Machine Learning Approach for the Prediction of Biomass Waste Pyrolysis Kinetics from Preliminary Analysis Authors: Kai Xiao, Xianghui Zhu Published: 2024 ABSTRACT In the present work, artificial neural network (ANN)-based machine learning models are developed to predict biomass pyrolysis kinetics. Data sets of thermogravimetric analysis and feedstock characterization from a diverse range of biomasses were used to build and test the networks. The composition of the raw biomass material was classified and used as input parameters of ANN models. Three models, which use ultimate analysis, proximate analysis, and three components as input parameters, were developed in this study. A total of 32 types of biomass raw materials were used, and 270 sets of kinetic data were obtained according to different pyrolysis conversion rates ranging from 0.1 to 0.9. Results show that increasing the number of neurons can improve the prediction accuracy. The optimized neuron number is 7–11. The largest relative deviation between experimental and modeling results for the three models are 20.80%, 14.06% and 12.85%, respectively, which proves that using cellulose, hemicellulose, and lignin as input parameters of the neural network model can better predict the activation energy of pyrolysis at each reaction stage. The particle swarm optimization algorithm could significantly improve the prediction accuracy of the BP-ANN model. The largest deviation for activated energy prediction decreases from 12.85% to 6.72%. Introduction Biomass is the only renewable organic carbon resource on earth, which comes from a wide variety of sources, including crop straw, fuelwood, domestic sewage and industrial wastewater, livestock manure, and so on. The development and utilization of biomass energy have received continuous attention from governments and researchers around the world. The World Biomass Association believes that biomass will continue to be the world’s largest renewable energy source in the future and estimates that global biomass supply could reach 150 EJ by 2035, of which 43% will come from agriculture, 52% from forestry, and about 5% from other waste.1 In this context, increasing the development and utilization of biomass energy will help improve the global energy consumption structure and help alleviate the depletion of traditional fossil resources, and biomass energy is expected to become the backbone of achieving the goal of “carbon neutrality”.2 Biomass energy conversion technology mainly includes thermal chemical methods, physical–chemical methods, biochemical methods, and chemical methods, among which thermal chemical conversion is the most common utilization method of biomass energy. The main thermochemical conversion methods of biomass include direct combustion, gasification, and pyrolysis. Pyrolysis is a technology that uses heat energy to break chemical bonds in biomass macromolecules and transform them into low-molecular-weight products under isolated air conditions. It is not only a clean and effective method for producing bio-oil, combustible gas, and biochar but also the initial stage of direct combustion and gasification.3,4 Therefore, an in-depth understanding of biomass pyrolysis characteristics and exploration of biomass pyrolysis laws are helpful to find more efficient biomass thermochemical conversion methods. From the perspective of energy utilization, lignocellulosic biomass has the greatest potential for development and is abundant and cheap. This kind of biomass mainly comes from crop straw and is mainly composed of cellulose, hemicellulose, and lignin in different proportions. The pyrolysis process involves numerous simultaneous complex reactions and the influence of the reaction conditions. The kinetic performance of the pyrolysis process is often very different from the proportion and characteristics of the obtained three-phase products. In the study of pyrolysis technology, a lot of work has been done to calculate the pyrolysis kinetic parameters of various lignocellulosic biomass through thermogravimetry (TG) analysis (TGA) and to obtain the proportion of three-phase products, product composition, and characteristics through pyrolysis experiment and analysis. However, all of these have undergone a strict kinetic calculation process and time-consuming experiment and analysis. It is not conducive to quickly and accurately understanding the characteristics of biomass and pyrolysis products. The rapid development of artificial intelligence, especially machine learning methods, which can establish correlations between input features and output targets based on large amounts of data, has been recognized as an excellent means to discover hidden relationships behind data and has been well applied in the field of competence. Therefore, based on the existing research results on the pyrolysis kinetics and pyrolysis product analysis of lignocellulosic biomass, the relationship between the composition and structure characteristics of lignocellulosic biomass and the pyrolysis process can be found by means of machine learning, which is helpful in selecting appropriate reaction conditions for different biomass and conduct directed regulation of the pyrolysis process, which can greatly simplify the experimental and computational analysis process. Reducing the cost of experiments will be a very meaningful work. At present, machine learning models for predicting