Time series prediction method for mycelium concentration, ethanol concentration and glucose concentration in fuel ethanol fermentation process

Abstract

The present application is a method for predicting the time series of mycelium concentration, ethanol concentration and glucose concentration in the fermentation process of fuel ethanol. The method constructs a fermentation end time prediction model based on a polynomial regression algorithm, and establishes a mathematical relationship between the fermentation end time and the initial conditions. Through the scaling of time scale, the time series of different batch data is transformed to achieve a unified standard end time. The method establishes a mycelium concentration prediction model based on the XGBoost algorithm. The obtained mycelium concentration and growth rate are added as supplementary mechanism knowledge to the prediction model of ethanol and glucose concentration. The method establishes a dynamic parameter model of artificial neural network, which can provide dynamic parameters according to time variables and environmental variables. By integrating it into the previous mycelium concentration, ethanol concentration and glucose concentration prediction model, the time series prediction of mycelium concentration, ethanol concentration and glucose concentration under different initial conditions can be realized.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of biological fermentation engineering, and relates to a dynamic modeling method for predicting mycelium concentration, ethanol concentration and glucose concentration in different fermentation batches of fuel ethanol based on mechanism and data driving. BACKGROUND

[0002] In recent years, developing new fuels that are pollution-free and low-cost to replace traditional fuels that are seriously polluting has become an important issue. Not only that, but as the world's reserves of fossil fuels such as crude oil and coal are gradually decreasing, it is imperative to find new fuels as alternatives. Fuel ethanol, as a gasoline additive and a new fuel, has attracted much attention in this field. Fuel ethanol refers to anhydrous ethanol with a volume fraction of 99.5% or more without the addition of denaturants, and is a clean high-octane fuel. Its renewability, oxygenation, universality of raw materials and green environmental protection closely match the purpose of "carbon neutralization and carbon peak". The industrial production of the ethanol fermentation process has undergone a long history of development, and its rich production experience has laid a solid foundation for the development of ethanol as a new fuel, and has great potential to become a major fuel energy in the future. At present, the main production route of fuel ethanol is the biomass fermentation method, and its process is mostly time-varying, highly complex and uncertain. Therefore, it is of great significance to establish an accurate model to describe the fermentation process. Not only can it help researchers more accurately study the ethanol fermentation process, but also can guide the transformation and optimization of process and operating conditions.

[0003] The success of hybrid modeling is due to the complementarity between different types of models. In traditional industrial modeling methods, the mechanism model established usually needs to study the theoretical knowledge of the industrial process or phenomenon in depth. Then try to abstract this process or phenomenon into mathematical form, and further implement it in the computer. When the mechanism model can accurately describe the mechanism, its accuracy, stability and interpretability are unparalleled. But this modeling method also has some disadvantages. The mechanism model established is usually single, and the mathematical process or mechanism it contains is limited. The mechanism of some processes is also difficult to describe, for example: the model has high dimensionality of variables, time-varying dynamics parameters, multiple collinearity between variables, and complex nonlinearity. Therefore, the advantages of mechanism model are good accuracy and good interpretability, but the modeling difficulty is high and the model does not have flexibility. In the modeling of machine learning, usually a fixed algorithm is used to mine the information in the data and does not need the support of mechanism knowledge. The algorithm will train the existing data, and adjust the parameters in the model according to the error between the prediction result and the true value. But the data-driven model also has some problems. When the collected data has the problem of sample imbalance, the model is easy to overfit to the majority of samples. Some models face the challenge of very small amount of data, and the algorithm is difficult to mine the complete model from the simple data. If the data has outliers or noise, the model may learn the wrong knowledge, which will reduce the prediction accuracy. Industrial production has high requirements for model stability and robustness, and pure data-driven model sometimes cannot meet the requirements. The advantage of data-driven model is simple modeling, but the requirement for collected data is high. In the hybrid model established in the invention, the data-driven model converts the problem of finding a suitable mechanism in the mechanism model into the problem of obtaining high-quality data and finding model parameters. The data-driven model extracts the nonlinear complex relationship that the mechanism model cannot reflect. And the mechanism model is used to reduce the influence of data noise, distribution and other factors on the data-driven model. The mechanism model as the basis of the hybrid model reduces the influence of noise and redundant information in the sample on the data-driven model, so as to avoid the hybrid model established to violate the mechanism and common sense. SUMMARY

