Probabilistic modeling and dynamic optimization for performance improvement and risk management of plant-wide operation

This study presents a novel algorithm for constructing a probabilistic model based on historical operation data and performing dynamic optimization for plant-wide control applications. The proposed approach consists of applying a self-organizing map (SOM) for identifying representative plant operation modes based on a discounted infinite horizon cost and approximate dynamic programming techniques for learning an optimal policy. A quantitative measure for risk is defined in terms of transition probability, and a systematic guideline for striking balance between risk and profit in decision making is provided with a mathematical proof. The efficacy of the proposed approach is illustrated on an integrated plant consisting of a reactor, a storage tank, and a separator with a recycle loop and Tennessee Eastman challenge problem. The algorithm is useful for learning an improved policy and reducing risk in plant operation when a plant-wide model is difficult to obtain and uncertainties affect operation performance significantly.