Lignocellulose biomass glucose production prediction method based on machine learning

Abstract

The invention relates to a lignocellulose biomass glucose production prediction method based on machine learning, and aims to improve the yield of glucose and optimize a biomass conversion process. According to the method, machine learning models such as a multi-layer perceptron, support vector regression, a random forest, a gradient boosting decision tree and a convolutional neural network are adopted to predict the LCB glucose yield processed by different preprocessing modes. Experimental results show that the convolutional neural network model is optimal in performance, the decision coefficient Rreaches up to 0.95, and the method has excellent generalization ability and prediction precision. The prediction precision is remarkably improved by feature extraction and principal component analysis in the model establishment process, and the Rof the convolutional neural network model is improved from 0.76 to 0.95. The efficient and accurate prediction method is provided for optimization of the biomass saccharification process, the method can be widely applied to the field of biological energy, and technical support is provided for improving the resource utilization efficiency of lignocellulose biomass.

Description

Technical Field

[0001] This invention belongs to the field of biomass energy, specifically to a method for predicting glucose production from lignocellulosic biomass based on machine learning. This technology can be widely applied in biomass energy conversion, environmental protection, and related industrial production fields. Background Art

[0002] Lignocellulosic biomass is primarily composed of cellulose, hemicellulose, and lignin. Cellulose and hemicellulose together form the primary components of plant cell walls, while lignin imparts strength and stability to these walls. The presence of these components gives lignocellulosic biomass a high degree of structural stability and resistance to degradation, posing challenges to its conversion.

[0003] The bioconversion and utilization of lignocellulosic biomass typically involves two steps: pretreatment and enzymatic hydrolysis. Pretreatment aims to disrupt the biomass structure and increase the accessibility of cellulose and hemicellulose to improve subsequent enzymatic hydrolysis. However, the presence of lignin makes the pretreatment process complex and costly. Existing pretreatment methods include physical, chemical, and biological methods. The effectiveness of each method is affected by multiple factors such as temperature, time, and pH, making the optimization process very complex.

[0004] Traditional optimization methods rely heavily on laboratory testing and experience, typically determining optimal conditions through single-variable experiments or traditional experimental designs (such as response surface methodology). While these methods can identify certain optimal conditions, they often struggle to comprehensively consider the complex interactions of multiple variables. The resulting predictions and selections may not be optimal, and they require extensive experimental work, which is time-consuming and laborious.

[0005] In recent years, machine learning technology has shown great potential in optimization problems. By analyzing a large amount of experimental data, machine learning can automatically identify the complex relationships between variables, thereby optimizing multivariate conditions and improving conversion efficiency. In the conversion process of lignocellulosic biomass, machine learning is expected to help predict and optimize pretreatment and enzymatic hydrolysis conditions, thereby improving sugar yield and conversion efficiency. However, although studies have attempted to apply machine learning to the optimization of biomass conversion, practical applications still face challenges such as low prediction accuracy and difficulty in model selection. Therefore, the present invention aims to select the most suitable model for enzymatic hydrolysis of lignocellulosic materials after pretreatment of glucose by comparing a variety of advanced machine learning techniques; and to improve the accuracy of glucose yield prediction through feature extraction and PCA dimensionality reduction. Summary of the Invention

[0006] The present invention provides a machine learning-based method for predicting glucose production from lignocellulosic biomass, aiming to address the low prediction accuracy and model selection difficulties inherent in existing technologies. This method addresses the difficulties in selecting LCB pretreatment methods and the inaccurate saccharification yield predictions in existing technologies by proposing a machine learning-based method for optimizing pretreatment and predicting saccharification yield.

[0007] An embodiment of the present invention provides a method for predicting glucose production from lignocellulosic biomass based on machine learning, the method comprising the following steps: S1, collect data on different biomass raw materials, various pretreatment methods and enzymatic hydrolysis parameters, build a multi-factor database, and form a set; S2. Preprocess the collected data, including data cleaning and normalization. Data cleaning is achieved by removing outliers and noise data. Statistical analysis methods are used to calculate the mean and standard deviation of the data. Data that deviates from the mean by more than a set multiple of the standard deviation is considered an outlier and is removed. When normalizing the data, the minimum-maximum normalization method is used to normalize the data to the range of [0, 1]. S3. Performing feature engineering on the pre-processed data, wherein the feature engineering includes feature extraction and PCA dimensionality reduction; the core of feature extraction is to establish a relationship between raw materials and pre-treatment methods and glucose yield; S4. Divide the dataset and determine hyperparameters; divide it into training and test sets in proportion to ensure a balanced distribution of different raw materials and preprocessing methods, and improve the generalization ability of the model; S5. Model training and comparative analysis to explore the complex effects of pretreatment and enzymatic hydrolysis parameters on glucose yield; S6. Model evaluation: Use the final application model to predict the experimental raw materials, pretreatment and enzymatic hydrolysis parameters, and divide the prediction results into data before and after PCA dimensionality reduction; compare and evaluate the two sets of data with the final experimental results.

