Abstract: In order to realize the real-time interaction between virtual space and real space, the digital twin system of nuclear reactor was built according to the actual nuclear reactor, with a large number of sensor devices and a large amount of data, resulting in problems of large storage space demand, low transmission efficiency and high complexity of data analysis. To solve the above problems, it is necessary to optimize the sensor arrangement and achieve efficient data storage, transmission and analysis. Few studies have been made to find feature parameters while obtaining their physical meaning and maintaining high efficiency simultaneously, which is the purpose of this paper. In this study, a parameter inversion model was built for the digital twin system of space thermionic reactors by combining clustering algorithm and artificial neural network based on random search optimization, so as to realize the inversion of core temperature data under heat pipe failure conditions. The main process of the model can be grouped into two parts, feature parameter extraction and parameter inversion. In the part of feature extraction, data were reduced from high dimension to low dimension, which constituted a data reduction. The temperature data were extracted from four areas of the space thermionic reactors core from the inside out, which were fuel, emitter, receiving pole and coolant, under 20% heat pipe failure condition. Temperature parameters can be regarded as numerical vectors with the same dimension as the number of time nodes, so the clustering method can divide these temperature parameters into different categories, and the parameter closest to the center of each category was taken as the parameter with typical change characteristics. Silhouette Score is usually used for evaluating the effect of clustering, which can be used to find the proper number of clusters. Therefore, K-means clustering algorithm and Silhouette Score were used as evaluation indicators to select the feature parameters of different regions of the core. In the second part, the feature parameters were inverted to other parameters by a fully connected artificial neural network, its hyper parameters optimized by random search configuration. The optimizing process was executed by setting a range for different hyper parameters like the number of hidden layers, the nodes of each layer and the learning rate, and then choosing a combination that had the smallest loss. Then the feature parameters were used as input and inverted to the rest as output. The inversion model was then verified with the temperature data of 15% heat pipe failure condition. The results show that by using the parameter inversion of fuel, emitter, receiving pole and coolant, the relative error of temperature inversion value is 0.55%, 0.41%, 0.19% and 0.18% respectively, and the feature parameters used for inversion are no more than 8% of the total parameters. Therefore, the parameter inversion model established in this study can obtain the physical meaning of the feature parameters and reverse the space reactor core temperature parameters with high accuracy, therefore resisting the influence of probe damage and data loss. The application of this method in the digital twin system of nuclear reactor can also provide an evaluation basis for the future installation of detectors in small nuclear reactors.