A method for predicting gasoline octane number

Abstract

This disclosure relates to a method for predicting the octane number of gasoline. A first calibration model is established by performing partial least squares multiple regression analysis on the first near-infrared spectrum of a blended gasoline sample and its standard octane number. Then, based on the construction process of the first calibration model, the spectral fitting residual matrix and the octane number fitting residual matrix are selected and used as the input signal and teacher signal of an extreme learning machine, respectively, to obtain a second calibration model. The combined application of the first and second calibration models comprehensively considers both the linear and nonlinear relationships between the first near-infrared spectrum and the octane number of the blended gasoline sample, providing a wider range of octane number prediction capabilities and improving the model's prediction accuracy. Furthermore, selecting the absorbance in the characteristic spectral region and the standard octane number for regression analysis during the first model construction process can further improve the accuracy of the octane number prediction results.

Description

Technical Field

[0001] This disclosure relates to the field of octane number detection, and more specifically, to a method for predicting the octane number of gasoline. Background Technology

[0002] Gasoline is one of the most profitable products for oil refineries, accounting for 60% to 70% of total profits. Gasoline production involves two main stages: the production and storage of component oils and blending. Generally, gasoline blending is the final stage before gasoline product delivery. This process is very complex, but the profit margin is huge, and it was once considered a key process for the successful operation of oil refineries. In the gasoline blending process, octane number is the most important quality indicator for gasoline; therefore, rapid measurement of octane number is crucial to the entire blending process.

[0003] Near-infrared spectroscopy (NIRS) has attracted much attention due to its advantages of rapid and non-destructive analysis. NIRS determination of gasoline octane number is one of the most widely and deeply studied testing methods in the petrochemical field. Octane number has a strong correlation with the molecular structure of gasoline; for example, an increase in the number of methyl groups, olefins, and aromatic hydrocarbons (CH4) leads to a higher octane number, while an increase in the number of methylene groups (CH4) leads to a lower octane number. Therefore, for the same type of gasoline (with a narrow range of octane number variation), there is a good linear relationship between near-infrared spectral absorbance and octane number. A highly practical analytical model can be established using partial least squares methods.

[0004] However, during gasoline blending, the mixing of different types of gasoline components (such as reforming components, cracking components, alkylation components, etc.) or the addition of antiknock agents such as MTBE increases the octane number variation range of the finished gasoline, weakens the linear relationship between near-infrared absorbance and octane number, and strengthens the nonlinear relationship. If a partial least squares linear analysis model is directly established using this data, the resulting model may suffer from reduced octane number prediction accuracy due to its weak ability to interpret nonlinear information. The existing solution to this problem is to divide the octane number range into narrower intervals and establish corresponding partial least squares models. In predicting the octane number of a test sample, a pattern recognition method is used to select the appropriate partial least squares model for the sample's spectrum. This model is then used to predict the octane number based on the sample's spectrum. Essentially, this method leverages the strong linear relationship between near-infrared absorbance and octane number within a narrow octane number range, using interval modeling to minimize the strong nonlinear relationship between near-infrared absorbance and octane number in blended gasoline with a wide octane number range. However, in practical applications, spectral identification and excessive model parameters often increase the time cost of prediction and the maintenance cost of the model, especially in operating conditions with a wide octane number range, thus limiting its practical application value. Of course, there are also near-infrared spectral prediction models for blended gasoline with a wide octane number range based on partial least squares methods, but these models typically improve robustness at the expense of prediction accuracy, with typical prediction standard deviations exceeding 0.3. Summary of the Invention

[0005] The purpose of this disclosure is to provide a method for predicting the octane number of gasoline. This method comprehensively considers the linear and nonlinear relationships between the near-infrared spectral absorbance of blended gasoline and its octane number, and has a wider range of octane number prediction capabilities and better prediction accuracy. It is especially suitable for predicting the octane number of gasoline with a large range of octane number variations (with strong nonlinearity).