[0008] In the above embodiment of the present invention, optionally, in step S1, data are collected from reviewed literature in the field of glucose production from lignocellulosic biomass published after 2010.

[0009] In the above embodiment of the present invention, optionally, in the step S1, the reviewed literature in the field of glucose production from lignocellulosic biomass published after 2010 is 539 data items that can be found on the web of science.

[0010] In the above embodiment of the present invention, optionally, in the S1 step, the different biomass raw materials include 23 raw materials such as sugarcane bagasse, corn straw, and switchgrass to ensure the wide applicability of the data; the different biomass raw materials are subjected to physical, chemical, and biological pretreatment, and the parameters recorded include pH value, temperature, time, and enzyme load; the composition of the materials before and after pretreatment, including cellulose, hemicellulose, lignin, and their glucose yield, are experimentally determined to establish a data set.

[0011] In the above embodiment of the present invention, optionally, in step S2, data preprocessing includes preliminary preprocessing in an Excel spreadsheet and subsequent data standardization in Python to eliminate the influence of feature dimensions and units.

[0012] In the above embodiment of the present invention, optionally, in step S3, the feature extraction also includes converting text language into data language and inputting it into the model; in the PCA dimensionality reduction process, the variance contribution rate of the feature is calculated, and the principal component with a cumulative contribution rate of more than 80% is selected to reduce the dimension and reduce the computational complexity.

[0013] In the above embodiment of the present invention, optionally, in step S4, the data set is divided into a ratio of 80% training set and 20% test set to ensure a balanced distribution of different raw materials and pretreatment methods, thereby improving the generalization ability of the model; For XGB, CNN, SVR, RF and / or MLP models, optimize the hyperparameters to improve the prediction performance; The XGB uses grid search to adjust the learning rate, maximum depth and subsampling rate; CNN optimizes the structure by adjusting the convolution kernel size, number of layers, learning rate and batch size; SVR uses grid search to optimize the kernel function, penalty parameter C, ε-insensitive interval and γ value; RF optimizes the model by adjusting the number of decision trees, maximum depth and minimum sample split number; MLP uses random search to optimize the number of hidden layer neurons, activation function and learning rate; Finally, the feature importance was evaluated through SHAP analysis, and the generalization ability of each model was verified on the test set to ensure the reliability of the prediction results.

[0014] In the above embodiment of the present invention, optionally, step S5 includes five supervised learning models, namely XGB, CNN, SVR, RF, and MLP, to explore the complex effects of pretreatment and enzymatic hydrolysis parameters on glucose yield. These models are adopted due to their excellent performance; the model performance is evaluated using the coefficient of determination, represented by R², and the root mean square error, represented by RMSE, where R² measures the fit of the model and RMSE reflects the prediction error. R² and RMSE are commonly used indicators to measure the predictive performance of regression models. R² is used to evaluate the model's ability to explain the variation in actual data. Its value ranges from 0 to 1. The closer it is to 1, the better the model fit and the closer the predicted value is to the true value. When R² is 1, it indicates a perfect fit. RMSE is used to measure the average deviation between the predicted value and the true value. The smaller the value, the lower the model error and the more accurate the prediction. Unlike R², RMSE has the same unit as the prediction target and therefore more intuitively reflects the absolute size of the prediction error. In practical applications, R² and RMSE are often used in conjunction: a model with a high R² but large RMSE may be susceptible to outliers or have bias concentrated in certain intervals, while a model with a small RMSE but low R² may not adequately account for variable variation. Therefore, the two complement each other and are essential metrics for evaluating regression model performance. In biofermentation or enzymatic hydrolysis prediction studies, models with high R² and low RMSE generally indicate more reliable predictions of product yield.

