A Real-Time Optimization Control Method for Underground Coalbed Methane SOFC Combined Heat and Power System

Abstract

This invention discloses a real-time optimization control method for an underground coalbed methane SOFC combined heat and power system. The method utilizes a distributed fiber optic temperature measurement network and electrochemical impedance spectroscopy to acquire real-time three-dimensional temperature field distribution data and electrode impedance data within the fuel cell stack. Based on this, a dual-channel hybrid neural network is used to fuse temporal and transient characteristics, outputting a quantitative carbon deposition risk index and a sulfur poisoning risk index to characterize the electrode health status. This dynamically corrects the anode kinetics and mass transfer parameters in the digital twin model. With system efficiency, temperature uniformity, and health indicators as optimization objectives, a hierarchical collaborative control architecture is employed for multi-objective rolling optimization to generate optimal operating setpoints. Finally, the fuel pretreatment, fuel cell stack operation, and waste heat recovery processes are coordinated and regulated. This invention also supports incremental model updates and predictive maintenance warnings. This method improves system efficiency, safety, and lifespan, and is applicable to SOFC combined heat and power scenarios using complex fuels such as sulfur-containing coalbed methane.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary fields of energy and power engineering, electrochemistry and intelligent control, and relates to a real-time optimization control method for an underground coalbed methane SOFC combined heat and power system. Background Technology

[0002] Underground coalbed methane, as an important unconventional natural gas resource, is of great significance for coal mine safety, greenhouse gas emission reduction, and clean energy supply. Solid oxide fuel cells (SOFCs), due to their high efficiency, low emissions, and wide fuel adaptability, are considered an ideal technology for distributed combined heat and power (CHP) using coalbed methane. However, coalbed methane has a complex composition, typically containing high concentrations of methane and trace amounts of highly toxic hydrogen sulfide (H2S), posing a serious challenge to the long-term stability of SOFC anode materials. In actual operation, H2S easily leads to irreversible sulfur poisoning of nickel-based anodes, while localized fuel insufficiency or uneven temperature can cause carbon buildup; both significantly accelerate stack performance degradation and may even cause permanent damage.

[0003] Furthermore, SOFC systems inherently possess characteristics such as strong nonlinearity, strong multiphysics coupling, and slow dynamic response. Traditional control strategies based on fixed models and static setpoints struggle to balance system efficiency, safety, and lifespan. Existing technologies sometimes attempt to adjust through offline calibration or simple feedback loops, but lack real-time sensing capabilities of the stack's internal state (such as three-dimensional temperature fields and electrode health), making precise feedforward-feedback coordinated control impossible. Simultaneously, the lack of online quantitative assessment of electrode degradation processes means maintenance largely relies on post-incident repairs, hindering predictive maintenance.

[0004] Therefore, there is a need for an intelligent control method that can deeply integrate mechanistic models and data-driven approaches, and possess the capabilities of real-time internal state perception, online model correction, multi-objective rolling optimization, and early fault warning, in order to ensure the efficient, safe, and long-life operation of underground coalbed methane SOFC cogeneration systems under complex operating conditions. Summary of the Invention

[0005] To address the problems existing in the background technology, this invention proposes a real-time optimization control method for underground coalbed methane SOFC combined heat and power systems.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A real-time optimization control method for an underground coalbed methane SOFC combined heat and power system, comprising the following steps: Establish a fully physical field coupled digital twin model for simulating SOFC combined heat and power systems; Real-time acquisition of direct measurement data reflecting the internal state of SOFC stacks, including at least three-dimensional temperature field distribution data and electrode electrochemical impedance spectroscopy data; Based on the direct measurement data and real-time system operating parameters, a machine learning algorithm is used to perform fusion calculations and output an evaluation index characterizing the current health status of the stack electrodes. Based on the electrode health status assessment index, the preset parameters in the all-physics field coupled digital twin model are dynamically corrected; Based on the modified digital twin model, multi-objective rolling optimization is performed with at least one of the following as optimization objectives: overall system efficiency, internal temperature uniformity of the fuel cell stack, and electrode health status assessment index, to obtain a set of optimal operating settings. The optimal operating setpoints are converted into control commands to coordinate and regulate the fuel pretreatment process, the fuel cell stack operation process, and the waste heat recovery process in the combined heat and power system.