[0006] To achieve the above objectives, this disclosure provides a method for predicting the octane number of gasoline, comprising the following steps:

[0007] Obtain the first near-infrared spectra of multiple blended gasoline samples and the standard octane number of each of the blended gasoline samples;

[0008] Obtain the absorbance of each characteristic spectral region in the first near-infrared spectrum;

[0009] Partial least squares multiple regression analysis was performed on the absorbance of the characteristic spectral region of all the blended gasoline samples and the standard octane number of all the blended gasoline samples to obtain the first correction model;

[0010] Based on the first calibration model, the spectral fitting residual matrix and the octane number fitting residual matrix corresponding to the f-th principal factor are selected respectively. The spectral fitting residual matrix is ​​used as the input signal of the extreme learning machine, and the octane number fitting residual matrix is ​​used as the teacher signal of the extreme learning machine to obtain the second calibration model.

[0011] Collect the second near-infrared spectrum of the gasoline sample to be tested; obtain the absorbance of the characteristic spectral region of the second near-infrared spectrum;

[0012] The predicted value of the first octane number is determined based on the absorbance of the characteristic spectral region of the second near-infrared spectrum and the first correction model.

[0013] The spectral fitting residual matrix corresponding to the f-th principal factor of the gasoline sample to be tested is selected according to the first correction model, and the second octane number prediction value is determined according to the spectral fitting residual matrix and the second correction model.

[0014] The predicted octane number of the gasoline to be tested is determined based on a linear combination of the first predicted octane number and the second predicted octane number.

[0015] Optionally, the characteristic spectral region is 5780-7700 cm⁻¹. -1 .

[0016] Optionally, the method further includes:

[0017] The spectral fitting residual matrix Ex is determined according to the following equation (1) or equation (2):

[0018] Ex = Ex(f) - Ex(f) best (1); Where Ex(f) represents the spectral residual matrix when the first correction model takes the f-th principal factor, Ex(f) best () indicates that the first calibration model takes the optimal principal factor f. best The spectral residual matrix at time f ≤ f best ;

[0019] Ex = Ex(f) - Ex(f) best+1～n (2); Ex(f best+1～n () indicates that the first calibration model takes the optimal principal factor f. best The spectral residual matrix of any one of the 1-n principal factors, where n is an integer from 2 to 15;

[0020] Optionally, the number of spectral fitting residual matrices Ex can be one or more.

[0021] Optionally, the method further includes:

[0022] The octane number fitting residual matrix Ey is determined according to the following equation (3) or (4):

[0023] Ey=Ey(f)-Ey(f best (3); Where Ey(f) represents the octane number residual matrix when the first correction model takes the f-th principal factor, Ey(f) best () indicates that the first calibration model takes the optimal principal factor f. best The octane number residual matrix at time f ≤ f best ;

[0024] Ey=Ey(f) (4);

[0025] Optionally, the number of octane number fitting residual matrices Ey can be one or more.

[0026] Optionally, the linear combination includes:

[0027] The first octane number prediction value and the second octane number prediction value are linearly summed.

[0028] Optionally, the method further includes:

[0029] Based on the second spectral fitting residual matrix and the second correction model, multiple predictions are made, and the average of the multiple prediction results is taken as the second octane number prediction value.

[0030] Optionally, the method includes:

[0031] Perform second-order differentiation on each of the first near-infrared spectra to obtain the first differential spectrum;

[0032] Obtain the first differential absorbance of the characteristic spectral region in each of the first differential spectra;

[0033] The first calibration model is obtained by performing partial least squares regression analysis on the first differential absorbance of all the blended gasoline samples and the standard octane number of all the blended gasoline samples.

[0034] Optionally, the method includes:

[0035] The second near-infrared spectrum is processed by second-order differentiation to obtain the second differential spectrum;

[0036] Obtain the absorbance of the characteristic spectral region of the second differential spectrum;

[0037] The predicted octane number is determined based on the absorbance in the characteristic spectral region of the second differential spectrum and the first correction model.

[0038] Optionally, the blended gasoline sample includes one or more of reformed gasoline, cracked gasoline, and alkylated gasoline.

[0039] Optionally, the window width for the second-order differential processing is 25.

[0040] Through the above technical solution, this disclosure provides a method for predicting the octane number of gasoline. A first calibration model is established by performing partial least squares multiple regression analysis on the first near-infrared spectrum of a blended gasoline sample and the standard octane number. Then, based on the construction process of the first calibration model, the spectral fitting residual matrix and the octane number fitting residual matrix are selected and used as the input signal and teacher signal of an extreme learning machine, respectively, to obtain a second calibration model. The combined application of the first and second calibration models can comprehensively consider the linear and nonlinear relationships between the first near-infrared spectrum and the octane number of the blended gasoline sample, providing a wider range of octane number prediction capabilities and improving the model's prediction accuracy. This is particularly suitable for predicting the octane number of gasoline with a large octane number variation range (strong nonlinearity). Furthermore, this disclosure selects the absorbance of the characteristic spectral region and the standard octane number for regression analysis during the first model construction process, which can further improve the accuracy of the octane number prediction results.

