基于MIC‒NNG‒LSTM的有机废液焚烧SCR入口NO x 浓度动态预测

李艳 ,  史艳华 ,  戴庆瑜 ,  刘嫣 ,  马晓燕

工程科学与技术 ›› 2025, Vol. 57 ›› Issue (03) : 21 -30.

PDF (3956KB)
工程科学与技术 ›› 2025, Vol. 57 ›› Issue (03) : 21 -30. DOI: 10.12454/j.jsuese.202400195
智能交叉科学与工程

基于MIC‒NNG‒LSTM的有机废液焚烧SCR入口NO x 浓度动态预测

作者信息 +

Dynamic Prediction of NO x Concentration at SCR Inlet for Organic Waste Liquid Incineration Based on MICNNGLSTM

Author information +
文章历史 +
PDF (4050K)

摘要

针对高盐有机废液焚烧及烟气处理过程中普遍存在的延迟、非线性和动态特性,提出一种基于自适应变量选择与长短期记忆神经网络(MIC‒NNG‒LSTM)的动态预测方法,对选择性催化还原法(SCR)脱硝塔入口NO x 浓度进行预测,解决当工况发生变化时,脱硝系统不能及时调整喷氨量的问题。该预测方法在传统长短期记忆神经网络(LSTM)的基础上,利用最大互信息系数(MIC)法确定相关辅助变量的延迟时间,以全面捕捉变量之间的动态关系。其次,利用MIC可以反映各输入变量相对于目标变量的重要程度,结合能够收缩变量系数的非负绞杀(NNG)算法,设计MIC‒NNG算法来压缩LSTM网络的输入节点数,剔除冗余变量,实现辅助变量的自适应选择。最后,将包含延迟时间的辅助变量集作为LSTM网络的输入变量集,从而建立SCR入口NO x 浓度动态预测模型。并与LSTM、MIC‒LSTM、NNG‒LSTM 3种预测SCR入口NO x 浓度的方法进行实验对比,浓度预测精度可达到93%,均方根误差减小为约1.5 mg/Nm3。结果表明:通过引入输入变量的延迟时间特性,能够更好地体现各变量之间的动态关系;MIC‒NNG算法相对于NNG算法能更准确地选择输入变量以缩短模型预测时间,提高预测精度和泛化能力。基于MIC‒NNG算法和LSTM网络的动态预测模型综合考虑了有机废液焚烧过程中变量的延迟特性和各参数之间的动态时序关系,可以为降低NO x 排放量提供新思路。