[0007] Compared with the prior art, the present invention has the following advantages: First, by synchronously acquiring distributed optical fiber temperature measurement and electrochemical impedance spectroscopy, high-precision, non-invasive direct sensing of the three-dimensional temperature field inside the fuel cell stack and the state of the electrode interface can be achieved.

[0008] Secondly, a dual-channel hybrid neural network model integrating time-series dynamic features and transient operating parameters is constructed, which can quantify the carbon deposition risk index and sulfur poisoning risk index online, accurately characterize the electrode health status, and provide a reliable basis for predictive maintenance.

[0009] Furthermore, the health assessment results are used to dynamically adjust the key anode parameters of the digital twin model, improving the model's ability to track and predict stack aging behavior. Based on this, a hierarchical collaborative control architecture is adopted, combining fast feedforward response and multi-objective rolling optimization, to achieve coordinated regulation of the entire process of fuel pretreatment, stack operation, and waste heat recovery. This improves overall system efficiency and temperature uniformity while ensuring stack safety.

[0010] Finally, through periodic incremental training and long-term trend prediction, the system has continuous adaptive capabilities, which can trigger protective measures in advance under disturbance conditions such as sudden changes in H2S concentration, and issue maintenance warnings before failure occurs, thereby extending the service life of the fuel cell stack, reducing operation and maintenance costs, and ensuring the long-term stable and efficient operation of the underground coalbed methane SOFC cogeneration system. Attached Figure Description

[0011] Figure 1 This is a flowchart of a real-time optimization control method for an underground coalbed methane SOFC combined heat and power system according to the present invention; Figure 2This is an information flow diagram of an underground coalbed methane SOFC combined heat and power system according to the present invention. Detailed Implementation

[0012] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0013] like Figures 1-2 As shown, the technical solution adopted in this invention is as follows: A real-time optimization control method for an underground coalbed methane SOFC combined heat and power system, comprising the following steps: S1. Establish a fully physical field coupled digital twin model for simulating SOFC combined heat and power systems.

[0014] Establishing a fully physical field coupled digital twin model for simulating a solid oxide fuel cell (SOFC) cogeneration system refers to constructing a fully coupled mathematical model on a multiphysics simulation software platform (such as COMSOL Multiphysics) that covers the entire process from the downhole in-situ catalytic purification module, the staged SOFC stack module, the ground-based cascaded heat recovery module to the exhaust gas treatment module.

[0015] This fully physical field coupled digital twin model needs to solve the mass conservation equation, momentum conservation equation, energy conservation equation, charge transport equation and component transport equation simultaneously, and coupled with an electrochemical reaction kinetics sub-model.

[0016] The mass conservation equation is expressed as follows: ; The momentum conservation equation is in the Navier-Stokes form: ; The energy conservation equation is: ; The charge transport equation is: ; The component transport equations adopt the Fick diffusion-convection form: ; In the above equation, Indicates gas density, Represents the velocity vector. Indicates pressure, Indicates dynamic viscosity. This represents the specific heat capacity at constant pressure. Indicates temperature. Indicates the effective thermal conductivity. This represents the volumetric heat source released by an electrochemical reaction. and These represent ionic conductivity and electronic conductivity (unit: S / m), respectively. and These represent ionic potential and electronic potential, respectively. Indicates the first Mass fraction of each component (dimensionless, range [0,1]). Indicates the first Effective diffusion coefficient of each component (unit: m² / s). Indicates the first Net formation rate of each component (unit: kg / (m³·s)).

[0017] The role of this fully physical field coupled digital twin model is to provide high-fidelity prediction capability of the internal state of the system (including temperature field, concentration field, current density distribution, local reaction rate, etc.), and to provide a mechanistic basis for subsequent state observation, model correction and optimized control.

[0018] The spatial discretization of the model adopts the finite element method, and the temporal discretization adopts the implicit backward difference scheme. The number of mesh elements is no less than 50,000 to ensure the resolution of key areas (such as the three-phase interface of electrodes).