[0004] The purpose of the present application is to provide a fuel ethanol fermentation process mycelium concentration, ethanol concentration and glucose concentration time series prediction method based on mechanism and data driving.

[0005] The specific technical scheme of the present application is: a fuel ethanol fermentation process mycelium concentration, ethanol concentration and glucose concentration time series prediction method, the specific steps comprising:

[0006] First, based on the initial conditions of different batches, a fermentation end time prediction model based on polynomial regression algorithm is established, and the fermentation data is transformed into a unified time scale.

[0007] Second, in order to more accurately predict the trend of mycelium concentration, a mycelium concentration prediction model based on XGBoost algorithm is established, and the predicted value of the mycelium growth rate in this period is calculated.

[0008] Third, a dynamic parameter model of artificial neural network is established, which can provide dynamic parameters according to time variables and environmental variables, and integrate them into the previous mycelium concentration, ethanol concentration and glucose concentration prediction model, so as to realize the time series prediction of mycelium concentration, ethanol concentration and glucose concentration under different initial conditions. The structure of the dynamic hybrid model is as shown in Figure 1

[0009] The advantages of the present application are: (1) the mathematical relationship between the initial conditions of fermentation and the end time of fermentation is established by using the polynomial regression algorithm, the model structure is simple, and the interpretability is strong. The model can predict the end time of fermentation with high accuracy, which has important guiding significance for ethanol production. (2) Based on the fermentation batch data of different initial glucose concentrations, the complementary characteristics of mechanism modeling and data-driven modeling are combined to establish a dynamic model considering the influence of fermentation time factor. The model helps to more truly simulate and predict the dynamic behavior of the fermentation process.

[0010] 1. Fermentation end time prediction model based on polynomial regression algorithm

[0011] In the production of fuel ethanol, the fermentation of ethanol is a complex process. After abstracting the process into a mathematical model, a complex multivariate system will be formed. For example, the end time of fermentation of different batches is different. This non-uniformity in time brings certain difficulties to the training of the fermentation process prediction model. Therefore, a fermentation end time prediction model based on polynomial regression algorithm is established to describe the mathematical relationship between initial glucose concentration, initial mycelium concentration and fermentation end time. Because the initial glucose concentration and the initial mycelium concentration have large order of magnitude difference, the two need to be standardized:

[0012]

[0013] wherein, represents the standardized initial glucose concentration; represents the initial glucose concentration; represents the mean of the initial glucose concentration; represents the standard deviation of the initial glucose concentration; represents the standardized initial mycelium concentration; ​Represents the initial mycelial concentration; The mean value representing the initial mycelial concentration; The standard deviation represents the initial mycelial concentration.

[0014] Choosing a quadratic polynomial regression model, discarding the squared terms and retaining only the interaction terms to increase the dimensionality of the features, the features of the fermentation end time prediction model are: the standardized initial glucose concentration. Standardized initial mycelial concentration and the product between the two, c s 0* ×c x 0* The expression for a multinomial regression model (with only interaction terms) is:

[0015] t pre =αc s 0* +βc x 0* +γc s 0* c x 0* +ε (2)

[0016] Among them, t pre The predicted fermentation end time is represented by α, β, and γ; α, β, and γ represent regression coefficients; and ε represents the intercept. The parameters of the model are determined by fitting experimental data.

[0017] 2. Data transformation with a unified time scale

[0018] Different batches have different time scales, resulting in varying sampling intervals and fermentation completion times. This inconsistency in the timeframes of the sample data poses challenges to model building during fermentation. Therefore, it is necessary to convert the time series of the training data into a uniform fermentation completion time and time interval through time scaling. This temporal uniformity is beneficial for model training.

