CN105045951A - Soft-measurement method for volatile fatty acid in effluent of anaerobic wastewater treatment system - Google Patents

Abstract

The present invention relates to a soft-measurement method for volatile fatty acid in effluent of an anaerobic wastewater treatment system. The method uses principal component analysis to realize dimensionality reduction of high-dimensional auxiliary variables and uses a least squares support vector machine which is particularly suitable for solving problems like small samples, nonlinearity and local minimum points and is high in training speed. The combination of the two realizes the soft-measurement of volatile fatty acid in the effluent of the anaerobic wastewater treatment system, and is of great significance to online monitoring and subsequent process optimization controlling of the anaerobic wastewater treatment system. Experiments have proved that, the model of the soft-measurement method is correct in theory and the application of an LS-SVM intelligent algorithm enables real-time prediction and estimation of the important control parameter, the volatile fatty acid, of the effluent of the anaerobic wastewater treatment system. The method can solve the problems of expensive measuring apparatuses, difficult maintenance and measuring lags, and is of great importance to the control and optimization of the anaerobic wastewater treatment system.

Description

A kind of flexible measurement method of anaerobic wastewater treatment system water outlet volatile fatty acid

Technical field

The present invention relates to technical field of waste water processing, relate in particular to a kind of flexible measurement method of anaerobic wastewater treatment system water outlet volatile fatty acid.

Background technology

Energy crisis and environmental pollution are two challenging greatly of being faced with of the current mankind; complicated organic material decomposition in waste water can be become biogas equal energy source material by Anaerobic Biotechnology, and therefore Anaerobic Biotechnology provides a practicable approach for solving energy and environment problem.Anaerobic fermentation methane phase technology, as the representative of Anaerobic Biotechnology, all has broad application prospects in high concentrated organic wastewater and organic solid waste.Anaerobic fermentation methane phase is a complicated biochemical process, particularly methanogen is harsher to requirement for environmental conditions, only be difficult to by microorganism spontaneous fermentation efficient, quick, the stable operation maintaining anaerobic fermentation process, normally carry out for making sweat and obtain higher biogas yield just to carry out monitor and forecast to anaerobic fermentation process.Although anaerobic fermentation methane phase technology is widely applied in China's Industrial Wastewater Treatment, but still there is more problem on the on-line monitoring and control of anaerobic fermentation process.The in-situ measurement equipment of the existing maturations such as Traditional materialized parameter p H value, temperature and oxidation-reduction potential and physico-chemical parameter such as volatile fatty acid (VFA, VolatileFattyAcids) and physiological parameter such as the biomass gas generation process to material impact are but difficult to realize on-line measurement.

VFA is the intermediate product of organic matter degradation, is also methanogenic direct substrate.In industrial methane-producing reactor runs, the generation often occurring to cause because of the accumulation of non-Timeliness coverage VFA " is become sour ".Become sour and be often referred to because methanogen activity sharply reduces, but acidifying bacterium still can continue decomposing organic matter and cause the accumulation of VFA, pH can be made to drop to 3-5, and now pH value and molecular state organic acid all can cause suppression to acidifying bacterium, and cause disposal system to lose processing power.The generation of " becoming sour " is catastrophic often to anaerobic reactor, and reactor, once there is " becoming sour ", is difficult to recover at short notice or be difficult at all to recover methanogen activity in reactor.Easily accumulate just because of VFAs in anaerobic methane production reactor and suppression is caused to methanobacteria, the detection method of the concentration of VFA being subject to the attention of height always.The method of current VFA concentration determination mainly contains the way of distillation, titrimetry, chromatography, colourimetry etc., and in order to realize the on-line monitoring of VFA, scholar has carried out large quantity research.The infrared spectrum energy of the people such as ZhangYS research carry out measuring for acetate propionate etc. but its accuracy and sensitivity not good enough.The gas chromatography of band self-actuated sampler of people's designs such as DiamantisV and the connection of reactor achieve the online of VFA and accurately measure, but gas chromatography costliness is difficult to realize industrial applications.The on-line determination VFA of Zhao Quanbao design and the automatic Titration system of basicity just achieve the robotization of titration process, and not by computation model and Automated library system, 6 constructed point Titration titration are accurate, but complex operation calculation of complex.Above VFAs offline inspection technology needs manually to carry out chemical examination operation, effluent quality quality fluctuation can be caused large, there is energy consumption large, the problems such as expense is high, and on-line monitoring technique many places are in laboratory stage, rarely have and be applied in actual industrial, and there is the delayed length of Measuring Time in most of detecting sensor, instrument, problem expensive and difficult in maintenance, is therefore necessary the on-line monitoring technique studying VFA further.