biomass pyrolysis activation energy (Ea), pre-exponential factor (A), and reaction order (n) have been established based on ultimate analysis, proximate analysis, and content of three components of biomass. Sunphorka et al.5 used artificial neural network (ANN) models to predict the pyrolysis kinetic parameters based on three components. They found that all components had an effect on the activation energy. Cellulose had a strong effect on the pre-exponential factor, while hemicellulose had a strong effect on the reaction order. Çepelioğullar et al.6 conducted activation energy prediction of biomass wastes based on different neural network topologies. The experimental thermal data fitted quite well to the ANN predicted data. These models can be used to predict the kinetic parameters of biomass pyrolysis to a certain extent. However, most machine learning models obtained a single pyrolysis kinetic parameter. However, the composition of biomass is complex, the pyrolysis reactions are different in each stage of the pyrolysis process, and there are large differences.3 Therefore, a large temperature range of pyrolysis reaction cannot be described by a single kinetic parameter.7,8 Therefore, it is necessary to combine different pyrolysis reaction stages to obtain kinetic parameters at different temperature reaction intervals or at different conversion stages so as to provide more accurate guidance for the entire pyrolysis process. However, this study has not yet been reported. In this paper, a machine learning model based on an ANN was constructed to investigate the structural characteristics of different biomass, such as elemental analysis, industrial analysis, three-component characteristics, etc., and to simulate and predict the pyrolysis kinetic characteristics of a large number of different lignocellulosic biomass materials. The changes of biomass pyrolysis kinetics under different conversion conditions were investigated, which provides important methodological support for the development of the biomass pyrolysis technology. Pyrolysis Mechanism of Lignocellulosic Biomass Pyrolysis Mechanism of Cellulose, Hemicellulose, and Lignin Domestic and foreign scholars have carried out a large number of experimental studies and proposed dynamic models targeting apparent weight loss of biomass pyrolysis, such as the distributed activation energy model.9 Recently, some scholars have proposed micronetwork biomass models, such as the Monte Carlo model,10 to predict the detailed component information on pyrolysis. However, it should be noted that in the above-mentioned model studies, the types of biomass targeted and the heating conditions are very limited, so the kinetic parameters of these models do not have the universality of biomass types and heating conditions. In fact, there is a very complex nonlinear relationship between the kinetics of biomass pyrolysis and the type of biomass (chemical composition) and the heating conditions to which it is subjected to. This nonlinear relationship is essential for the development of universal biomass pyrolysis models, but it has not been fully recognized so far and needs further study. TG Experimental Study of Lignocellulosic Biomass Biomass is mainly composed of cellulose, hemicellulose, and lignin, and the pyrolysis behavior of the three components reflects the whole pyrolysis law of the birth material to a large extent. Therefore, it is necessary to study the influence of the three components on the pyrolysis of biomass. In order to study the influence of three components on the weight loss characteristics of biomass pyrolysis, a thermal balance analyzer was used to study the weight loss characteristics of the three components and biomass, and on this basis, the pyrolysis kinetic model of biomass and the three components was established. Cellulose, hemicellulose, and lignin were purchased from Sigma-Aldrich: cellulose was microcrystalline cellulose (CAS: 0009004346); lignin was alkali lignin, with low sulfonate content; and due to the complex composition of hemicellulose, it is difficult to purchase directly. Therefore, xylan from pine wood was used as the model compound of hemicellulose in this paper. The basic composition is O-acetyl-4-O-methylglucuronic acid xylose. Pine wood was selected as the typical biomass as raw material in the experiment, which was crushed by the sample preparation equipment (KER-1/100A sealed sample preparation crusher). The particle size was less than 100 μm, and the composition analysis is shown in Table 1. Table 1Composition of Biomasssamplesultimate analysis %proximate analysis CHObNashvolatilefixed carbonPine wood46.725.8643.370.761.2877.6618.28 cellulosehemicelluloselignin 45.3421.9628.76 Composition of Biomass The thermal balance analyzer was NETZSCH STA 409PC synchronous thermal analyzer. The NETZSCH STA 409C synchronous thermal analyzer performs TGA at operating temperatures up to 1300 °C with a temperature accuracy of <1 K. In this paper, the sample was heated from 50 to 800 °C at a constant rate of 10 K/min, the carrying gas was Ar, the carrying gas flow rate was 50 mL/min, and the sample mass was (10 ± 0.5) mg. TG Results of Biomass Pyrolysis In this section, the detailed pyrolysis of cellulose, hemicellulose, lignin, and pine wood is studied using the TGA method. The TG and differential thermogravimetry (DTG) curves of four kinds of materials are summarized in Figure 1 as