[0041] Other features and advantages of this disclosure will be described in detail in the following detailed description section. Attached Figure Description

[0042] The accompanying drawings are provided to further illustrate the present disclosure and form part of the specification. They are used together with the following detailed description to explain the present disclosure, but do not constitute a limitation thereof. In the drawings:

[0043] Figure 1 This is a flowchart illustrating a method for predicting the octane number of gasoline according to one embodiment of this disclosure. Detailed Implementation

[0044] The specific embodiments of this disclosure will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit this disclosure.

[0045] This disclosure provides a method for predicting the octane number of gasoline, such as Figure 1 As shown, the process includes the following steps S101-S108:

[0046] S101. Obtain the first near-infrared spectrum of multiple blended gasoline samples and the standard octane number of each blended gasoline sample;

[0047] S102. Obtain the absorbance of the characteristic spectral region in each of the first near-infrared spectra;

[0048] S103. Perform partial least squares multiple regression analysis on the absorbance of the characteristic spectral region of all the blended gasoline samples and the standard octane number of all the blended gasoline samples to obtain the first correction model.

[0049] S104. Based on the first correction model, select the spectral fitting residual matrix and the octane number fitting residual matrix corresponding to the f-th principal factor respectively, use the spectral fitting residual matrix as the input signal of the extreme learning machine, and use the octane number fitting residual matrix as the teacher signal of the extreme learning machine to obtain the second correction model.

[0050] S105. Collect the second near-infrared spectrum of the gasoline sample to be tested; obtain the absorbance of the characteristic spectral region of the second near-infrared spectrum;

[0051] S106. Determine the predicted value of the first octane number based on the absorbance of the characteristic spectral region of the second near-infrared spectrum and the first correction model.

[0052] S107. Select the spectral fitting residual matrix corresponding to the f-th principal factor of the gasoline sample to be tested according to the first correction model, and determine the second octane number prediction value according to the second spectral fitting residual matrix and the second correction model.

[0053] S108. Determine the predicted octane number of the gasoline to be tested based on a linear combination of the first predicted octane number and the second predicted octane number.

[0054] This disclosure provides a method for predicting the octane number of gasoline. A first calibration model is established by performing partial least squares multiple regression analysis on the first near-infrared spectrum of a blended gasoline sample and the standard octane number. Then, based on the construction process of the first calibration model, the spectral fitting residual matrix and the octane number fitting residual matrix are selected and used as the input signal and teacher signal of an extreme learning machine, respectively, to obtain a second calibration model. The combined application of the first and second calibration models comprehensively considers both the linear and nonlinear relationships between the first near-infrared spectrum and the octane number of the blended gasoline sample, providing a wider range of octane number prediction capabilities and improving the model's prediction accuracy. This method is particularly suitable for predicting the octane number of gasoline with a large octane number variation range (strong nonlinearity). Furthermore, this disclosure selects the absorbance of the characteristic spectral region and the standard octane number for regression analysis during the first model construction process, which can further improve the accuracy of the octane number prediction results.

[0055] In this disclosure, partial least squares regression (PLS) analysis can be performed using conventional selection methods and software in the art. Specifically, this disclosure uses an interactive test method to determine the optimal number of principal factors f. best .

[0056] In this disclosure, Extreme Learning Machine (ELM) is a type of machine learning system or method built on Feedforward Neuron Network (FNN).

[0057] This disclosure corrects the nonlinear relationship between the absorbance of a blended gasoline sample in the characteristic spectral region and the standard octane number measured by a standard method using an extreme learning machine (ELM). The spectral fitting residual of partial least squares regression is used as the input signal of the extreme learning machine (ELM), and its octane number fitting residual is used as the teacher signal (i.e., the desired output value, also known as the supervision signal). The extreme learning machine is subjected to supervised learning (also known as supervised learning) to adjust the parameters of the ELM so that the ELM meets the performance requirements of predicting the nonlinear relationship between the near-infrared spectrum and the octane number of gasoline samples.