[0019] Boundary conditions are set based on the actual system interface, including the inlet fuel gas composition (methane mole fraction). hydrogen sulfide concentration Flow fluctuation range Ambient pressure (101.325 kPa) and cooling water temperature The initial conditions are set as steady-state startup conditions, with an average stack temperature of [temperature value missing]. This fully physical field coupled digital twin model must include a virtual sensor output node and an electrode microstate evolution sub-model, which together constitute the data interface for subsequent state perception and model correction.

[0020] S2. Real-time acquisition of direct measurement data reflecting the internal state of the SOFC stack, wherein the direct measurement data includes at least three-dimensional temperature field distribution data inside the stack and electrode electrochemical impedance spectroscopy data.

[0021] Real-time acquisition of direct measurement data reflecting the internal state of a solid oxide fuel cell stack refers to the continuous acquisition of raw sensing signals that characterize the internal working state of the stack during system operation using a physical sensor array embedded in the SOFC stack structure at a sampling frequency of not less than 1 Hz.

[0022] The direct measurement data includes at least three-dimensional temperature field distribution data inside the stack and electrode electrochemical impedance spectroscopy data.

[0023] The three-dimensional temperature field distribution data inside the fuel cell stack is provided by a distributed fiber optic temperature measurement system or a miniature thermocouple array, with a spatial resolution of 5mm×5mm×5mm, a temperature measurement range of 600°C to 1000°C, and a measurement accuracy of ±1.5°C. This data is represented in three-dimensional tensor form. ,in , , These represent the internal spatial coordinates of the fuel cell stack (unit: m). Time (unit: ), , , These represent the number of discrete sampling points along the three coordinate axes, with values ​​determined based on the fuel cell geometry and sensor density; a typical value is... , , .

[0024] Electrochemical impedance spectroscopy data were obtained by applying a sinusoidal perturbation voltage of 10 mV to an electrochemical workstation, with a frequency scan range of 0.1 Hz to 10 mV. 5 Hz, logarithmically spaced sampling at no less than 50 frequency points, the resulting complex impedance data is expressed as ,in Indicates the first Test frequencies (unit: Hz) , , Represents the real part impedance (unit: Ω). Represents the imaginary part of the impedance (unit: Ω). It is the imaginary unit.

[0025] The acquisition cycle of electrode electrochemical impedance spectroscopy data is no more than 300 seconds to ensure the capture of dynamic deterioration processes such as electrode aging, carbon deposition, or sulfur poisoning.

[0026] The two types of direct measurement data mentioned above are transmitted in real time to the digital twin model operation platform via industrial Ethernet for subsequent state correction and parameter identification.

[0027] Specifically, the real-time acquisition of three-dimensional temperature field distribution data inside the fuel cell stack includes: continuously acquiring temperature profile data along the airflow direction and stacking direction of the fuel cell stack at a preset spatial resolution through a distributed optical fiber temperature sensing network deployed on the surface of the metal connector of the solid oxide fuel cell stack.

[0028] The distributed fiber optic temperature sensing network employs Optical Time Domain Reflectometry (OTDR) technology based on Raman scattering, with a spatial resolution set at 5 mm. This value is determined based on the minimum resolvable distance requirement for high-temperature industrial temperature sensing fibers. Temperature profile data is generated along the airflow direction of the fuel cell stack (defined as...). (axial direction) and stacking direction (defined as) (Axial direction) synchronous acquisition, Axial direction covers the entire length of the anode flow channel of a single cell Number of sampling points , This refers to spatial resolution. Axial coverage of the total height of the fuel cell stack The corresponding stacking layer consists of 30 single cells, and the number of sampling points is [number missing]. , .

[0029] The collected temperature profile data is represented as a two-dimensional matrix. ,in Indicates the first The coordinates of the airflow direction (unit: ), , Indicates the first The stacking direction position coordinates (unit: ), , Time (unit: ).

[0030] The temperature measurement range of this distributed fiber optic temperature sensing network is: to Temperature measurement accuracy is With a response time of less than 2 seconds, it meets the performance requirements for online monitoring of high-temperature energy equipment.

