Fan blade pyrolysis process parameter optimization method based on pinn and MPC
By constructing a CDAEM-PINN model and combining it with MPC, the problem of neglecting the coupling relationship between frequency factor and activation energy and the damping effect during the pyrolysis of decommissioned wind turbine blades was solved. This enabled precise control of the pyrolysis process and efficient resource recovery, improving the recycling efficiency and performance protection of glass fiber.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NORTH CHINA ELECTRIC POWER UNIV
- Filing Date
- 2026-04-14
- Publication Date
- 2026-07-14
Smart Images
Figure CN122389573A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of wind turbine blade waste treatment technology, and in particular to a method for optimizing parameters of the wind turbine blade pyrolysis process based on PINN and MPC. Background Technology
[0002] With the rapid development of the global wind energy industry, the decommissioning and disposal of wind turbine blades composed of non-degradable glass fiber and carbon fiber reinforced composite materials has become increasingly prominent. In recent years, pyrolysis has gained widespread attention as an effective method for treating decommissioned wind turbine blades due to its high resource recovery and low pollutant emissions. However, existing modeling methods for the pyrolysis process of decommissioned wind turbine blades have significant limitations. They fail to consider the relationship between the frequency factor and activation energy during pyrolysis, nor the influence of damping during pyrolysis. Furthermore, research on modeling the pyrolysis process of decommissioned wind turbine blades using machine learning mainly focuses on data fitting to establish a model of the relationship between system input and output, without addressing models that embed the physical laws of the pyrolysis process into neural networks. Simultaneously, in terms of process control, existing blade pyrolysis treatments primarily rely on fixed process parameters determined based on experience or offline experiments, lacking online closed-loop control strategies that can collect data based on the reaction progress and dynamically optimize process statistics, leading to difficulties in responding to the impacts of raw material supply and environmental disturbances. To address these issues, we have invented a novel parameter optimization method for the pyrolysis process of wind turbine blades based on PINN and MPC, solving the aforementioned technical problems. Summary of the Invention
[0003] The purpose of this invention is to provide a method for optimizing the parameters of the pyrolysis process of wind turbine blades based on PINN and MPC, which can achieve accurate prediction and control of the distribution of pyrolysis products of decommissioned wind turbine blades and effectively improve the efficiency of resource utilization of decommissioned wind turbine blades through pyrolysis recovery.
[0004] The objective of this invention can be achieved through the following technical measures: a method for optimizing the parameters of a wind turbine blade pyrolysis process based on PINN and MPC, which includes:
[0005] Step 1: Construct a complementary distributed activation energy model that integrates the coupling relationship between damping factor, frequency factor, and activation energy;
[0006] Step 2: Determine the objective function and calculate the parameter information of the complementary distributed activation energy model using the particle swarm optimization algorithm;
[0007] Step 3: Use the solid product yield calculated by the complementary distributed activation energy model as the physical constraint of the physical information neural network, train the physical neural network including pyrolysis characteristics, and obtain the prediction result.
[0008] Step 4: Embed the trained physical information neural network containing pyrolysis characteristics into the model predictive control framework to establish a nonlinear model control suitable for the pyrolysis process of decommissioned wind turbine blades.
[0009] Step 5: Based on the prediction results, online rolling optimization and feedback correction are performed using a nonlinear model control to control the recovery and pyrolysis process of the pyrolysis product glass fiber.
[0010] The objective of this invention can also be achieved through the following technical measures:
[0011] In step 1, a damping term for the pyrolysis of wind turbine blades is established. An exponential function is used to characterize the coupling relationship between the frequency factor and the activation energy. A complementary distributed activation energy model suitable for the pyrolysis of decommissioned wind turbine blades is proposed, and the damping term of this complementary distributed activation energy model is defined. It is expressed by the following formula:
[0012] in, The damping factor, Reaction time;
[0013] Assuming the pyrolysis process of decommissioned wind turbine blades involves If there are two parallel reactions, then the coupling relationship between the frequency factor and the activation energy is:
[0014] in, It is a constant. and The first The frequency factor and activation energy of a reaction E0 is the mean value of the activation energy, which is a fitting parameter characterizing the correlation between the frequency factor and the activation energy.
[0015] The formula for the conversion rate α of the complementary distributed activation energy model is:
[0016] in, For heating rate, For temperature, This is the universal gas constant. Let be the activation energy distribution function; A be the frequency factor; T0 be the initial temperature; and E be the activation energy of the reaction. Satisfy variance A Gaussian distribution with standard deviation, the formula is as follows:
[0017] In summary, the conversion rate α of the complementary distributed activation energy model is expressed as follows: (5)
[0018] In step 2, by setting the objective function To evaluate the predictive performance of the complementary distributed activation energy model, it is assumed that the predictive results of the complementary distributed activation energy model are... Given a set of data points, the objective function is expressed as follows:
[0019] in, The model's prediction results, The original dataset obtained from the experiment is used to calculate the parameter information of the complementary distributed activation energy model by using the experimental data to minimize the objective function.
[0020] Step 3 includes:
[0021] Step 31: Use the standard feedforward neural network structure as the forward propagation method of the physical information neural network to train the forward propagation of the physical information neural network;
[0022] Step 32: The solid product yield calculated by the complementary distributed activation energy model is embedded as a physical constraint term into the loss function of the backpropagation of the physical information neural network, and the backpropagation of the physical information neural network including pyrolysis characteristics is trained.
[0023] Step 33: During neural network training, the parameters of the physical information neural network are adaptively adjusted and the loss function containing physical constraints is minimized using the Adam optimizer.
[0024] Step 34: Repeat step 33 until the iteration termination condition is met, then output the prediction results of the CDAEM-PINN model, i.e., the yield of the three-phase product.
