Research and improvement of deep learning tool chain for electric power applications

With the promotion of artificial intelligence technology in various industries in China, deep learning technology driven by electric power data has begun to be applied on a large scale in the electric power industry, but in actual application we have also found some problems, for example, the current process and tools, seemingly clear and independent, but are actually interdependent and not flexible enough; for common requirements such as retrieval, functions are not flexible enough and often re-developed. Accumulated algorithm exploration will cause the same sample to appear repeatedly in multiple training sets, The waste of resources is serious. This article elaborated on the causes and solutions to solve these problems, redesigned the process and architecture, and developed software tools. First of all, this article believes that deep learning models have a life cycle, and proposes to reorganize the deep learning tool chain with the model warehouse as the center. Secondly, combined with the actual application of electric power AI, we decoupled the original data, labeled data, and sample definitions. Finally, we designed an expression of public functions such as retrieval, taking into account flexibility and versatility.