[0015] In the above embodiment of the present invention, optionally, in step 5, the relationship between R² and RMSE is expressed by the following formula:

[0016] Ypred,i: Glucose yield predicted by the model, Yact,i: Glucose yield obtained from the experiment, Ym: the average glucose yield obtained from all experiments, N: number of samples, i: loop variable.

[0017] After the present invention adopts the above technical solution, it has the following main effects: This method collects data on different biomass raw materials, multiple pretreatment methods and enzymatic hydrolysis conditions, constructs a multi-factor database, and uses machine learning technology to accurately predict saccharification yield. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0019] Figure 1 It is the overall flow chart of the present invention; Figure 2 It is a schematic diagram of the process of the present invention; Figure 3 This is a graph showing the effects of different raw materials and pretreatment methods on glucose yields according to the present invention; Figure 4 is a data set distribution diagram of the present invention; Figure 5 is the PCA scree plot of the present invention; Figure 6 The regression error graphs of the first five PCA models of the present invention are shown in FIG. Figure 7 Graphs showing the regression errors of the five models after PCA of the present invention. DETAILED DESCRIPTION

[0020] In order to better understand the technical solution of the present invention, it should be clear that the embodiments described are only some of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0021] The terms used in the embodiments of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The singular forms "a", "an", "the" and "the" used in the embodiments of the present invention and the appended claims are also intended to include plural forms unless the context clearly indicates otherwise.

[0022] To make the purpose, technical solutions, and advantages of the embodiments of the present invention clearer, the embodiments of the present invention are described in further detail. However, the embodiments described are only some embodiments of the present invention, not all. Based on the embodiments of the present invention, other embodiments that are not creatively created by those skilled in the art are within the scope of protection of the present invention.

[0023] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present invention more clearly understood, the following specific embodiments are provided, but are not intended to limit the scope of the present invention.

[0024] The present invention provides a method for predicting glucose production from lignocellulosic biomass (LCB) based on machine learning, specifically a method for converting lignocellulose into glucose, comprising the following steps, see Figure 1-2 , data collection, feature engineering and PCA dimensionality reduction, model training, model evaluation, etc. Table 1 shows the meaning of the names. The following is a detailed description of the implementation method with reference to the accompanying drawings.

[0025] Table 1 Name meaning comparison table

[0026] 1. Collect data on different raw materials and pretreatment methods In order to build a widely applicable model for different lignocellulose substrates and processing processes, this study collected experimental data under various raw materials and different pretreatment methods. The factors affecting glucose yield are relatively complex, involving multiple variables such as raw material composition, pretreatment method, and enzymatic hydrolysis conditions. The diversity and breadth of data are crucial for establishing an efficient prediction model, such as Figure 4 As shown in Figure e, the glucose yield varies from 1.5% to 99.8%, indicating that the dataset covers the distribution of glucose yields under different raw material properties and reaction conditions, which helps to deeply explore the contribution of influencing factors.

[0027] The lignocellulosic raw materials involved in the experiment had large compositional differences, with varying degrees of variation in the content of glucan, xylan, and lignin. The average glucan content of the raw materials was 38.05%, with a median of 38.06%; the average xylan content was 21.07%, with a median of 20.6%; and the average lignin content was 21.02%, with a median of 20.4%. Figure 4 These components have an important impact on the subsequent pretreatment and enzymatic hydrolysis efficiency.

[0028] Figure 4 The effects of different raw materials and pretreatment methods of the present invention on glucose yield are shown. The main purpose of pretreatment is to remove xylan and lignin, thereby enriching glucan and improving the accessibility and digestibility of cellulose. The data show that after pretreatment, the glucan content increased significantly, with an average value of 54.94%, an increase of 16.89% over the raw material, and a median of 55.37%, an increase of 17.31%. In contrast, the contents of xylan and lignin decreased, with the average content of xylan falling to 18.43%, a decrease of 2.64%, and the median falling to 14.96%, a decrease of 5.64%; the average content of lignin fell to 17.12%, a decrease of 3.9%, and the median fell to 14.88%, a decrease of 5.52%. These changes indicate that there are certain differences in the removal effects of pretreatment on xylan and lignin, and different raw materials and pretreatment methods lead to different removal efficiencies.