The measurement problem of general solution industrial process has two approach: one is follow traditional detection technique thinking, and the direct-on-line of implementation procedure parameter is measured as mentioned above in the form of hardware; Another kind is exactly the thinking adopting indirect inspection, utilizes other metrical informations easily obtained, by calculating the estimation realized measured variable.The soft-measuring technique emerged in process control and detection field is in recent years exactly the concentrated reflection of this thought.Hard measurement theoretical source is that the deduction that 20 century 70 Brosilow propose controls.So-called hard measurement is exactly according to the process variable (i.e. auxiliary variable) can surveyed, easily survey and the mathematical relation being difficult to the variable to be measured (i.e. leading variable data) directly obtained, according to certain optiaml ciriterion, adopt various computing method, with software approach realization to the measurement of variable to be measured or estimation, therefore soft-measuring technique is also called soft instrument technology, and oneself through being widely used in process control and optimization at present.

In general soft-measuring technique mainly comprises: auxiliary variable is chosen, data prediction, soft sensor modeling and modelling verification four parts.The selection of auxiliary variable does not generally have the guidance method of versatility, often passes through theoretical and empirical analysis according to concrete object, chooses the variable relevant to leading variable data as auxiliary variable.Auxiliary variable is chosen and follow-up modeling too much can be made comparatively complicated, reduces auxiliary variable and lost part information may reduce model accuracy.Principal component analysis (PCA) (PrincipalComponentAnalysis, PCA) be for reducing one of the most direct means of higher-dimension numerical example in statistics, simultaneously also maximum to the greatest extent all information that may keep original sample, thus be widely used, therefore can realize higher-dimension auxiliary variable variable dimensionality reduction, retain raw information to greatest extent simultaneously.Data prediction is the steps necessary of soft sensor modeling, often comprises the methods such as missing values process, abnormality value removing, data smoothing and standardization, selects wherein one or more methods to carry out pre-service to the raw data collected as required.The main body of soft-measuring technique and core set up soft-sensing model, and the conventional method of Modling model mainly contains: modelling by mechanism method, regression analysis, neural network and support vector machine (SVM) method etc.Wherein neural network and SVM these two kinds belong to black-box modeling method together, do not require the determination internal mechanism of object, are therefore relatively applicable to complicated sewage disposal process, apply also extensive.SVM is the more a kind of new technology that receives publicity in machine learning field in recent years, is Corpus--based Method principle, and compare the discovery learning mechanism of neural network, the experience composition of SVM is very few, has more strict mathematic(al) argument.Meanwhile, SVM is less for the dependence of provided sample data, and generalization ability is comparatively strong, and locally optimal solution must be globally optimal solution, avoids generation dimension disaster, is specially adapted to solve the problems such as small sample, non-linear, local minimum point.On SVM basis, SuykensJ.A.K proposes least square method support vector machine (LS-SVM), is mainly incorporated in SVM by least square line sexual system, with the secondary quadratic term e of training error ²instead of the slack variable in optimization aim, and instead of inequality constrain by equality constraint, problem arises is for solving a system of linear equations the most at last, greatly reduces working time, improves the speed of training.

Have not yet to see document and the patent report of the flexible measurement method of the anaerobic wastewater treatment system water outlet volatile fatty acid based on principal component analysis (PCA)-least square method supporting vector machine.External M.Dixon is for waste water fermentation processing procedure, neural network is used to carry out data mining to VFA, the domestic Yao Chong tinkling of pieces of jade proposes a kind of soft-measuring modeling method based on fuzzy neural network on this basis and predicts VFA value, the two confirms the feasibility of VFA hard measurement, but on auxiliary variable variables choice, do not carry out systematic research with modeling method, there is some problems.

Summary of the invention

The object of this invention is to provide a kind of flexible measurement method of anaerobic wastewater treatment system water outlet volatile fatty acid.Principal component analysis (PCA) realizes the dimension-reduction treatment of higher-dimension auxiliary variable variable, least square method supporting vector machine solves the problems such as small sample, non-linear, local minimum point to being specially adapted to, and training speed is fast, the two combines the hard measurement realizing anaerobic wastewater treatment system volatile fatty acid, to the on-line monitoring of anaerobic wastewater treatment system and subsequent technique optimal control significant, the index that other are difficult to on-line monitoring to anaerobic wastewater treatment system simultaneously has directive significance.

To achieve these goals, technical scheme of the present invention is as follows.