Abstract

Objective This study proposes a dynamic prediction method based on MIC‒NNG‒LSTM that addresses the typical delay, nonlinearity, and dynamic characteristics present in the incineration and flue gas treatment of high-concentration saline organic waste liquids. The method predicts the NO x concentration at the inlet of the selective catalytic reduction (SCR) denitration tower and solves the problem of the denitration system’s inability to adjust the ammonia injection amount in a timely manner when operating conditions change. Methods Firstly, given the complex operation of the waste liquids incineration process, which involved strong coupling and high correlation among operating parameters, the prediction method examined in this study was based on the traditional Long Short‒Term Memory (LSTM) neural network as the foundational model. This model captured the temporal dependency in the input data and memorized and utilized information over extended periods, ensuring the dynamic temporal relationship between auxiliary variables and target variables throughout the modeling process. Secondly, variables that affected the NO x concentration at the inlet of the Selective Catalytic Reduction (SCR) system, such as furnace temperature, air supply, and natural gas flow rate, exhibited nonlinearity and time lag. The Maximal Information Coefficient (MIC), being relatively insensitive to time lags, automatically identified the optimal delay time between variables in time series. Accordingly, this study adopted the MIC method to determine the delay time of relevant auxiliary variables, comprehensively capturing the dynamic relationships among variables. The variables affecting the NO x concentration at the SCR inlet belonged to high-dimensional multi-feature variable data that contained a substantial amount of redundant information. This study applied MIC to reflect the importance of each input variable relative to the target variable, improved the Non-negative Garrote (NNG) algorithm capable of shrinking variable coefficients, and designed the MIC‒NNG algorithm to reduce the input node count of the LSTM network, eliminating redundant variables and achieving adaptive selection of auxiliary variables. Finally, the set of auxiliary variables, including delay time, was used as the input variable set for the LSTM network to establish a dynamic prediction model for the NO x concentration at the SCR inlet. Experimental comparisons were conducted with three other methods for predicting NO x concentration at the SCR inlet: LSTM, MIC‒LSTM, and NNG‒LSTM. Results and Discussions Once the experimental results were analyzed and compared, the basic LSTM prediction model, which did not consider the time lag between input variables and the impact of redundant variables, exhibited significant fluctuations in its prediction fitting curve. When the NO x concentration underwent drastic fluctuations, the prediction results displayed lags, indicating lower prediction accuracy. The MIC‒LSTM model, which applied the MIC algorithm to screen out nine relevant variables and estimate their delay times, achieved a better fit to the actual NO x concentration values compared to the LSTM model, with reduced prediction lags. This confirmed that considering the delay time of auxiliary variables captured the fluctuation characteristics of NO x concentration more effectively, and selecting highly correlated auxiliary variables contributed to improving prediction accuracy. The NNG‒LSTM model, which incorporated the NNG algorithm, reduced the number of auxiliary variables participating in the prediction from 19 to 13, though this was still four more variables than those selected by MIC. This indicated that neglecting the importance of auxiliary variables relative to the target variable increased the number of weakly correlated variables included in the prediction. Although the NNG‒LSTM model demonstrated better prediction fitting ability than the MIC‒LSTM model, its failure to consider the time delay of auxiliary variables on SCR inlet NO x concentration resulted in prediction lags, demonstrating that considering the time delay of auxiliary variables affected prediction accuracy. The proposed MIC‒NNG‒LSTM model, which integrated the improved NNG algorithm, reduced the number of auxiliary variables participating in the prediction to 7, with more accurately selected relevant variables compared to the NNG‒LSTM model. This led to a fitting curve that closely matched the actual values, with improved prediction lags. The combination of MIC’s time delay estimation and MIC‒NNG’s variable selection allowed the prediction model to achieve higher accuracy and lower complexity. The prediction error distributions revealed that the LSTM model’s errors were relatively dispersed, while the MIC‒LSTM model’s errors were more concentrated. The NNG‒LSTM model’s error distribution was narrower than those of the LSTM and MIC‒LSTM models, and the MIC‒NNG‒LSTM model’s error distribution was the most concentrated. The standard deviations of the four models were 4.007 0, 2.679 3, 1.826 8, and 0.812 5, respectively, with the MIC‒NNG‒LSTM model showing the smallest standard deviation, further confirming its superior prediction capability. The performance indicators RMSE for the LSTM, MIC‒LSTM, NNG‒LSTM, and MIC‒NNG‒LSTM models were 3.552 2, 2.492 1, 1.875 6, and 1.567 4 mg/Nm3, respectively. The MAPE values were 0.007 9%, 0.004 7%, 0.004 2%, and 0.003 4%, and the A·R2(adjusted R-squared) values were 74%, 86%, 88%, and 93%, respectively. The MIC‒NNG‒LSTM model demonstrated the best performance among the four models in all indicators, confirming that it accounted for the impact of time delay on prediction results and eliminated redundant variables, achieving optimal prediction performance. This further highlighted the accuracy and effectiveness of the MIC‒NNG‒LSTM prediction model. Conclusions The results indicate that considering the delay time of input variables ensures the dynamic performance of the LSTM network while accurately expressing the nonlinear relationship between SCR inlet NO x concentration and related auxiliary variables. The MIC‒NNG algorithm more accurately selects input variables than the NNG algorithm, shortening model prediction time and improving both prediction accuracy and generalization ability. The dynamic prediction model based on the MIC‒NNG algorithm and LSTM neural network comprehensively considers the delay characteristics of variables and the dynamic time-series relationships between parameters in the incineration process of organic waste liquid, which provides a new approach for reducing NO x emissions.

Graphical abstract

关键词

有机废液 / 动态预测 / 变量选择 / 长短期记忆神经网络 / MIC‒NNG算法

Key words

organic waste liquid / dynamic prediction / variable selection / LSTM / MIC‒NNG algorithm

引用本文

引用格式 ▾
李艳,史艳华,戴庆瑜,刘嫣,马晓燕. 基于MIC‒NNG‒LSTM的有机废液焚烧SCR入口NO x 浓度动态预测[J]. 工程科学与技术, 2025, 57(03): 21-30 DOI:10.12454/j.jsuese.202400195

登录浏览全文

4963

注册一个新账户 忘记密码

本刊网刊
高盐有机废液焚烧会产生大量NO x,而《危险废物焚烧污染控制标准》规定,NO x 排放每小时均值需低于300 mg/Nm3,但超低排放要求不高于50 mg/m3,甚至部分地区要求低于30 mg/Nm3。为满足排放标准,工程上通常采用选择性催化还原(SCR)脱硝技术[12],将NH3注入烟气中,在催化剂作用下将NO x 还原为N2和H2O,因此,及时、准确地控制喷氨量是保证NO x 排放达标的关键。喷氨量的计算公式如下所示:
QNH3=f(α,Q,Cin,eout)
式中:函数f由不同控制方案决定;α为氨氮摩尔比,通常为定值;Q为烟气流量,正常工况下其变化幅度较小;eout为SCR出口NO x 浓度测量值与设定值偏差,为满足国家排放要求,其辅值不能过大;Cin为SCR入口NO x 浓度,受废液流量、炉膛温度等因素影响,具有较大的时变性[3],是影响喷氨量的关键因素。
目前,工程上的连续排放检测系统(CEMS)在测量入口NO x 浓度时存在较大的滞后(采样、传输和处理延迟)[4],导致喷氨量调整不及时,从而使得NO x 排放超标或氨逃逸。因此,实时、准确地测量入口NO x 浓度至关重要[56]。为此,国内外学者提出了利用机器学习和数据驱动的方法对入口NO x 浓度进行预测,来改善脱硝控制效果。
Matsuzaki等[7]借助数据缩减法(ISSA)减少辅助变量,进而运用极端随机树(ET)回归预测入口NO x 浓度。郭瑞君等[8]采用主元分析法选择相关变量,基于非线性多元数据拟合方法建立入口NO x 浓度预测模型。金秀章等[9]在运用最小冗余最大相关性(mRMR)选择特征变量后,将在线序列极限学习机(OSELM)和自回归积分滑动平均模型(ARIMA)组合对入口NO x 浓度预测。刘延泉[10]和董威[11]等均采用互信息法选择特征变量,结合被寻优算法优化的数据驱动模型,准确地预测了入口NO x 浓度。刘长良等[12]提出了条件互信息法(CMI)的变量选择方法,结合BP神经网络预测入口NO x 浓度。上述研究均属于静态预测模型,先选择相关辅助变量,在已定变量集的情况下结合数据驱动方法进行预测,没有考虑变量之间的动态关系。Gao[13]和李影[14]等提出采用长短期记忆神经网络(LSTM)来捕捉NO x 浓度的动态变化趋势,实现了预测模型的动态性。但上述预测模型依然既不能在预测过程中自适应选择当时最相关的变量,也忽略了输入变量的时滞对目标变量的影响,难以处理高维多特征的时序数据,使得其在应对复杂的有机废液焚烧工况时,难以持续保持较高精度。
针对上述问题,本文提出自适应变量选择与长短期记忆神经网络(MIC‒NNG‒LSTM)的动态预测模型对入口NO x 浓度进行预测。其中,LSTM能够捕捉变量间的时序依赖关系,并记忆较长时间内的信息,从而保证预测模型的动态性。非负绞杀算法(NNG)[1516]可以嵌入到LSTM中压缩输入权重,实现变量的动态自适应选择。同时,利用最大信息系数算法(MIC)[1718]可以反映各辅助变量对目标变量的重要程度以及对时滞的不敏感的特点,估计辅助变量的时延,并改进NNG算法的约束项,从而提高自适应选择的准确性。

1 初始辅助变量选择和时延估计

1.1 初始辅助变量选择

高盐有机废液焚烧后烟气脱硝SCR系统的布置在烟道尾部(脱硫装置与烟囱之间),其工艺流程如图1所示。

图1中,有机废液通过喷枪系统送入焚烧炉,在燃烧室内,利用高温氧化原理,在助燃空气的作用下与天然气发生反应,废液迅速干燥、热解、氧化(燃烧),产生较多的热力型NO x,在烟囱出口处或含氧量较高区域生成更多NO x。对于烟气中的NO x,该工艺采用三级脱硝,即选择性非催化还原(SNCR)炉内脱硝+脱硫塔碱液吸收少量NO2+SCR脱硝,使烟气排放时,NO x 小于50 mg/Nm3。其中,SNCR是在900~1000 ℃的环境下,氨水还原燃烧生成的部分NO x。然后,对烟气进行除尘降温送入脱硫塔,碱液吸收少量NO2形成少量的硝酸盐。最后,将烟气通过脱白塔和烟道高温升温后送入脱硝塔,与还原剂在315~400 ℃的温度下,通过催化剂作用,还原烟气中的NO x

通过分析整个有机废液焚烧工艺过程中NO x 生成机理和影响SCR入口NO x 浓度的因素,可以初步筛选出与SCR入口NO x 浓度相关的19个初始辅助变量:炉内温度(x1)、送风量(x2)、天然气流量(x3)、炉膛氧含量(x4)、废液流量(x5)、SNCR脱硝效率(x6)、烟气停留时间(x7)、引风量(x8)、蒸汽压力(x9)、一次风温度(x10)、SCR入口温度(x11)、二次风温度(x12)、给水温度(x13)、脱硫塔脱硝效率(x14)、蒸汽流量(x15)、蒸汽温度(x16)、炉膛压力(x17)、炉膛出口温度(x18)、给水压力(x19)。

1.2 基于MIC算法的时延估计

1.2.1 时滞描述

定义建立SCR入口NO x 浓度预测模型的样本数据集D={ X (t); y(t)},其中, X (t)=[x1(t), x2(t), …, xp (t)]为t时刻(t=1, 2, …, TT为时间戳总数)候选辅助变量,y(t)为t时刻目标变量SCR入口NO x 浓度,p为候选辅助变量个数。

依据现场经验,建模的辅助变量与目标变量之间存在时间滞后[19],因此,与当前t时刻SCR入口NO x 浓度y(t)相关性最大的并不是各辅助变量当前时刻的信息xi (t),而是tτi 时刻的变量信息xi (tτi )(i=1, 2, …, p),τi 为辅助变量xi 的时延参数。由于燃烧炉正常运行周期为10 min左右,故将影响SCR入口NO x 浓度的辅助变量的最大延迟时间τmax限制在600 s以内,所以包含辅助变量xi (t)延迟信息的辅助变量序列为{xi (t), xi (t‒1),…, xi (tτi ), …, xi (tτmax)}。

综上分析可知,确定各辅助变量时滞的目的是使预测模型的输入变量包含更多对目标变量有用的信息,能够对目标变量具有最强的解释能力。故延迟时间的估计可以转换为确定{xi (t), xi (t‒1),…, xi (tτi ),…, xi (tτmax)}与y(t)的相关性,相关性越大,该时刻辅助变量携带的有用信息越多。考虑到废液焚烧工况存在非线性特性,本文采用MIC来分析辅助变量在历史时刻的值与y(t)的相关性,即可得到涵盖延迟时间的序列数据集D={x1(tτ1), x2(tτ2),…, xP (tτp ); y(t)},过程如图2所示。