follows. TG and DTG curves of four kinds of biomass materials. As shown in Figure 1a, the weight loss process of cellulose pyrolysis can be composed of four stages. The first stage is from room temperature to 140 °C, which is the water release stage, and the weight loss is 1.2%, which is the vaporization process of cellulose by absorbing heat in water. The second stage is 140–250 °C, which corresponds to the depolymerization reaction and “glass transition” of cellulose, and the polymerization degree is rapidly reduced, producing active cellulose, which is composed of low polymerization degree cellulose fragments. Since the formation of active cellulose does not cause weight loss, the weight loss of this stage of the reaction is very small. The third stage is the volatile analysis stage from 250 to 370 °C, in which the weight loss is 86.12%. The temperature corresponding to the peak point of the DTG curve is 338 °C, and the weight loss rate is −18.95%/min. A series of reactions, including dehydration, decarboxylate reaction, glucosidic bond breaking of active cellulose, ring opening, and splitting of the pyran ring, take place subsequently, with volatiles and char produced. The fourth stage is from 370 °C to the end of the reaction (800 °C), in which a small amount of residual volatiles continues to be precipitated. In the char polycondensation stage, the weight loss is 5.12%, including the aromatization reaction to produce glycolaldehyde and other alcohols, aldehydes, and acids. As shown in Figure 1b, the weight loss of hemicellulose pyrolysis with xylan as the model can be divided into three stages. The first stage is the water release stage from room temperature to 150 °C, and the weight loss is 5.52%, which is the water evaporation process. The second stage is the volatilization release stage from 150 to 328 °C, in which the weight loss is 53.14%. The temperature corresponding to the peak point of the DTG curve is 295 °C, and the weight loss rate is −8.17%/min. In this stage, the xylan side chain and glucoside bond begin to break. Dehydration reactions, depolymerization, and aromatic polycondensation reactions occur at the same time. The products include fragments of acid anhydride, water-soluble acid, coke, water, l-glucan, furan, and so on. In the third stage, from 328 °C to the final temperature (800 °C), residual volatiles continue to release, and in the char polycondensation stage, the final solid residue of the reaction reduces to 23.6%, which is very close to the result of 24% solid residue when Gao-jin Lv et al.11 studied the pyrolysis of hemicelluloses extracted from corn stalks at the final temperature of 700 °C. As shown in Figure 1c, there are three obvious weight loss peaks in the lignin pyrolysis process, corresponding to three stages: The first stage is the phase of water precipitation from room temperature to 175 °C, with a weight loss of 5.68%. And the peak temperature is 95 °C with a weight loss rate of −0.83%/min. The external water content of lignin is higher than that of cellulose and xylan, which is related to the special moisture absorption of lignin.12 The second stage is the volatilization stage from 175 to 565 °C, with the weight loss of 31.26%. The corresponding temperature of the weight loss peak is 326 °C. The dehydration reaction mainly occurs when the temperature is lower than 500 °C, and as the temperature continues to rise, lignin monomer products are generated.13 The third stage is the continuous release of residual volatiles and the process of char polycondensation from 565 °C to the final temperature (800 °C). In this stage, the weight loss is 7.63%, and there is an obvious weight loss peak, corresponding to 683 °C. The weight loss curve of pine wood during the temperature rise from 50 to 800 °C at 10 K/min is shown in Figure 1d. It can be seen that the pyrolysis process of pine is divided into three stages. The first stage of pyrolysis is from room temperature to 140 °C, and the weight loss is 1.1%. From 140 to 390 °C is the second stage of pine pyrolysis, corresponding to the volatilizing analysis stage. The weight loss is 67.36%, and the maximum weight loss rate occurs at 360 °C, which is −7.22%/min. The third stage is 390–800 °C, in which the residual volatiles continue to release from char and the char condensation process enhances. The final solid residual amount is 22.21%. Kinetic Study of Biomass Pyrolysis Many researchers have done a great deal of research on the pyrolysis mechanism of biomass. At present, there are many methods used to analyze the kinetic data of the non-isothermal solid phase reaction, which can be generally divided into the mechanism model method and iso-conversional method.14 The Dynamics Committee of the International Confederation for Thermal Analysis (ICTAC) does not recommend the use of the model method due to the inherent uncertainties in the determination of kinetic parameters.15 When the iso-conversional method is used, it is not constrained by the fixed model, and the uncertainty of estimating dynamic parameters can be avoided. For example, Wu et al.16 used iso-conversional methods to reveal the pyrolysis mechanism of tobacco waste and inferred the pyrolysis reactions under different conversion ranges according to the changing trend of activation energy. Some iso-conversional methods, such as KAS (Kissinger–Akahira–Sunose) and FWO (Flynn–Wall–Ozawa), which are based on the