CN106407648A - Rapid batch forecast method for key property of gasoline - Google Patents

Abstract

The invention discloses a rapid batch forecast method for key property of gasoline. The method is used for a refinery enterprise and comprises the steps of firstly, performing spectrum pre-processing by employing a baseline correction, cutout and vector normalization method on the basis of near-infrared spectrum of the gasoline; secondly, obtaining a coefficient matrix by employing fast Fourier transform, and performing BP neural network training; and finally, performing batch forecast on the key property of the gasoline. Compared with a general nonlinear modeling method, the method has the advantages that the batch forecast of a plurality of properties is achieved, the forecast accuracy is ensured, meanwhile, the real-time performance of forecasting is improved, real-time control on on-line blending of the gasoline of the refinery enterprise is facilitated, and the economic benefit of the enterprise is further improved.

Description

A kind of rapid batch Forecasting Methodology of gasoline key property

Technical field

The present invention relates to the gasoline property context of detection of petrochemical industry, specifically a kind of based on nerual network techniqueGasoline key property Forecasting Methodology.

Background technology

During oil refining processing and gasoline concoction, need to analyze multiple properties of gasoline product, such as research octane number (RON)(RON), anti-knock index, boiling range etc..Quick and precisely prediction gasoline property plays an important role in oil refining processing and gasoline concoction.

For improve gasoline property prediction rapidity, current someone on the basis of conventional offset minimum binary (PLS), using manyThe PLS of dependent variable is predicted, can the multiple property of batch forecast.However, this method based on PLS is only to PIn matter, the prediction of the linearity preferable property has preferable precision.

In the key property of gasoline, property and the near infrared spectrum such as 10% evaporating temperature, 50% evaporating temperature, end point of distillationBetween there is stronger non-linear relation, the non-linear modeling method therefore with artificial neural network as representative have started to applyQuick detection in gasoline property.This method improves the precision of prediction of model to a certain extent, but what the method was set upModel is normally only used for predicting single gasoline property.The neutral net of this single output, needs when predicting multiple property to buildFound multiple models, the operation and maintenance work of each model is very numerous and diverse, brings difficulty to engineering real-time application.

Content of the invention

For solving the problems, such as prior art, the present invention proposes a kind of rapid batch Forecasting Methodology of gasoline key property,Gasoline sample in library of spectra is carried out conventional pretreatment Fast Fourier Transform (FFT) by the method first, then sets up the BP god of multi outputThrough network analysis model, finally according to this model, batch forecast is carried out to multiple properties of sample to be tested.Specifically include following steps：

(1) it is based on gasoline atlas of near infrared spectra, first Pretreated spectra is carried out to library of spectra sample and gasoline sample to be measured；

(2) pretreated spectroscopic data is carried out Fast Fourier Transform (FFT), obtain the coefficient matrices A of Fast Fourier Transform (FFT)；

(3) choose the front m row of A as the input of neutral net and Configuration network parameter；

(4) pretreated library of spectra gasoline sample, as training sample, carries out neural metwork training；

(5) using the neural network model training, batch forecast is carried out to multiple properties of sample to be tested.

The key property that this method is predicted include research octane number (RON), anti-knock index, density, 10% evaporating temperature,50% evaporating temperature and the end point of distillation.

Preferably, Pretreated spectra includes baseline correction, intercepting and vector normalization.

Preferably, this method chooses 6400cm^-1And 9200cm^-12 wave number points are as two basic points of baseline correction.

Baseline correction is passed through formula (1) and is calculated：

In formula, x_iFor gasoline near infrared spectrum wave number；kx_i+ b is through 6400cm^-1And 9200cm^-12 points straightLine equation, wherein k are this straight slope, and b is this Linear intercept；y_iRepresent former spectrogram in wave number x_iUnder absorbance；Represent baseAfter line correction, spectrogram is in wave number x_iUnder absorbance.

Preferably, this method chooses 4000cm^-1～4800cm^-1Spectrogram modeling in wave number section.

This method, when spectrogram is carried out with vector normalization, is calculated using formula (2)：

In formula, X_ijRefer to i-th sample absorbance under wave number j；Refer to the absorbance values of i-th sample；M isThe number of wave number point；X_ij^*Absorbance under wave number j for i-th sample after expression vector normalization.

After pretreatment, gasoline near-infrared spectrogram high fdrequency component is less, and the coefficient matrices A after FFT typically existsAmplitude very little after 20th Frequency point, therefore this method choose the input as neutral net for front 20 row of A, i.e. m=20.

The parameter of neutral net is configured using following：The number of hidden nodes is 30, and output node number is 6, i.e. gasoline to be measuredThe number of key property, hidden layer excitation function tansig, output layer excitation function purelin, train function trainlm, performanceFunction mse, performance arget value 0.05, learning coefficient 0.02.

The present invention adopts root-mean-square error for the evaluation of property j final result, i.e. RMSE_j.Root-mean-square error is in engineeringIt is widely adopted in measurement, it is sensitive to the especially big or special little error reflection in one group of measurement, can reflect measurement wellPrecision.RMSE_jCalculated by formula (3)：

In formula, n is the number of gasoline to be measured；Refer to the predicted value of property j of i-th gasoline to be measured；x_ijRefer to i-th to be measuredThe actual value of property j of gasoline.RMSE_jValue is less, illustrates higher to the accuracy of property j prediction, prediction effect is better.

Beneficial effect：

Detection method provided by the present invention is based on gasoline near infrared spectrum, combines nerve net using Fast Fourier Transform (FFT)Network technology, realizes the batch quick detection of gasoline key property.The present invention is to the baseline correction of spectrogram, intercepting and neutral netInput has carried out being directed to Sexual behavior mode, ensure that precision of prediction while reducing amount of calculation.With general non-linear modeling methodCompare, this method energy is quick, Accurate Prediction gasoline key property, contributes to the real-time control of the gasoline on-line blending of Petrochemical EnterprisesSystem, and then improve the economic benefit of enterprise.

In conjunction with Fig. 2, it is the typical atlas of near infrared spectra of 92# product oil, it can be found that in 6000cm^-1～10000cm^-1RippleIn section, spectrogram is relatively steady, 6400cm^-1And 9200cm^-12 points of absorbance is relatively low, and therefore this method chooses 6400cm-¹With9200cm^-1, as two basic points of baseline correction, amount of calculation is few and accuracy is high for 2 wave number points.

Because gasoline near-infrared spectrogram contains much noise in high frequency region, the spectrogram information of low frequency range is less, therefore can notUsing whole near infrared spectrums as modeling wave band, need to carry out spectrogram intercepting.This method finds through test, 4000cm^-1~4800cm^-1Spectrogram modeling effect in wave number section is best.

Brief description

Fig. 1 predicts procedural block diagram for gasoline property rapid batch

Fig. 2 is the typical atlas of near infrared spectra of 92# product oil

Specific embodiment

The invention will be further described with case study on implementation below in conjunction with the accompanying drawings.

The present invention, introduces the gasoline key property Forecasting Methodology based on nerual network technique taking certain 92# product oil as a example.Table1 is the numbering of certain all sample of 92# product oil and its corresponding property.

Certain 92# product oil sample number of table 1 and corresponding property

In Table 1, the sample of numbering 92#-1～99 is Sample Storehouse sample, and the sample of numbering 92#-100～108 is to be measuredSample.The near infrared spectrum data of all gasoline samples is carried out, after conventional pretreatment, completing using Matlab function fft ()Fast Fourier Transform (FFT), obtains transform coefficient matrix A.To obtaining atlas of near infrared spectra after matrix A delivery under each Frequency pointAmplitude, intercepts front 20 row.Table 2 gives the coefficient amplitude of part sample (92#-1～8).

The amplitude of the Fast Fourier Transform (FFT) coefficient of table 2 part 92# product oil sample

Complete the training of multi output BP network, first Configuration network parameter using Matlab Neural Network Toolbox：Hidden layerNodes 30, output node number 6, hidden layer excitation function tansig, output layer excitation function purelin, train functionTrainlm, performance function mse, performance arget value 0.05, learning coefficient 0.02.Then the gasoline sample being 92#-1～99 by numberingFront 20 row of this coefficient matrices A, as network inputs, carry out BP network training.

After the completion of training, before the gasoline sample coefficient matrix that numbering is 92#-100～109,20 row are defeated as networkEnter, carry out the prediction of BP network, predicting the outcome of 92#-100～109 is as shown in table 3.

The predicting the outcome of table 3 92# product oil sample to be tested property

Table 4 is the root-mean-square error of 92# product oil sample to be tested property.

The predicated error of table 4 92# product oil sample to be tested

In order to contrast, the test experiments that conventional neural networks predict the single property of gasoline, test result such as table 5 institute are carried outShow.