[0029] In addition, the average recovery rate of glucan reached 88.42% ( Figure 4 b), indicating that most of the cellulose was effectively retained during the pretreatment process. However, the removal rates of xylan and lignin varied greatly, reflecting that different pretreatment methods had different effects on these components. For example, the ratio of glucan to lignin (Glucan / Lignin ratio) after pretreatment varied between 0.29 and 15 ( Figure 4 This range is larger than the results reported in the literature. This may be because this study used multiple preprocessing methods, while the literature only selected one preprocessing method, resulting in a wider range of data. Figure 4 The enzymatic hydrolysis time in d also covers the time required for the enzymatic hydrolysis of most lignocellulosic materials, which will help in the subsequent exploration of the effect of enzymatic hydrolysis time on glucose yield.

[0030] Different pretreatment methods improve the enzymatic digestibility of raw materials by disrupting lignin structure and altering cellulose crystal structure, thereby affecting glucose yield. However, due to the diversity of pretreatment methods, the effectiveness of xylan and lignin removal varies significantly. Therefore, integrating data from different raw materials and pretreatment methods can help develop a more universal model and provide in-depth analysis of the key factors affecting glucose yield.

[0031] 2. Data Preprocessing Data preprocessing steps are crucial to ensure data quality. First, the collected data is recorded in an Excel spreadsheet. Subsequently, data cleaning is performed to deal with outliers, missing values, and errors. Next, the data is further processed, including encoding to convert categorical variables into numerical format, normalization to standardize the data, correlation testing to identify the relationship between features, and feature scaling to adjust the numerical range. These operations help in accurate and reliable analysis of the data. Data processing methods such as Figure 1 shown.

[0032] 3. In the data preprocessing stage, feature extraction is first performed on the lignocellulosic raw materials and pretreatment methods in order to convert the text information into numerical inputs acceptable to the model. Traditional encoding methods such as label encoding, one-hot encoding, and binary encoding are often difficult to capture the deep relationship between input variables and glucose yields. Therefore, this study proposed a feature extraction method based on ratio conversion to establish a direct connection between raw material type and pretreatment method and glucose yield, and enhance the comparability of input variables so that the model can effectively analyze their impact. Specifically, the raw materials are divided into six categories: bamboo, straw, grass, trees, grain residues, and sugarcane bagasse, while the pretreatment methods are divided into eight categories: no pretreatment, physical, acid, alkali, inorganic salt, organic solvent, deep eutectic solvent (DES), and ionic liquid. Subsequently, the average glucose yields of different categories of raw materials and pretreatment methods were counted ( Figure 3 a in Figure 3 b in the figure), taking the category with the highest glucose yield as the benchmark unit 1, calculate the ratios of other categories relative to the benchmark and use them as model input. For example, the glucose yield of cereal residue is the highest (72.4%), so it is defined as unit 1, and the values ​​of other raw materials are calculated according to their ratios to the yield of cereal residue: bamboo is 0.56, trees are 0.77, grasses are 0.89, straw is 0.94, and sugarcane bagasse is 0.97. Similarly, among the pretreatment methods, the glucose yield of alkali pretreatment is the highest (79.3%), which is defined as unit 1, and the values ​​of other pretreatment methods are converted into corresponding ratios and input into the model, such as no pretreatment is 0.32, physical pretreatment is 0.84, acid pretreatment is 0.75, inorganic salt pretreatment and DES pretreatment are both 0.95, and ionic liquid pretreatment is 0.88 ( Figure 3 b) in the above example.

[0033] In order to optimize the data structure and reduce redundant information, this study further used principal component analysis (PCA) to reduce the dimensionality of the data. PCA calculates the covariance matrix of the data and performs eigenvalue decomposition to extract the principal components with the largest variance contribution, thereby mapping high-dimensional data to a low-dimensional space while retaining key information as much as possible. In the specific implementation process, IBM SPSS Statistics software was used to perform PCA analysis on the data to identify the main influencing factors and reduce the variable dimensions. In the data set of this study, the glucose yield was affected by many factors, including the lignin content of the raw material, the glucose content after pretreatment, the lignin removal rate, the xylose removal rate, and the glucose / lignin ratio. The cumulative contribution rate of the principal components calculated by PCA exceeded 85% (see Figure 5 ), ensuring that the core features of the data are retained after dimensionality reduction.

