<p>In the process of modern industrial production, it is difficult to realize real-time monitoring effectively because of economic or technical constraints. To mitigate these problems, soft sensing techniques are used to predict target variables that are difficult to measure directly. Zebra optimization algorithm based on chaotic convergence factor and Gaussian variation is proposed to optimize the parameters of MLP soft sensor model of debutanizer column. Firstly, the designed chaotic convergence factor is introduced into the foraging behavior stage of zebras, and it is used to search near each potential solution. The introduction of chaotic convergence factor is conducive to the dynamic development and exploration of the algorithm, which makes the search strategy of the algorithm more flexible, and can accelerate the convergence speed while ensuring the solution accuracy. Then, at the end of each iteration of the algorithm, Gaussian variation is carried out for each individual in the population to improve the diversity of the population and increase the probability of the algorithm jumping out of the local optimal. The simulation experiment is divided into three parts. Firstly, CEC2022 test function is used to test the performance of ZOA with different chaotic convergence factors, and the best variant c5SinZOA is determined. Then, c5SinZOA is compared with other intelligent optimization algorithms (golden sine algorithm, whale optimization algorithm, coatis optimization algorithm, human evolution optimization algorithm, goose optimization algorithm) to verify the effectiveness of the improved strategy and the superiority of c5SinZOA. Finally, c5SinZOA and the above comparison algorithm are used to optimize the model parameters of multi-layer perceptron (MLP) to obtain the soft-sensing model of the production process of butane tower. From the quantitative index of simulation results, compared with other algorithms, MSE of c5SinZOA is lower than 0.0008–0.0059, RMSE is lower than 0.0038–0.0256, MAE is lower than 0.0008–0.0129, R² is higher than 0.0255–0.1864. It has obvious advantages in prediction accuracy and model fitting effect, and can better complete the prediction task of the soft sensor model in the production process of debutanizer column.</p>