电力信息与通信技术2026，Vol.24Issue(1)：34-44,11.DOI:10.16543/j.2095-641x.electric.power.ict.2026.01.04
基于混合训练模型的增强异构算力互联方法研究
Research on Enhanced Heterogeneous Computing Power Interconnection Method Based on Hybrid Training Model
摘要
Abstract
To address the challenges of heterogeneous computing power synergy in the power industry's efforts to build self-controllable large-model infrastructure,this paper proposes an intelligent synergistic network(ISN)model for grid computing power interconnection.The model achieves unified scheduling and communication adaptation for multi-vendor chips through the self-developed communication library SynCCL,designs the ISN-LB dynamic load balancing algorithm to precisely guide ECMP hashing results and avoid path conflicts,and introduces the ISN-CC global congestion control mechanism for proactive traffic regulation.Validated on a real dataset containing minute-level load data of a provincial grid from 2020-2023,provided by a state grid provincial company,the experimental results show that the ISN solution performs excellently across three typical GPU heterogeneous combinations(Biren+MetaX,Biren+NVIDIA,NVIDIA+MetaX).It achieves a maximum synergistic communication efficiency of 98.2%,a training efficiency of 94.1%,and a significant improvement in network load balancing degree.Particularly regarding power grid load prediction accuracy,the ISN solution reduces the normalized root mean square error(NRMSE)and mean absolute percentage error(MAPE)to 0.070 and 5.55%respectively,significantly outperforming comparative schemes.This effectively demonstrates the practical value of ISN in enhancing the training efficiency and business accuracy of power AI models.关键词
ISN/异构算力互联/负载均衡/集合通信/模型训练Key words
ISN(intelligent synergistic network)/heterogeneous computing power interconnection/load balancing/collective communication/model training分类
信息技术与安全科学引用本文复制引用
高昆仑,叶青河,王岳,甘津瑞,王晓辉,汪洋..基于混合训练模型的增强异构算力互联方法研究[J].电力信息与通信技术,2026,24(1):34-44,11.基金项目
国家电网有限公司总部科技项目资助(5700-202458331A-2-1-ZX). （5700-202458331A-2-1-ZX）
