Deep learning-based enhancement of integrated process control performance in nuclear reactor operations

This study proposes a hybrid intelligent Proportional Integral Derivative (PID) control approach that dynamically adjusts proportional (K_p), integral (K_i), and derivative (K_d) gains based on the transient severity level to enhance control performance in a nuclear reactor. The control approach is demonstrated through reactivity insertion and primary pump failure transients in a thermal-spectrum molten salt reactor (MSR) operating at 1 MW_th. It utilized deep learning-based predictive models to estimate power peak, time to peak, and power ramp rate, which served as indicators of transient severity and guided real-time control adjustments. To ensure reliability and accuracy, uncertainty quantification (UQ) methods, including deep ensembles, Monte Carlo (MC) dropout, and Bayesian Neural Networks (BNN) last layer, were applied to assess prediction confidence. Results obtained from the predictive models were aligned with results obtained from simulations using the System Dynamics Analysis Tool (SDAT). Only minor uncertainties were observed across all models. By integrating predictive models, the proposed deep learning-based PID controller method outperformed the genetic algorithm-tuned PID controller, particularly in mitigating power fluctuations during pump failure transients. The findings demonstrate the feasibility and effectiveness of using a deep learning-based PID control approach in maintaining reactor stability and improving reactor safety under transient conditions.