typical integral method, are widely used for biomass pyrolysis.17,18 Ceylan et al.19 used KAS, FWO, and Coats–Redfern methods to calculate the activation energy, pre-exponential factor, and reaction order of waste hazelnut shell biomass. However, these studies obtained pyrolysis kinetic parameters through the analysis of TGA data and then inferred the pyrolysis mechanism of biomass, and they did not further reveal the reaction mechanism according to the changes of functional groups, element content, free radicals, and other physicochemical properties of biochar during pyrolysis. The effects of cellulose, hemicellulose, and lignin on the physicochemical properties of biomass during pyrolysis were ignored. The general expression of non-isothermal kinetics of biomass decomposition can be expressed as1Biomass conversion ratio can be defined as2The heterogeneous function is related to the reaction mechanism and can be expressed as3In which, k(T) refers to the reaction kinetic constants; f(a) refers to the reaction mechanism function; T refers to the temperature; t refers to the time; α represents the conversion ratio; and m0, mt, and m∞ represent the initial mass, real time mass, and final mass of the solid residual, respectively. The rate constant is a function of temperature and can be expressed as4 Combined with the above expression, the basic expression of the analytical method for calculating kinetic parameters based on TGA data can be obtained5 The TGA is performed at a constant heating rate, and the heating rate is defined below6 The equal conversion rate can be expressed as a function of temperature, but temperature also depends on the rate of heating, so it can be expressed as7In the above formula, k is the rate constant and β is the heating rate. As an integral equivalent conversion model, FWO is a more accurate method of estimating the activation energy. ln(β) has a linear relationship with 1/T at different heating rates β. The simplified FWO model can be expressed as (Wang et al.20)8 Global kinetic models of biomass pyrolysis are often determined using TGA, which aims to understand the apparent kinetics of biomass pyrolysis by measuring mass loss and pyrolysis temperature in relation to the reaction time. Generally, the mass loss observed during biomass pyrolysis can be divided into three stages. The first stage corresponds to the evaporation of free water in the biomass. The second stage corresponds to the release of volatile products, which is the main stage of biomass pyrolysis. The final stage is the formation of biochar, in which there is a slight loss of biomass weight. Based on the TG results in Figure 1, the detailed kinetic parameters are calculated using the Coats–Redfern method. The results are summarized in Table 2 as follows. Table 2Kinetic Parameters of Biomass and the Three Componentsmaterialheating rate (K/min)temperature range (°C)Ea (kJ/mol)cellulose10250–370236.62xylan10150–328128.78lignin10175–56568.07pine wood10145–390176.22 Kinetic Parameters of Biomass and the Three Components As shown in Table 2, the activation energies for cellulose, xylan, and lignin are 236.62, 128.78, and 68.07 kJ/mol, respectively. The results correspond to the work of Zhou et al.21 The activation energy of pine wood is about 176.22 kJ/mol, which is the result of the superposition of three components.22 However, TG studies can obtain only lumped product yields, so global kinetic models are usually represented by lumped reactions. With the development of pyrolysis research, many global models with multistep reactions have been proposed, which can aggregate the yield of the final pyrolysis process. Among them, the global mechanism of cellulose pyrolysis has been studied extensively. The most popular example is the B–S model, which states that cellulose first breaks down into active cellulose, representing a liquid phase, which then degrades through two competing reactions: one to produce bio-oil and the other to produce coke and gas. This model can predict the experimental data well and has a wide application prospect. Based on the B–S model, other models of the cellulose pyrolysis mechanism have been developed. There are fewer global kinetic models of hemicellulose and lignin pyrolysis than linear models because their structures and pyrolysis processes are more complex. Most of these are built on the basis of the classic B–S model. The summary of the exiting mechanism for the three components is shown in Figure 2.3,23 Global reaction for the three components of lignocellulosic biomass. It can be seen that due to the different pyrolysis characteristics of the three components of biomass and the large differences in reaction characteristics at different temperature ranges, the iso-conversional model based on different heating rates can more clearly analyze the pyrolysis and conversion mechanism of biomass at each stage. The kinetic parameters of the lumped reaction in the global model include activation energy and pre-exponential factors, which are usually obtained through TG experiments at different heating rates and through the iso-conversional model.24 Modeling Method ANN Model Building Back propagation (BP) ANN is proposed by Rumelhart and McClelland.25,26 It is a multilayer feed-forward neural network that is trained and optimized based on the error reverse propagation algorithm. BP-ANN is one of the most widely used neural networks in different