上海交通大学学报

上海交通大学学报

• •    下一篇

计及算力需求响应的神经分支电-算网快速优化方法（网络首发）

  

  1. 1. 上海交通大学国家电投智慧能源创新学院；2. 上海非碳基能源转换与利用研究院； 3. 上海交通大学电力传输与功率变换控制教育部重点实验室；4. 西安交通大学电气工程学院；5. 国网四川省电力公司
  • 基金资助:
    国家电网公司总部科技项目(5108-202226031A-1-1-ZN)

Neural branching power-computation network fast optimization method considering computing demand response

  1. 1. College of Smart Energy, Shanghai Jiao Tong University, Shanghai 200240, China；2. Shanghai Non-Carbon Energy Conversion and Utilization Institute, Shanghai 200240, China；3. Key Laboratory of Control of Power Transmission and Conversion of the Ministry of Education, Shanghai Jiao Tong University, Shanghai 200240, China；4. School of Electrical Engineering, Xi 'an Jiaotong University, Xi’an 710049, China；5. State Grid Sichuan Electric Power Company, Chengdu 610041, China

PDF (PC)

14

摘要： 数据中心的快速发展使其可以作为需求响应参与电力系统调度，通过在区域间调度数据中心内算力资源能够实现节能减排、节约成本的目的，但在电力系统调度中考虑数据中心算力资源的需求响应面临计算速度不足的问题，基于此提出了计及算力需求响应的神经分支电-算网快速优化方法。首先建立考虑算力资源需求响应的电-算网双层优化模型，其次结合图卷积神经网络与分支定界法，应用于双层模型中。通过历史数据训练，计及算力需求响应的神经分支电-算网快速优化方法具备快速确定分支定界变量顺序、最小化迭代次数的能力，显著提高求解速度，实现考虑数据中心算力资源的机组组合需求响应高速求解。在“东数西算”工程仿真场景中验证所提方法性能，与伪成本分支算法相比，平均缩短了求解时间39.1%；与商用求解器CPLEX相比，平均缩短了求解时间38.1%；与基于机器学习的优化加速算法Extratrees相比，平均缩短了求解时间13.5%。此外若将其用于日内调度，系统协同调度频率从1次/h提高到了4次/h，平均可提高可再生能源消纳量17.4%。

关键词: 混合整数规划, 机组组合, 需求响应, 数据中心, 神经分支

Abstract: The rapid development of data centers makes it possible for them to participate in power system dispatching as demand response. By dispatching the computing resources in data centers between regions, it is possible to achieve the goal of energy saving and emission reduction and cost saving. However, considering the demand response of data center computing resources in power system dispatching faces the problem of insufficient computing speed. Based on this, a neural branching power-computation network fast optimization method considering computing demand response is proposed. First, a unit commitment double-layer model considering computing power resource demand response is established. Then, graph convolutional neural network and branch and bound method are combined and applied to the double-layer model. Through historical data training, the neural branching power-computation network fast optimization method considering computing demand response has the ability to quickly determine the order of branch and bound variables and minimize the number of iterations, which significantly improves the solution speed and realizes the fast solution of unit combination demand response considering data center computing resources. The performance of the proposed method is verified in the simulation scenario of "East Data West Computing" engineering. Compared with the pseudo cost branching algorithm, the average solution time is reduced by 39.1%. Compared with the commercial solver CPLEX, the average solution time is reduced by 38.1%. Compared with the machine learning-based optimization acceleration algorithm Extratrees, the average solution time is reduced by 13.5%. In addition, the system coordinated dispatching frequency is increased from 1 time/h to 4 times/h, and the average renewable energy absorption is increased by 17.4%.

Key words: MILP, unit commitment, demand response, IDC, neural branching

中图分类号: