| 引用本文: | 赵嘉豪,周赣,黄莉,陆春艳,陶晓峰,冯燕钧.CPU-GPU异构计算框架下的高性能用电负荷预测[J].电力自动化设备,2021,41(11): |
| ZHAO Jiahao,ZHOU Gan,HUANG Li,LU Chunyan,TAO Xiaofeng,FENG Yanjun.High-performance electricity load forecasting under CPU-GPU heterogeneous computing framework[J].Electric Power Automation Equipment,2021,41(11): |
|
| 摘要: |
| 随着电网的快速发展，用电信息采集系统的数据计算业务面临着巨大挑战。近年来，图形处理器(GPU)因其在浮点计算速度和存储带宽方面的优势成为高性能计算问题中的研究热点，也被成功应用在电力系统计算分析等科学计算领域。在基于人工智能方法的电力负荷预测问题中，以往大部分研究仅考虑了使用GPU加速预测模型的训练，而并未应用在数据集的获取和计算上。提出了一种基于中央处理器-图形处理器(CPU-GPU)异构计算框架下全流程加速的高性能用电负荷预测方案。首先结合统一计算架构(CUDA)和多线程技术实现了使用多台GPU完成用电负荷的并行预处理，随后在聚类分析后基于XGBoost算法完成了多台区负荷预测，并利用GPU加速了模型的训练计算。最后通过对深圳市43 254个台区用电信息的实例分析，验证了所提方法的高效性与适用性。 |
| 关键词: 用电信息采集系统 负荷预测 GPU 异构计算 XGBoost |
| DOI：10.16081/j.epae.202106008 |
| 分类号:TM715 |
| 基金项目:国家自然科学基金资助项目(51877038)；国家电网公司总部科技项目(5400-202018421A-0-0-00) |
|
| High-performance electricity load forecasting under CPU-GPU heterogeneous computing framework |
|
ZHAO Jiahao1, ZHOU Gan1, HUANG Li1, LU Chunyan2, TAO Xiaofeng2, FENG Yanjun1
|
|
1.School of Electrical Engineering, Southeast University, Nanjing 210096, China;2.NARI Group Corporation/State Grid Electric Power Research Institute, Nanjing 211106, China
|
| Abstract: |
| With the rapid development of power grid, the electricity information acquisition system faces great challenges in the data computing business. Recently, GPU(Graphics Processing Unit) has become important issues in high-performance computing problems due to its superior performance on floating-point computing speed and memory bandwidth. GPU has been successfully applied to scientific computing fields such as power system caculating analysis. When facing power load forecasting problem based on artificial intelligence methods, most of the past researches only considered training of GPU-accelerated forecasting model, but not applied to data acquisition and calculation. A high-performance electricity load forecasting solution under the CPU-GPU(Central Processing Unit-Graphics Processing Unit) heterogeneous computing framework is proposed. Firstly, with the help of CUDA(Compute Unified Device Architecture) and multi-threading technology, power data is computed in parallel by multi-GPUs. Afterwards, with the help of cluster analysis, multi-station load forecasting is completed based on XGBoost algorithm, where GPU accelerates the model training calculation. Finally, through the case analysis of the electricity information of 43 254 stations in Shenzhen, the efficiency and applicability of the proposed method are verified. |
| Key words: electricity information acquisition system electric load forecasting GPU heterogeneous computing XGBoost |