1.2.2 时延估计

对历史时刻辅助变量值构成的(xi (tτi ), y(t))散点图进行ab行网格化,通过计算区间样本数在Q网格(大小为a×b)样本容量的占比得出联合概率密度p(xQ, yQ )和边缘概率密度p(xQ )、p(yQ ),并计算出此网格下的互信息(MI),记为I(x; y|Q),表达式为:

I(x;y|Q)=xQx yQyp(xQ,yQ)lb p(xQ,yQ)p(xQ)p(yQ)

由于同一规模(大小为a×b)下存在其他划分方式,计算不同情况下最大的MI值,将其归一化后,组成最大归一化互信息特征矩阵 M (D) a, b,表达式为:

M(D)a,b=maxQΩ(a,b)(I(x;y|Q))lb min(a,b)

式中:Ω(a,b)表示大小为a×b的网格划分集合;lb min(a, b)为归一化因子,确保结果在[0, 1]之间。求得特征矩阵中最大值作为最终MIC值,记为MIC(x, y|D)。计算方式如下:

MIC(x,y|D)=maxa×bBM(D)a,b

式中,B为网格Q划分a×b的上限值,取数据总量n的0.6次方[16,20]

利用式(2)~(4)获得各辅助变量xi 在一个周期内每个采样时刻下与SCR入口NO x 浓度之间的MIC值,将MIC值最大的时刻作为该辅助变量的迟延时间τi。由于数据采样周期为1 min,所以预估的延迟时间如图3所示,应为60 s的整数倍,最终所得涵盖延迟时间的辅助变量数据集为D={x1(t‒4), x2(t‒1),…, x19(t); y(t)}。

2 MIC‒NNG‒LSTM网络预测模型

2.1 LSTM网络预测模型

每个LSTM单元[21]包含3个输入,即上一时刻记忆细胞状态 C (t‒1)、上一时刻隐含状态 h (t‒1)和当前时刻输入变量信息 X (t),其中, C (t‒1)=[C1(t‒1),C2(t‒1),…,Cp (t‒1)], h (t‒1)=[h1(t‒1),h2(t‒1),…,hq (t‒1)], X (t‒1)=[x1(t‒1),x2(t‒1),…,xp (t‒1)],q为隐含层节点数;3个门控单元分别是遗忘门 f (t)、输入门 I (t)、输出门 O (t),其作用是通过相应的激活函数来控制细胞状态的信息保留程度;两个输出分别为当前时刻记忆细胞状态 C (t)和当前时刻隐含状态 h (t)。图4为LSTM内部运算展开图,其工作原理为:首先,分别将 h (t‒1)和 X (t)的信息传送到 f (t)、 I (t)、 O (t)和候选细胞 C'(t)中,通过 f (t)将 C (t‒1)中的信息选择性遗忘;同时, C'(t)中 h (t‒1)和 X (t)信息通过 I (t)选择性保存,并更新 C (t)中的信息;最后, O (t)将 C (t)中的信息更新到 h (t)中,作为下一时刻的单元输入。图4中,by为输出层的偏置顶,LSTM的信息更新和动态记忆规则如下:

Wxf=w11xfw12xf...w1pxfw21xfw22xf...w2pxfwq1xfwq2xf...wqpxf
fk(t)=σ((i=1pxi(t)wkixf+j=1qhj(t-1wkjhf))+bkf
Ik(t)=σ((i=1pxi(t)wkixI+j=1qhj(t-1wkjhI))+bkI
Ck'(t)=tanh((i=1pxi(t)wkixC+j=1qhj(t-1wkjhC))+bkC
Ok(t)=σ((i=1pxi(t)wkixO+j=1qhj(t-1wkjhO))+bkO
C(t)=f(t)C(t-1)+I(t)C'(t-1)
h(t)=O(t)tanh(C(t))

式(5)~(11)中:wkixfwkixIwkixCwkixO分别为矩阵WxfWxIWxCWxO的元素,WxfWxIWxCWxO分别为输入到遗忘门 f (t)、输入门 I (t)、候选记忆单元 C '(t)、输出门 O (t)的权重矩阵;wkjhfwkjhIwkjhCwkjhO分别为矩阵WhfWhIWhCWhO的元素,WhfWhIWhCWhO分别为隐藏状态到遗忘门 f (t)、输入门 I (t)、候选记忆单元 C '(t)、输出门 O (t)的权重矩阵;fk (t)、Ik (t)、Ck'(t)、Ok (t)分别为 f (t)、 I (t)、 C '(t)、 O (t)第k个元素;bkfbkIbkCbkO分别为偏置向量 bfbIbCbO 的元素;为逐元素相乘;tanh和σ用作激活函数,σ为Sigmoid函数,输出值区间为[0,1],0表示完全遗忘,1表示完全保留;k为隐含层 h (t)节点索引,k=1,2,,qjh (t‒1)的维度索引,j=1,2,,q