[0031] The sensing optical fiber is fixed within a pre-machined micro-groove on the surface of the metal connector, or attached to the surface of the metal connector via a high-temperature resistant thermally conductive adhesive layer to ensure thermal contact. The metal connector is made of Crofer 22APU alloy and has an operating temperature range of [temperature range missing]. to .

[0032] The micro guide groove has a cross-sectional dimension of 1.0 mm wide and 0.8 mm deep, and is continuously opened along the airflow direction. The bottom of the groove and the metal connector body form a continuous heat conduction path. The machining tolerance of the guide groove meets the medium precision level.

[0033] If a high-temperature resistant thermally conductive adhesive layer is used, then this high-temperature resistant thermally conductive adhesive layer is composed of Composed of a filled siloxane-based ceramic precursor, the thermal conductivity after curing is not less than [value missing]. The upper limit of long-term use temperature is Shear bond strength ≥2.0MPa.

[0034] The sensing fiber is a polyimide-coated quartz fiber with an outer diameter of [missing information]. The coating has a high temperature resistance. When placed inside a micro guide groove, the gap between it and the surface of the metal connector is less than [the required value]. .

[0035] When using a high-temperature resistant thermally conductive adhesive layer for bonding, the adhesive layer thickness should be controlled within... To minimize thermal resistance. Thermal contact quality is determined by interfacial thermal resistance. Quantization is defined as: ,in This indicates the actual surface temperature of the metal connector (unit: K). The sensor fiber optic temperature is represented in Kelvin (K), and q represents the heat flux density per unit area (W / m²). Requirements during engineering implementation... If this condition is not met, the thermal contact is deemed to have failed, and the sensing fiber optic cable must be reinstalled.

[0036] Specifically, the real-time acquisition of electrode electrochemical impedance spectroscopy data includes: continuously calculating the rate of change of total system current and total fuel flow rate within the most recent time window. Wherein, the total system current... This represents the DC current (in A) measured at the SOFC stack output terminal, acquired by a Hall effect current sensor at a sampling frequency of 1 kHz. Total fuel flow rate. This represents the mass flow rate (kg / s) of the reforming gas supplied to the anode of the fuel cell stack, collected by a thermal mass flow meter at a sampling frequency of 100 Hz. (Most recent time window) The value is set to 60 seconds, which is determined based on the minimum observation period for judging steady-state operation.

[0037] Total system current rate of change Defined as: ; Total fuel flow rate of change Defined as: ; The above integration is implemented in real time on a digital signal processor (DSP) using the trapezoidal numerical integration method, with a time step of 1 ms. Rate of change and rate of change Used for subsequent steady-state determination.

[0038] When the rates of change are all below their respective first steady-state determination thresholds, the determination system enters a quasi-steady-state window that can be used for measurement. The first steady-state determination threshold is related to the total system current. rate of change The value is set to 0.5 A / s, which is determined based on typical load fluctuation test data of SOFC stacks at rated power of 5 kW, to ensure that the electrochemical reaction interface is in a local equilibrium state.

[0039] The first steady-state determination threshold is based on the total fuel flow rate. rate of change The value was set at 0.001 kg / (s²), which was determined based on the requirements for fuel supply stability and the experimental results of the critical flow rate gradient for SOFC anode carbon deposition.

[0040] The decision logic is a Boolean condition: if the rate of change And the rate of change If the system reaches a certain threshold, it is determined that it has entered a quasi-steady-state window suitable for measurement; otherwise, monitoring continues. The start time of the quasi-steady-state window is denoted as [start time of the window]. The duration should be no less than 300 seconds to meet the timing requirements of subsequent multi-frequency disturbance injection and data acquisition.