[0019] Different fermentation processes are mapped to the same standardized time according to equation (3), i.e., the total fermentation time is 34 hours. Based on the converted standardized time, the time series of mycelial concentration, ethanol concentration and glucose concentration with 2-hour intervals are used for interpolation training.

[0020]

[0021] Where t* represents the standard time of the current actual sampling time after unifying the time scale; t represents the current actual sampling time point; t pre This represents the fermentation end time predicted using a multinomial regression model.

[0022] 3. Mycelial Concentration Prediction Model Based on XGBoost Algorithm

[0023] After pure mechanistic modeling, it was found that the Logistic equation could neither adequately fit the trend of mycelial concentration changes nor reflect the negative feedback effect between the environment and yeast in the mid-to-late stages of fermentation. Therefore, in the hybrid modeling process, the mechanistic model of mycelial growth was discarded, and a data-driven model was built specifically for it. A mycelial concentration prediction model based on the XGBoost algorithm was constructed, with t as the input variable. * Mycelial concentration c at time x t* ethanol concentration c P t* glucose concentration c s t* and fermentation standard time t * Its output is t * Mycelial concentration c at time +1 x t*+1 Choose the coefficient of determination R. 2 As an evaluation metric for the XGBoost model:

[0024]

[0025] In the formula, c x Experimental values ​​representing mycelial concentration; The predicted mycelial concentration represents the XGBoost model. represents the average value of the experimental data on mycelial concentration; m represents the sample size of the training data. Based on the experimental data, the parameters are determined with the goal of optimizing the evaluation index. After determining the optimal combination of hyperparameters, the predicted value of mycelial growth rate within this time period is calculated using the difference formula:

[0026]

[0027] This model can obtain relatively accurate predictions of mycelial concentration and growth rate, providing supplementary mechanistic knowledge for ethanol and glucose prediction models.

[0028] 4. Artificial Neural Network Model for Ethanol Kinetic Parameters

[0029] When the initial glucose concentrations differ significantly, pure mechanistic modeling suffers from mismatch. Applying the trained mechanistic model to other fermentation batches yields unsatisfactory results. To address the mismatch issue caused by pure mechanistic models, this invention establishes an artificial neural network-based ethanol kinetic parameter model. The goal of this model is to find a fixed set of weights and thresholds to derive the parameters that change with environmental variables (c...). x t* c Pt* and c s t* The dynamic parameters that change with time. The input variable of the neural network model is t. * Mycelial concentration c at time x t* ethanol concentration c P t* glucose concentration c s t* and fermentation standard time t * Its output is t * The three kinetic parameters k1 used in the ethanol mechanism model at time +1 t*+1 k2 t*+1 and K3 t*+1 In the neural network model, the activation function for the hidden layers is the Sigmoid function, and the activation function for the output layer is the Purelin linear transfer function. The improved ethanol generation mechanism model is as follows:

[0030]

[0031] By adding the mycelial concentration and growth rate predicted by the XGBoost model as supplementary mechanistic knowledge to the differential equation (6), t can be calculated. * The ethanol concentration c at time +1 P t*+1 The problem is now transformed into an optimization problem, with the objective function being to minimize the root mean square error between the predicted and actual ethanol concentrations. The variables to be optimized are the weights and thresholds of the artificial neural network model. The simplex method is used to optimize this problem, and the expression for its objective function is as follows:

[0032]

[0033] In the formula, c P Experimental values ​​representing ethanol concentration; The predicted value represents the ethanol concentration.