将涵盖时滞的19个辅助变量作为LSTM的输入变量,将SCR入口NO x 浓度作为LSTM的输出,利用有机废液处理过程的实际历史数据对LSTM进行训练,得到一个预测能力较好的LSTM预测模型,且模型输出可以表示为:

y(t)=g(k=1qwkyh(t)+by)=g((k=1qwky(O(t)tanh(f(t)C(t-1)+I(t)C'(t))))+by)

式中,wky为输出层中隐藏节点到输出节点的权重参数,g(·)为激活函数。

2.2 基于MICNNG算法自适应变量选择

NNG算法是通过将收缩算子 λλ =[λ1, λ2,…, λp ])作为LSTM网络输入权重的约束项[22],构成了非线性约束优化,其表达式为:

λi*(s)=argminm=1n(y(m)-i=119λiw^kixi(m))2,λi0,i=1pλis

式中,n为样本数量,m为样本编号,y(m)为第m个样本的目标变量值,xi(m)为第m个样本的第i个辅助变量值,w^ki为输入权重的估计值,s为绞杀参数。通过上式求得的优化收缩算子λ*,得到优化的系数估计w˜ki

w˜ki=λi*(s)w^ki

则优化后的输出y˜可以表示为:

y˜=i=119w˜kixi

变量选择的原理是通过调整s值来控制对变量的约束强度。当sp时,约束项失效,意味着所有辅助变量被保留;更多的收缩算子随着s值的减小趋于0,即更多的变量被剔除,降低模型复杂度。但由于约束项对所有变量施加的约束力度一样,无法有效保护相关变量和剔除相关性弱的变量。本文利用MIC算法得到各辅助变量对目标变量的重要程度[23],改进约束项的约束力度。

首先,利用式(2)~(4)获得各辅助变量与目标变量的MIC值 υυ =[υ1, υ2,, υ19],将MIC值与MIC平均值的比值作为相对MIC值 ζζ =[ζ1, ζ2,, ζ19],表示各辅助变量相对于目标变量的重要程度,计算式如下:

ζi=υii=1pυi/p

然后,将各辅助变量的相对MIC值的倒数定义为自适应收缩权重 μμ =[μ1, μ2, , μp ],计算公式如下:

μi=1/ζi

最后,将收缩权重嵌入NNG算法的约束项中,构成MIC‒NNG约束项:

i=119μiλis

辅助变量对目标变量影响越大,则ζi 越大;μi 越小,绞杀的变量越多,反之亦然。

对本文的19个辅助变量采用NNG算法和MIC‒NNG算法分别进行筛选测试,图5为各辅助变量对应的收缩算子λ数值统计。可以看出,MIC‒NNG算法较NNG算法对变量压缩更明确,且λ值大于0.95时MIC‒NNG算法对变量的选择与专家对影响入口NO x 浓度关键因素分析相符合。变量选择结果以及改进后的NNG算法在变量选择方面的准确性将在后文讨论。

2.3 MIC‒NNG‒LSTM预测模型

2.3.1 模型建立

利用式(12)所示的预测模型,训练出一组收缩算子 λ =[λ1, λ2,…, λp ],其表达式为:

λ=t=1T(y(t)-g((k=1qwky(Ok(t)tanh(fk(t)Ck(t-1)+Ik(t)Ck'(t))))+by))2

λi 和LSTM输入权重的估计值w^ik嵌入LSTM模型,更新网络输入权重,获得NNG‒LSTM模型的输出表达式:

y^(t)=g(k=1qwkyh^(t)+by)=g((k=1qwky(O^(t)tanh(f^(t)C(t-1)+I^(t)C^'(t))))+by)

式中,^表加入收缩算子的相应结果,各门控单元更新表达式为:

f^k(t)=σ((i=1pxi(t)(λiw^kixf)+j=1qhj(t-1wkjhf))+bkf
I^k(t)=σ((i=1pxi(t)(λiw^kixI)+j=1qhj(t-1wkjhI))+bkI
O^k(t)=σ((i=1pxi(t)(λiw^kixO)+j=1qhj(t-1wkjhO))+bkO
C^k'(t)=tanh((i=1pxi(t)(λiw^kixC)+j=1qhj(t-1wkjhC))+bkC

通过NNG‒LSTM模型,得到MIC‒NNG算法的收缩算子表达式为:

λi0,i=1pμiλis;λ*(s)=argmin(t=1n(y^(t)-g((k=1qwky(O^k(t)tanh(f^k(t)Ck(t-1)+I^k(t)C^k'(t))))+by))2)

对于上述非线性约束最优问题中,s可以利用训练集的预测值与实际值之间的均方根误差,求得误差最小时的ss=6.3),作为最优绞杀参数s*。由于遗传算法(GA)[24]能够在参数空间中进行全局搜索,且不要求目标函数是连续可导的,适用于解决各类实际问题。故采用GA求解在s*情况下优化的收缩算子λ*,将其代入式(13)可得到新的输入权重系数w˜ik,则MIC‒NNG‒LSTM预测模型的输出为:

y˜(t)=g(k=1qwkyh˜(t)+by)=g((k=1qwky(O˜(t)tanh(f˜(t)C(t-1)+I˜(t)C˜'(t))))+by)

式中,~表该变量为MIC‒MG‒LSTM的输出。

预测模型的训练算法采用梯度反向传播算法[25],利用半均方根误差(HMSE,记为LHMSE)计算损失值:

LHMSE=12nm=1n(ym-y^m)2

式中,ymy^m分别为真实值和预测值。通过HMSE可以看出模型的训练状态,图6为优化后模型的HMSE值变化曲线,值越小表明训练结果越接近预期效果。

2.3.2 模型参数调优

由于随机搜索法(RS)[26]具有简单易实现、并行性和高效性等优点,本文采用RS对LSTM的参数进行组合寻优。同时,利用Adam优化器[21,27]可以为不同参数设计独立自适应性学习率,故在设置初始学习率为0.005后,加入Adam更新学习率。表1为预测模型中绞杀参数、LSTM和Adam参数的设定。

3 实验设置与结果分析

为了验证所建模型的有效性,选取某有机废液处理过程实际历史数据运行状态下的1 400组样本数据,包括19个辅助变量和1个目标变量(SCR入口NO x 浓度)。对采集到的数据进行降噪和归一化处理后,其中:80%作为训练、验证集,将包含时延信息的辅助变量作为候选输入变量集,建立符合该有机废液焚烧SCR入口NO x 浓度的预测模型;20%作为测试集,用于验证算法的性能。

3.1 模型评价指标

采用平均绝对百分比误差(MAPE,记为eMAPE)、均方根误差(RMSE,记为eRMSE)、校正决定系数(A·R2,记为dAR)来评价模型的优劣。MAPE反映误差值和真实值之间的比例,RMSE反映预测值和真实值之间的误差大小,A·R2反映模型数据预测的追踪能力,通过这3个指标可以全面反映模型的预测精度。从模型筛选出的输入变量数p可以反映MIC‒NNG算法在剔除冗余变量方面的优越性。各指标对应的计算公式如下:

eMAPE=1nm=1nym-y^mym
eRMSE=1nm=1n(ym-y^m)2
dAR=1-m=1n(ym-y^m)2m=1n(ym-y¯m)2

3.2 消融实验

为体现MIC‒NNG‒LSTM模型预测SCR入口NO x 浓度的优越性,设计消融实验,在LSTM的基础上加入改进模块。MIC‒LSTM模型利用MIC算法估计其时延参数,将含有时延信息的特征变量集作为LSTM的输入;NNG‒LSTM模型不考虑辅助变量时延对预测结果的影响,只对LSTM的输入权重进行压缩,筛除冗余变量;MIC‒NNG‒LSTM模型在考虑了辅助变量时延的同时,利用本文设计的MIC‒NNG算法对辅助变量自适应选择。

由于神经网络的每次训练和测试的结果会略有差异,因此对每个模型在同一运行环境下,采用相同的数据进行15次仿真,将仿真结果的平均值作为最终的预测结果。图78分别为4种模型预测拟合曲线、带高斯拟合的预测误差频率分布,表2为不同模型的预测性能指标对比。

3.3 结果分析

从图78表2可以看出,由于LSTM模型未考虑输入变量之间的延迟和冗余变量对入口NO x 的影响,在输入变量出现较大波动时,预测结果存在明显滞后,且预测精度也很低。MIC‒LSTM模型考虑了时滞的影响,可以捕捉变量之间的动态关系,因此,性能指标MAPE和RMSE值明显比LSTM模型的小,拟合曲线也没有明显的滞后情况,但在波动较大时预测依然不准确(图7(b)),这是由于相关性弱的辅助变量没有被剔除而影响预测结果。NNG‒LSTM模型虽然优化了LSTM网络的输入权重,减少了冗余变量的影响,但因没有考虑输入变量之间的延迟对入口NO x 的影响,故该模型相较于MIC‒LSTM模型MAPE和RMSE减小了,但变量个数更多(p值多4)(表2),使得其拟合效果虽有提升,但预测结果仍存在滞后(图7(c))。同时,由图8可知,相较于LSTM预测模型,MIC‒LSTM和NNG‒LSTM模型的预测误差分布明显收窄且集中在较小区间内,标准差明显减小,表明模型预测误差集中度提高,预测精度有所改善。

由图78可知:MIC‒NNG‒LSTM模型的预测值基本可以拟合真实值,预测误差最小且基本控制在很小的范围内;标准差为0.812 5,也是最小的。由表2可知,MIC‒NNG‒LSTM模型选择的输入变量最少,且各项性能指标明显优于其他几种模型。这是因为其结合了MIC和NNG的优点,同时考虑辅助变量的个数和时滞对入口NO x 的影响。

综上所述,MIC‒NNG‒LSTM模型在预测SCR入口NO x 浓度方面展现出了更强的学习能力、更高的预测精度和更低的复杂度,为SCR系统的实时控制和NO x 排放达标提供了新的技术思路。

4 结 论

1)本文结合MIC和NNG算法的优势,设计MIC‒NNG算法对辅助变量自适应选择,并确定其延迟时间,解决了预测模型输入变量的冗余和时滞问题。

2)为后续研究优化改进脱硝过程控制系统,使SCR系统实现在NO x 达到超低排放的同时降低企业成本提供技术支撑。

3)虽然本文模型提高了预测精度,但未考虑到模型的修正,并且实验所采集的样本不一定可以完全代表所有时间的工况,使得测量精度可能会随时间的变化而下降。后续可以采用在线自校正和模型不定期更新的方式,进一步提高动态预测模型的精度。

参考文献

[1]

Yang Tingting, Bai Yang, You Lyu,et al.Study on multi-obj-ective optimal control of SCR denitrification system[J].Pr-oceedings of the CSEE,2021,41(14):4905‒4911.