fields. Similar to other machine learning algorithms, there is no need for the ANN model to uncover the detailed mechanism of the real process in advance; it only needs to learn some rules between conditions and results, through its network training, and gets the results closest to the expected output values from the given input values.27 During BP-ANN modeling, several layers (one layer or multiple layers) of neurons are added between the input layer and the output layer, which are called hidden layers. The neurons in the hidden layers are called hidden nodes. The hidden nodes have no direct connection or information interaction with the outside world, but the change in their state can affect the relationship between the input and output variables. Each layer can have several nodes.28 During the ANN modeling, the prediction performance largely depends on the input layer parameters and network parameters. The input layer is discussed above. As for the BP-ANN model, the major network parameters, including the transfer function type of network layers, the weights and bias of the network, and the neuron number of the hidden layer, have crucial influence on model performance. Transfer function plays a key role in BP-ANN modeling. The transfer function for the input layer and output layer is usually chosen as a linear function, namely pureline function. With respect to the hidden layer, nonlinear transfer functions are usually used. Sigmoid functions are the most widely used nonlinear transfer functions in BP-ANN modeling. There are two typical sigmoid functions, namely, Log-Sigmoid function and Tan-Sigmoid function. The equations of the Log-Sigmoid function and Tan-Sigmoid function are shown as follows.910 Log-Sigmoid function changes from 0 to 1, while the value range of the Tan-Sigmoid function changes from −1 to 1. The comparison of the prediction performance between logsig and tansig functions were conducted. In this paper, since the kinetic parameters of biomass pyrolysis are always positive, the Log-sigmoid transfer function is more appropriate. Further, the number of hidden layers is another important adjusting factor. From the basic ANN theory,29 some empirical formulas could be used for estimating the neuron number in the hidden layer. The following empirical formula could be used to evaluate the neuron number of the hidden layer.11In which, n represents the neuron number of the hidden layer and i and o represent the numbers of the input layer and output layer, respectively. α is an adjusting constant, the value of which could be varied from 1 to 10. During the modeling, the optimized neuron number was obtained by trial and error analysis.29,30 The mean square error (MSE) was used as the indicator for the evaluation of the training performance of the network. The MSE value was calculated by the following equation12in which, n refers to the number of validated samples; Yi represents the actual results of samples; and represents the modeling results. wi represents the weight of each sample. To enhance the predictive capability of the ANN model, the data sets were randomly partitioned into training (70%), validation (15%), and test (15%) subsets. The network was trained using the Levenberg–Marquardt back-propagation algorithm, which is a widely recognized optimization technique in machine learning. The learning rate and training target minimum error were set to 0.01 and 0.0001 respectively. Data Sets Collection for Lignocellulosic Biomass Pyrolysis in Different Conditions In this wok, the non-isothermal pyrolysis kinetics of lignocellulosic biomass was studied. As mentioned in the second part, biomass will show different reaction characteristics in different pyrolysis reaction stages during the heating process, so its reaction kinetic parameters will also change with the pyrolysis reaction. Therefore, in this paper, an equal conversion reaction model is adopted, and the conversion rate of the pyrolysis reaction process is taken as an important input parameter. At the same time, the composition of the biomass is investigated, including ultimate analysis, proximate analysis, and the proportion of the three components (cellulose, hemicellulose, and lignin). The data parameters are derived from the data published in the literature. A total of 32 types of biomass raw materials were used, and 270 sets of kinetic data were obtained according to different pyrolysis conversion rates ranging from 0.1 to 0.9. The detailed data are summarized in the attached file. Before the data can be used, they need to be normalized or standardized. The so-called normalization is to change the input and output data of the network to limit the processing to [0 1] or [−1 1]. Since the kinetic parameters and biomass composition and structure parameters in this paper are positive numbers, the normalization range is 0–1. The purpose of normalization is to prevent different types of parameters from producing different weights in the model prediction process, which will affect the overall prediction error. The BP-ANN model in this paper uses the preprocessing function mapminmax to normalize the data. After the prediction results are obtained in the model, the function is also used to reverse normalize the prediction results, and the actual prediction value is obtained. In this work, three BP-ANN models are built. The structural configurations of the models are summarized