[0058] Specifically, this disclosure may also include a process for training the ELM to improve its effectiveness. The training set samples are gasoline samples with known standard octane numbers, and the number of training set samples can be set according to actual needs (e.g., target accuracy). In this disclosure, the training set contains 70 samples. The characteristic spectral absorbance of the near-infrared spectra of the training set samples and their corresponding standard octane numbers are used as the input signal and teacher signal, respectively (the same method as in this disclosure for constructing the ELM calibration model). The specific training process follows methods conventionally used in the art. Through the training process, the activation function and the number of neurons in the hidden layers of the ELM model are determined.

[0059] In one specific embodiment, the activation function of the Extreme Learning Machine is a sig function or a sin function; when the activation function is a sig function, the number of hidden layer neurons in the Extreme Learning Machine is 600; when the activation function is a sin function, the number of hidden layer neurons in the Extreme Learning Machine is 800.

[0060] In one embodiment, the characteristic spectral region is 5780-7700 cm⁻¹. -1 In this embodiment, selecting absorbance and standard octane number within an appropriate characteristic spectral range to construct the model can improve the correlation between absorbance and standard octane number and simplify the computational workload in the model construction process.

[0061] In one embodiment, the method further includes:

[0062] The spectral fitting residual matrix Ex is determined according to the following equation (1) or equation (2):

[0063] Ex = Ex(f) - Ex(f) best (1); Where Ex(f) represents the spectral residual matrix when the first correction model takes the f-th principal factor, Ex(f) best () indicates that the first calibration model takes the optimal principal factor f. best The spectral residual matrix at time f ≤ f best ;

[0064] Ex = Ex(f) - Ex(f) best+1～n (2); Ex(f best+1～n () indicates that the first calibration model takes the optimal principal factor f. best The spectral residual matrix of any one of the 1-n principal factors, where n is an integer from 2 to 15.

[0065] In this disclosure, two spectral fitting residual matrices are obtained using equations (1) and (2) above, respectively. Both spectral fitting residual matrices are used to construct a second correction model, resulting in two second correction models. Based on the verification analysis results of the two second correction models, such as dispersion and mean square error, the formula corresponding to the second correction model with better prediction performance is selected.

[0066] In one specific implementation, the number of spectral fitting residual matrices Ex is one or more. Specifically, multiple principal factors f can be selected, resulting in one or more spectral fitting residual matrices Ex. For example, f best When the value is 15, f can be one or more of the values ​​from 1 to 15.

[0067] In one embodiment, the method further includes:

[0068] The octane number fitting residual matrix E is determined according to the following equation (3) or equation (4). y :

[0069] Ey=Ey(f)-Ey(f best (3); Where Ey(f) represents the octane number residual matrix when the first correction model takes the f-th principal factor, Ey(f) best () indicates that the first calibration model takes the optimal principal factor f. best The octane number residual matrix at time f ≤ f best ;

[0070] Ey=Ey(f) (4).

[0071] In one specific embodiment, the number of octane number fitting residual matrices Ey is one or more. In this disclosure, the number of octane number fitting residual matrices Ey corresponds to the number of spectral fitting residual matrices Ex. Each principal factor f yields a corresponding set of spectral fitting residual matrices Ex and octane number fitting residual matrices Ey.

[0072] In this disclosure, the selection methods of equations (3) and (4) are similar to those of equations (1) and (2) mentioned above, and are determined based on the verification analysis results of the second correction model obtained from different equations (3) and (4). The equation corresponding to the second correction model with better prediction effect is selected.

[0073] In a more specific embodiment, this disclosure may select one or more principal factors to obtain one or more sets of corresponding spectral fitting residual matrices E through equation (1) or (2) and (3) or (4) above. x The residual matrix E is fitted with the alkyl number. y , each group of E x and E y The input signal and teacher signal are used respectively to calibrate the Extreme Learning Machine, resulting in a second calibration model.

[0074] In one specific implementation, this disclosure uses an interactive testing method to determine the optimal number of principal factors f. best .

[0075] In one specific implementation, the present disclosure selects the f-th principal factor using the following method:

[0076] The residual matrix E corresponding to the spectrum of each principal factor is fitted. x The residual matrix E is fitted with the alkyl number. y The input signal and teacher signal are used respectively to correct the extreme learning machine, and the second correction model corresponding to each principal factor is obtained.