[0025] In step 31, the forward propagation of the physical information neural network is a process of mapping data from the input layer to the output layer step by step through linear transformation, ultimately producing a prediction result at the output layer; each hidden layer consists of multiple neurons, and each neuron produces an output. The output of a neuron is:
[0026] in, As weight, For the network layer, For the first The output of each neuron For the first The first in the layer A bias term related to each neuron; Let the three-phase product index of the pyrolysis of decomposition of retired wind turbine blades be... Where 1, 2, and 3 correspond to solid, liquid, and gaseous products, respectively; the predicted yield of the pyrolysis products is:
[0027] in, For the output layer The predicted value of each neuron, This serves as input data for actual pyrolysis experiments. This represents the actual pyrolysis experiment yield of the dataset. For the number of neurons, For the last hidden layer Weights associated with each neuron This is the corresponding output value of the neuron. This is a bias term.
[0028] In step 32, the backpropagation of the physical information neural network involves continuously adjusting the weights and biases of each neuron to minimize the error between the model's predicted and actual values, thereby optimizing the performance of the physical information neural network. The weights of each neuron are represented by the following parameters:
[0029] in, This represents the network parameters after iterative updates, which can be either weights or biases. The exponential shift direction of the gradient, i.e., the first moment estimate, is equivalent to the momentum mechanism and is used to smooth the update direction of parameters. The exponential moving precision of the squared gradient, also known as the second moment estimate, is used to adaptively and dynamically adjust the learning rate for each parameter based on the historical gradient magnitude. This represents the global learning rate, which determines the basic magnitude of parameter updates. It is a very small, smooth spot designed to prevent instability in the denominator and ensure the stability of numerical calculations; Indicates the first The weight or bias value updated in the next iteration. This represents the actual parameter update amount determined by the adaptive learning rate; the total loss function of the physical information neural network consists of two parts: data loss and physical constraint loss. To quantify the difference between the predicted yield and the experimental yield based on the mean squared error, the calculation formula is as follows:
[0030] in, It is the number of samples. It is the first The first sample The true yield of the phase, The first one, predicted from the forward propagation The first sample Predicted yield of phases; physical constraint loss This refers to the cumulative loss derived from the physical equations applied to the three-phase products of the pyrolysis process; the physical constraint formula is:
[0031] in The first is calculated based on the complementary distributed activation energy model. Conversion rate of individual data yes The corresponding yield of solid products in the pyrolysis products of decomposed wind turbine blades. The solid product yield predicted solely through forward propagation of the physical information neural network (CDAEM-PINN) does not consider the influence of the complementary distributed activation energy model. By establishing a weighted coupling between data loss and physical constraint loss using adjustable weight parameters, the total loss function of the CDAEM-PINN model, which incorporates pyrolysis characteristics, is expressed as follows:
[0032] in and These are the weighting coefficients for data loss and physical constraint loss, respectively. Both are adjustable weights and their sum is 1. The final prediction result obtained based on the CDAEM-PINN model, The forward mapping function of the physical information neural network that has been trained through backpropagation and reached a convergent state; Input the actual pyrolysis experimental data for the CDAEM-PINN model; The dataset contains the actual pyrolysis yield data corresponding to the output. To optimize the performance of the CDAEM-PINN model, the network parameters are continuously iterated and updated using the backpropagation algorithm, based on the idea of minimizing the loss function, until the total loss function tends to stabilize and converge. The final prediction result of the constructed CDAEM-PINN model is then obtained. It is expressed as follows:
[0033] This serves as input data for actual pyrolysis experiments. This represents the actual pyrolysis experimental yield of the dataset.
[0034] In step 4, the trained physical information neural network containing pyrolysis characteristics, namely the CDAEM-PINN model, is embedded into the model predictive control framework as the internal predictive model of the model predictive control framework. With the goal of maximizing product yield prediction and optimizing product performance, control increment constraints are introduced to smooth the heating action. Constraints containing the pyrolysis dynamics of retired wind turbine blades are set to establish a nonlinear model control suitable for the pyrolysis process of retired wind turbine blades.
[0035] Step 4 includes:
[0036] Step 41: Construct a model predictive control optimization model, and based on the pyrolysis process of decommissioned wind turbine blades, identify the optimization variables; pyrolysis temperature. Heating rate Both the heating rate and reaction time t have a significant impact on the product yield, with the heating rate being the most important factor. It is the main controllable input variable, while pyrolysis temperature Let the reaction time t be a state variable that varies with the control; The optimization variables at time step are:
[0037] Where M represents the control time domain, The control increment representing the heating rate, This represents the sequence of control increments within the control time domain M;
[0038] The objective function focuses on maximizing product yield, minimizing energy consumption, and optimizing product performance, thus constructing a multi-objective optimization model. Simultaneously, the physical constraints of the complementary distributed activation energy model are embedded to ensure that the optimization results conform to the kinetics of the pyrolysis reaction. Let the objective function be:
[0039] in, To predict the yield of the product, The target quantity for the theoretical maximum yield. To predict the cumulative heating energy consumption of the system in the time domain, Represents the heat multiplication function of glass fiber. The norm of the heating rate increment is adjusted to prevent drastic fluctuations in heating power. , and These are the weighting coefficients for each optimization objective; Based on the following empirical formula:
[0040] Where s is the system coefficient; This is the current pyrolysis control sampling time; The prediction time domain length of MPC, For time integration variables, This is used to represent the energy loss at this moment; the constraints include two parts: the parameter range and the complementary distributed activation energy model.
[0041] Step 42: Perform nonlinear rolling time-domain optimization solution using a sequential quadratic programming algorithm for online iterative solution. First, at the sampling time... At that time, obtain the current online state estimate of the system. This includes reaction conversion rate, reaction temperature, and heating rate at the previous moment. Next, using the trained CDAEM-PINN model as the kernel of model predictive control, the future state trajectory is iteratively calculated in the prediction time domain based on all current real control sequences. and temperature trajectory .
[0042] In step 5, online state estimates of reaction temperature and reaction conversion rate are obtained at each sampling time. The actual physical state of the pyrolysis furnace at this time is continuously read, and the actual physical state is used as the start of the next prediction of the CDAEM-PINN model to achieve closed-loop feedback correction.
