Data-Driven Predictive Control of a Nonlinear CSTR Process

Continuous stirred tank reactors (CSTRs) are challenging to control because of their nonlinear dynamics and the strong interaction between concentration and temperature. These challenges become more pronounced when operating conditions vary and reaction kinetics are uncertain, which often limits the effectiveness of conventional control strategies. In this study, a data-driven predictive control approach is developed for a generic nonlinear CSTR by integrating a Long Short-Term Memory (LSTM) neural network within an MPC framework. A benchmark exothermic CSTR described by coupled mass and energy balance equations with Arrhenius-type kinetics and jacket heat exchange is used as a reference process. Reactor concentration and temperature are selected as the state variables, while the coolant temperature serves as the manipulated input. The first-principles model is employed to generate operational data and to evaluate closed-loop performance. Dynamic simulation data are generated over 300 min with a sampling time of 0.5 min and cover multiple operating regions. Disturbances in feed concentration (±10%) and feed temperature (±5 K), together with 1% measurement noise, are introduced to reflect realistic operating conditions. An LSTM network with two hidden layers of 32 units each is trained to perform multi-step prediction of reactor states. On unseen test data, the model achieves root-mean-square errors of approximately 0.02 kmol/m³ for concentration and 2.0 K for temperature. The trained LSTM is embedded into an MPC scheme with a prediction horizon of 10 steps and explicit input and temperature constraints. Closed-loop simulations indicate that the proposed LSTM-based MPC improves set-point tracking and disturbance rejection compared with conventional PID control and nominal model-based MPC, while achieving reduced overshoot and faster stabilization. The results suggest that data-driven predictive control provides a practical alternative for nonlinear CSTR systems when accurate mechanistic models are difficult to obtain.