[0077] The predicted gasoline octane number for each principal factor is obtained based on the first and second correction models, and the mean square error between the predicted gasoline octane number and the actual octane number is calculated. The principal factor with the smallest mean square error or the principal factor with a stable mean square error is selected as f in the above formulas (1)-(4).

[0078] Specifically, the grid search method can be used to select f in equations (1)-(4) above. Specifically, the grid search method is used to list all possible principal factors f, and then each one is verified to select the principal factor f with the smallest mean square error or the principal factor f with the mean square error tending to be stable.

[0079] In this disclosure, appropriate principal factors are obtained by back-calculating the mean square error value as the above equations (1)-(4)f, which can fully compare the spectral fitting residual matrix E of multiple principal factors. x The residual matrix E is fitted with the alkyl number. y The model construction results improve the accuracy of octane number prediction.

[0080] In one implementation,

[0081] The linear combinations include:

[0082] The first octane number prediction value and the second octane number prediction value are linearly summed.

[0083] In one specific embodiment, the octane number y of the sample to be tested... un The prediction process includes:

[0084] (1) Determine the predicted value of the first octane number

[0085] x un The absorbance of the characteristic spectral region of the sample to be tested is input into the established first calibration model to obtain the predicted value of the first octane number, y1.

[0086] (2) Determine the predicted second octane number

[0087] Based on the above equation (1) or (2) determined during the establishment of the second model, the spectral fitting residual matrix corresponding to the f-th principal factor of the sample to be tested is obtained; the spectral fitting residual matrix of the sample to be tested is input into the second correction model to obtain the second octane number prediction value, y2.

[0088] (3) Solve for the final octane number prediction result

[0089] The first and second octane number predictions are linearly summed to obtain the final octane number prediction y. un , that is, y un = y1 + y2.

[0090] In one embodiment, the method further includes: performing multiple predictions based on the second spectral fitting residual matrix and the second correction model, and taking the average of the multiple prediction results as the predicted second octane number. This avoids errors and improves the accuracy of the prediction results.

[0091] In one embodiment, the method further includes:

[0092] Perform second-order differentiation on each of the first near-infrared spectra to obtain the first differential spectrum;

[0093] Obtain the first differential absorbance of the characteristic spectral region in each of the differential spectra;

[0094] The first correction model was obtained by performing regression analysis on the first differential absorbance of all the blended gasoline samples and the standard octane number of all the blended gasoline samples.

[0095] In one embodiment, the method further includes:

[0096] The second near-infrared spectrum is processed by second-order differentiation to obtain the second differential spectrum;

[0097] Obtain the absorbance of the characteristic spectral region of the second differential spectrum;

[0098] The predicted octane number is determined based on the absorbance in the characteristic spectral region of the second differential spectrum and the first correction model.

[0099] This disclosure eliminates baseline shift, drift, and background interference in the spectrum by performing second-order differential processing on the first or second near-infrared spectrum.

[0100] In one specific embodiment, the window width for the second-order differential processing is 25. In this disclosure, the second-order differential processing method can be differentiated using the Savitzky-Golay (SG) method.

[0101] In one embodiment, the blended gasoline sample includes one or more of reformed gasoline, cracked gasoline, and alkylated gasoline.

[0102] In this disclosure, a first model and a second model with relatively accurate prediction results can be constructed using a small number of blended gasoline samples, for example, 70 blended gasoline samples.

[0103] In one specific embodiment, the blended gasoline samples can be of different types, with a minimum of 100 samples, to further improve the accuracy and reliability of the established model.

[0104] In one specific embodiment, the near-infrared spectroscopy detection conditions include: a scanning range of 4000-10000 cm⁻¹. -1 In this disclosure, the resolution can be selected based on the model construction accuracy or instrument accuracy, for example, a resolution of 4cm. -1 8cm -1 wait.

[0105] In this disclosure, the standard octane number is tested using methods conventionally chosen in the art, such as according to the standard method GB / T5487.

[0106] In this disclosure, gasoline samples are divided into a calibration set and a test set. The calibration set is used for model building, while the test set is used to verify the predicted values ​​of the built model against the actual values. This disclosure uses a relatively large number of calibration set samples with good representativeness; that is, the octane ratings of the calibration set samples should cover the octane ratings of all predicted gasoline samples to ensure the established model has good generalization ability. Preferably, the number of calibration set samples should account for 2 / 3 to 3 / 4 of the total sample size, with the remaining 1 / 3 to 1 / 4 used as the test set. Preferably, the distribution patterns of the calibration set and the test set samples are approximately the same.