in Figure 3 as follows. BP-ANN structures for different models. Model-1 uses the ultimate analysis and conversion ratio as the input variables. Carbon (C), hydrogen (H), and oxygen (O) are three main elements. The content of nitrogen and sulfur elements is relatively small and has little effect on TGA characteristics; therefore, it is not used as input data. Model-2 uses proximate analysis results, namely, moisture, volatile matter, and fixed carbon and ash content, as the input variables. Model-3 uses three components, namely, cellulose, hemicellulose, and lignin content, as the input parameters. Results and Discussion Prediction of Kinetic Parameters Based on ANN Modeling In this section, the BP-ANN models are built to predict the kinetic parameters, mainly the activation energies for different kinds of lignocellulosic biomass materials. As discussed above, under non-isothermal thermogravimetric conditions, biomass pyrolysis has different thermogravimetric rates at different temperature ranges. At the same time, in different temperature ranges, the reaction of the pyrolysis process also varies with temperature and conversion rate. Therefore, for biomass, a single pyrolysis kinetic parameter cannot be used to accurately describe the characteristics of the entire pyrolysis stage. Using the equal conversion model, the activation energy of different biomasses at different conversion stages can be obtained, and the pyrolysis process of complex biomass can be further analyzed. The pyrolysis process can then be precisely regulated. As discussed in Section 3, the detailed description for the BP-ANN models for prediction of activation energy is obtained. First, the influence of neuron number in the hidden layer on the prediction performance is studied. The neuron number is varied from 2 to 13. The change of the MSE with neuron number is summarized in Figure 4 as follows. Influence of neuron number on the modeling performance (single hidden layer with transfer function: Logsig). It can be seen that MSE decreases rapidly with the increase in the number of neurons, indicating that increasing the number of neurons can improve the prediction accuracy. With respect to Model-1, when the number of neurons reaches 11, MSE reaches a minimum value of 0.0211. With the increase of the number of hidden layer neurons, the MSE increases, indicating that the excessive number of neurons may lead to overfitting of the prediction, which will lead to the decline of the prediction effect. At the same time, the simulation results show that when the number of neurons reaches 7, the decline rate of MSE begins to slow down. Model-2 and Model-3 have a similar tendency. Therefore, for these models, it is reasonable to choose 7–11 as the optimal number of neurons. The regression analysis of the training, validation, and testing results are shown in Figure 5 below. Regression analysis of the correlation results. It can be seen that the R-squared values for three models are 0.82349, 0.83051, and 0.88242, respectively. It shows that the prediction accuracy for activation energy evaluation during the biomass pyrolysis process is reasonable. Model-3 seems to show better performance for network training. Although the pyrolysis process of waste is very complicated and there quite are many influencing factors, the ANN model can avoid the understanding of the mechanism and achieve quite good prediction performance. Next, the prediction performance for pyrolysis activated energy at different conversion ratios is studied. Different input parameters are used as the data in the input layer. The comparison between experiments and modeling results are summarized in Figure 6 as follows. Activated energy prediction based on different input parameters. As shown in Figure 6, the prediction performance for activated energy at different conversion ratios by different input parameters are compared. With respect to ANN-1, the largest deviation between experiment (308.7 kJ/mol) and model results (244.5 kJ/mol) is 64.2 kJ/mol at a conversion ratio of 0.4. As for ANN-2, the largest deviation is 40 kJ/mol at a conversion ratio of 0.7. Three components are used as input data in ANN-3. The largest deviation is 36 kJ/mol at a conversion ratio of 0.7. The three neural network models can predict the activation energy calculated by the conversion model to a certain extent. In comparison, better prediction accuracy can be obtained by using three components as input data to predict the activation energy of the reaction at different stages. Furthermore, the relative differences between experimental and modeling results are calculated for the three models. The results are summarized in Table 3 as follows. Table 3Relative Difference between Modeling and Experimental Results (Relative Difference Value, Unit %)conversion ratio0.10.20.30.40.50.60.70.80.9ANN-1/%8.871.7915.5220.809.573.9713.777.884.17ANN-2/%3.625.515.467.583.830.6514.062.782.49ANN-3/%2.6412.853.561.655.761.9512.622.424.08 Relative Difference between Modeling and Experimental Results (Relative Difference Value, Unit %) As shown in Table 3, the largest relative differences between experimental and modeling results for the three models are 20.80%, 14.06%, and 12.85% respectively. It is obvious that using cellulose, hemicellulose, and lignin as input parameters of the neural network model can better predict the activation energy of pyrolysis at each reaction