[0043] The objective of this invention can also be achieved through the following technical measures: a wind turbine blade pyrolysis process parameter optimization system based on PINN and MPC, which uses a PINN and MPC-based wind turbine blade pyrolysis process parameter optimization method to control the recovery and pyrolysis process of the pyrolysis product glass fiber.
[0044] This invention relates to a parameter optimization method for the pyrolysis process of wind turbine blades based on PINN and MPC. Specifically, it involves a Physical Information Neural Network (PINN) incorporating a Complementary Distributed Activation Energy Model (CDAEM) and Model Predictive Control (MPC) for optimizing the parameters of a decommissioned wind turbine blade pyrolysis control system. Considering the damping phenomenon occurring during blade pyrolysis, a CDAEM that considers the pyrolysis characteristics of decommissioned wind turbine blades is constructed by fusing the complementary relationship between the frequency factor and activation energy and the damping factor. This CDAEM is then embedded into the PINN as a physical constraint. The network is trained using the minimum sum of data fitting loss and physical constraint loss as the objective function, achieving accurate prediction of the product distribution during the pyrolysis process of decommissioned wind turbine blades. Simultaneously, the trained CDAEM-PINN model is used as the kernel of the MPC to construct an online rolling optimization strategy. Parameters are dynamically adjusted online according to the reaction process, achieving accurate prediction and optimized control of the pyrolysis process of decommissioned wind turbine blades. Compared with existing technologies, this invention has the following technical advantages: First, it constructs a CDAEM that integrates the damping factor and the frequency factor-activation energy coupling relationship, accurately describing the energy dissipation and kinetic characteristics during the pyrolysis reaction process. Secondly, this model is embedded into PINN as a physical constraint, achieving a jointly driven effect of physics and data. Then, the model parameters are determined through the network's total loss function. Finally, a multi-objective optimization model is constructed, encompassing maximizing glass fiber yield, minimizing energy consumption, and protecting glass fiber performance. Based on the rolling optimization principle of MPC, a set of nonlinear optimal heating trajectories is obtained. Compared with existing machine learning methods, the proposed CDAEM-PINN combined with MPC not only improves the prediction accuracy of pyrolysis products from decommissioned wind turbine blades but also provides an effective heating solution for the problem of glass fiber performance damage caused by heating temperature in traditional fixed-heating processes, thus optimizing the parameters of the pyrolysis control system for decommissioned wind turbine blades. Attached Figure Description
[0045] Figure 1 This is a process flow diagram of the pyrolysis of decommissioned wind turbine blades in this invention;
[0046] Figure 2 This is a diagram of the CDAEM-PINN framework in this invention;
[0047] Figure 3 This is a correlation analysis diagram of the frequency factor and activation energy obtained based on thermogravimetric data in this invention;
[0048] Figure 4 These are conversion rate and temperature distribution diagrams of various activation energy models under different heating rates in the embodiments of the present invention;
[0049] Figure 5 This is a scatter plot of carbon yield predictions for each method in the embodiments of the present invention;
[0050] Figure 6 This is a scatter plot of tar yield predictions for each method in the embodiments of the present invention;
[0051] Figure 7 This is a scatter plot of the syngas yield prediction for each method in the embodiments of the present invention;
[0052] Figure 8 This is a graph showing the results of temperature gauges and solid product yields in an MPC system incorporating machine learning methods, as described in this embodiment of the invention.
[0053] Figure 9 This is a comparison chart of the heating strategies and glass fiber yield of the MPC optimization strategy and other control methods in the embodiments of the present invention;
[0054] Figure 10 This is a flowchart of the wind turbine blade pyrolysis process parameter optimization method based on PINN and MPC according to the present invention. Detailed Implementation
[0055] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.
[0056] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments of the present invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, and / or combinations thereof.
[0057] like Figure 10 As shown, Figure 10 This is a flowchart of the wind turbine blade pyrolysis process parameter optimization method based on PINN and MPC according to the present invention. The method includes:
[0058] Step 1: Construct a complementary distributed activation energy model (CDAEM) that integrates the coupling relationship between damping factor, frequency factor and activation energy.
[0059] A damping term for wind turbine blade pyrolysis is established, and an exponential function is used to characterize the coupling relationship between the frequency factor and the activation energy. A complementary distributed activation energy model (CDAEM) suitable for the pyrolysis of decommissioned wind turbine blades is proposed.
[0060] CDAEM comprehensively considers the kinetic mechanism of the pyrolysis reaction of decommissioned wind turbine blades and the thermal damping effect. It constructs the exponential coupling relationship between the damping factor and the frequency factor-activation energy during the pyrolysis process of decommissioned wind turbine blades in the form of formulas. The model's damping term... It is expressed by the following formula:
[0061] in, The damping factor, For the reaction time. Assume the pyrolysis process of the decommissioned wind turbine blades involves... If there are two parallel reactions, the coupling relationship between the frequency factor and the activation energy can be expressed as follows:
[0062] in, It is a constant. and The first The frequency factor and activation energy of a reaction E0 is the mean value of the activation energy, which is a fitting parameter characterizing the correlation between the frequency factor and the activation energy.
[0063] The above ideas are then applied to the Distributed Activation Energy Model (DAEM), overcoming the shortcomings of traditional DAEM in its insufficient description of energy dissipation during the reaction process and its neglect of the intrinsic coupling relationship between the frequency factor and activation energy as independent parameters. The conversion rate α is the proportion of the mass of reacted feedstock to the total initial feedstock mass, and the yield of solid products is obtained by subtracting the total conversion rate (α). The formula for the conversion rate α in DAEM is as follows:
[0064] in, For heating rate, For temperature, This is the universal gas constant. Let A be the activation energy distribution function, A be the frequency factor, T0 be the initial temperature (default is 0), and E be the activation energy of the reaction. Satisfy variance A Gaussian distribution with standard deviation is represented as follows:
[0065] In summary, according to formulas 1-4, the damping ratio is temperature-dependent. Formula 1 is modified by multiplying the damping ratio by the temperature, and the result is obtained using the integration principle. The core idea of DAEM is to divide the pyrolysis process into multiple smaller stages, and the summation of each stage allows for the analysis of the entire pyrolysis process. Therefore, the conversion rate α of the constructed CDAEM is expressed as follows: (5)
[0066] Step 2: Determine the objective function and calculate the parameter information of the complementary distributed activation energy model (CDAEM) using the particle swarm optimization algorithm.
