基于双深度输入凸神经网络多模型的中间点过热度预测控制
Intermediate point superheat predictive control based on double-depth input convex neural network with multi-model
钟信 1冯磊华 1何金奇 2杨锋3
作者信息
- 1. 长沙理工大学能源与动力工程学院,湖南 长沙 410114
- 2. 陕西高业能源科技有限公司,陕西 西安 710061
- 3. 华自科技股份有限公司,湖南 长沙 410006
- 折叠
摘要
新能源大量并网,超临界火电机组参与调峰容易造成中间点过热度较大波动,从而导致过热蒸汽超温等问题.为较好控制中间点过热度达到稳定,提出了一种基于双深度输入凸神经网络多模型(muti-DDICNN model)的中间点过热度预测方法,分别训练了不同预测步长下子模型,构建了中间点过热度状态预测网络(SPNN)和误差预测网络(EPNN).利用此预测网络凸性质,设计了一种基于双深度输入凸神经网络多模型预测控制器(DDICNN-MPC),将控制问题转化为凸优化问题,求取控制矩阵对目标函数的雅可比矩阵,采用梯度下降法计算控制矩阵最优解.仿真结果表明,DDICNN-MPC能快速平稳地跟踪中间点过热度设定值,且稳态误差较小,具有较好的调节能力.
Abstract
As a large number of new energy is connected to the grid,the participation of supercritical thermal power units in peak regulation tends to cause the superheat of intermediate points to fluctuate greatly,resulting in superheated steam over temperature and other problems.In order to better control the intermediate point superheat to achieve stability,a prediction method of intermediate point superheat based on double-depth input convex neural network multi-model(muti-DDICNN model)was proposed.Sub-models with different prediction step sizes were trained respectively,and the intermediate point superheat state prediction network(SPNN)and error prediction network(EPNN)were constructed.Based on the convex property of prediction network,a multi-model predictive controller(DDICNN-MPC)based on convex neural network with double-depth input is designed.The control problem is transformed into a convex optimization problem,the Jacobian matrix of control matrix to objective function is obtained,and the optimal solution of control matrix is calculated by gradient descent method.The simulation results show that,the DDICNN-MPC can track the intermediate point superheat setting quickly and stably,and the steady-state error is small,so it has good adjustment ability.
关键词
中间点过热度/输入凸神经网络/模型预测控制/梯度下降法/凸优化Key words
intermediate point superheat/input convex neural network/model predictive control/gradient descent algorithm/convex optimization引用本文复制引用
基金项目
湖南省自然科学基金(2018JJ3552)
出版年
2024
国家科技期刊平台
NSTL
维普
万方数据