stage. Based on the individual relative difference values, the average relative error (ARE) could be obtained by the following equation.13in which, the deviation refers to relative difference values given in Table 2 and n refers to the number of deviation numbers, which equals 9 in this work. ARE values for the three models are calculated and summarized as follows. It can be seen from Figure 7 that ARE values are 9.59, 5.11, and 5.28 for the three models, respectively. Therefore, ANN-2 and ANN-3 seem to have a comparable performance and show better performance than ANN-1. Combined with the individual difference values, the largest relative difference between the experimental and modeling results for ANN-3 is 12.85%, which is lower than those of ANN-2 and ANN-3. In terms of accuracy and stability of the model prediction, ANN-3 has better prediction performance. ARE values for the three models. Three components constitute the main structure of the lignocellulosic biomass. To some extent, the pyrolysis of biomass is the superposition of the three components. Of course, there are also interactions between the three components, and their synergistic effect has a significant impact on the change in activation energy and product generation during the pyrolysis. Therefore, the simple superposition of the three components cannot be used to replace the actual biomass. However, from the perspective of the prediction of activation energy, three components as input parameters have better results than ultimate and proximate analyses, which is understandable. Because ultimate analysis and proximate analysis cannot reflect the actual composition of the biomass, the ratio of the three components can better reflect the internal structural characteristics of the biomass, which is also the reason for using it as an input parameter to achieve a better performance. PSO-ANN Modeling for Biomass Pyrolysis The BP neural network can predict the pyrolysis kinetic parameters to a certain extent. However, for complex multidimensional training data or setting different parameters, there are some defects such as overfitting, low convergence speed, falling into local optimal, and poor prediction effect.31 This greatly limits the application of neural networks in the complex pyrolysis process of biomass. In recent years, swarm intelligence optimization ANNs have gradually become a research hotspot in the field of modern optimization, and the particle swarm optimization (PSO) algorithm is a major research hotspot in the field of optimization algorithm. PSO was first proposed by Eberhart and Kennedy,32 and its basic concept is derived from the study of the foraging behavior of birds.33 The particle speed can be dynamically adjusted according to the historical optimal position of the particle and the historical optimal position of the population. A particle has only two properties: velocity, which represents how fast it is moving, and position, which represents the direction it is moving. The optimal solution searched by each particle individually is called the individual extreme value, and the optimal individual extreme value in the particle swarm is the current global optimal solution. Through continuous iteration, the particle velocity and position are updated. Finally, the optimal solution satisfying the termination conditions is obtained. In this section, the PSO method is used to optimize the BP-ANN model. The detailed configuration of the PSO-ANN modeling process could be summarized as follows. As shown in Figure 8, the process of PSO-ANN is essentially to use PSO to determine the initial weight and threshold of the BP neural network. Fitness function (objective function) is the prediction error of the BP neural network. According to the fitness function, the PSO algorithm seeks the optimal position and then initializes the optimal BP weight and threshold. Logical configuration of PSO-ANN modeling for biomass pyrolysis. In the PSO algorithm, the dimension D of the space where the particle is located is determined by the number of undetermined parameters. The number of particles is N. When applying the PSO algorithm, the particle swarm is first initialized and the positions and velocities of particles are randomly generated in the D-dimensional space,32 where the position and velocities of the jth particle are, respectively, expressed by eqs 5 and 6 as follows271415 During parameter optimization, the position and speed values of the next iteration step are calculated by eqs 7 and 8 according to the historical optimal position of each particle1617in which, k is the number of evolutions; ε is the inertia weight; and c1 and c2 are the accelerated constants. r1 and r2 are random variables in the optimization process. PjD and GD are the local optimal solution of a single particle and the global optimal solution of all particles, respectively.27,34 In this work, the particle population is chosen as 50; the update parameters for particle velocity and position are c1 = 1.8, c2 = 1.8; and the generation of evolutionary is chosen as 10. The vibration of fitness of the PSO-ANN model with evolution generation is summarized in Figure 9 as follows. Fitness of the PSO-ANN model. As shown in Figure 9, with the increasing evolution generation, the PSO-ANN fitness decreases. When the evolution generation reaches 5, the fitness value decreases to 36.36. Upon further increasing evolution generation, the