With the acceleration of the construction of new power systems, the degree of informatization of power grids has continued to deepen. Only relying on traditional task scheduling-type replica management strategies, the huge data center composed of tens of thousands of virtual machines can no longer meet the needs of big data processing latency of new power business such as precise load control in the future. In this regard, based on the full consideration of network traffic and data center location distribution, this paper constructs a stochastic configuration network (SCN)-based adaptive replica management system of power big data. Meanwhile, a C-means clustering-based classification of underlying equipment and network traffic prediction is proposed to complete the real-time evaluation of database network resources. On this basis, in order to effectively improve the efficiency of power big data replica management and reduce data processing latency, a novel power system-oriented data storage and selection replica management algorithm is proposed to realize the flexible storage and optimal selection of power big data copies. Finally, a pilot test was carried out in the corresponding provincial companies. The algorithm can effectively handle the storage of power big data, reduce the data processing latency, and reduce the completion time of field operations across distributed DCs by an average of 12.19%.