Based on a flexible measurement method for the anaerobic wastewater treatment system water outlet volatile fatty acid of principal component analysis (PCA)-least square method supporting vector machine, comprise the following steps:

(1) build anaerobic wastewater treatment system, obtain auxiliary variable data sample A and the leading variable data sample B of water outlet volatile fatty acid;

(2) pre-service is carried out to the auxiliary variable data sample A collected and leading variable data sample B;

(3) principal component analysis (PCA) is carried out to pretreated auxiliary variable data sample A, set up the new auxiliary variable data sample based on principal component analysis (PCA) and new data sample is divided into training sample and test sample book;

(4) training sample that divides of corresponding step (3), sets up the water outlet volatile fatty acid soft-sensing model based on least square method supporting vector machine, and constantly training is until meet specification error;

(5) corresponding to the test sample book that step (3) divides, applying step (4) is set up and the water outlet volatile fatty acid soft-sensing model based on least square method supporting vector machine after training is tested, the validity of checking soft-sensing model;

(6) will to train and the water outlet volatile fatty acid soft-sensing model based on least square method supporting vector machine after testing embeds in industrial computer, utilize the configuration software MCGS in operating mode machine to build human-computer interaction interface;

(7) OPC technology is adopted to realize based on the exchanges data between the water outlet volatile fatty acid soft-sensing model of least square method supporting vector machine and the configuration software MCGS of industrial computer, the data that MCGS collects are delivered to the water outlet volatile fatty acid soft-sensing model based on least square method supporting vector machine, calculate the predicted value of water outlet water outlet Vfa Concentration, then this predicted value is back to the display of industrial computer human-computer interaction interface;

(8) constantly repeat step (6), thus realize the on-line real time monitoring of anaerobic wastewater treatment system water outlet volatile fatty acid.

Described step (1) anaerobic wastewater treatment system comprises anaerobic reactor, the monitoring relevant to anaerobic effluent volatile fatty acid and measuring instrument, industrial computer, A/D and D/A module; Described monitoring and measuring instrument are for detecting anaerobic reaction actuator temperature T, oxidation-reduction potential ORP, flow of inlet water Qin, reactor pH value, gas production rate Q _gas, methane number percent CH ₄%, carbon dioxide number percent CO ₂%, hydrogen percentages H ₂%, can obtain hydraulic detention time HRT by described flow of inlet water Qin.

Described step (1) auxiliary variable data sample A={A ₁, A ₂..., A _n, wherein A _j=[a _j1, a _j2..., a _jm] ^t(j≤n), n is auxiliary variable data dimension, and in the present invention, auxiliary variable data are by temperature T, oxidation-reduction potential ORP, hydraulic detention time HRT, reactor pH value, gas production rate Q _gas, methane number percent CH ₄%, carbon dioxide number percent CO ₂%, hydrogen percentages H ₂% forms, therefore n=8; Leading variable data sample B=[b ₁, b ₂... b _m] ^t, the Vfa Concentration corresponding by auxiliary variable sample forms, and m is number of samples, m>n.

First described step (2) preprocess method uses Pauta criterion to auxiliary variable sample A and the leading variable data sample B together paired rejecting abnormalities value of pre-service, then adopt formula (1) to be normalized respectively A and B, obtain auxiliary variable normalization matrix X _{m × n}with leading variable data normalization matrix Y _{m × 1}:

$x_{ij}=\frac{a_{ij}-a_{\min ,j}}{a_{\max ,j}-a_{\min ,j}},(i=1,2,...,m,j=1,2,...,n)$ (1) x in formula _ijfor variate-value after standardization, a _ijfor variate-value before standardization, a _{min, j}for the minimum value of a jth variable, a _{min, j}for the maximal value of a jth variable.

Described step (3) is to pretreated normalization matrix X _{m × n}carry out principal component analysis (PCA) to build new auxiliary variable data sample and refer to first to normalization sample matrix X _{m × n}carry out zero-mean standardization and obtain normalized matrix Z _{m × n}, the then covariance matrix R of Criterion matrix _{n × n}, obtain n eigenwert of correlation matrix and by descending sort from big to small, calculate the pivot contribution rate ρ that jth is individual _j, described step pivot contribution rate ρ _jcomputing formula be:

${\mathrm{\ρ}}_j=\frac{{\mathrm{\λ}}_j}{\underset{j=1}{\overset{n}{\Σ}}{\mathrm{\λ}}_j}$

λ in formula _jeigenvalue contribution rate of accumulative total ρ for correlation matrix R is greater than k the pivot of 85% as new auxiliary variable data sample, and the computing formula of described step pivot contribution rate ρ is:

$\mathrm{\ρ}=\frac{\underset{j=1}{\overset{k}{\Σ}}{\mathrm{\λ}}_j}{\underset{j=1}{\overset{n}{\Σ}}{\mathrm{\λ}}_j}$

λ in formula _jfor the eigenwert of correlation matrix R, j, k≤n.As follows particularly:

(3.1) first X is calculated by formula (2) _{m × n}average and variance, then utilize formula (3) calculate to X _{m × n}carry out zero-mean standardization and obtain normalized matrix Z _{m × n}:

$\stackrel{\‾}{X_j}=\frac{1}{m}{\mathrm{\Σ}}_{i=1}^m{x}_{ij},S_j=\sqrt{\frac{1}{m-1}{\mathrm{\Σ}}_{i=1}^m{(x_{ij}-\stackrel{\‾}{X_j})}^2},j=1,2,...n,n=8---\left(2\right)$

$Z_{ij}=\frac{x_{ij}-X_j}{S_j},i=1,2,...,m,j=1,2,...,n,m>n---\left(3\right)$

(3.2) formula (4) and (5) are utilized to ask normalized matrix Z _{m × n}covariance matrix R _{n × n}:

$R_{n\×n}=\frac{z^{T}z}{m-1}={\left(r_{ij}\right)}_{n\×n}---\left(4\right)$

$r_{ij}=\frac{1}{m-1}{\mathrm{\Σ}}_{k=1}^m{Z}_{ki}\×Z_{kj},i,j=1,2,...,n---\left(5\right)$

(3.3) n the eigenvalue λ of R is solved according to formula (6) _j(j=1,2 ..., n), and by order arrangement from big to small, λ ₁>=λ ₂>=...>=λ _n>=0, the unit character vector b of individual features value is separated according to formula (7) _j(j=1,2 ..., n), b _j=(b _1j, b _2j..., b _nj):

|R-λ _jE|＝0(6)

Rb＝λ _jb(7)

(3.4) calculate accumulative variance contribution ratio by formula (8), determine major component number k, front k major component of accumulative variance contribution ratio >=85% contains most information, and other compositions below can be given up:

$\frac{{\Σ}_{j=1}^k{\mathrm{\λ}}_j}{{\Σ}_{j=1}^n{\mathrm{\λ}}_j}\≥85\%---\left(8\right)$

(3.5) formula (9) is utilized to be projected in k dimension by normalized matrix Z, the new data sample matrix U of composition k pivot, U ₁be called first principal component, U ₂be called Second principal component... U _kbe called kth major component, such original sample achieves will be tieed up to k from n dimension:

U _i,j＝Z _i ^Tb _j,i＝1,2,…m,j＝1,2,…k(9)

(3.6) the new sample after pivot constituent analysis is divided into training sample and test sample book, in order to follow-up use

Training sample is used to set up and train the soft-sensing model based on least square method supporting vector machine in described step (4), support vector machine (SupportVectorMachine, SVM) be a kind of new intelligent algorithm developed on the VC dimension theory and Structural risk minization principle of Statistical Learning Theory, first put forward by people such as Vapnik, its basic thought is exactly the concept introducing kernel function, inseparable for lower dimensional space problem is mapped to higher dimensional space linear separability by kernel function, do at higher dimensional space and classify or return process, SVM it solution small sample, many distinctive advantages are had in Nonlinear Modeling and high dimensional pattern identification.Least square method supporting vector machine (LeastSquareSupportVectorMachine, LSSVM) proposed by SuykensJAK in calendar year 2001, it is the expansion of standard SVM, adopt least square line sexual system as loss function, transfer inequality constrain condition to equality constraint, greatly the complexity of shortcut calculation.

For given sample set D{ (x _i, y _i), i=1,2 ..., l}, wherein x _i∈ R ⁿfor n ties up auxiliary variable vector, y _i∈ R is target-dominant variable data, and LS-SVM can be described as following optimization problem:

${\min }_{\mathrm{\ω},b,\mathrm{\ξ}}J(\mathrm{\ω},\mathrm{\ξ})=\frac{1}{2}\left|\right|\mathrm{\ω}|{|}^2+\frac{1}{2}{\mathrm{\γ\Σ}}_{i=1}^l{\mathrm{\ξ}}_i^2,\mathrm{\γ}>0---\left(10\right)$

Wherein ξ _ibe the training error of i-th sample point, for empiric risk, in order to weigh the complicacy of machine learning, γ >0 be penalty factor also known as regularization parameter, in order to balancing machine in training study complicacy and empiric risk, formula (10) meets constraint condition:

y _i＝ω ^TΦ(x _i)+b+ξ _i,i＝1,2,…,l(11)

Introduce Lagrange function:

$L(\mathrm{\ω},b,\mathrm{\ξ},\mathrm{\α})=J(\mathrm{\ω},\mathrm{\ξ})-{\mathrm{\Σ}}_{i=1}^l{\mathrm{\α}}_i\[{\mathrm{\ω}}^T\mathrm{\Φ}\left(x_i\right)+b+{\mathrm{\ξ}}_i-y_i\]---\left(12\right)$

α in formula _ibe Lagrange multiplier, utilize Karush-Kuhn-Tucker ' s (KKT) optimal condition to be optimized above formula, ask local derviation to obtain to ω, b, ξ, α:

$\left\{\begin{array}{c}{}^{\frac{\∂L}{\∂\mathrm{\ω}}=\mathrm{\ω}-{\Σ}_{i=1}^l{\mathrm{\α}}_i\mathrm{\Φ}\left(x_i\right)=0}\\ \frac{\∂L}{\∂b}=-{\Σ}_{i=1}^l{\mathrm{\α}}_i=0\\ \frac{\∂L}{\∂{\mathrm{\ξ}}_{i,}}=\mathrm{\γ}{\Σ}_{i=1}^l{\mathrm{\ξ}}_i-{\Σ}_{i=1}^l{\mathrm{\α}}_i=0\\ \frac{\∂L}{\∂{\mathrm{\α}}_i}={\mathrm{\ω}}^T\mathrm{\Φ}\left(x_i\right)+b+{\mathrm{\ξ}}_i-y_i=0\end{array}\right.---\left(13\right)$

Eliminate ω, ξ, the problem of optimization just can be converted into linear equation below:

K (x in formula _i, x _j)=Φ (x _i) ^tΦ (x _j), i, j=1,2 ..., l, definition Ω _ij=K (x _i, x _j) be kernel function, the linear kernel function of conventional kernel function, Polynomial kernel function, Radial basis kernel function (RBF) and Sigmoid kernel function.The present invention sets up soft-sensing model to adopt Radial basis kernel function RBF, and this kernel function form is:

$K(x_i,x_j)=\exp (-\frac{{||x_i,x_j||}^2}{2{\mathrm{\σ}}^2})---\left(15\right)$

In formula, σ is core width, makes Ω={ Ω _ij| i, j=1,2 ..., l}, I=[1,1 ..., 1] ^t, α=[α ₁, α ₂..., α _l] ^t, y=[y ₁, y ₂, y _l] ^t, then above formula can abbreviation be:

$\left[\begin{array}{cc}0& I^T\\ I& \mathrm{\Ω}+\frac{I}{\mathrm{\γ}}\end{array}\right]\left[\begin{array}{c}b\\ \mathrm{\α}\end{array}\right]=\left[\begin{array}{c}0\\ y\end{array}\right]---\left(16\right)$

Utilize the above-mentioned system of linear equations of least square solution can try to achieve the estimation of α and b, then estimate that the soft-sensing model of gained is:

$y\left(x\right)={\mathrm{\Σ}}_{i=1}^l{\mathrm{\α}}_{i}K(x,x_i)+b---\left(17\right)$

In formula, x is auxiliary variable sample data, α _ifor Lagrange multiplier, b is side-play amount, K (x, x _i) be kernel function.

In described step (4) when utilizing LS-SVM modeling, suitable kernel function, nuclear parameter σ ²decisive role is played with the performance of regularization parameter γ to model.Because anaerobic digestion process is complicated nonlinear system, we choose radial basis function RBF as kernel function and adopt grid data service definite kernel parameter σ ²with regularization parameter γ optimized scope, then finally select nuclear parameter σ with cross-validation method ²with regularization parameter γ optimal value.

The test data using in described step (5) step (3) to divide, to set up and the LS-SVM soft-sensing model trained is tested, assesses its estimated performance.Estimated performance index comprises following:

(1) absolute error (AbsoluteError, AE)

AE＝y _p,i-y _i,i＝1,2,…m(18)

Y in formula _p,ithe predicted value of least square method supporting vector machine, y _irepresent actual value, AE represents that obtained result deducts measured true value, and it has size and symbol, represents the degree of measurement result deviation true value.

(2) relative error (RelativeError, RE)

$RE=\frac{y_{p,i}-y_i}{y_i}\×100\%,i=1,2,...m---\left(19\right)$

RE represents the ratio of relative error magnitudes and the actual value of tested value, and relative error more can reflect the degree of reliability of prediction.

(3) mean absolute percentage error (MeanAbsolutePercentError, MAPE)

$MAPE=\frac{1}{m}{\mathrm{\Σ}}_{i=1}^m\left|\frac{y_{p,i}-y_i}{y_i}\right|\×100\%,i=1,2,...m---\left(20\right)$

MAPE is the mean value of the absolute value summation of all relative errors, can better reflect the actual conditions of predicted value.

(4) root-mean-square error (RootMeanSquareError, RMSE)

$RMSE=\sqrt{\frac{1}{m}{\mathrm{\Σ}}_{i=1}^m{(y_{p,i}-y_i)}^2},i=1,2,...m---\left(21\right)$

RMSE is mainly in order to illustrate the dispersion degree of sample.The value of RMSE is less, illustrate that forecast model describes experimental data and has better levels of precision, otherwise model prediction accuracy is poor.