[0107] The present invention will be further described in detail below with examples, but the present invention is not limited thereto.

[0108] In the following examples and comparative examples, the test conditions for near-infrared spectroscopy testing of the calibration set samples and the test set samples are the same.

[0109] Example 1

[0110] (1) Determining the octane number of crude oil using standard methods

[0111] 459 samples of blended gasoline from a factory were collected, and their research octane number (RON) was determined using the GB / T5487 method (as the standard octane number). 100 representative samples of finished gasoline were selected as blended gasoline samples and formed a calibration set. The remaining samples were used as the test set, with a sample ratio of 7:3 between the calibration set and the test set.

[0112] (2) Construct the first and second correction models

[0113] Near-infrared spectroscopy was performed on all calibration set samples to obtain the first near-infrared spectra of all calibration set samples. The scanning range of the near-infrared spectrometer was 4000-10000 cm⁻¹. -1 The resolution is 8cm. -1 The scanning interval is set to 4cm. -1 .

[0114] The first near-infrared spectrum of each calibration set sample is processed by second-order differentiation with a window width of 25 to obtain the first differential spectrum.

[0115] For each first differential spectrum, select 5780-7700 cm⁻¹ -1 The absorbance in the spectral region will be the 5780-7700 cm⁻¹ of all calibration set samples. -1 The absorbance within the spectral region forms the absorbance matrix X, and the standard octane values ​​of all calibration set samples measured in step (1) form the octane value matrix Y.

[0116] Partial least squares multiple regression analysis (PLS regression analysis) was performed on the absorbance matrix X and the octane number matrix Y to obtain the first calibration model with linear fit.

[0117] In the PLS regression analysis, all principal factors were identified, and the optimal number of principal factors f was determined using the cross-validation method. best The spectral fitting residual matrix and octane number fitting residual matrix corresponding to each principal factor (f) are used as the input signal and teacher signal of the extreme learning machine, respectively, to obtain the second correction model corresponding to each principal factor (f). The spectral fitting residual corresponding to each principal factor (f) is determined according to equation (2): Ex = Ex(f) - Ex(f) best+1～n (2), where f best+1～n Take f best+1 In this embodiment 1, f best If f is the 11th principal factor, then best+1 This is the 12th principal factor number. The octane number fitting residual matrix corresponding to each principal factor (f) is determined according to equation (4): E y =E y (f) (4), that is, directly using the octane residual matrix corresponding to the f-th principal factor.

[0118] The first calibration model and the second calibration model corresponding to each principal factor are validated using test set samples. The specific process is the same as the following step (3). The mean squared error values ​​of the first calibration model and the second calibration model corresponding to each principal factor are obtained, and the principal factor with the smallest mean squared error value is selected as the number of principal factors to construct the second calibration model. This process uses a global search algorithm.

[0119] In this embodiment, the number of principal factors selected is 7.

[0120] The spectral fitting residual matrix (E) is calculated according to the above formula (2). X =E X(7) -E X(12) According to equation (4), the octane number fitting residual matrix (E) is obtained. Y =E Y(7) Then, the spectral fitting residual matrix and the octane number fitting residual matrix are used as the input signal and teacher signal of the Extreme Learning Machine, respectively. The ELM algorithm is used to perform nonlinear correction on the spectral fitting residual matrix and the octane number fitting residual matrix to obtain the second correction model of nonlinear fitting. The ELM algorithm used in this embodiment is a pre-trained ELM algorithm, and the parameters of the ELM algorithm are listed in Table 1 below.

[0121] (3) The prediction results of the first and second correction models are verified using test set samples (the first and second correction models are collectively referred to as "PLS-ELM").

