Online autonomous calibration of digital twins using machine learning with application to nuclear power plants

As a near-zero carbon emission energy source, nuclear energy plays an important role in the current world energy decarbonization scenario. Digital twin is a key technology for the continued development of nuclear energy applications. The digital twin requires real-time, high-precision simulations that are beyond the capabilities of current nuclear energy system simulation programs. Therefore, this study proposes an autonomous calibration method for the digital twin of nuclear power plants to compensate for the error in the results of the low accuracy digital twin that can run quickly to obtain higher accuracy results to meet both high accuracy and real-time requirements. The proposed method consists of offline and online stages. In the offline stage, digital twin simulations are first performed. The simulated data and corresponding measurements data (or real data) are used to build an error database, which will be used for the next step of data-driven model training. To reduce the complexity of calibration model, the error database samples are then grouped by clustering. Data-driven calibration models are built on each group based on the simulated data and errors. In the online stage, the digital twin runs in parallel with the nuclear power plant and receives real-time data. The calibration model is continuously updated using dynamic error database. The feasibility of the new proposed method has been demonstrated on measured data from the PKLIII B3.1 steam generator pipe rupture (SGTR) experiment. The results showed that the physical quantities such as pressure, temperature and mass flow rate were well calibrated during the 1000 s of parallel running. The R2 of all physical quantities including temperature, flow rate, and pressure are above 0.99.