[0041] Within the quasi-steady-state window, a small-amplitude, multi-frequency AC current disturbance signal of a preset waveform is injected into the fuel cell stack by the excitation source; wherein, the excitation source is a programmable bidirectional electronic load with current disturbance injection function, and its bandwidth is not less than 100kHz; the small-amplitude, multi-frequency AC current disturbance signal of the preset waveform It consists of the superposition of K equally spaced logarithmic sine components, and its expression is: ; in, Indicates the first The current amplitude (in A) of each frequency component is uniformly set to 0.1A. This value satisfies the small amplitude condition, that is, it does not exceed the total system current. 2% to avoid nonlinear response; Indicates the first The disturbance frequencies (in Hz) range from [number] to [number]. , The frequency points are distributed at logarithmic intervals, that is ; in, Indicates the first The initial phase (in rad) of each frequency component is randomly generated. The interval is designed to reduce harmonic interference; this multi-frequency alternating current disturbance signal The current is injected into the SOFC stack anode-cathode circuit through the current control loop of the excitation source, with the injection start time being... This is to avoid the initial transients of the window.

[0042] The voltage response signals across the fuel cell stack are synchronously and rapidly acquired, and the complex impedance spectrum corresponding to the multi-frequency current disturbance is calculated using frequency domain analysis. The voltage response signals across the fuel cell stack are... Data was acquired using a differential voltage probe at a sampling frequency of 1MHz, with a vertical resolution of 16 bits, and a total acquisition time of [duration missing]. ( The lowest disturbance frequency is used to ensure sufficient cycle count at the lowest frequency point; the acquired data is stored in high-speed memory after passing through an anti-aliasing filter (cutoff frequency 50kHz); the frequency domain analysis method uses Discrete Fourier Transform (DFT) to analyze the current disturbance signal. and voltage response signal Calculate its value at each disturbance frequency. The complex spectral components at that location.

[0043] Complex impedance spectrum Solve using the following formula: ;in, Represents the real part impedance (unit: Ω). The complex impedance spectrum represents the imaginary part of the impedance (unit: Ω). As output as electrode electrochemical impedance spectroscopy data, it is used to characterize the electrode interface charge transfer resistance, gas diffusion resistance, and electrolyte ion conduction characteristics.

[0044] Specifically, the frequency domain analysis method is either Fast Fourier Transform (FFT) or lock-in amplification (LSA). The Fast Fourier Transform (FFT) is an efficient implementation of the Discrete Fourier Transform based on the Cooley-Tukey algorithm, used to convert a time-domain signal into a frequency-domain complex spectrum. Its input is the voltage response signal across the fuel cell stack. and multi-frequency alternating current disturbance signals The equally spaced sampling sequence is output as each perturbation frequency. Complex voltage components at the point With complex current components .

[0045] Lock-in amplification is a narrowband filtering method based on phase-sensitive detection of a reference signal, which uses the voltage response signal... Orthogonal reference signals of the same frequency and Multiply and low-pass filter to extract each frequency. The real and imaginary response components are then combined to form a complex impedance.

[0046] Both methods are implemented in firmware on a digital signal processor (DSP) or field-programmable gate array (FPGA); if a fast Fourier transform is used, a sampling length of... It is an integer power of 2 and satisfies To ensure frequency resolution ( Sampling frequency (unit: Hz)

[0047] If lock-in amplification technology is used, then each frequency A separate digital phase-locked loop needs to be configured, with an integral time constant. To suppress noise, the system automatically selects one method based on the frequency density of the disturbance signal and the real-time requirements; the two methods cannot be used simultaneously.

[0048] After obtaining the complex impedance spectrum, the process further includes a data validity verification step: calculating the coherence function or total harmonic distortion of the impedance spectrum; if its value is lower than the second steady-state determination threshold, the measurement data is deemed valid and adopted. The coherence function... Defined as: ;

[0049] in, Indicates voltage response signal With current disturbance signal In frequency Cross-power spectral density at (unit: ), Indicates voltage response signal Self-power spectral density (unit: ), Indicates current disturbance signal The self-power spectral density (unit: A²); coherence function The range of values ​​is The closer the value is to 1, the stronger the linear correlation.