(5) related coefficient (correlationcoefficient, R)

$R=\frac{{\mathrm{\Σ}}_{i=1}^m(y_i-\stackrel{\‾}{y})(y_p-\stackrel{\‾}{y_p})}{\sqrt{{\mathrm{\Σ}}_{i=1}^m{(y_i-\stackrel{\‾}{y})}^2{\mathrm{\Σ}}_{i=1}^m{(y_p-\stackrel{\‾}{y_p})}^2}},i=1,2,...m---\left(22\right)$

In formula for actual value average, for predicted value average, m is number of samples, and R reflects the power of predicted value and actual value linear relationship, R more close to 1 predicted value and actual value more close.

Soft-sensing model set up and after testing the least square method supporting vector machine model insertion trained in industrial computer configuration software MCGS, employing OPC technology realizes the exchanges data between volatile fatty acid least square method supporting vector machine soft-sensing model and configuration software, the data gathered by configuration software MCGS are delivered to water outlet VFA least square method supporting vector machine soft-sensing model, calculate the predicted value of water outlet VFA, again this value is back to industrial computer human-computer interaction interface, continuous repetition above-mentioned steps, thus realize the on-line real time monitoring of anaerobic wastewater treatment system water outlet VFA.

The described least square method supporting vector machine soft-sensing model of described step (4) is:

$y\left(x\right)=\underset{i=1}{\overset{l}{\Σ}}{\mathrm{\α}}_{i}K(x,x_i)+b$

In formula, x is auxiliary variable sample data, α _ifor Lagrange multiplier, b is side-play amount, K (x, x _i) be kernel function.

The test sample book that described step (5) corresponding step (3) divides, the least square method supporting vector machine model performance index that testing procedure (4) is set up comprises error (ERR), relative error (RE), mean absolute percentage error (MAPE), root-mean-square error (RMSE) and coefficient R.

In sum, the anaerobic wastewater treatment system water outlet volatile fatty acid soft-sensing model of the Based PC A-LSSVM set up is theoretical correct, meet better to anaerobic effluent volatile fatty acid predicted value and True Data, error is relative also less, meet practical application completely, the present invention has the following advantages and beneficial effect relative to prior art tool:

(1) principal component analysis (PCA) combines with least square method supporting vector machine and sets up PCA-LSSVM soft-sensing model by the present invention, PCA technology can be avoided owing to selecting auxiliary variable variable by experience and miss important information, can reduce auxiliary variable dimension simultaneously and reduce model complexity.

(2) problem of measuring equipment length time lag can be solved to the hard measurement of water outlet volatile fatty acid, real realization is to the comprehensively real-time monitoring of wastewater effluent water quality, prevent pop-up threat, may be used for the optimal control that FEEDBACK CONTROL realizes anaerobic wastewater treatment system simultaneously.

(3) the present invention may be used for replacing the expensive measuring equipment of part, maintenance cost saving, reduces cost for wastewater treatment, easily applies in waste water treatment engineering, have good Social benefit and economic benefit.

Accompanying drawing explanation

Fig. 1 is Based PC A-LSSVM volatile fatty acid soft-sensing model design flow diagram.

Fig. 2 is the Pareto figure that principal component analysis (PCA) major component adds up variance contribution ratio.

The load diagram (biplot) of Fig. 3 principal component analysis (PCA).

The comparison diagram of Fig. 4 VFA training data and test data actual value and predicted value

Fig. 5 soft-sensing model predicated error curve

Fig. 6 soft-sensing model relative error curve.

Fig. 7 soft-sensing model checking related coefficient figure.

Exchanges data flow process between Fig. 8 MCGS and MATLAB

Embodiment

Below in conjunction with specific embodiment and accompanying drawing thereof, the present invention is further elaborated.Embodiment is only for explaining instead of limitation of the present invention the present invention.If have the process of special detailed description or symbol (as suspension points) to be all that those skilled in the art can refer to prior art understanding or realize it is noted that following.

Anaerobic wastewater treatment system based on principal component analysis (PCA)-least square method supporting vector machine goes out a flexible measurement method for volatile fat sour water, and whole hard measurement flow process is shown in Fig. 1.The present embodiment chooses the IC process system built in laboratory as anaerobic wastewater treatment system, uses MCGS as configuration software, uses MATLAB as soft sensor modeling instrument, is comprised the following steps by simulating, verifying performance of the present invention and feasibility:

(1) build anaerobic wastewater treatment system, adopt IC anaerobic reactor as anaerobic treatment process, setting hydraulic detention time HRT is 24,15.36,12.29 and 9.83 hours, gathers T, pH, ORP, Q under each hydraulic detention time _gas, CH ₄%, CO ₂%, H ₂% and corresponding volatile fatty acid VFA.

(2) carry out pre-service to the data collected, first use Pauta criterion to reject outlier, finally obtain 90 groups of data and be then normalized between data compression to [0,1] sample data, normalization is undertaken by following formula.