[0067] By setting the objective function To evaluate the prediction performance of CDAEM, assuming that the prediction results of CDAEM are... Given a set of data points, the objective function is expressed as follows:
[0068] in, The model's prediction results, The original dataset was obtained from the experiment. To minimize the objective function, the experimental data was used to calculate the parameters of CDAEM, preparing for the establishment of a neural network model. The experimental data was obtained from pyrolysis experiments. The calculated CDAEM parameters represent glass fiber yield performance indicators.
[0069] Step 3: Use the solid product yield calculated by the complementary distributed activation energy model as the physical constraint of the physical information neural network (PINN) to train the physical information neural network and obtain the prediction results.
[0070] A CDAEM-PINN-based system was constructed to accurately predict the pyrolysis process of decommissioned wind turbine blades. CDAEM was used to calculate the yield of solid products conforming to physical laws, serving as the physical constraint for PINN. PINN, a standard feedforward neural network, was responsible for learning the complex nonlinear mapping relationship between the input data and the three-phase product yields. Deep fusion was performed using a joint loss function that included data fitting loss and physical constraint loss to ensure that the PINN prediction results both fit the experimental data and strictly follow the pyrolysis reaction kinetics, thus achieving high-precision and interpretable prediction of the pyrolysis process of decommissioned wind turbine blades. The specific training steps included:
[0071] Step 31: The standard feedforward neural network structure is used as the forward propagation method for PINN. The training of PINN forward propagation is shown below:
[0072] PINN's forward propagation refers to the process of mapping data from the input layer to the output layer through linear transformations, ultimately producing a prediction result at the output layer. Each hidden layer consists of multiple neurons, and each neuron produces an output. The output of a neuron can be represented as:
[0073] in, As weight, For the network layer, For the first The output of each neuron For the first The first in the layer The bias term is related to each neuron. Let the index of the three-phase products from the pyrolysis of decommissioned wind turbine blades be... Where 1, 2, and 3 correspond to the solid, liquid, and gaseous products, respectively. The yield prediction of the pyrolysis products can be obtained as follows:
[0074] in, For the output layer The predicted value of each neuron, This serves as input data for actual pyrolysis experiments. This represents the actual pyrolysis experiment yield of the dataset. For the number of neurons, For the last hidden layer Weights associated with each neuron This is the corresponding output value of the neuron. This is the bias term (deviation term).
[0075] Step 32: Embed CDAEM as a physical constraint term into the loss function of the backpropagation of the neural network model. The training of PINN backpropagation is shown below:
[0076] PINN's backpropagation continuously adjusts the weights and biases of each neuron to minimize the error between the model's predictions and actual values, thereby optimizing the model's performance. The weights of each neuron are represented by the following parameters:
[0077] in, This represents the network parameters (weights or biases) after iterative updates. It represents the exponential shift direction of the gradient (first-order moment estimation), which is equivalent to the momentum mechanism and is used to smooth the update direction of parameters; The exponential moving precision (second moment estimate) of the squared gradient is used to adaptively and dynamically adjust the learning rate for each parameter based on the historical gradient magnitude. This represents the global learning rate, which determines the basic magnitude of parameter updates. For a very small, smooth spot (usually The purpose of this is to prevent the denominator from becoming unstable in order to ensure the stability of numerical calculations. Indicates the first The weight values after the next iteration. This represents the actual parameter update amount determined by the adaptive learning rate. The total loss function of this model consists of two parts: data loss and physical constraint loss. Data loss To quantify the difference between the predicted yield and the experimental yield based on the mean squared error, the calculation formula is as follows:
[0078] in, It is the number of samples. It is the first The first sample The true yield of the phase, The first one, predicted from the forward propagation The first sample Predicted yield for each phase. Physical constraint loss. This refers to the cumulative loss derived from the physical equations applied to the three-phase products of the pyrolysis process. The physical constraint formula can be expressed as:
[0079] in It is the number calculated based on CDAEM. Conversion rate of individual data yes The corresponding yield of solid products in the pyrolysis products of decomposed wind turbine blades. The solid product yield is predicted solely through PINN forward propagation, without considering CDAEM. By establishing a weighted coupling between data loss and physical constraint loss using adjustable weight parameters, the total loss function of the physical information neural network incorporating pyrolysis characteristics, i.e., the CDAEM-PINN model, is expressed as follows:
[0080] in and These are the weighting coefficients for data loss and physical constraint loss, respectively. Both are adjustable weights and their sum is 1. The final prediction results obtained from CDAEM-PINN (such as solid-liquid-gas three-phase yield). The forward mapping function of the physical information neural network that has been trained through backpropagation and reached a convergent state; Input the actual pyrolysis experimental data of the model (such as reaction temperature, heating rate, conversion rate, and reaction time). The dataset contains the actual pyrolysis yield data corresponding to the output. To optimize model performance, the network parameters are iteratively updated using the backpropagation algorithm, based on the idea of minimizing the loss function, until the total loss function stabilizes and converges. The final prediction results of the constructed CDAEM-PINN are shown below. It is expressed as follows:
[0081] This serves as input data for actual pyrolysis experiments. This represents the actual pyrolysis experimental yield of the dataset.
[0082] Step 33: During neural network training, the Adam optimizer adaptively adjusts the neural network parameters and minimizes the loss function that includes physical constraints. Specifically, this includes optimizing the weights of the neural network. and bias And to minimize the loss function (Equation 12) to approach zero.
[0083] Step 34: Repeat step 33 until the iteration terminates when the objective function approaches zero. Output the network's prediction result, which is the yield of the three-phase product.
