Adaptive sliding mode load-following control of a small modular reactor via reinforcement learning, nonlinear extended state observer, and neural network

The research on small modular reactors (SMRs) is one of the recent primary trends in the nuclear technology fields. Unfortunately, the published research on the fundamental load-following control problem of the SMR is mostly based on explicit models, which are highly restricted by model mismatches, parameter perturbations, and disturbances. To this end, this paper aims to develop a novel model-free near-optimal adaptive sliding mode load-following control (ASMLFC) strategy for the SMR, invoking the state-of-the-art sliding mode control techniques, reinforcement learning (RL), nonlinear extended state observer (NESO), and radial basis function neural network (RBFNN). The dynamic model of the SMR is first transformed into a nonlinear second-order cascade system, where all unavailable system states, nonlinear dynamics, and model uncertainties are aggregated into a fully uncertain lumped function. This function, together with the unmeasured system state, is then estimated by the proposed NESO, utilizing only the available output measurement and control input from the SMR system. After that, the model-free ASMLFC strategy, integrated with the estimates of the unknown lumped function and unmeasured system state provided by the NESO, actor–critic architecture of the RL, and approximation capability of the RBFNN, is developed, such that the near-optimal load-following control quality can be guaranteed, independently of any unpredictable model-related factors. By employing Lyapunov’s direct method, the Routh–Hurwitz criterion, and the Hamilton–Jacobi–Bellman equation, the adaptive update laws for the RL-based optimal value function and control action, along with the RBFNN weight vector, are theoretically derived. Moreover, the uniform ultimate boundedness of both the NESO estimation error dynamic and the SMR load control error under the developed ASMLFC strategy is rigorously proven. In the end, comparative simulation studies further confirm that the developed ASMLFC strategy outperforms the pre-existing optimal proportional–integral–derivative controller, nonlinear dynamic inversion controller, and model-free fuzzy neural network controller.