[0034] 4. Dataset division and hyperparameter determination This study used one or a combination of five supervised machine learning models (XGB, CNN, SVR, RF, and MLP) to explore the effects of pretreatment and enzymatic hydrolysis parameters on glucose yield. To ensure the reliability and generalization ability of the model, the dataset was divided into an 80% training set and a 20% test set. The training set was used to train the model to learn the relationship between the input variables and glucose yield, while the test set was used to evaluate the model's generalization ability to prevent overfitting or underfitting. Before model training, all data were standardized to eliminate dimensional and unit differences between different variables, thereby improving the model's learning ability and computational stability. In addition, since the values ​​of the data after normalization are between 0 and 1, it helps to improve the convergence speed of model training and ensure that the effects of different variables on the prediction results can be reasonably compared.

[0035] In order to optimize the performance of the model, appropriate hyperparameter settings were further determined. This study selected five models, namely RF, SVR, XGB, MLP and CNN, and optimized hyperparameters according to the characteristics of different models. Hyperparameters are parameters set before training, which are different from the internal parameters learned by the model through training data. Reasonable hyperparameter optimization can significantly improve the predictive ability of the model while preventing overfitting or underfitting. Optimizing hyperparameters can not only effectively control the complexity of the model, but also improve computational efficiency, reduce training time and computing resource consumption. This study systematically adjusted the key hyperparameters of different models and used cross-validation to select the optimal parameter combination during training to ensure that the model can perform best on the test set. The final hyperparameter types and value ranges are detailed in Table 2. Through reasonable data set division and hyperparameter optimization, it is ensured that the model can fully learn the data characteristics and have good generalization ability, thereby accurately predicting glucose yield.

[0036] Table 2 Model hyperparameters

[0037] 5. Model training and comparative analysis This study trained and compared five supervised machine learning models for glucose yield prediction: XGB, CNN, SVR, RF, and MLP. To explore the impact of feature extraction and PCA dimensionality reduction on model performance, the models were first trained without feature extraction, then trained after feature extraction, and finally evaluated the improvement in the models' predictive capabilities by optimizing the input features through PCA dimensionality reduction.

[0038] Without feature extraction, as shown in Table 3, the model directly uses textual input variables for raw material type and pretreatment method, resulting in limited information representation and impacting model prediction accuracy. Test set R² results show that the MLP model achieved an R² of only 0.53 and an RMSE of 17.8. SVR performed even weaker, with an R² of only 0.45 and an RMSE of 18.3. The RF and XGB models performed relatively well, but their R²s were only 0.75 and 0.78, with RMSEs of 13.5 and 12.8, respectively. The CNN model achieved the best performance, achieving a test set R² of 0.76 and an RMSE of 13.8. However, overall, the predictive power of all models remained limited, primarily due to the inability of direct textual input to effectively capture the impact of raw material and pretreatment method on glucose yield. Therefore, to optimize model performance, feature extraction methods were implemented to convert textual information into numerical values, enhancing correlations between variables and improving comparability of input variables.

[0039] Table 3 Model fitting without feature extraction

[0040] After feature extraction, Figure 6 As shown in Figure 2, the prediction performance of all models has been significantly improved. The R² of MLP has increased from 0.53 to 0.66, and the RMSE has decreased from 17.8 to 14.8 ( Figure 6 d in ); SVR R² increased from 0.45 to 0.57, and RMSE decreased from 18.3 to 14.8 ( Figure 6 b in the figure); RF and XGB both achieved an R² of 0.83 on the test set, and their RMSE dropped to 9.28 and 11.2 respectively ( Figure 6 a and Figure 6 The CNN model showed the most significant improvement, with R² increasing from 0.76 to 0.89 and RMSE significantly decreasing from 13.8 to 7.18 ( Figure 6 e in the figure). Overall, CNN performed best among the five models, demonstrating its strong learning and generalization capabilities, better capturing the complex relationship between input variables and glucose yield. These results demonstrate that feature extraction methods effectively address the information loss associated with direct text data input and improve the model's predictive capabilities.