[0084] Step 4: Embed the trained physical information neural network containing pyrolysis characteristics into the MPC framework as the internal prediction model of the model predictive control framework; aiming at maximizing product yield prediction and optimizing product performance, introduce control increment constraints to smooth the heating action; set constraints including the pyrolysis dynamics of decommissioned wind turbine blades, thereby establishing a nonlinear model control suitable for the pyrolysis process of decommissioned wind turbine blades. Specifically, this includes:
[0085] S41, MPC optimization model construction:
[0086] Based on the pyrolysis process of decommissioned wind turbine blades, the optimization variables are identified. Pyrolysis temperature ( ), heating rate ( Both the heating rate (t) and reaction time (t) have a significant impact on the product yield, with the heating rate (t) being the most important factor. ) is the main controllable input variable, while pyrolysis temperature ( Let and reaction time (t) be state variables that vary with control. The optimization variables at time step are:
[0087] Where M represents the control time domain, The control increment representing the heating rate, This represents the sequence of control increments within the control time domain M.
[0088] The objective function focuses on maximizing product yield, minimizing energy consumption, and optimizing product performance, thus constructing a multi-objective optimization model. Simultaneously, CDAEM physical constraints are embedded to ensure that the optimization results conform to the kinetics of the pyrolysis reaction. Let the objective function be:
[0089] in, To predict the yield of the product, The target quantity for the theoretical maximum yield. To predict the cumulative heating energy consumption of the system in the time domain, Represents the heat multiplication function of glass fiber. The norm of the heating rate increment is adjusted to prevent drastic fluctuations in heating power. , and These are the weighting coefficients for each optimization objective. Based on the following empirical formula:
[0090] Where s is the system coefficient (usually related to the physical hardware properties of the pyrolysis reactor, such as specific heat capacity, discharge quality, and heat transfer efficiency). This is the current pyrolysis control sampling time; The prediction time domain length of MPC, For time integration variables, This is used to represent the energy loss at this moment. The constraints include two parts: the parameter range and CDAEM. The constraints can be expressed as follows: (17)
[0091] In the formula, This indicates that the system is in the prediction time domain for the th time. The predicted temperature at any given time is kept within a safe pyrolysis temperature range, where 1 < <P;
[0092] Indicates the system's first time in the control time domain Heating rate at time 0 < < , and This indicates the minimum and maximum heating rates that the heating device can provide under the current conditions;
[0093] Is the system in The heating rate control increment at any given time;
[0094] To introduce CDAEM physical constraints;
[0095] It is the yield of the solid product obtained by the CDAEM method at time i.
[0096] The solid product yield prediction part of CDAEM is incorporated as a soft constraint into the objective function to ensure that the optimization results are consistent with physical laws and to prevent bias caused by purely data-driven approaches. This is to ensure the allowable temperature range for the performance of glass fiber.
[0097] Step 42, Nonlinear rolling time-domain optimization solution:
[0098] The pyrolysis of decommissioned wind turbine blades is a typical nonlinear problem. This step uses a sequential quadratic programming algorithm for online iterative solution. First, at the sampling time... At that time, obtain the current online state estimate of the system. This includes reaction conversion rate, reaction temperature, and heating rate at the previous moment. Next, using the trained CDAEM-PINN as the kernel of the model predictive control, the future state trajectory is iteratively calculated in the prediction time domain based on all current real control sequences. and temperature trajectory .
[0099] Search within the constraints to find the objective function. (Equation 15) The minimum optimal control sequence is expressed as follows:
[0100] This refers to the current sampling time. The calculated optimal control increment sequence (the optimal solution output by the optimization algorithm is obtained from this formula). The algorithm dynamically evaluates whether the temperature at each step meets the constraints, thereby minimizing situations that could damage the glass fiber. Finally, based on the rolling optimization principle predicted by the model, the calculated... This yields a set of optimal heating schemes, enabling precise control of the pyrolysis process.
[0101] Step 5: Based on the prediction results obtained from CDAEM-PINN, online rolling optimization and feedback correction are performed using MPC to achieve efficient recovery of the pyrolysis product glass fiber and precise control of the pyrolysis process. Feedback correction essentially introduces the system's actual measured values to correct the model's prediction errors. In step 42, the "online state estimate" of the reaction temperature and conversion rate is obtained at each sampling time. This means that during the closed-loop control process, the system continuously reads the actual physical state of the pyrolysis furnace and uses this actual state as the starting point for the next CDAEM-PINN prediction, thus achieving closed-loop feedback correction.
[0102] This invention constructs a CDAEM (Conditional Design Model) that incorporates the complementary relationship between frequency factor and activation energy, as well as a damping factor. This model is then integrated into PINN (Pulse-Induced Numerical Model) as a loss function to predict the product distribution of rapidly pyrolyzed decomposing wind turbine blades. The established CDAEM-PINN model is embedded into MPC (Multi-Process Control) to optimize the parameters of the pyrolysis control system. This invention comprehensively considers the physical characteristics and data information of the pyrolysis of decomposing wind turbine blades. PINN is used to achieve joint driving force from the pyrolysis mechanism and data, enabling accurate prediction of product distribution. Finally, PINN incorporating pyrolysis characteristics is introduced into MPC to achieve online rolling optimization of the pyrolysis control system parameters, thereby achieving precise control of the pyrolysis process of decomposing wind turbine blades under complex pyrolysis environments.
[0103] In a specific embodiment of the present invention, the decommissioned wind turbine blades used for pyrolysis are mainly composed of a composite material containing 30-40 wt% glass fiber and 60-70 wt% epoxy resin. High-purity nitrogen is used as an inert carrier gas to purge the reaction system, thereby ensuring that the blade pyrolysis process is carried out under anoxic conditions. Key input parameters for the experiment include temperature, ranging from 50°C to 800°C, and heating rates of 5 K / min, 20 K / min, and 40 K / min. The main output parameters are the yields of solid, liquid, and gaseous products. The conversion rate of the experimental data is calculated using the following formula: (19)
[0104] in Indicates the initial mass of the reactants. Indicates temperature The remaining mass below, and Indicates temperature Ash mass was measured. Multiple thermogravimetric and pyrolysis experiments were conducted on decommissioned wind turbine blades, recording data at three different heating rates in 5°C increments. The thermogravimetric dataset used to simulate CDAEM consists of experimental data at these three heating rates, with 150 preprocessed data points for each rate. Similarly, the pyrolysis dataset used for CDAEM-PINN modeling includes 450 experimental data points collected at the same three heating rates. To better demonstrate the reliability of the established model, all datasets were divided into two parts: 75% for training and 25% for testing. The main output parameters of the pyrolysis process are the yields of the solid, liquid, and gas phases. The solid phase yield mainly consists of glass fibers, while the liquid phase yield includes condensates of water and organic liquids. The gas phase yield consists of non-condensable gases such as hydrogen, methane, and carbon monoxide. These input and output parameters are crucial for evaluating the pyrolysis reaction characteristics under different experimental conditions.