The experimental result of the single property of table 5 neural network prediction

Contrast table 4 and table 5 are it is found that in addition to indivedual property such as 10% evaporating temperature, the organon of this method prediction is pungentThe root-mean-square error of the properties such as alkane value, anti-knock index is superior to the neural net prediction method of routine.This shows, this method forThe prediction effect of gasoline key property is preferable, and energy batch forecast, and real-time is more excellent.

Claims (9)

1. a kind of rapid batch Forecasting Methodology of gasoline key property is it is characterised in that the method comprises the following steps：

(1) it is based on gasoline atlas of near infrared spectra, first Pretreated spectra is carried out to library of spectra sample and gasoline sample to be measured；

(2) pretreated spectroscopic data is carried out Fast Fourier Transform (FFT), obtain the coefficient matrices A of Fast Fourier Transform (FFT)；

(3) choose the front m row of A as the input of neutral net and Configuration network parameter；

(4) pretreated library of spectra gasoline, as training sample, carries out neural metwork training；

(5) using the neural network model training, batch forecast is carried out to multiple properties of sample to be tested and obtain the final of propertyEvaluation of result.

2. a kind of rapid batch Forecasting Methodology of gasoline key property according to claim 1 is it is characterised in that step (1)Described in Pretreated spectra include baseline correction, intercepting and vector normalization.

3. a kind of rapid batch Forecasting Methodology of gasoline key property according to claim 2 is it is characterised in that described baseTwo basic points of line correction choose 6400cm^-1And 9200cm^-12 wave number points, baseline correction is calculated by following formula：

$y_i^{*}=y_i-({\mathrm{kx}}_i+b)$

In formula, x_iFor gasoline near infrared spectrum wave number；kx_i+ b is 6400cm in spectrogram^-1And 9200cm^-1Two wave numbersThe linear equation of point, wherein k is this straight slope, and b is this Linear intercept；y_iRepresent former spectrogram in wave number x_iUnder absorbance；After representing baseline correction, spectrogram is in wave number x_iUnder absorbance.

4. a kind of rapid batch Forecasting Methodology of gasoline key property according to claim 2 is it is characterised in that intercept4000cm^-1～4800cm^-1Interior spectrogram is modeling.

5. a kind of rapid batch Forecasting Methodology of gasoline key property according to claim 2 is it is characterised in that to spectrogramWhen carrying out vector normalization, calculated by following formula：

${X_{ij}}^{*}=\frac{X_{ij}-{\stackrel{\‾}{X}}_i}{\sqrt{\underset{j=1}{\overset{m}{\Σ}}{(X_{ij}-{\stackrel{\‾}{X}}_i)}^2}}$

In formula, X_ijRefer to i-th sample absorbance under wave number j；Refer to the absorbance values of i-th sample；M is wave numberThe number of point；X_ij^*Absorbance under wave number j for i-th sample after expression vector normalization.

6. the rapid batch Forecasting Methodology of a kind of gasoline key property according to any one of Claims 1 to 5, its feature existsParameter in neutral net is configured using following：The number of hidden nodes is 30, and output node number is 6, hidden layer excitation function tansig,Output layer excitation function purelin, trains function trainlm, performance function mse, performance arget value 0.05, learning coefficient0.02.

7. the rapid batch Forecasting Methodology of a kind of gasoline key property according to any one of Claims 1 to 5, its feature existsBefore the coefficient matrix of the method selection Fast Fourier Transform (FFT), 20 row are as the input of neutral net, i.e. m=20.

8. the rapid batch Forecasting Methodology of a kind of gasoline key property according to any one of Claims 1 to 5, its feature existsIn the method, root-mean-square error is adopted for the final result evaluation of property j, i.e. RMSE_j, calculated by following formula：

${\mathrm{RMSE}}_j=\sqrt{\frac{\underset{i=1}{\overset{n}{\Σ}}{({\hat{x}}_{ij}-x_{ij})}^2}{n}}$

In formula, n is the number of gasoline sample to be measured；Refer to the predicted value of property j of i-th sample to be tested；x_ijRefer to i-th to be measuredThe actual value of property j of sample.

9. the rapid batch Forecasting Methodology of a kind of gasoline key property according to any one of Claims 1 to 5, its feature existsInclude in described key property：Research octane number (RON), anti-knock index, density, 10% evaporating temperature, 50% evaporating temperature and endEvaporate a little.