[0122] For each sample in the test set: near-infrared spectroscopy was performed to obtain the second near-infrared spectrum. After second-order differential processing with a window width of 25, the second differential spectrum was obtained. The wavenumber range of 5780-7700 cm⁻¹ was selected from the second differential spectrum. -1 The absorbance in the spectral region is taken as the absorbance to be measured; the absorbance to be measured is input into the first calibration model to obtain the predicted value of the first octane number; and the spectral fitting residual matrix corresponding to the 7th principal factor of the test set sample is selected according to the first calibration model, and the second spectral fitting residual matrix is ​​calculated using the following formula (2): Ex um =Ex um (7)-Ex um (f 12 (2) Input the second spectrum fitting residual matrix into the second correction model to obtain the second octane number prediction value (repeated prediction 100 times, and the average value of 100 times is taken as the second octane number prediction value); linearly add the first octane number prediction value and the obtained second octane number prediction value to obtain the gasoline octane number prediction value of each test set sample.

[0123] The relevant statistical parameters used to establish the model are shown in Table 1. The comparison results of the predicted octane numbers of the test set samples with the standard octane numbers determined by the GB / T5487 method are shown in Table 2 (taking 30 data points as an example).

[0124] The formulas for calculating the relevant statistical parameters are as follows: (5) and (6):

[0125]

[0126] in, Let y be the predicted value of the i-th test set sample. i (i = 1, 2, ..., n) represents the true value (standard octane number) of the i-th sample, where n is the number of samples in the test set. The calculation results are shown in Table 1.

[0127] Table 1

[0128]

[0129]

[0130] Table 2

[0131]

[0132] Where, deviation = predicted value - actual value.

[0133] Example 2

[0134] The first and second calibration models were established and verified according to the method in Example 1. The difference was that, in establishing the second calibration model, the octane number fitting residual was modified to the octane number residual at the f-th principal factor minus the optimal principal factor f. best The residual matrix obtained from the octane number residual at that time, i.e., E Y =E Y(7) -E Y(11) The relevant statistical parameters of the calibration set and the test set are shown in Table 3, and the comparison results of the predicted values ​​of the test set samples and the values ​​measured by the GB / T5487 method are shown in Table 4. The ELM algorithm used in this embodiment is a pre-trained ELM algorithm, and the parameters of the ELM algorithm are listed in Table 3 below.

[0135] Table 3

[0136]

[0137] Table 4

[0138]

[0139]

[0140] Comparative Example 1

[0141] The difference from Example 1 is that only the first step of partial least squares regression analysis (PLS) is performed to construct the first calibration model, without performing the step of constructing the second calibration model, and the optimal principal factor f is not performed. best The model was used for prediction under the given conditions. The relevant statistical parameters for the calibration set and the test set are shown in Table 5, and the comparison results between the predicted values ​​of the test set samples and the values ​​measured by the GB / T5487 method are shown in Table 6.

[0142] Table 5

[0143]

[0144] Table 6

[0145]

[0146] Comparative Example 2

[0147] The difference from Example 1 is that the Extreme Learning Machine (ELM) algorithm is used alone to build the calibration model. The ELM algorithm used in this example is a pre-trained ELM algorithm, and its parameters are listed in Table 7 below. The model is then used for prediction under optimal parameter conditions. The relevant statistical parameters of the calibration set and the test set are shown in Table 7, and the comparison results of the predicted values ​​of the test set samples and the values ​​measured by the GB / T5487 method are shown in Table 8.

[0148] Table 7

[0149]

[0150]

[0151] Table 8

[0152]

[0153] Based on the relevant data from Examples 1-2 and Comparative Examples 1-2 above, it can be seen that the mean square error of the gasoline octane number prediction model obtained using the method provided in this disclosure is smaller. This indicates that compared with the standalone linear partial least squares gasoline octane number prediction method (Comparative Example 1) and the standalone nonlinear limit learning machine gasoline octane number prediction method (Comparative Example 2), the method provided in this disclosure has better prediction accuracy for gasoline octane number. Furthermore, the R-squared value of the gasoline octane number prediction model obtained in this disclosure is significantly higher than that of the standard method. 2 The larger absorbance indicates that the absorbance of the model constructed in this application is more interpretable for octane numbers and is more practical.

[0154] The preferred embodiments of this disclosure have been described in detail above with reference to the accompanying drawings. However, this disclosure is not limited to the specific details of the above embodiments. Within the scope of the technical concept of this disclosure, various simple modifications can be made to the technical solutions of this disclosure, and these simple modifications all fall within the protection scope of this disclosure.

[0155] It should also be noted that the various specific technical features described in the above embodiments can be combined in any suitable manner without contradiction. To avoid unnecessary repetition, this disclosure will not describe the various possible combinations separately.

