TRANSFER LEARNING FOR MONITORING OF THERMAL HYDRAULIC PROCESSES

Thermal hydraulic systems are important and useful in several engineering applications and disciplines, especially those related to energy production, as well as in industrial processes. These systems are used for heat transfer, energy conversion, cooling, heating, ventilation, etc. They are especially critical for the safe and efficient operation of nuclear reactors. Thermal hydraulic systems can help manage heat generated during nuclear fission, ensuring that the reactor operates within safe temperature and pressure ranges. Proper thermal hydraulic design is crucial to prevent accidents and maintain the integrity of a reactor. In recent and past years, related research has focused on the enhancement of nuclear reactor safety with the use of sensors and intelligent systems. Sensors and other instrumentation and control (I&C) entities are used for protection, monitoring and control processes, all of which play an important role for a system’s safety and security. Lately, the incorporation of artificial intelligence (AI) and machine learning (ML) methods for monitoring of systems, including nuclear reactors, has been explored by the research community. AI and ML techniques for automation of monitoring tasks in reactors have shown to lead to a plethora of advantages, including but not limited to, reduction in operation and maintenance (O&M) costs, early detection of faults, or recalibration of equipment before the scheduled maintenance. AI and ML approaches are data-driven, meaning that they rely on data produced by a system to learn and produce results. Since they are only limited to input or output data, they do not require detailed knowledge of a system, which preserves the privacy of any facility. Sensors and other equipment installed in systems such as nuclear reactors record huge amounts of data, which can be used to train ML models. The data-driven nature of AI and ML models moreover makes them crosscutting, which means that they are possibly applicable to different systems or equipment. A scientific term to describe this is called transfer learning (TL), which pertains to the transfer of knowledge between domains. TL has been explored by the scientific community in recent years. However, its use has not been thoroughly investigated for thermal hydraulic processes. This dissertation explores the application of TL in several nuclear-related thermal hydraulic facilities. First, we establish the feasibility of TL by performing real-time monitoring of thermocouple sensors, and determine the limits of TL applicability in this context by studying the correlations between prediction error and flow. Measurement data are obtained, in two separate experiments, in a flow loop filled with water and with liquid metal Galinstan. We develop long short-term memory (LSTM) recurrent neural networks (RNN) for sensor predictions by training on the sensor’s own prior history, and transfer learning LSTM (TL-LSTM) by training on a correlated sensor’s prior history. Sensor cross-correlations are identified by calculating the Pearson correlation coefficient of the time series. The accuracy of LSTM and TL-LSTM predictions of temperature is studied as a function of Reynolds number (Re). The root mean squared error (RMSE) for the test segment of time series of each sensor is shown to linearly increase with Re for both water and Galinstan fluids. Using linear correlations, we estimate the range of values of Re for which RMSE is smaller than the thermocouple measurement uncertainty. For both water and Galinstan fluids, we show that both LSTM and TL-LSTM provide reliable estimations of temperature for typical flow regimes in a nuclear reactor. The LSTM runtime is shown to be substantially smaller than the data acquisition rate, which allows for performing estimation and validation of sensor measurements in real time. In another work, we benchmark the performance of LSTM network ML model and Autoregressive Integrated Moving Average (ARIMA) statistical model in temporal forecasting of distributed temperature sensing (DTS), to determine training data selection strategies for improving a model’s performance. Data in this study consists of fluid temperature transient measured with two co-located Rayleigh scattering fiber optic sensors (FOS) in a forced convection mixing zone of a thermal tee. We treat each gauge of a FOS as an independent temperature sensor. We first study prediction of DTS time series using Vanilla LSTM and ARIMA models trained on prior history of the same FOS that is used for testing. Next, we investigate zero-shot forecasting (ZSF) with LSTM and ARIMA trained on history of the co-located FOS only, which is advantageous when limited training data is available. The ZSF MaxAPE and RMSPE values for ARIMA are comparable to those of the Vanilla use case, while the error values for LSTM increase. We show that in ZSF, performance of LSTM network can be improved by training on most correlated gauges between the two FOS, which are identified by calculating the Pearson correlation coefficient. Performance of ZSF LSTM can be further enhanced through TL, where LSTM is re-trained on a subset of the FOS that is the target of forecasting. In the last part of this dissertation, we propose to use TL to compensate for lack of training data in advanced reactors, and examine whether TL can be deployed for monitoring tasks in these systems. TL can be used to create pre-trained ML models with data from small-scale research facilities, which can then be fine-tuned to monitor GenIV reactors. In this work, we develop pre-trained Transformer and LSTM networks by training them on temperature measurements from thermal hydraulic flow loops operating with water and Galinstan fluids at room temperature at Argonne National Laboratory. The pre-trained models are then fine-tuned and re-trained with minimal additional data to perform predictions of time series of high temperature measurements obtained from the Engineering Test Unit (ETU) at Kairos Power. Performance of LSTM and Transformers is investigated by varying the size of the lookback window, to determine the optimal range for error minimization, and forecast horizon, to assess anticipatory capabilities of the models. Results of this study show that LSTM have lower prediction errors than Transformers, but LSTM errors increase more rapidly with increasing lookback window size and increasing forecast horizon compared to the Transformers errors.