Abstract: The switching performance of the solenoid valve in a 1 N thruster assembly is a critical factor influencing the reliability of satellite propulsion systems. Excessively long closing times observed during low-temperature tests can result in project delays and cost overruns. Since conventional testing methods are unable to detect anomalies in advance, a high-accuracy prediction model is essential for early risk identification. In this study, production test data from 194 sets of 1 N thruster assemblies were used to develop back propagation (BP) and convolutional neural network (CNN) models. Twenty-one key parameters of the solenoid valve were selected as inputs, with the low-temperature closing time as the output. The results indicate that the BP neural network model achieved significantly higher prediction accuracy than the CNN model, suggesting that its fully connected structure is better suited for small-scale, well-characterized engineering data. These findings demonstrate the strong potential of neural network algorithms for predicting the performance of aerospace components, providing data support for parameter optimization and risk mitigation, thereby improving development efficiency and product reliability.