[0034] 5. Artificial Neural Network Model for Glucose Kinetics

[0035] The input variable of this neural network model is t. * Mycelial concentration c at time x t* ethanol concentration c P t* glucose concentration c s t* and fermentation standard time t * Its output is t * The three kinetic parameters k4 used in the glucose mechanism model at time +1t *+1 k5 t*+1 and k6 t*+1 The improved glucose consumption mechanism model is as follows:

[0036]

[0037] The c x , and data obtained by using the above model are added to the differential equation (8), and then the glucose concentration c * at t s +1 can be calculated. t*+1 The expression of the objective function of the optimization problem is as follows:

[0038]

[0039] In the formula, c s represents the experimental value of the glucose concentration; and c represents the predicted value of the glucose concentration.

[0040] Finally, the time series of the mycelium concentration, the ethanol concentration and the glucose concentration predicted by the dynamic hybrid model of the fuel ethanol fermentation process at the standard time are scaled and transformed back to the real fermentation time, and the specific formula is as follows:

[0041]

[0042] In the formula, t* represents the standard time of the current real sampling time in the unified time scale; t represents the current real sampling time point; t pre represents the end time of the fermentation predicted by using the polynomial regression model. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 is a structure diagram of a method for predicting the time series of the mycelium concentration, the ethanol concentration and the glucose concentration of the fuel ethanol fermentation process based on mechanism and data driving. DETAILED DESCRIPTION

[0044] The application is further described by the following examples, but the scope of the application is not limited by the examples:

[0045] 1. Fermentation data acquisition and pretreatment

[0046] The fermentation data used in the present application can be divided into three categories according to the initial glucose concentration: about 220 g / L, about 250 g / L, and above 285 g / L. Six batches of fermentation data were randomly selected as the training set, and two batches of fermentation data were selected as the test data. The sampling time interval and the fermentation end time of each batch of fermentation data are different, including the corresponding mycelium concentration, ethanol concentration and glucose concentration at different sampling times. The expression for predicting the fermentation end time by polynomial regression algorithm is:

[0047] t pre = 4.2424c s 0* - 3.0377c x 0* + 2.5902c s 0* c x 0* + 32.2870 (11)

[0048] According to formula (3), the fermentation end time is standardized, and the mycelium concentration, ethanol concentration and glucose concentration in the training set are interpolated to a uniform time scale time series according to the conversion time. The piecewise cubic Hermite interpolation "pchip" in the interpolation function "interp1" on the MATLAB platform is used. That is, each batch of data is 18 groups of samples with a fermentation time of 0-34 hours and a time interval of 2 hours. The training set contains 108 groups of data, corresponding to the time series of mycelium concentration, ethanol concentration and glucose concentration of 6 batches. The test set contains data of two batches of fuel ethanol fermentation process.

[0049] 2. Mycelium concentration prediction model based on XGBoost algorithm

[0050] The present application selects the hyperparameters in the XGBoost framework through 5-fold cross-validation. Some parameters are selected as follows: objective ='reg: linear', max_depth = 6, eta = 0.3, n_estimators = 18, subsample = 1, gamma = 0, lambda = 1, alpha = 0, colsample_bytree = 1, colsample_bylevel = 1. After training the model, the predicted value of the mycelium concentration at time t * + 1 can be obtained. The predicted value of the growth rate in this time period is calculated by difference formula (5).

[0051] 3. Artificial neural network ethanol kinetic parameter model

[0052] Based on the performance of the ethanol prediction model in the training set, the number of hidden layer neurons in the neural network is selected to be 4. The connection mode is chosen to be fully connected, resulting in a total of 35 weights and thresholds required. After initializing the weights and thresholds of the neural network model, the three kinetic parameters k1 used by the ethanol mechanism model at the next time step can be calculated. t*+1 k2 t*+1 and K3 t*+1 By adding the mycelial concentration and growth rate predicted by the XGBoost model as supplementary mechanistic knowledge to the improved differential equation (6), t can be calculated. * The ethanol concentration c at time +1 P t*+1 The optimal weights and thresholds for the ethanol kinetic parameter model obtained using the simplex method are as follows:

[0053]

[0054] b = [b1 b2 b3 b4] T =[0.6752 0.2489 0.3174 0.2685] T

[0055]

