适用于电力场景的人工智能中台技术研究与探索

Research and Exploration of Artificial Intelligence Middle Platform Technology Suitable for Power Scene

摘要:

目前电力人工智能技术在电力各业务领域都有一定的应用成果，但大多在业务应用层面，缺少对人工智能技术系统级的解决方案。文中对人工智能在电力行业应用落地存在的问题进行探讨，给出了解决办法。针对样本收集面临数据分散、收集困难的情况，一方面建设统一平台进行样本收集，使得各地样本收集快速、简便；另一方面引入数据回流思想，将推理侧检测的数据传回样本收集平台，实现样本筛选、收集流程自动化。对于数据标注工作量大的问题，提出了主动交互式标注技术，实现样本数据智能标注。对于模型训练样本量少的问题，引入迁移学习的思想，采用预训练模型，在不影响模型效果的同时，还减少模型训练时间。对于模型迁移至边端设备，因边端设备架构、模型框架造成模型移植性差的问题，基于开放神经网络交换（ONNX）实现不同目标架构的模型转换，解决硬件兼容的问题，提升模型的复用性。

Abstract:

Recently, artificial intelligence technology in power system has some application achievements in various electric business fields, but most of them are at the application level, lacking system level solutions. The problems existing in the application of artificial intelligence in power industry are discussed, and the solutions are given. Aiming at the situation that the sample collection is faced with data dispersion and collection difficulties, on the one hand, a unified platform is built for sample collection to make it fast and simple,on the other hand, the idea of data backflow is introduced to collect the data detected on the reasoning side to the sample collection platform, which realizes the automation of sample screening and collection process. Since data annotation is a laborintensive work, an active interactive annotation technology is proposed to realize the intelligent annotation of sample data. For the problem of small sample size of model training, the idea of transfer learning is introduced, and the pretraining model is adopted, which not only does not affect the effect of the model, but also reduces the training time of the model. For the model migration to edge devices, the poor portability of the model is caused by the edge device architecture and model framework. The model conversion of different target architectures is realized based on open neural network exchange(ONNX) to solve the problem of hardware compatibility and improve the reusability of the model.