[0050] Total Harmonic Distortion (THD) is defined as the ratio of the effective values ​​of all harmonic components except the fundamental frequency to the effective value of the fundamental frequency, for each disturbance frequency. In calculating the voltage response , The expression for isoharmonic components is: ; in, To consider the highest harmonic order, take , For frequency Voltage amplitude at (unit: The second steady-state determination threshold is for the coherence function. The value is set to 0.95, which is determined based on the minimum requirements for the reliability of power quality measurement signals; the second steady-state judgment threshold is set to 5% for total harmonic distortion (THD), which is determined based on experimental data of the nonlinear response of SOFC electrodes under small disturbances.

[0051] The Boolean logic for data validity verification is: if a coherence function is used... As a criterion, then when When the frequency point is determined to be valid, the data is considered valid; if the total harmonic distortion (THD) is used as the criterion, then when When this happens, the data at that frequency point is deemed valid; the system defaults to prioritizing the use of coherent functions. Verification was conducted only when the signal-to-noise ratio was too low. If reliable calculation is not possible, switch to Total Harmonic Distortion (THD). If all K frequency points meet the corresponding criteria, the measurement data is deemed valid and adopted; otherwise, the measurement results are discarded, and the measurement process is restarted in the quasi-steady-state window monitoring stage.

[0052] S3. Based on the direct measurement data and the real-time operating parameters of the system, a machine learning algorithm is used to perform fusion calculations and output an evaluation index characterizing the current health status of the stack electrodes.

[0053] Specifically, the fusion calculation through machine learning algorithms includes: constructing a hybrid neural network model with a dual-channel input structure.

[0054] This hybrid neural network model is a customized deep learning architecture deployed on an embedded AI acceleration chip (such as NVIDIA Jetson AGX Orin) and implemented as firmware. Its overall structure includes two independent input channels (a first channel and a second channel), a feature concatenation layer, and a fully connected neural network layer with a shared output. All weight parameters of the model were determined through supervised learning training offline. The training dataset was generated based on accelerated aging experiments involving carbon deposition and sulfur poisoning, and the labels included a quantified carbon deposition risk index. And quantitative sulfur poisoning risk index Where 0 represents no risk and 1 represents severe failure; the input of this hybrid neural network model is strictly limited to direct measurement data and real-time system operating parameters, without introducing external prior information.

[0055] The first channel is a time-series processing channel, configured to input the historical sequence of the three-dimensional temperature field distribution data inside the fuel cell stack into a long short-term memory network layer to extract the implicit temporal dynamic features. The three-dimensional temperature field distribution data inside the fuel cell stack... It has been defined as a three-dimensional tensor with spatial dimensions of . , , Historical sequence length The corresponding time window is 600 seconds (sampling period 10 seconds), forming a four-dimensional input tensor. The tensor is first converted into a two-dimensional matrix through a flattening operation. ,in This represents the temperature feature dimension for a single frame. Subsequently... The data is fed into a single-layer Long Short-Term Memory (LSTM) network layer containing 128 memory units. The hidden state update formula is as follows:

[0056] ; in, For the first The flattened temperature vector at time t, In hidden state, In cellular state, For the Sigmoid function, For element-wise multiplication, and The LSTM layer outputs the trainable weight matrix and bias vector, and the final hidden state. These are the implicit temporal dynamic features used to characterize the evolution trend of the temperature field (such as local hot spot migration, thermal gradient accumulation, etc.), which are closely related to the electrode carbon deposition process.

[0057] The second channel is a transient feature channel, configured to input the characteristic frequency impedance values ​​from the current electrode electrochemical impedance spectroscopy data, along with the real-time operating parameters of the system, into a fully connected neural network layer for feature mapping and fusion. The electrode electrochemical impedance spectroscopy data... ,Include Each frequency point.

[0058] The characteristic frequency impedance value is selected from the real part of the complex impedance in three key frequency bands: low frequency. Corresponding gas diffusion resistance (Unit: Ω), Intermediate Frequency Corresponding charge transfer impedance (Unit: Ω), High frequency Corresponding ohmic impedance (Unit: Ω); The above three impedance values ​​are related to the real-time operating parameters of the system (including the total system current). Total fuel flow Oxygen concentration in anode tail gas Cathode air flow Average operating temperature of fuel cell stack Together they constitute the transient feature vector. .