[0056] b' = [b1' b2' b3'] T =[0.2204 0.2504 0.5253] T

[0057] In the formula, V represents the weight matrix from the input layer to the hidden layer in the neural network; v 21 b represents the weights between the second neuron in the input layer and the first neuron in the hidden layer, and so on; b represents the threshold vector of the hidden layer neurons; b1 represents the threshold of the first neuron in the hidden layer, and so on; W represents the weight matrix from the hidden layer to the output layer; w 21 b' represents the weight between the second neuron in the hidden layer and the first neuron in the output layer, and so on; b' represents the threshold vector of the neuron in the output layer; b1' represents the threshold of the first neuron in the output layer, and so on.

[0058] 4. Artificial Neural Network Model for Glucose Kinetics

[0059] Based on the fitting performance of the glucose prediction model training set, the number of hidden layer neurons in the neural network was selected to be 3. The c obtained using the above model will be... x , and By adding the data to the improved differential equation (8), t can be calculated. *Glucose concentration c at time t s t*+1 The optimal weight values and threshold values of the glucose kinetic parameter model found by using the simplex method are:

[0060]

[0061] b = [b1 b2 b3] T = [4.8630 0.5804 2.4842] T

[0062]

[0063] b' = [b1' b2' b3'] T = [0.1367 -0.3971 0.2127] T .

Claims

1. A method for time series prediction of mycelium concentration, ethanol concentration and glucose concentration in a fuel ethanol fermentation process, characterized by the steps of Comprise: First, the fermentation end time prediction model based on polynomial regression algorithm, the mathematical relationship between the initial conditions and the fermentation end time is established, which provides the prediction value of the fermentation end time for the subsequent time scale unification; through the scaling of time scale, the current real sampling time point of different batches is transformed to reach the unified fermentation end time; Second, the mycelium concentration prediction model based on XGBoost algorithm adds the obtained prediction value of mycelium concentration and growth rate as supplementary mechanism knowledge to the prediction model of ethanol concentration and glucose concentration; Third, the dynamic parameter model of artificial neural network is established, which provides dynamic parameters according to time variables and environmental variables; it is fused into the previous mycelium concentration, ethanol concentration and glucose concentration prediction model to realize the time series prediction of mycelium concentration, ethanol concentration and glucose concentration under different initial conditions; the steps are as follows: (1) find the mechanism model of fuel ethanol fermentation process as the basis of the hybrid model; (2) construct the fermentation end time prediction model; (3) data transformation of unified time scale; (4) construct the prediction model of mycelium concentration and growth rate to realize the fitting of Logistic equation in the mechanism model to mycelium concentration; (5) according to the mechanism knowledge of mycelium concentration and growth rate provided by the model, combine the dynamic parameter model to construct the prediction model of ethanol and glucose concentration; (6) scale the mycelium concentration, ethanol concentration and glucose concentration predicted by the dynamic hybrid model under the standard time to the real fermentation time to realize the time series prediction.

2. The fuel ethanol fermentation process mycelium concentration, ethanol concentration, and glucose concentration time series prediction method according to claim 1, characterized by, The fermentation end time prediction model establishes the mathematical relationship between the initial glucose concentration, initial mycelium concentration and fermentation end time: due to the large difference in the order of magnitude of the initial glucose concentration and the initial mycelium concentration, the two are standardized: ， ， wherein, represents the standardized initial glucose concentration; represents the initial glucose concentration; represents the mean of the initial glucose concentration; represents the standard deviation of the initial glucose concentration; represents the standardized initial mycelium concentration; represents the initial mycelium concentration; represents the mean of the initial mycelium concentration; represents the standard deviation of the initial mycelium concentration; a quadratic polynomial regression model is selected, the square term is discarded, and only the interaction term is retained to upgrade the features; the features of the fermentation end time prediction model are: the standardized initial glucose concentration , the standardized initial mycelium concentration , and the product therebetween; the expression of the polynomial regression model only has the interaction term is: ， where t pre represents the predicted value of the end time of fermentation; a, β, γ represent the regression coefficients, and ε represents the intercept. The parameters of the model are determined by fitting the experimental data.

