Soft sensing technology is an effective way to solve the problem that important quality indicators of processing industries cannot be detected online, especially in the chemical industry. Owing to the complex working conditions, strong nonlinearity, strong coupling, and time-varying characteristics of chemical production processes, how to establish a soft sensing model with good prediction performance has become a valuable research topic. A soft sensing model based on a single-model method cannot guarantee global prediction accuracy, and the model stability is poor. A hybrid modeling method can integrate different modeling methods to describe the process characteristics of an object more comprehensively, so as to significantly improve the prediction accuracy and stability of the soft sensing model. In this paper, the key process parameter (solid-liquid ratio) in the evaporation salt (ES)-making process is taken as an example to carry out the following research. Firstly, aiming at the problems of production data obtained from the chemical industry, such as missing values, data inconsistency, high dimensions, high correlation, and time-series characteristics of features, an effective feature extraction method is proposed. On this basis, two data-driven models, the deep neural network (DNN) model for non-temporal regression prediction and the long short-term memory neural network (LSTM) model for temporal regression prediction, are established, and the regression performance of these two soft sensing models is evaluated. Secondly, another feature selection method based on prior domain knowledge, expert experience, and data mining is proposed. On this basis, a hybrid soft sensing model, the LightGBM model, is constructed for key process parameter prediction under different feature inputs, and the regression performance is evaluated. Simulation results demonstrate that introducing domain knowledge and expert experience to the modeling can enhance the interpretability of models, simplify the molding process, and further improve model performance.