[0059] The vector The input is fed into two fully connected neural network layers: the first layer contains 64 neurons (ReLU activation), and the second layer contains 128 neurons (ReLU activation). The output is a fused feature vector. The role of this fully connected neural network layer is to map the physically heterogeneous transient parameters to a unified semantic space, thereby enhancing the coupling characterization ability of sulfur poisoning (related to low-frequency impedance and H2S concentration) and operating conditions (such as fuel utilization rate).

[0060] The time-series dynamic features are concatenated with the fused features output from the second channel and input into a subsequent fully connected neural network layer for calculation. The final output is the electrode health status assessment index, which includes a quantified carbon deposition risk index and a quantified sulfur poisoning risk index.

[0061] Wherein, the time-series dynamic features are The fusion feature of the second channel output is The two are concatenated along the feature dimension to form a joint feature vector. The joint feature vector The signal is passed through three fully connected neural network layers: the first layer contains 256 neurons (ReLU activation), the second layer contains 128 neurons (ReLU activation), and the third layer contains 2 neurons (no activation function). The output of the third layer is the final evaluation metric, while the output of the first neuron is the quantitative carbon deposition risk index. The output of the second neuron is a quantitative sulfur poisoning risk index. These two indicators together constitute the electrode health status assessment index.

[0062] The hybrid neural network model is trained through supervised learning, and the sample labels in the training dataset are known true values ​​of carbon deposition risk. True value of sulfur poisoning risk The model training uses a multi-task loss function:

[0063] ; in The task weights are defined by MSE, which stands for mean squared error. If any input sub-item is invalid (e.g., the impedance spectrum fails the data validity verification), the inference process terminates and no evaluation index is output.

[0064] S4. Based on the electrode health status assessment index, dynamically correct the preset parameters in the all-physical field coupled digital twin model.

[0065] Specifically, the dynamic correction of preset parameters in the digital twin model involves establishing a mapping relationship between the electrode health status assessment index and the anode kinetic parameters and mass transfer parameters in the digital twin model. The electrode health status assessment index includes a quantified carbon deposition risk index. And quantitative sulfur poisoning risk index Both have been defined as dimensionless scalars taking values ​​in the interval [0,1]; in the digital twin model, the anode dynamics parameter refers to the anode exchange current density. The unit is A / m², which characterizes the electrochemical reaction activity of the electrode. In the digital twin model, the mass transfer parameter refers to the effective gas diffusion coefficient of the anode. The unit is m² / s, representing the mass transfer capacity of fuel gas in the porous anode medium; the mapping relationship is achieved through two independent monotonically decreasing functions, namely:

[0066] ; ; in, Anode exchange current density The nominal initial value, The effective gas diffusion coefficient of the anode The nominal initial value is determined based on experimental data of typical SOFC Ni-YSZ anode materials under H2 fuel conditions at 800°C. Coefficient , , , The weighting factor is obtained by least-squares fitting of impedance change and performance degradation data from 50 sets of accelerated aging experiments, and satisfies the following conditions: , This ensures that the corrected parameters are non-negative; the mapping relationship is embedded in the digital twin model solver in the form of a lookup table or analytic function and is called each time the parameters are updated.

[0067] Based on the mapping relationship, the quantified carbon deposition risk index at the current moment is used. and the quantitative sulfur poisoning risk index The anode exchange current density in the digital twin model is adjusted in real time. and the effective gas diffusion coefficient of the anode The value. The adjustment operation is triggered after each valid evaluation metric is output, with a time interval of 10 seconds;

[0068] The specific execution process is as follows: First, read the quantitative carbon deposition risk index at the current moment. And quantitative sulfur poisoning risk index Then, substituting the above mapping formula, the updated anode exchange current density is calculated. and the effective gas diffusion coefficient of the anode If the calculation result satisfies and If the new value is correct, it will be written into the material property database of the digital twin model and used for the full physics coupling solution in the next time step; otherwise, it will be judged as an abnormal state, the previous valid parameter value will remain unchanged, and an abnormal health state alarm signal will be triggered. This adjustment mechanism ensures that the electrochemical and proton transfer models inside the digital twin model always reflect the true degradation state of the electrode, thereby improving the prediction accuracy of key state variables such as temperature field and current density distribution.