[0105] Figure 1This is a process flow diagram of the rapid pyrolysis of retired wind turbine blades in this invention; wherein equipment 1 is a raw material silo, and the outlet of the raw material silo is connected to the pyrolysis reactor 2, whose main function is to provide the pyrolysis reactor 2 with the retired wind turbine blades that have been cut and crushed. The outlet of the pyrolysis reactor 2 is connected to the carbon storage tank 3 and the condenser tower group 4. The gas outlet at the top of the condenser tower group is connected to the gas collection bottle 6, the bottom outlet of the condenser tower group is connected to the oil storage tank 5, the side of the condenser tower group is connected to the dry packed tower 7, the outlet of the dry packed tower 7 is connected to the adsorption tower 8, and the outlet of the adsorption tower 8 is connected to the exhaust fan 9; the bottom condensate collection parts of the condenser tower group 4 and the dry packed tower 8 are both connected to the oil collection tank 5.
[0106] Figure 2 This is a diagram of the physical information neural network framework based on complementary distributed activation energy (PIDN) in this invention. The network consists of two parts: forward propagation and backward propagation. Specifically, in the forward propagation stage, the input layer receives real-time data from the pyrolysis process of decommissioned wind turbine blades, which then flows into the hidden layer. The hidden layer, through the multiplication of weight matrices of numerous neurons and processing with a nonlinear activation function, enables PINN to fully learn the knowledge related to the pyrolysis of decommissioned wind turbine blades. Next, the predicted yield of the three-phase products is obtained through the output layer, and then the total loss function of CDAEM-PINN is calculated. In the backward propagation stage, the goal is to minimize the total loss. The gradient is calculated backward along the network, and the weights and bias parameters of all neurons in the hidden layer are continuously iteratively updated.
[0107] Figure 3 This is a correlation analysis diagram of the frequency factor and activation energy obtained based on thermogravimetric data in this invention; for example... Figure 3 As shown, the frequency factor and activation energy are not independent, but rather exhibit an exponential relationship, consistent with the content mentioned in the complementary activation energy model. The predictive performance of the proposed CDAEM is evaluated by formulating an objective function. Assuming that CDAEM has N prediction results, it can be expressed as follows: (20)
[0108] in This indicates the model's prediction results. This represents the original dataset obtained from the experiment. The root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination are used to express the results. Three performance metrics are used to evaluate the predictive accuracy of the CDAEM-PINN model.
[0109] Figure 4 These are the conversion rates and temperature distributions of various activation energy models at different heating rates in the embodiments of the present invention; such as... Figure 4 As shown, in , and At three different heating rates, compared to the traditional activation energy model, the correlation coefficient of the CDAEM model with damping term proposed in this invention is approximately 0.99. In the low-temperature region of the curve, the addition of the damping term significantly shifts the model downward, making it closer to the experimental data. This indicates that increasing thermal inertial damping can improve the accuracy of the CDAEM model in predicting the pyrolysis process.
[0110] Table 1. Glass fiber yield performance indicators for each method
[0111] Figure 5 This is a scatter plot of the predicted solid product yields for each method in the embodiments of the present invention;
[0112] Table 2. Glass fiber yield performance indicators for each method
[0113] Figure 6 This is a scatter plot of tar yield predictions for each method in the embodiments of the present invention;
[0114] Table 3. Tar yield performance indicators of each method
[0115] Figure 7 This is a scatter plot showing the predicted syngas yield for each method in the embodiments of the present invention.
[0116] Table 4 Syngas Yield Performance Indicators for Each Method
[0117] like Figure 5 , 6 As shown in Figures 7 and 8, compared with existing methods, the predicted values of CDAEM-PINN are closer to the ideal diagonal, indicating that the model performs better in characterizing the product distribution characteristics during the pyrolysis of decommissioned wind turbine blades. CDAEM-PINN demonstrates a stronger ability to capture product distribution characteristics during the pyrolysis of decommissioned wind turbine blades. Furthermore, the statistical indicators in Tables 2, 3, and 4 further confirm that the CDAEM-PINN model performs better in capturing product distribution characteristics compared to existing methods, and its higher indicator values indicate an excellent fit between the model and the actual data.
[0118] In this embodiment, CDAEM is superior to DAME. The value was nearly an order of magnitude lower, indicating that the damping factor significantly enhanced the thermal compensation effect in predicting the pyrolysis kinetics of decommissioned wind turbine blades. Furthermore, compared to existing methods, CDAEM-PINN can more accurately predict the distribution of rapid pyrolysis products in decommissioned wind turbine blades and is superior in capturing the nonlinear kinetic characteristics of decommissioned wind turbine blade pyrolysis. In summary, the proposed CDAEM-PINN reduced RMSE by approximately 54%, MAE by approximately 53%, and all product predictions... All exceeded 90%.
[0119] Figure 8 This is a graph showing the temperature gauge and solid product yield results of the MPC using machine learning methods in this embodiment of the invention; for example... Figure 8 As shown, under the model predictive control strategy, the temperature exhibits a smooth and nonlinear upward trend, while the yield of solid products eventually tends to equilibrium with increasing temperature. This indicates that using the trained CDAEM-PINN as the internal predictive model enables MPC to effectively perform rolling optimization based on the online state, thereby outputting a set of optimal heating curves.