[0041] Although feature extraction optimizes model performance, there are still strong correlations between some of the 15 input variables, which may lead to data redundancy and affect the prediction effect of the model. In order to further optimize the input features, PCA is used for dimensionality reduction to reduce the interference of redundant variables and improve the computational efficiency of the model. Figure 7As shown in the figure, the first eight principal components calculated by PCA explain 86.5% of the variance, meeting statistical standards. Therefore, the 15 input variables were reduced to eight principal components, and the models were trained based on the reduced features. After PCA processing, the predictive ability of all models was further improved. The R² of MLP increased from 0.66 to 0.73, and the RMSE decreased from 14.8 to 14.2 ( Figure 7 d in ); SVR R² increased from 0.57 to 0.71, and RMSE decreased to 13.6 ( Figure 7 b in the figure); R² of RF and XGB further improved to 0.85 and 0.86, and RMSE decreased to 11.12 and 10.18 ( Figure 7 a and Figure 7 The performance of the CNN model improved most significantly, with the test set R² increasing from 0.89 to 0.95 and the RMSE significantly decreasing from 10.8 to 5.51 ( Figure 7 This indicates that PCA dimensionality reduction effectively reduces data redundancy and optimizes the feature space, enabling the model to more efficiently extract key variable information and improve the prediction accuracy of glucose yield.

[0042] Further analysis of the R² results on the training and test sets revealed that XGB, RF, and CNN achieved R² values ​​of 0.99, 0.98, and 0.98, respectively, on the training set, indicating a high degree of fit to the training data. However, performance on the test set declined, with XGB and RF's R² values ​​dropping to 0.82 and 0.85, respectively, indicating limited generalization to new data. In contrast, CNN maintained strong predictive power, achieving the highest R² value of 0.95 on the test set, demonstrating the best generalization ability. MLP and SVR achieved R² values ​​of 0.82 and 0.85, respectively, on the training set, but relatively low R² values ​​of 0.73 and 0.71, respectively, on the test set, indicating relatively weak predictive power on new data. Overall, CNN performed best in this study. Its multi-layer convolutional and pooling architecture effectively extracts local features and forms a hierarchical feature representation, enhancing its ability to model complex variable relationships. Furthermore, CNN exhibits translation invariance, enabling it to adapt to data changes and improving model stability. Furthermore, parameter sharing and sparse connectivity reduce the number of model parameters, mitigate the risk of overfitting, and improve computational efficiency and generalization capabilities. Therefore, the CNN model demonstrates significant advantages over other machine learning models in predicting glucose yield.

[0043] 6. Experimental Verification Results The experiment selected two raw materials, corn straw and switchgrass. The raw materials of the model included 23 raw materials, including corn straw and switchgrass. The experiment selected two of the multiple raw materials for verification. After alkaline treatment, enzymatic hydrolysis experiments were carried out. The parameters of the corn straw during the experiment were: before pretreatment, the glucan content was 34.7%, the xylan content was 21.3%, and the lignin content was 21.7%. After pretreatment, the glucan content was 57.2%, the xylan content was 21.2%, and the lignin content was 7.8%. The glucan retention rate was 98.25%, and the hemicellulose and lignin removal rates were 40.68% and 78.58%, respectively. After pretreatment, the cellulose to lignin ratio was 7.333. The enzyme loading was 5 FPU / g, the temperature was 50°C, and the enzymatic hydrolysis time was 120h. The relevant parameters of switchgrass were as follows: glucan content of 32.1%, xylan content of 18.5%, and lignin content of 19.5% before pretreatment; glucan content of 56.7%, xylan content of 20.4%, and lignin content of 13.9% after pretreatment; glucan retention of 97.15%, hemicellulose and lignin removal rates of 39.35% and 39.21%, respectively; cellulose to lignin ratio of 4.079 after pretreatment; enzyme loading of 5 FPU / g, temperature of 50°C, and enzymatic hydrolysis time of 120 hours. The parameters of the two feedstocks were input into the model to predict glucose yield.

[0044] Table 4 Predicted and experimental values ​​of glucose yield

[0045] The results show that the data after PCA dimensionality reduction is closer to the experimental data.

[0046] The conclusions drawn from the experiment are as follows: The model was tested on 23 raw materials, including corn stover and switchgrass. For convenience, two raw materials, corn stover and switchgrass, were selected for this experiment. Under identical alkaline pretreatment and enzymatic hydrolysis conditions, their relevant parameters were input into the model to predict glucose yield. Comparisons were then made with experimental results. The results showed that the model's predictions after PCA dimensionality reduction were closer to the experimental values. For example, the experimental yield of corn stover was 58.4%, while the predicted value after dimensionality reduction was 57.8%, compared to 72.2% before dimensionality reduction. For switchgrass, the experimental yield was 51.6%, while the predicted value after dimensionality reduction was 51%, compared to 61.7% before dimensionality reduction. This result demonstrates that PCA dimensionality reduction can effectively reduce redundant features and mitigate the impact of multiple linear relationships between features on model prediction accuracy, thereby improving the reliability of the prediction results.

