Two dimensional chaotic mapping zebra optimization algorithm in polar coordinate system for debutanizer column feature selection and prediction model

Abstract

Soft-sensor technology predicts quality variables hard for hard instruments to measure directly, but input variables have redundancy. Thus, feature selection is needed to reduce data dimensionality and improve prediction accuracy. Aiming at data feature selection and prediction model solution in debutanizer column production, a two-dimensional chaotic mapping zebra optimization algorithm (ZOA) in polar coordinate system is proposed. Firstly, coupled and uncoupled polar coordinate two-dimensional chaotic convergence factors are designed to optimize foraging stage for dynamic balance of exploitation and exploration. Then, the position information of attacked individuals and a sinusoidal increment term are introduced to improve the defense stage, avoiding blind search and breaking the limitation of the original algorithm’s declining late-stage search capability. In experiments, input feature signals are first decomposed by CEEMDAN, then the improved ZOA selects the optimal feature subset for prediction. Simulation experiments include three parts: the first two determine optimal variants ZOA-cρ2 and ZOA-ucρ2 via CEC2022 test functions; the last compares the proposed algorithm with other intelligent optimization algorithms on the debutanizer column dataset. Results show that compared with predictions using only original input features, the prediction results of the feature subset selected by ZOA-cρ2 decrease by 99.86 %, 96.24 % and 96.26 % in MSE, RMSE and MAE, respectively, and increase by 84.63 % in R², proving ZOA-cρ2 has great advantages in solving feature selection and prediction model tasks.