[0156] Furthermore, various different embodiments of this disclosure can be combined in any way, as long as they do not violate the spirit of this disclosure, they should also be regarded as the content disclosed in this disclosure.

Claims

1. A method for predicting the octane number of gasoline, characterized in that, Includes the following steps: Obtain the first near-infrared spectra of multiple blended gasoline samples and the standard octane number of each of the blended gasoline samples; Obtain the absorbance of each characteristic spectral region in the first near-infrared spectrum; Partial least squares multiple regression analysis was performed on the absorbance of the characteristic spectral region of all the blended gasoline samples and the standard octane number of all the blended gasoline samples to obtain the first correction model; Based on the first calibration model, the spectral fitting residual matrix and the octane number fitting residual matrix corresponding to the f-th principal factor are selected respectively. The spectral fitting residual matrix is ​​used as the input signal of the extreme learning machine, and the octane number fitting residual matrix is ​​used as the teacher signal of the extreme learning machine to obtain the second calibration model. Collect the second near-infrared spectrum of the gasoline sample to be tested; Obtain the absorbance of the characteristic spectral region of the second near-infrared spectrum; The predicted value of the first octane number is determined based on the absorbance of the characteristic spectral region of the second near-infrared spectrum and the first correction model. The spectral fitting residual matrix corresponding to the f-th principal factor of the gasoline sample to be tested is selected according to the first correction model, and the second octane number prediction value is determined according to the spectral fitting residual matrix and the second correction model. The predicted octane number of the gasoline to be tested is determined based on a linear combination of the first predicted octane number and the second predicted octane number. The method also includes: The spectral fitting residual matrix is ​​determined according to the following equation (1) or equation (2). Ex : Ex = Ex ( f )- Ex ( f best (1); where Ex ( f () represents the spectral residual matrix when the first correction model takes the f-th principal factor. Ex ( f best () indicates that the first calibration model takes the optimal principal factors. f best The spectral residual matrix at time, where f ≤ f best ; Ex = Ex ( f )- Ex ( f best+1~n (2); Ex ( f best+1~n () indicates that the first calibration model takes the optimal principal factors. f best The spectral residual matrix of any one of the 1 to n principal factors, where n is an integer from 2 to 15; The octane number fitting residual matrix is ​​determined according to the following equation (3) or equation (4). Ey : Ey = Ey ( f )- Ey ( f best (3); among which Ey ( f () represents the octane number residual matrix when the first calibration model takes the f-th principal factor. Ey ( f best () indicates that the first calibration model takes the optimal principal factors. f best The octane number residual matrix at time, where f ≤ f best ; Ey = Ey ( f ) （4）。 2. The method according to claim 1, characterized in that, The characteristic spectral region is 5780-7700 cm⁻¹. -1 .

3. The method according to claim 1, characterized in that, The spectral fitting residual matrix Ex The number of elements can be one or more.

4. The method according to claim 1, characterized in that, The octane number fitting residual matrix Ey The number of elements can be one or more.

5. The method according to claim 1, characterized in that, The linear combinations include: The first octane number prediction value and the second octane number prediction value are linearly summed.

6. The method according to claim 1, characterized in that, The method also includes: Based on the spectral fitting residual matrix and the second correction model, multiple predictions are made, and the average of the multiple prediction results is taken as the second octane number prediction value.

7. The method according to claim 1, characterized in that, The method includes: Perform second-order differentiation on each of the first near-infrared spectra to obtain the first differential spectrum; Obtain the first differential absorbance of the characteristic spectral region in each of the first differential spectra; Partial least squares regression analysis was performed on the first differential absorbance of all the blended gasoline samples and the standard octane number of all the blended gasoline samples to obtain the first correction model.

8. The method according to claim 7, characterized in that, The method includes: The second near-infrared spectrum is processed by second-order differentiation to obtain the second differential spectrum; Obtain the absorbance of the characteristic spectral region of the second differential spectrum; The predicted octane number is determined based on the absorbance in the characteristic spectral region of the second differential spectrum and the first correction model.

9. The method according to claim 1, characterized in that, The blended gasoline sample includes one or more of reformed gasoline, cracked gasoline, and alkylated gasoline.

10. The method according to claim 7 or 8, characterized in that, The window width for the second-order differential processing is 25.

Images