavy water research reactor. The core simulator is coupled to the CRFR and they interchange data in every time step resulting in a real time prediction of DNBR rather than a static estimation as in previous methods.

COBRA III-C code is used for data acquisition and generalisation tests but the critical heat flux routine has been modified to be used for the author’s special low pressure heavy water case. Several surveys have been conducted to achieve the best correlation for heavy water properties. Results are indicating a good prediction in many transients, varying from multiple failure of a heat exchanger and pressuriser as well as LOCA accident to a simple positive reactivity insertion

The proposed learning algorithm has the advantage that it does not require a precisely defined interval and operates whilst the network is running. This powerful architecture could be used for other online modeling of real time complex and nonlinear systems.