[0120] Figure 9 This is a comparison chart of the heating strategies and glass fiber yield of the method in this embodiment of the invention with other control methods; as shown. Figure 9 As shown, compared with single MPC and fuzzy PID control methods, the PINN-MPC strategy proposed in this invention can quickly reach the required reaction temperature via a more optimized path. Furthermore, in the comparison of glass fiber yield, this strategy can strictly control the reaction temperature within a safe threshold, thereby maximizing yield while effectively protecting the performance of the glass fiber and ensuring its stability under high-temperature conditions.
[0121] Under the constraints of key process parameters such as temperature, heating rate, and reaction time, this invention aims to maximize glass fiber yield while minimizing energy consumption and protecting glass fiber performance. This invention uses a trained CDAEM-PINN model as the internal prediction model for MPC (Multi-Process Control) to analyze the nonlinear dynamic characteristics of the pyrolysis process of decommissioned wind turbine blades. Based on the principle of rolling time-domain optimization, the heating rate is dynamically adjusted at each sampling time. The model ultimately outputs a set of optimized dynamic heating trajectories that change over time. Based on the prediction results of CDAEM-PINN, the temperature change trend during the pyrolysis process of decommissioned wind turbine blades is monitored, and the reaction temperature is controlled within a safe threshold that does not degrade glass fiber performance. Ultimately, this achieves efficient recovery of the pyrolysis system and dynamic and precise control of the pyrolysis process of decommissioned wind turbine blades.
[0122] Finally, it should be noted that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
[0123] Except for the technical features described in the specification, all other technologies are known to those skilled in the art.
Claims
1. A method for optimizing parameters in the pyrolysis process of wind turbine blades based on PINN and MPC, characterized in that, The method for optimizing wind turbine blade pyrolysis process parameters based on PINN and MPC includes: Step 1: Construct a complementary distributed activation energy model that integrates the coupling relationship between damping factor, frequency factor, and activation energy; Step 2: Determine the objective function and calculate the parameter information of the complementary distributed activation energy model using the particle swarm optimization algorithm; Step 3: Use the solid product yield calculated by the complementary distributed activation energy model as the physical constraint of the physical information neural network, train the physical neural network including pyrolysis characteristics, and obtain the prediction result. Step 4: Embed the trained physical information neural network containing pyrolysis characteristics into the model predictive control framework to establish a nonlinear model control suitable for the pyrolysis process of decommissioned wind turbine blades. Step 5: Based on the prediction results, online rolling optimization and feedback correction are performed using a nonlinear model control to control the recovery and pyrolysis process of the pyrolysis product glass fiber.
2. The method for optimizing wind turbine blade pyrolysis process parameters based on PINN and MPC according to claim 1, characterized in that, In step 1, a damping term for the pyrolysis of wind turbine blades is established. An exponential function is used to characterize the coupling relationship between the frequency factor and the activation energy. A complementary distributed activation energy model suitable for the pyrolysis of decommissioned wind turbine blades is proposed, and the damping term of this complementary distributed activation energy model is defined. It is expressed by the following formula: in, The damping factor, Reaction time; Assuming the pyrolysis process of decommissioned wind turbine blades involves If there are two parallel reactions, then the coupling relationship between the frequency factor and the activation energy is: in, It is a constant. and The first The frequency factor and activation energy of a reaction E0 is the mean value of the activation energy, which is a fitting parameter characterizing the correlation between the frequency factor and the activation energy. The formula for the conversion rate α of the complementary distributed activation energy model is: in, For heating rate, For temperature, This is the universal gas constant. Let be the activation energy distribution function; A be the frequency factor; T0 be the initial temperature; and E be the activation energy of the reaction. Satisfy variance A Gaussian distribution with standard deviation, the formula is as follows: In summary, the conversion rate α of the complementary distributed activation energy model is expressed as follows: （5）。 3. The method for optimizing wind turbine blade pyrolysis process parameters based on PINN and MPC according to claim 1, characterized in that, In step 2, by setting the objective function To evaluate the predictive performance of the complementary distributed activation energy model, it is assumed that the predictive results of the complementary distributed activation energy model are... Given a set of data points, the objective function is expressed as follows: in, The model's prediction results, The original dataset obtained from the experiment is used to calculate the parameter information of the complementary distributed activation energy model by using the experimental data to minimize the objective function.
4. The method for optimizing wind turbine blade pyrolysis process parameters based on PINN and MPC according to claim 1, characterized in that, Step 3 includes: Step 31: Use the standard feedforward neural network structure as the forward propagation method of the physical information neural network to train the forward propagation of the physical information neural network; Step 32: The solid product yield calculated by the complementary distributed activation energy model is embedded as a physical constraint term into the loss function of the backpropagation of the physical information neural network, and the backpropagation of the physical information neural network including pyrolysis characteristics is trained. Step 33: During neural network training, the parameters of the physical information neural network are adaptively adjusted and the loss function containing physical constraints is minimized using the Adam optimizer. Step 34: Repeat step 33 until the iteration termination condition is met, then output the prediction results of the CDAEM-PINN model, i.e., the yield of the three-phase product.
5. The method for optimizing wind turbine blade pyrolysis process parameters based on PINN and MPC according to claim 4, characterized in that, In step 31, the forward propagation of the physical information neural network is a process of mapping data from the input layer to the output layer step by step through linear transformation, ultimately producing a prediction result at the output layer; each hidden layer consists of multiple neurons, and each neuron produces an output. The output of a neuron is: in, As weight, For the network layer, For the first The output of each neuron For the first The first in the layer A bias term related to each neuron; Let the three-phase product index of the pyrolysis of decomposition of retired wind turbine blades be... Where 1, 2, and 3 correspond to solid, liquid, and gaseous products, respectively; the predicted yield of the pyrolysis products is: in, For the output layer The predicted value of each neuron, This serves as input data for actual pyrolysis experiments. This represents the actual pyrolysis experiment yield of the dataset. For the number of neurons, For the last hidden layer Weights associated with each neuron This is the corresponding output value of the neuron. This is a bias term.