[2]

杨婷婷,白杨,吕游,.SCR脱硝系统多目标优化控制研究[J].中国电机工程学报,2021,41(14):4905‒4911.

[3]

Ma Zenghui, Wang Lian’gang, Liang Feng.An engineering control method of SCR denitration system[C]//Proceedings of the 2023 IEEE 3rd International Conference on Data Science and Computer Application(ICDSCA).Dalian:IEEE,2023:1243‒1247. doi:10.1109/icdsca59871.2023.10392714

[4]

Yan Laiqing, Dong Ze.Dynamic prediction model of SCR denitrification system based on k-nearest neighbor mutual information and WKOPLS[J].Proceedings of the CSEE,2019,39(10):2970‒2980.

[5]

闫来清,董泽.基于k-近邻互信息和WKOPLS的SCR脱硝系统动态预测模型[J].中国电机工程学报,2019,39(10):2970‒2980.

[6]

Lin Weijun, Li Min, Zhao Chang,et al.NO x concentration soft measurement and ammonia injection optimization based on delay features[J].Journal of Chinese Society of Power Engineering,2023,43(10):1352‒1362.

[7]

林伟俊,李敏,赵畅,.基于延迟特性的NO x 浓度软测量与喷氨优化[J].动力工程学报,2023,43(10):1352‒1362.

[8]

Tang Zhenhao, Li Yanyan, Chai Xiangying,et al.Adaptive nonlinear model predictive control of NO x emissions under load constraints in power plant boilers[J].Journal of Chemical Engineering of Japan,2020,53(1):36‒44. doi:10.1252/jcej.19we142

[9]

Liu Guofu, Zhang Yu, Shen Dekui,et al.Anticipatory NH3 injection control for SCR system based on the prediction of the inlet NO x concentration[J].Journal of the Energy Institute,2021,94:167‒175. doi:10.1016/j.joei.2020.07.002

[10]

Matsuzaki A, Kiribuchi D, Shimizu K.Machine learning approach to NO x prediction for SCR process of thermal power plant[J].IFAC-Papers OnLine,2023,56(2):2858‒2864. doi:10.1016/j.ifacol.2023.10.1401

[11]

Guo Ruijun, Zhou Lei, Zhang Guobin,et al.Denitration control system based on SCR inlet NO x concentration and flue gas flow prediction[J].Journal of Engineering for Thermal Energy and Power,2020,35(9):141‒147.

[12]

郭瑞君,周磊,张国斌,.基于SCR入口NO x 浓度和烟气流量预测的脱硝控制系统[J].热能动力工程,2020,35(9):141‒147.

[13]

Jin Xiuzhang, Chen Jiazheng, Li Yangfeng.SCR inlet of th-ermal power plant based on ARIMA-OSELM prediction of NO x concentration[J].Acta Metrologica Sinica,2023,44(9):1458‒1466.

[14]

金秀章,陈佳政,李阳峰.基于ARIMA-OSE-LM的火电厂SCR入口NO x 浓度预测建模研究[J].计量学报,2023,44(9):1458‒1466.

[15]

Liu Yanquan, Yang Kun, Wang Rubei.Prediction of NO x co-ncentration at the entrance of denitration reactor based on mutual information and PSO-LSSVM[J].Electric Power Science and Engineering,2018,34(9):67‒72.

[16]

刘延泉,杨堃,王如蓓.基于互信息和PSO-LSSVM的脱硝反应器入口NO x 浓度预测[J].电力科学与工程,2018,34(9):67‒72.

[17]

Dong Wei, Lin Zijie, Wang Yayun.A CIFE-FOA-DELM method for predicting NO x concentration at the inlet of SCR denitration system[J].Electric Power Technology and Environmental Protection,2024,40(3):313‒320.

[18]

董威,林子杰,王雅昀.基于CIFE-FOA-DELM的SCR脱硝入口NO x 浓度预测方法研究[J].电力科技与环保,2024,40(3):313‒320.

[19]

Liu Changliang, Zhang Cong, Wang Ziqi.Joint selection met-hod of variables and time-delay in NO x production model ba-sed on conditional mutual information[J].Journal of North China Electric Power University(Natural Science Edition),2019,46(6):58‒65.

[20]

刘长良,张丛,王梓齐.基于条件互信息的NO x 生成量模型变量-时滞联合选择方法[J].华北电力大学学报(自然科学版),2019,46(6):58‒65.

[21]

Gao Chunyu, Xie Ming, Jiang Wei,et al.Prediction of NO x in SCR inlet of thermal power plant based on mutual information and long short-term memory network[C]//Proceedings of the 2020 5th International Conference on Po-wer and Renewable Energy (ICPRE).Shanghai:IEEE,2020:512‒516. doi:10.1109/icpre51194.2020.9233111

[22]

Li Ying, Zhuo Jiankun, Wu Yifan,et al.Interpretable prediction model for NO x mass concentration at SCR reactor inlet in coal-fired power plants under flexible operating conditions[J].Thermal Power Generation,2024,53(7):119‒128.

