Abstract: The deformation data of surrounding rock in underground caverns of hydropower stations have the characteristics of uncertain changes and short sequence samples, the traditional abnormal data recognition method has high missed recognition rate and misjudgment rate.With the development of intelligent technology, it is a hot topic to establish a more reliable abnormal data recognition method through neural network.However, the traditional neural network has some problems, such as weak temporal correlation and complex calculation.Therefore, an abnormal data recognition algorithm for surrounding rock deformation of underground caverns based on time-domain convolutional neural network(TCN) and criterion adaptation was proposed in this paper.The algorithm considered the relationship between the front and back of the monitoring data sequence, and used TCN technology to establish a more reliable sequence model.At the same time, according to the characteristics of monitoring data of underground caverns, the optimal recognition criterion of adaptive matching was realized by considering three aspects of error median, data fluctuation and instrument accuracy.The algorithm was applied to recognition of abnormal data of surrounding rock deformation of underground cavern in Yebatan Hydropower Station.It was proved that the algorithm can effectively avoid the problems of gradient explosion, disappearance and time-consuming, which greatly improved the efficiency and recognition rate of abnormal value analysis.Relevant experiences can be used as reference in the recognition of abnormal monitoring data of similar projects.