(695b) Reinforced Learning Based Control Algorithm for Anarobic Digestion Under Feedstock Uncertaint

Essential to creating a sustainable future is implementing and optimizing alternative technologies that can help mitigate the current carbon crisis. However, many biological systems are plagued by uncertainty due to the complexity of organic feedstocks. Anaerobic digestion systems experience this overriding uncertainty in the feedstock, especially when multiple feedstocks are present. A reinforced learning algorithm can be used to control and maintain this process to solve this issue. First, several hundred data points were collected from literature to use in this algorithm. Then, due to inconsistencies in feedstock analysis throughout literature, a linear optimization algorithm was used to determine the biomolecular composition of the feedstocks to implement in an anaerobic digestion model. From this data, random samples were taken to establish the uncertainty in our model. To control this process, two different learning algorithms are analyzed for comparison. A Markov decision process (MDP) that optimizes and controls anaerobic digestion under the established uncertainty. Also, a reinforced learning algorithm where the uncertainty is not established. Next, these algorithms are implemented into a kinetic model based on anaerobic digestion model no. 1(ADM1) to determine the target methane output for our process. To apply these algorithms, ADM1 is simulated, determining the efficiency under several conditions. These conditions are a steady-state model, a step model, and a continuously changing model. This model will allow for an alternative automated control system for this delicate process, ensuring stability in the reactor. Furthermore, these results will lead to understanding how scientists and engineers can apply reinforced learning to complex biological systems like anaerobic digestion.