fitness value maintains unchanged. Therefore, using five generations of evolution for the particles is suitable in this model. Next, the activated energy prediction performance is evaluated. The comparison of modeling and experimental results is summarized in Figure 10 as follows. Activated energy prediction by PSO-BP-ANN modeling. It can be seen that the PSO algorithm could significantly improve the prediction performance of the BP-ANN model. The largest relative deviation for activated carbon is about 6.72% at a conversion ratio of 0.2. With respect to the BP-ANN model, the largest relative deviation is 12.85%. It should be noted that the relative difference value of BP-ANN is 2.42% when conversion rate = 0.8, which is lower than that of PSO-BP-ANN (3.36%). However, this does not mean that the BP-ANN model has better predictive performance. The prediction of the neural network model is for the whole data, which is to build a neural network through a large amount of data. The difference in individual data does not reflect the accuracy of the model as a whole. With respect to these two models, it is obvious that the PSO-BP-ANN model shows better performance for the activation energy evaluation on the whole. Besides, the relative difference values for conversion ratios of 0.2 and 0.7 are much larger than those for other conversion ratios. The reason may be that these two points are in the transition stage of the significant pyrolysis reaction process. The conversion ratio of 0.2 represents the initial stage of the pyrolysis reaction, and at this stage, the three components show different reaction properties. Therefore, there are large fluctuations in the activation energy. With respect to the conversion ratio of 0.7, the point represents the final stage of the pyrolysis reaction. The rapid devolatilization reaction gradually changes to a slow polycondensation stage with a small amount of gas release. These two stages largely depend on the raw materials. Therefore, the relative difference values are much larger than those of other stages. Biomass is mainly composed of three components, namely, cellulose, hemicellulose, and lignin. The pyrolysis of lignocellulosic biomass is a complicated process. In addition to the respective pyrolysis characteristics of the three components, the interaction between the three components also affects the pyrolysis process. A single kinetic equation cannot accurately describe the characteristics of the whole pyrolysis process. Using an isoconversional model, the reaction characteristics of different pyrolysis stages can be revealed. In general, the proportion of the three components of biomass has a key effect on the overall pyrolysis reaction characteristics of biomass. This paper compares ultimate analysis, proximate analysis, and three components as input parameters of the ANN model and finds that three components as input parameters can get better prediction of dynamic parameters. However, the individual BP-ANN model is prone to overfitting or local convergence problems in the process of network training. The PSO optimization algorithm can optimize the key weights of the neural network in advance, thus avoiding the adverse effects of random weights on the neural network. In this paper, the prediction effects of PSO-BP-ANN and BP-ANN models on biomass pyrolysis kinetic parameters were compared, and the results proved that the PSO optimization algorithm could significantly improve the prediction accuracy of the BP-ANN model. Conclusions The pyrolytic behavior of lignocellulosic biomass is highly complex, and its kinetic behavior varies with operating conditions and the type of biomass. In order to reduce the time and cost of the experimental process, it is necessary to build a precise model to quickly and accurately predict the pyrolysis process of biomass. In the present work, ANN-based machine learning models are developed to predict the biomass pyrolysis kinetics. Data sets of TGA and feedstock characterization from a diverse range of biomass were used to build and test the networks. The composition analysis of raw biomass material was classified and used as input data for ANN models. Three models, which use ultimate analysis, proximate analysis, and three components as input parameters, were developed in this study. A total of 32 types of biomass raw materials were used, and 270 sets of kinetic data were obtained according to different pyrolysis conversion rates ranging from 0.1 to 0.9. The main conclusions can be summarized as follows. BP-ANN modeling results show that increasing the number of neurons can improve the prediction accuracy. The optimized neuron number ranges from 7 to 11, in which the MSE reaches a lower level. Comparing different parameters as input layer, the largest relative deviation between experimental and modeling results for the three models are 20.80%, 14.06% and 12.85%, respectively, which proves that using cellulose, hemicellulose, and lignin as input parameters of the neural network model can achieve better prediction performance for the activation energy of pyrolysis at each reaction stage. Individual BP-ANN models are prone to overfitting or local convergence problems in the process of network training. The PSO algorithm could significantly improve the prediction accuracy of the BP-ANN model. The largest deviation for activated energy prediction decreases from 12.85% to 6.72%.