$x_{ij}=\frac{a_{ij}-a_{min,j}}{a_{\max ,j}-a_{\min ,j}},(i=1,2,...,m,j=1,2,...,n)$

(3) to sample data principal component analysis (PCA) after normalization, specific formula for calculation is shown in that instructions no longer repeats here.First to sample data standardization, normalized matrix Z is asked _{m × n}covariance matrix R _{n × n}, solve n the eigenvalue λ of R _j(j=1,2 ..., n), and by order arrangement from big to small, λ ₁>=λ ₂>=...>=λ _n>=0, separate the unit character vector b of individual features value _j(j=1,2 ..., n), b _j=(b _1j, b _2j..., b _nj), calculate each major component variance contribution ratio ρ _jas shown in table 1, Figure 2 shows that each major component adds up the Pareto figure of variance contribution ratio, Fig. 3 is the load diagram (biplot) of each variable in major component 1 and major component 2, variable positive correlation spatially nearer is in the figure stronger, from initial point more away from variable by major component 1 and major component 2 explain more complete.The accumulative variance contribution ratio of front 4 major components is 85.33%, and met the requirement of accumulative variance contribution ratio >=85%, but front 5 major components add up variance contribution ratio reaches 91.89%, therefore determine major component number 5, other compositions are below given up.Finally extract the modeling training data of the front 70 groups of data after principal component analysis (PCA) as least square method supporting vector machine, rear 20 groups of data make test data.

The variance contribution ratio of each major component of table 1

(4) utilize training data in MATLAB, set up least square method supporting vector machine soft-sensing model, Selection of kernel function radial basis function RBF, adopt grid data service definite kernel parameter σ ²with regularization parameter γ optimized scope, then finally select the parameter value of leading variable data optimum with cross-validation method, last arrives σ ², γ optimal value is 129.6434 and 0.03470267.

(5) utilize test data test model performance, Fig. 4 is the comparison diagram of training data and test data VFA actual value and predicted value.The degree that absolute error AE reflects the deviation true value that predicts the outcome is shown in Fig. 5, relative error RE relative error more can reflect that the degree of reliability of prediction is shown in Fig. 6, mean absolute percentage error MAPE is the mean value of the absolute value summation of all relative errors, the actual conditions of predicted value can better be reflected, train MAPE to be 0.647% in this example, prediction MAPE is 0.619%.Root-mean-square error RMSE is mainly in order to illustrate the dispersion degree of sample, and training and predicted root mean square error are respectively 0.0049 and 0.0040.Coefficient R reflects the power of predicted value and actual value linear relationship, and this routine related coefficient reaches 0.99995 and sees Fig. 7.As can be seen from above performance index, the soft-sensing model based on the anaerobic wastewater treatment system water outlet volatile fatty acid of principal component analysis (PCA)-least square method supporting vector machine predicts the value of anaerobic wastewater treatment system volatile fatty acid comparatively accurately.

(6) the exchanges data flow process between MCGS and MATLAB is shown in Fig. 8.Installation MCGS configuration software and senior kit thereof are to correct path.Startup MCGS then opc server function starts automatically.According to the good human-computer interaction interface of engine request configuration.In the engineering that configuration is good, define HRT, T, pH, ORP, Q _gas, CH ₄%, CO ₂%, H ₂% and water outlet volatility hard measurement rreturn value VFA_SOFT9 parameter, carry out read-write operation for MCGS.Start MATLAB, auxiliary variable order opcregister (' install ') in command window, be used for installing provided by OPCFoundation a set ofly can browse other computing machines on network and the core component that can communicate.Auxiliary variable order again: a=opcda (' localhost ', ' MCGS.OPC.Server '); Connect (da), then MCGS and MATLAB connects.After the PCA-LSSVM soft-sensing model trained is embedded industrial computer, during system cloud gray model, computing machine is the waterpower HRT collected, T, pH, ORP, Qgas, CH ₄%, CO ₂%, H ₂the data of % pass to software, run PCA-LSSVM soft-sensing model and carry out prediction water outlet VFA value, hard measurement value is turned back to human-computer interaction interface.

Above-described embodiment is the present invention's preferably embodiment; but embodiments of the present invention are not restricted to the described embodiments; change, the modification done under other any does not deviate from Spirit Essence of the present invention and principle, substitute, combine, simplify; all should be the substitute mode of equivalence, be included in protection scope of the present invention.

