An Optimized Control Strategy Based on Multidimensional Feature Operation Pattern

The popular data-driven control algorithm such as iterative learning control (ILC) and reinforcement learning control (RLC) is inefficient in the continuous chemical processes for the batchwise feature or dependence on the process mechanism. To quickly and accurately realizing the real-time control completely based on historical data, a data-driven optimized control strategy based on multidimensional feature operation pattern is proposed. To extracting the depth information of historical data, a multidimensional feature operation pattern presented by a 2-D matrix with seven trend indicators is defined and thus characterizing the running states of the process more accurately. The optimized multidimensional feature operation pattern library is then constructed by screening the historical optimal control performance. Furthermore, an improved Levenshtein distance (ILD) is proposed to indicate the similarity of two multidimensional feature operation patterns. By the LD between the current feature operation pattern and the one in operation pattern library, the optimized operation strategy is matched and found. The feasibility of the proposed control strategy is verified by the simulation of Tennessee Eastman (TE) process.