基于多层感知机的超超临界火电机组煤耗建模

Modeling coal consumption of supercritical thermal power units based on multilayer perceptron neural network

提高燃煤火力发电机组的运行灵活性成为目前消纳新能源入网的可靠手段,然而燃煤机组频繁变负荷、变煤种等特点增加了机组运行优化控制的难度,迫切需研究和开发更智能、更先进的运行优化方法。建立准确的能耗动态模型,是实现优化运行的基础,为提高机组运行优化效率,提出了一种基于多层感知机神经网络(MLP)的煤耗建模方法,该方法统筹考虑锅炉侧和汽机侧参数对煤耗的耦合影响关系。以一台660 MW超超临界火电机组为建模对象,选择了54个主要的运行参数作为模型的输入特征,输入特性包含锅炉侧主要运行参数、汽机侧主要运行参数以及电气侧主要运行参数,选取2 a运行数据作为建模数据,预处理后最终确定306234组用于模型的训练与验证,其中80%数据用于模型训练,剩余20%用于模型验证。模型训练采用EarlyStopping技术监控训练集的均方根误差M_(SE),模型训练过程中还需对网络层数、隐藏层节点数及训练批次数据量大小(batch_size)优化。神经网络的初始系数为随机生成,计算结果可能存在随机性,因此每个参数的调参次数均为5次,并采用其平均数作为最终值,最终确定模型隐藏层数为1,隐藏层节点数为64,激活函数选用relu,每次训练抓取的数据量为256。在训练集和测试集表现效果方面,分别有97.69%和97.67%的煤耗预测值与实际值绝对误差值±0.325 g/kWh,其中分别有62.22%和61.93%的煤耗预测值与实际值绝对误差值位于±0.125 g/kWh,同时训练集和测试集在不同误差时出现频率基本相同,说明MPL模型的泛化能力较好。MLP可替代机理模型进行煤耗计算,为实现在线实时优化奠定基础。

Improving the operational flexibility of coal-fired thermal power generating units has become a reliable means to consume new energy into the grid at present,however,the characteristics of the frequent load and coal changes of coal-fired units have greatly increased the difficulty of unit operation optimization control,and there is an urgent need to develop a more intelligent and advanced operation optimization method.To improve the efficiency of unit operation optimization,a coal consumption modeling method based on multilayer perceptron neural network(MLP)was proposed,which integrated the coupling influence relationship between boiler-side and steam engine side parameters on coal consumption.A 660 MW supercritical thermal power unit was used as the modeling object,and 54 main operating parameters were selected as the input features of the model,including boiler-side main operating parameters,steam-engine side operating parameters and electrical-side operating parameters.80%of the data were used for model training and the remaining 20%were used for model validation.Early Stopping technique was used to monitor the root mean square error(M_(SE))of the training set,the number of network layers,the number of nodes in the hidden layer and the size of the training batch(batch_size)need to be optimized during the model training.In terms of the performance of the training and testing sets,97.69%and 97.67%of the absolute errors between the predicted and actual coal consumption values are between±0.325 g/kWh,in which 62.22%and 61.22%are in the range of±10.25 g/kWh,respectively.The MLP can replace the mechanism model for coal consumption calculation and lay the foundation for online real-time optimization.