Model predictive control of anaerobic digestion processes using a long short-term memory network predictor

Area of Science:

• Biochemical Engineering
• Process Control
• Artificial Intelligence

Background:

• Anaerobic digestion (AD) processes are crucial for biogas production but challenging to control due to their complexity and nonlinear dynamics.
• Traditional control methods often rely on detailed mechanistic models that require extensive parameterization and may not be robust to process variations.
• There is a need for data-driven control strategies that can utilize readily available online measurements for efficient AD process management.

Purpose of the Study:

• To develop and validate a Model Predictive Control (MPC) strategy integrated with Long Short-Term Memory (LSTM) networks for controlling anaerobic digestion (AD) processes.
• To evaluate the performance of the LSTM-MPC approach using readily available online measurements in simulated AD environments.
• To assess the computational feasibility of the proposed control strategy for real-time applications.

Main Methods:

• Implementation of a Model Predictive Control (MPC) framework.
• Utilizing Long Short-Term Memory (LSTM) networks as internal predictive models trained on simulated AD data.
• Testing the integrated LSTM-MPC strategy in two simulated environments: Anaerobic Model No. 2 (AM2) and Anaerobic Digestion Model No. 1 (ADM1).
• Introducing stochastic disturbances and nonlinear operating conditions to evaluate robustness.

Main Results:

• LSTM networks were successfully trained to predict methane flow rates, even with stochastic disturbances.
• The integrated LSTM-MPC framework demonstrated robust setpoint tracking for methane flow rates in both AM2 and ADM1 simulations.
• The control strategy maintained process stability under nonlinear operating conditions and influent disturbances.
• Computational requirements were found to be feasible for real-time implementation in AD processes.

Conclusions:

• The LSTM-MPC strategy presents a promising, data-driven, and computationally efficient alternative for controlling anaerobic digestion (AD) processes.
• This approach offers a practical solution compared to traditional mechanistic model-based methods, requiring less complex measurements.
• The validated strategy enhances AD process stability and performance through effective methane flow rate control.