3. The fuel ethanol fermentation process mycelial concentration, ethanol concentration, and glucose concentration time series prediction method according to claim 1, characterized by, The data transformation of unified time scale is to map the fermentation processes of different time lengths to the same standardized time according to the different time lengths of each fermentation process, and interpolate the mycelium concentration, ethanol concentration and glucose concentration in the training set to time series of several time intervals, and the standardized fermentation end time conversion formula is as follows: ， Wherein, t * Standard time representing current real adoption time after unified time scale; t represents current real sampling time point; t pre Represent the predicted end time of fermentation by using a polynomial regression model.

4. The fuel ethanol fermentation process mycelium concentration, ethanol concentration, and glucose concentration time series prediction method according to claim 3, characterized by, The total fermentation time is 34 hours; the time interval is 2 hours.

5. The fuel ethanol fermentation process mycelium concentration, ethanol concentration, and glucose concentration time series prediction method according to claim 1, characterized by, Construct the mycelial concentration prediction model based on the XGBoost algorithm: the input variable is t. * mycelial concentration at time ethanol concentration glucose concentration and fermentation standard time t * Its output is t * Mycelial concentration at time +1 ; Determination coefficient R 2 As an evaluation metric for the XGBoost model: ， wherein c x representing the experimental values of mycelium concentration; representing the predicted results of mycelium concentration by XGBoost model; representing the average values of mycelium concentration experimental data; m represents the sample size of the training data; Based on experimental data, the parameters are determined with the optimal evaluation index as the target; after determining the hyperparameter combination, the prediction value of the mycelium growth rate in this time period is calculated according to the difference formula: 。 6. The fuel ethanol fermentation process mycelium concentration, ethanol concentration, and glucose concentration time series prediction method according to claim 1, characterized by, The dynamic parameter model of the artificial neural network is established: the input variable of the neural network model is the mycelium concentration at time t * The mycelium concentration at time t The ethanol concentration at time t The glucose concentration at time t And the fermentation standard time t * The output is the three dynamic parameters used by the ethanol mechanism model at time t * +1 The activation function of the hidden layer in the neural network model is selected as the Sigmoid function, and the activation function of the output layer is selected as the linear transmission function Purelin; the objective of this model is to obtain a set of fixed weight values and threshold values, and to obtain the dynamic parameters changing with the environmental variables And The time variable; the improved ethanol generation mechanism model is as follows: ， where c p is the experimental value of the ethanol concentration, c x is the mycelium concentration; The hyphal concentration and growth rate predicted by the XGBoost model were added as supplemental mechanistic knowledge to the improved ethanol differential equations to calculate the ethanol concentration at time t * +1 .

7. The fuel ethanol fermentation process mycelium concentration, ethanol concentration, and glucose concentration time series prediction method according to claim 6, characterized by, The dynamic parameter model of artificial neural network is optimized, the objective function is to minimize the root mean square error between the predicted value and the experimental value of ethanol concentration, and the optimized variable is the weight and threshold value of the neural network model; simplex method is used to optimize the problem, and the objective function expression is as follows: ， where c P represents the experimental value of ethanol concentration; represents the predicted value of ethanol concentration.

8. The fuel ethanol fermentation process mycelium concentration, ethanol concentration, and glucose concentration time series prediction method according to claim 1, characterized by, The time series of mycelium concentration, ethanol concentration and glucose concentration predicted by the dynamic hybrid model of fuel ethanol fermentation process under the standard time is scaled and transformed back to the real fermentation time, and the specific formula is as follows: ， Wherein, t * Standard time representing current real adoption time after unified time scale; t represents current real sampling time point; t pre Represent the predicted end time of fermentation by using a polynomial regression model.

Images