[0047] Beneficial effects of this application: Using machine learning methods to predict glucose yield after enzymatic hydrolysis of lignocellulosic feedstocks offers significant technical advantages and practical value over traditional experimental methods, significantly improving prediction accuracy and efficiency. Traditional methods typically rely on the accumulation of extensive experimental data or the development of empirical models, which are not only time-consuming and costly, but also often struggle to accurately reflect the nonlinear coupling relationships between parameters when faced with complex multi-factor influences (such as feedstock type, pretreatment method, and enzymatic hydrolysis conditions). Machine learning technology, on the other hand, can fully exploit the underlying patterns in historical data to construct highly accurate prediction models, effectively addressing issues such as strong experimental uncertainty and complex variable correlations.

[0048] Furthermore, compared to existing techniques that use machine learning methods for glucose yield prediction on a single biomass or pretreatment, the present invention has broader applicability, adapting to different lignocellulosic feedstocks or pretreatment methods, and exhibiting excellent versatility. By introducing principal component analysis (PCA) to reduce the dimensionality of the raw feature data, the present invention reduces model complexity while retaining key information, improving generalization and stability. The model demonstrated a high degree of fit in actual predictions, with predicted glucose yields highly consistent with experimental data. The model also maintained good adaptability and accuracy across different feedstock types. In summary, compared to experimental determination of glucose yield and existing machine learning methods for glucose prediction, this method not only significantly saves experimental resources and optimizes process development, but also exhibits versatility and is applicable to common lignocellulosic biomass and pretreatment methods. The present invention represents a significant technological advancement and practical application value in improving the accuracy of glucose yield predictions, reducing the development costs of enzymatic hydrolysis processes, and promoting the efficient conversion of lignocellulosic resources.

[0049] The above is merely an embodiment of the present invention. Common knowledge such as the known specific structures and characteristics in the scheme is not described in detail here. Ordinary technicians in the relevant field are aware of all common technical knowledge in the technical field to which the invention belongs before the application date or priority date, can obtain all existing technologies in the field, and have the ability to apply conventional experimental means before that date. Ordinary technicians in the relevant field can improve and implement the present scheme based on their own abilities under the guidance of this application. Some typical known structures or methods should not become an obstacle for ordinary technicians in the relevant field to implement this application.

[0050] It should be noted that those skilled in the art may make various modifications and improvements without departing from the structure of the present invention, and these modifications and improvements should also be considered within the scope of protection of the present invention. These modifications and improvements will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application shall be based on the content of the claims, and the specific embodiments and other descriptions in the specification may be used to interpret the content of the claims.

[0051] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for predicting glucose production from lignocellulosic biomass based on machine learning, characterized in that: The method comprises the following steps: S1, collect data on different biomass raw materials, various pretreatment methods and enzymatic hydrolysis parameters, build a multi-factor database, and form a set; S2. Preprocess the collected data, including data cleaning and normalization. Data cleaning is achieved by removing outliers and noise data. Statistical analysis methods are used to calculate the mean and standard deviation of the data. Data that deviates from the mean by more than a set multiple of the standard deviation is considered an outlier and is removed. When normalizing the data, the minimum-maximum normalization method is used to normalize the data to the range of [0, 1]. S3. Performing feature engineering on the pre-processed data, wherein the feature engineering includes feature extraction and PCA dimensionality reduction; the core of feature extraction is to establish a relationship between raw materials and pre-treatment methods and glucose yield; S4. Divide the dataset and determine hyperparameters; divide it into training and test sets in proportion to ensure a balanced distribution of different raw materials and preprocessing methods, and improve the generalization ability of the model; S5. Model training and comparative analysis to explore the complex effects of pretreatment and enzymatic hydrolysis parameters on glucose yield; S6. Model evaluation: Use the final application model to predict the experimental raw materials, pretreatment and enzymatic hydrolysis parameters, and divide the prediction results into data before and after PCA dimensionality reduction; compare and evaluate the two sets of data with the final experimental results.

