Storage of large-scale and multivariate power data faces challenges such as low efficiency and insufficient memory capacity. Traditional data storage optimization methods like data index and data compress methods have their own advantages and disadvantages. How to effectively apply these methods to power data storage is a core problem in current research. To solve the problem, a storage optimization decision method for multivariate power data that integrates random forest algorithm and Q-learning algorithm was proposed. Key technologies of this method were these as follows. Firstly, a storage optimization strategy decision model based on an improved random forest algorithm was developed. By the introduction of information gain method, the model could evaluate the data access frequency, query time, storage speed and data redundancy rate during database data storage, and made choice on storage optimization strategies like direct data storage, data index storage, and data compress storage. Secondly, a data storage algorithm decision model based on the improved Q-learning algorithm was proposed. By the introduction of multi-scale learning mechanism, prioritized experience replay mechanism and positive/negative reward mechanism, the model could determine the suitable index algorithm for data index storage and the compress algorithm for data compress storage. The proposed method effectively integrates the technological advantages of data index and data compress storage, can improve data storage efficiency and save storage space significantly, and thus provide a new solution for the management of large-scale and multivariate power data.