6. The method for optimizing wind turbine blade pyrolysis process parameters based on PINN and MPC according to claim 5, characterized in that, In step 32, the backpropagation of the physical information neural network involves continuously adjusting the weights and biases of each neuron to minimize the error between the model's predicted and actual values, thereby optimizing the performance of the physical information neural network. The weights of each neuron are represented by the following parameters: in, This represents the network parameters after iterative updates, which can be either weights or biases. The exponential shift direction of the gradient, i.e., the first moment estimate, is equivalent to the momentum mechanism and is used to smooth the update direction of parameters. The exponential moving precision of the squared gradient, also known as the second moment estimate, is used to adaptively and dynamically adjust the learning rate for each parameter based on the historical gradient magnitude. This represents the global learning rate, which determines the basic magnitude of parameter updates. It is a very small, smooth spot designed to prevent instability in the denominator and ensure the stability of numerical calculations; Indicates the first The weight or bias value updated in the next iteration. This represents the actual parameter update amount determined by the adaptive learning rate; the total loss function of the physical information neural network consists of two parts: data loss and physical constraint loss. To quantify the difference between the predicted yield and the experimental yield based on the mean squared error, the calculation formula is as follows: in, It is the number of samples. It is the first The first sample The true yield of the phase, The first one, predicted from the forward propagation The first sample Predicted yield of phases; physical constraint loss This refers to the cumulative loss derived from the physical equations applied to the three-phase products of the pyrolysis process; the physical constraint formula is: in The first is calculated based on the complementary distributed activation energy model. Conversion rate of individual data yes The corresponding yield of solid products in the pyrolysis products of decomposed wind turbine blades. The solid product yield predicted solely through forward propagation of the physical information neural network (CDAEM-PINN) does not consider the influence of the complementary distributed activation energy model. By establishing a weighted coupling between data loss and physical constraint loss using adjustable weight parameters, the total loss function of the CDAEM-PINN model, which incorporates pyrolysis characteristics, is expressed as follows: in and These are the weighting coefficients for data loss and physical constraint loss, respectively. Both are adjustable weights and their sum is 1. The final prediction result obtained based on the CDAEM-PINN model, The forward mapping function of the physical information neural network that has been trained through backpropagation and reached a convergent state; Input the actual pyrolysis experimental data for the CDAEM-PINN model; The dataset contains the actual pyrolysis yield data corresponding to the output. To optimize the performance of the CDAEM-PINN model, the network parameters are continuously iterated and updated using the backpropagation algorithm, based on the idea of minimizing the loss function, until the total loss function tends to stabilize and converge. The final prediction result of the constructed CDAEM-PINN model is then obtained. It is expressed as follows: This serves as input data for actual pyrolysis experiments. This represents the actual pyrolysis experimental yield of the dataset.
7. The method for optimizing wind turbine blade pyrolysis process parameters based on PINN and MPC according to claim 1, characterized in that, In step 4, the trained physical information neural network containing pyrolysis characteristics, namely the CDAEM-PINN model, is embedded into the model predictive control framework as the internal predictive model of the model predictive control framework. With the goal of maximizing product yield prediction and optimizing product performance, control increment constraints are introduced to smooth the heating action. Constraints containing the pyrolysis dynamics of retired wind turbine blades are set to establish a nonlinear model control suitable for the pyrolysis process of retired wind turbine blades.
8. The method for optimizing wind turbine blade pyrolysis process parameters based on PINN and MPC according to claim 1, characterized in that, Step 4 includes: Step 41: Construct a model predictive control optimization model, and based on the pyrolysis process of decommissioned wind turbine blades, identify the optimization variables; pyrolysis temperature. Heating rate Both the heating rate and reaction time t have a significant impact on the product yield, with the heating rate being the most important factor. It is the main controllable input variable, while pyrolysis temperature Let the reaction time t be a state variable that varies with the control; The optimization variables at time step are: Where M represents the control time domain, The control increment representing the heating rate, This represents the sequence of control increments within the control time domain M; The objective function focuses on maximizing product yield, minimizing energy consumption, and optimizing product performance, thus constructing a multi-objective optimization model. Simultaneously, the physical constraints of the complementary distributed activation energy model are embedded to ensure that the optimization results conform to the kinetics of the pyrolysis reaction. Let the objective function be: in, To predict the yield of the product, The target quantity for the theoretical maximum yield. To predict the system's cumulative heating energy consumption in the time domain, Represents the heat multiplication function of glass fiber. The norm of the heating rate increment is adjusted to prevent drastic fluctuations in heating power. , and These are the weighting coefficients for each optimization objective; Based on the following empirical formula: Where s is the system coefficient; This is the current pyrolysis control sampling time; The prediction time domain length of MPC, For time integration variables, This is used to represent the energy loss at this moment; the constraints include two parts: the parameter range and the complementary distributed activation energy model. Step 42: Perform nonlinear rolling time-domain optimization solution using a sequential quadratic programming algorithm for online iterative solution. First, at the sampling time... At that time, obtain the online state estimate of the current system. This includes reaction conversion rate, reaction temperature, and heating rate at the previous moment. Next, using the trained CDAEM-PINN model as the kernel of model predictive control, the future state trajectory is iteratively calculated in the prediction time domain based on all current real control sequences. and temperature trajectory .
9. The method for optimizing wind turbine blade pyrolysis process parameters based on PINN and MPC according to claim 8, characterized in that, In step 5, online state estimates of reaction temperature and reaction conversion rate are obtained at each sampling time. The actual physical state of the pyrolysis furnace at this time is continuously read, and the actual physical state is used as the start of the next prediction of the CDAEM-PINN model to achieve closed-loop feedback correction.
10. A parameter optimization system for the pyrolysis process of wind turbine blades based on PINN and MPC, characterized in that, The PINN and MPC-based wind turbine blade pyrolysis process parameter optimization system uses the PINN and MPC-based wind turbine blade pyrolysis process parameter optimization method described in any one of claims 1-9 to control the recovery and pyrolysis process of the pyrolysis product glass fiber.