The natural gas light hydrocarbon recovery unit contains a number of key operation parameters, including the composition of the mixed-refrigerant, the temperature of cryogenic separator, plate pressure, etc. These parameters directly affect the energy consumption and product quality of the system. The relationship among these parameters are complex and interrelated, which makes it complicated to build theoretical models systematically. Based on the actual production data of a liquid hydrocarbon recovery unit, a BP neural network model for optimizing and predicting the mixed refrigerant composition and other key operation parameters was established to achieve the goal of saving energy and increasing efficiency. The model can adapt to the changes in natural gas feed and production requirements and the overall prediction accuracy was high. Most of the mean absolute percentage error (MAPE) of the output parameters was less than 5%, and the minimum error was as low as 0.118%. The output parameters with unsatisfactory prediction effects were optimized by genetic algorithm (GA). After the optimization, the error of the liquid phase flow rate of the refrigerant separator decreased from 9.208% to 3.321%, and the error of plate pressure of the demethanizer reduced from 9.602% to 4.051%. Based on the established GA-BP neural network model, the refrigerant components and liquid phase flow rate and pressure of the refrigerant separator were optimized under different feeding conditions in summer and winter. The optimization results showed that the molar fraction of methane, propane, and isobutane in mixed-refrigerant and the flow rate of refrigerant separator should be appropriately increased in summer, and the molar fraction of ethylene should be reduced. In winter, the molar fraction of isobutane and the pressure of liquid refrigerant should be properly reduced. Taking the summer feed conditions as an example, the optimization of the mixed refrigerant proportion and various operating parameters resulted in a reduction of the refrigeration system's energy consumption by 518.12 kW. The optimization of key operation parameters based on the neural network model can increase the ethane yield and reduce cross-section temperature difference of main cold box, which is of great significance for actual production process.