Claims (9)

1. a flexible measurement method for anaerobic wastewater treatment system water outlet volatile fatty acid, is characterized in that comprising the following steps:

(1) build anaerobic wastewater treatment system, obtain auxiliary variable data sample A and the leading variable data sample B of water outlet volatile fatty acid;

(2) pre-service is carried out to the auxiliary variable data sample A collected and leading variable data sample B;

(3) principal component analysis (PCA) is carried out to pretreated auxiliary variable data sample A, set up the new auxiliary variable data sample based on principal component analysis (PCA) and new data sample is divided into training sample and test sample book;

(4) training sample that divides of corresponding step (3), sets up the water outlet volatile fatty acid soft-sensing model based on least square method supporting vector machine, and constantly training is until meet specification error;

(5) corresponding to the test sample book that step (3) divides, applying step (4) is set up and the water outlet volatile fatty acid soft-sensing model based on least square method supporting vector machine after training is tested, the validity of checking soft-sensing model;

(6) will to train and the water outlet volatile fatty acid soft-sensing model based on least square method supporting vector machine after testing embeds in industrial computer, utilize the configuration software MCGS in operating mode machine to build human-computer interaction interface;

(7) OPC technology is adopted to realize based on the exchanges data between the water outlet volatile fatty acid soft-sensing model of least square method supporting vector machine and the configuration software MCGS of industrial computer, the data that MCGS collects are delivered to the water outlet volatile fatty acid soft-sensing model based on least square method supporting vector machine, calculate the predicted value of water outlet water outlet Vfa Concentration, then this predicted value is back to the display of industrial computer human-computer interaction interface;

(8) constantly repeat step (6), thus realize the on-line real time monitoring of anaerobic wastewater treatment system water outlet volatile fatty acid.

2. the flexible measurement method of a kind of anaerobic wastewater treatment system water outlet volatile fatty acid according to claim 1, it is characterized in that: described step (1) anaerobic wastewater treatment system comprises anaerobic reactor, the monitoring relevant to anaerobic effluent volatile fatty acid and measuring instrument, industrial computer, A/D and D/A module; Described monitoring and measuring instrument are for detecting anaerobic reaction actuator temperature T, oxidation-reduction potential ORP, flow of inlet water Qin, reactor pH value, gas production rate Q _gas, methane number percent CH ₄%, carbon dioxide number percent CO ₂%, hydrogen percentages H ₂%, can obtain hydraulic detention time HRT by described flow of inlet water Qin.

3. the flexible measurement method of a kind of anaerobic wastewater treatment system water outlet volatile fatty acid according to claim 1, is characterized in that: described step (1) auxiliary variable data sample A is by temperature T, oxidation-reduction potential ORP, hydraulic detention time HRT, reactor pH value, gas production rate Q _gas, methane number percent CH ₄%, carbon dioxide number percent CO ₂%, hydrogen percentages H ₂% forms, and leading variable data sample B is made up of the Vfa Concentration of correspondence.

4. the flexible measurement method of a kind of anaerobic wastewater treatment system water outlet volatile fatty acid according to claim 1, it is characterized in that: described step (2) pre-service comprises according to Pauta criterion rejecting abnormalities value, is then normalized sample.

5. the flexible measurement method of a kind of anaerobic wastewater treatment system water outlet volatile fatty acid according to claim 3, it is characterized in that: described step (3) is carried out principal component analysis (PCA) to the normalization matrix obtained after pre-service and comprised the data sample after to normalization and carry out zero-mean standardization and obtain normalized matrix Z, the correlation matrix R of Criterion matrix Z, obtain n the eigenwert of correlation matrix R and by descending sort from big to small, calculate the pivot contribution rate ρ of jth pivot _jwith the accumulative pivot contribution rate ρ of a front k pivot, choose accumulative pivot contribution rate and be greater than k the pivot of 85% as new data sample, and be divided into training sample and test sample book, wherein j, k≤n.

6. the flexible measurement method of a kind of anaerobic wastewater treatment system water outlet volatile fatty acid according to claim 4, is characterized in that: described pivot contribution rate ρ _jcomputing formula be:

λ in formula _jfor the eigenwert of correlation matrix R.

7. the flexible measurement method of a kind of anaerobic wastewater treatment system water outlet volatile fatty acid according to claim 4, is characterized in that: the computing formula of described step pivot contribution rate ρ is:

λ in formula _jfor the eigenwert of correlation matrix R, j, k≤n.

8. the flexible measurement method of a kind of anaerobic wastewater treatment system water outlet volatile fatty acid according to claim 1, is characterized in that: the described least square method supporting vector machine soft-sensing model of described step (4) is:

In formula, x is auxiliary variable sample data, α _ifor Lagrange multiplier, b is side-play amount, K (x, x _i) be kernel function.

9. the flexible measurement method of a kind of anaerobic wastewater treatment system water outlet volatile fatty acid according to claim 1, it is characterized in that: the test sample book that described step (5) corresponding step (3) divides, the least square method supporting vector machine model performance index that testing procedure (4) is set up comprises error (ERR), relative error (RE), mean absolute percentage error (MAPE), root-mean-square error (RMSE) and coefficient R.