基于长短时记忆神经网络的深度调峰火电机组一次调频能力在线估计  

On-line estimation of primary frequency regulation capability of deep peak regulation thermal power unit based on LSTM neural network

作　　者：张小科 王子杰 夏大伟 王建波 胡怀中[2] ZHANG Xiaoke;WANG Zijie;XIA Dawei;WANG Jianbo;HU Huaizhong(Electric Power Research Institute of Stata Grid Henan Electric Power Company,Zhengzhou 450052,China;School of Automation Science and Engineering,Xi’an Jiaotong University,Xi’an 710100,China;Stata Grid Henan Electric Power Company,Zhengzhou 450000,China)

机构地区：[1]国网河南省电力公司电力科学研究院,河南郑州450052 [2]西安交通大学自动化科学与工程学院,陕西西安710100 [3]国网河南省电力公司,河南郑州450000

出　　处：《热力发电》2023年第8期172-178,共7页Thermal Power Generation

摘　　要：随着“碳达峰、碳中和”战略目标的推进,火电机组更多地参与深度调峰运行。深度调峰工况下,火电机组蓄热量不足,一次调频能力下降,导致以额定工况标定的机组一次调频能力与实际调频能力之间出现很大偏差,威胁电网频率安全。对此,提出一种基于长短时记忆(LSTM)神经网络的深度调峰火电机组一次调频能力在线估计方法。利用LSTM神经网络的时序记忆能力和非线性特征提取能力将面向稳态机组设计的静态模型改进为考虑机组动态运行过程的动态模型,修正机组变负荷过程、一次调频历史动作等扰动因素造成的误差。通过分级建模方法,针对机组蓄热、汽轮机做功动态等调频能力影响因素的不同特性,设计具有不同神经网络结构的子模型,将锅炉侧影响纳入考虑范畴,提高调频估计结果精度。采用某电厂机组运行数据检验该方法,结果表明,与电力系统采用的传统方法相比,该方法估计结果具有更高精度,并且在稳态和变负荷等不同工况下均有较好的应用效果。With the promotion of China’s“carbon peaking and carbon neutral”strategy,thermal power units are more involved in deep peak regulation.Under the conditions of deep peak regulation,the thermal power unit is insufficient in heat storage,and the primary frequency regulation capability decreases,resulting in a large deviation between the unit’s primary frequency regulation capability calibrated under the rated operating condition and the actual frequency regulation capability,threatening the frequency security of the power grid.Aiming at this problem,an online estimation method of primary frequency regulation capability of deep peak regulation thermal power units based on LSTM neural network is proposed.The static model of steady-state unit design was improved to a dynamic model,considering the dynamic operation process of the unit by using the time sequence memory ability and nonlinear feature extraction ability of LSTM neural network,and the errors caused by the disturbance factors such as the load changing process and the historical action of primary frequency regulation were corrected.Based on the hierarchical modeling method,the sub-models with different neural network structures were designed for the different characteristics of the factors affecting the frequency regulation capacity,such as heat storage of the unit and steam turbine work performance,and the effects of furnace side were taken into account to improve the accuracy of frequency regulation estimation results.Compared with the traditional method used in the power system,the estimation result of this method has higher accuracy,and has better performance under different working conditions such as steady state and variable load.

关 键 词：深度调峰 一次调频 LSTM神经网络 在线估计 

分 类 号：TM621[电气工程—电力系统及自动化] TP183[自动化与计算机技术—控制理论与控制工程]