[23]

李影,卓建坤,吴逸凡,.可解释的变负荷下燃煤机组SCR反应器入口NO x 质量浓度预测模型[J].热力发电,2024,53(7):119‒128.

[24]

Sui Lin, Du Bosheng, Zhang Mengyan,et al.A new variable selection algorithm for LSTM neural network[C]//Procee-dings of the 2021 IEEE 10th Data Driven Control and Le-arning Systems Conference(DDCLS).Suzhou:IEEE,2021:571‒576. doi:10.1109/ddcls52934.2021.9455564

[25]

Sun Kai, Sui Lin, Zhang Fangfang,et al.Dynamic soft sensor algorithm based on nonnegative garrote and long short-term memory neural network[J].Control Theory & Applications,2023,40(1):83‒93.

[26]

孙凯,隋璘,张芳芳,.基于非负绞杀与长短期记忆神经网络的动态软测量算法[J].控制理论与应用,2023,40(1):83‒93.

[27]

Reshef D N, Reshef Y A, Finucane H K,et al.Detecting novel associations in large data sets[J].Science,2011,334(6062):1518‒1524. doi:10.1126/science.1205438

[28]

Wu Kangluo, Huang Jun, Li Zhenghui,et al.Dynamic prediction of the NO x concentration at SCR system outlet bas-ed on MIC-CFS-LSTM model[J].Clean Coal Technology,2023,29(6):142‒150.

[29]

吴康洛,黄俊,李峥辉,.基于MIC-CFS-LSTM的SCR出口NO x 浓度动态预测[J].洁净煤技术,2023,29(6):142‒150.

[30]

Zhao Zheng, Wang Jin, Ma Yijie,et al.Optimal control of SCR denitration system based on stepped generalized prediction[J].Journal of North China Electric Power University(Natural Science Edition),2022,49(6):68‒75.

[31]

赵征,王金,马毅杰,.基于阶梯式广义预测的SCR脱硝系统优化控制[J].华北电力大学学报(自然科学版),2022,49(6):68‒75.

[32]

Ren Jia, Sun Siyu, Bao Ke.Fault diagnosis method in industrial processes based on maximal information coefficient and depth residual graph convolution[J].Journal of Chemical Engineering of Chinese Universities,2023,37(1):111‒119.

[33]

任佳,孙思宇,鲍克.基于最大信息系数和深度残差图卷积的工业过程故障诊断方法[J].高校化学工程学报,2023,37(1):111‒119.

[34]

Wang Boqiao, Zhang Jinnan, Zhang Bin,et al.Flame propagation speed prediction model of premixed methane gas de-flagration experiments based on Adamax-LSTM for FLNG[J].Journal of Loss Prevention in the Process Industries,2024,91:105386. doi:10.1016/j.jlp.2024.105386

[35]

Kipruto E, Sauerbrei W.Comparison of variable selection pr-ocedures and investigation of the role of shrinkage in li-near regression-protocol of a simulation study in low-dim-ensional data[J].PLoS One,2022,17(10):e0271240. doi:10.1371/journal.pone.0271240

[36]

Luo Youxi, Xie Kunming, Hu Chaozhu,et al.MICCFS:A co-rrelation-based feature selection algorithm based on maxi-mum information coefficient[J].Journal of Central China Normal University(Natural Sciences),2023,57(6):777‒785.

[37]

罗幼喜,谢昆明,胡超竹,.基于最大信息系数的关联性特征选择算法:MICCFS[J].华中师范大学学报(自然科学版),2023,57(6):777‒785.

[38]

Altanany M Y, Badawy M, Ebrahim G A,et al.Modeling and optimizing linear projects using LSM and non-domin-ated Sorting Genetic Algorithm(NSGA-Ⅱ)[J].Automation in Construction,2024,165:105567. doi:10.1016/j.autcon.2024.105567

[39]

Li Penghua, Chai Yi, Xiong Qingyu.Quantum gate Elman ne-ural network and its quantized extended gradient back-pr-opagation training algorithm[J].Acta Automatica Sinica,2013,39(9):1511‒1522.

[40]

李鹏华,柴毅,熊庆宇.量子门Elman神经网络及其梯度扩展的量子反向传播学习算法[J].自动化学报,2013,39(9):1511‒1522.

[41]

Chai Xuqing, Li Shihao, Liang Fengwei.A novel battery SOC estimation method based on random search optimiz-ed LSTM neural network[J].Energy,2024,306:132583. doi:10.1016/j.energy.2024.132583

[42]

de Campos Souza P V, Dragoni M.IFNN:Enhanced interpr-etability and optimization in FNN via Adam algorithm[J].Information Sciences,2024,678:121002. doi:10.1016/j.ins.2024.121002

基金资助

陕西省重点研发计划项目(2023‒YBGY‒277)

陕西省技术创新引导专项项目(2023GXLH‒071)

陕西省自然科学基础研究计划一般项目(2022JM‒408)

AI Summary AI Mindmap
PDF (3956KB)

204

访问

0

被引

详细

导航
相关文章

AI思维导图

/