2. The method for predicting glucose production from lignocellulosic biomass based on machine learning according to claim 1, wherein: In step S1, data were collected from the reviewed literature published after 2010 in the field of glucose production from lignocellulosic biomass.

3. The method for predicting glucose production from lignocellulosic biomass based on machine learning according to claim 2, wherein: In the step S1, the reviewed literature in the field of glucose production from lignocellulosic biomass published after 2010 is 539 data available on the web of science.

4. The method for predicting glucose production from lignocellulosic biomass based on machine learning according to claim 1, wherein: In step S1, the different biomass raw materials include sugarcane bagasse and corn straw to ensure the wide applicability of the data; the different biomass raw materials are subjected to physical, chemical, and biological pretreatment, and parameters including pH value, temperature, time, and enzyme load are recorded; the composition of the materials before and after pretreatment, including cellulose, hemicellulose, lignin, and their glucose yield, are experimentally determined to establish a data set.

5. The method for predicting glucose production from lignocellulosic biomass based on machine learning according to claim 1, wherein: In step S2, data preprocessing includes preliminary preprocessing in an Excel spreadsheet and subsequent data standardization in Python to eliminate the influence of feature dimensions and units.

6. The method for predicting glucose production from lignocellulosic biomass based on machine learning according to claim 1, wherein: In step S3, the feature extraction also includes converting text language into data language and inputting it into the model; in the PCA dimensionality reduction process, the variance contribution rate of the feature is calculated, and the principal component with a cumulative contribution rate of more than 80% is selected to reduce the dimension and reduce the computational complexity.

7. The method for predicting glucose production from lignocellulosic biomass based on machine learning according to claim 1, wherein: In step S4, the data set is divided into 80% training set and 20% test set to ensure a balanced distribution of different raw materials and pretreatment methods and improve the generalization ability of the model; For XGB, CNN, SVR, RF and / or MLP models, optimize the hyperparameters to improve the prediction performance; The XGB uses grid search to adjust the learning rate, maximum depth and subsampling rate; CNN optimizes the structure by adjusting the convolution kernel size, number of layers, learning rate and batch size; SVR uses grid search to optimize the kernel function, penalty parameter C, ε-insensitive interval and γ value; RF optimizes the model by adjusting the number of decision trees, maximum depth and minimum sample split number; MLP uses random search to optimize the number of hidden layer neurons, activation function and learning rate; Finally, the feature importance was evaluated through SHAP analysis, and the generalization ability of each model was verified on the test set to ensure the reliability of the prediction results.

8. The method for predicting glucose production from lignocellulosic biomass based on machine learning according to claim 1, wherein: In step S5, one or a combination of five supervised learning models, including XGB, CNN, SVR, RF, and MLP, is used to explore the complex effects of pretreatment and enzymatic hydrolysis parameters on glucose yield. These models are adopted due to their excellent performance. Model performance is evaluated using the coefficient of determination (R²), which is the coefficient of determination, and the root mean square error (RMSE), which is the root mean square error. R² measures the degree of model fit, and RMSE reflects the prediction error. R² and RMSE are commonly used metrics to measure the predictive performance of regression models. R² is used to assess the model's ability to explain the variation in actual data. Its value ranges from 0 to 1. The closer it is to 1, the better the model fits, and the closer the predicted value is to the true value. When R² is 1, it indicates a perfect fit. RMSE is used to measure the average deviation between the predicted value and the true value. The smaller the value, the lower the model error and the more accurate the prediction. Unlike R², RMSE has the same unit as the prediction target, so it more intuitively reflects the absolute size of the prediction error. In practical applications, R² and RMSE usually need to be used in combination: models with higher R² but larger RMSE may be affected by outliers or have deviations concentrated in certain intervals, while models with smaller RMSE but lower R² may not explain variable variation well. Therefore, the two complement each other and are indispensable indicators for evaluating the performance of regression models. In biological fermentation or enzymatic hydrolysis prediction research, models with high R² and low RMSE usually mean more reliable predictions of product yield.

9. The method for predicting glucose production from lignocellulosic biomass based on machine learning according to claim 8, wherein: In step 5, the relationship between R² and RMSE is expressed by the following formula: ； Ypred,i: Glucose yield predicted by the model, Yact,i: Glucose yield obtained from the experiment, Ym: the average glucose yield obtained from all experiments, N: number of samples, i: loop variable.

Images