[0069] S5. Based on the modified digital twin model, multi-objective rolling optimization is performed with at least one of the following as optimization objectives: overall system efficiency, internal temperature uniformity of the fuel cell stack, and electrode health status assessment index, to obtain a set of optimal operating settings.

[0070] The revised digital twin model refers to one that has been based on a quantified carbon deposition risk index. And quantitative sulfur poisoning risk index Dynamically update anode exchange current density and the effective gas diffusion coefficient of the anode A fully physical field coupled model. Overall system efficiency. Defined as the ratio of the net output electrical power of the fuel cell stack to the input power of the lower calorific value fuel, the expression is:

[0071] ; in, Net output power of the fuel cell stack (unit: W). Total fuel flow rate (unit: kg / s). The lower calorific value of the fuel (unit: J / kg) is taken as 5.0 × 10⁻⁶. 7 J / kg (corresponding to typical components of reformed gas); internal temperature uniformity of the fuel cell stack Defined as three-dimensional temperature field distribution data inside the fuel cell stack The reciprocal of the standard deviation normalization index: ; in, The standard deviation of temperature (unit: K). Average temperature (unit: K). As a reference temperature difference threshold, This represents the total number of temperature sampling points. The electrode health status assessment indicators include a quantitative carbon deposition risk index. And quantitative sulfur poisoning risk index Multi-objective rolling optimization employs prediction time domain. Control Time Domain The Model Predictive Control (MPC) framework, in each Solve the following multi-objective optimization problems within a given period:

[0072] ; in, The control vector includes fuel distribution ratio, excess air coefficient, and heat recovery working fluid flow rate; weighting coefficients. And satisfy The operator selects the mode based on the operating mode (e.g., high-efficiency mode). Longevity model ).

[0073] The constraints include: a maximum stack temperature of 850°C and a minimum excess air coefficient. (Preventing anodizing), maximum fuel efficiency (To prevent carbon buildup); the optimization solver employs a Sequential Quadratic Programming (SQP) algorithm, implemented in software on an industrial PC; the output is a set of optimal operating settings. .

[0074] Specifically, the multi-objective rolling optimization is implemented using a hierarchical collaborative control architecture: a fast control layer and a rule-based feedforward controller running in parallel. The feedforward controller is configured to: when the rate of change of hydrogen sulfide concentration in the inlet fuel gas exceeds the mutation threshold, immediately adjust the power setting value of the desulfurization unit and the opening setting value of the fuel supply valve in front of the stack according to a preset ratio.

[0075] Among them, the hydrogen sulfide concentration in the inlet fuel gas Measured by an online sulfur analyzer at a frequency of 1 Hz, the unit is ppm; its rate of change Defined as: ; Calculated using first-order difference approximation: ; The mutation threshold is set at 50 ppm / s, determined based on experimental data of the transient poisoning of H2S by SOFC anodes. Exceeding this value will lead to irreversible performance degradation. The feedforward controller is deployed as firmware within the programmable logic controller (PLC). When this occurs, immediately perform the following actions: adjust the power setpoint of the desulfurization unit. Increase ,in This is the gain coefficient. Simultaneously, the opening setting value of the fuel supply valve at the front end of the fuel cell stack is also set. Decrease ,in The above adjustments are completed within 100ms to ensure protection is implemented before H2S reaches the fuel cell stack; if If so, the feedforward controller output will be zero increment.

[0076] The optimized control layer invokes a multi-objective optimization algorithm at a fixed period. This algorithm uses a modified digital twin model as a predictor to collaboratively optimize fuel allocation ratios, excess air coefficients, and heat recovery working fluid flow rates, generating global optimization setpoints. The fixed period is... This is consistent with the aforementioned rolling optimization cycle.

[0077] Fuel distribution ratio Defined as the proportion of fuel flow entering the main fuel cell stack to the total fuel flow, with a value ranging from [0.7, 0.95]. Excess air coefficient. Defined as the ratio of actual airflow to the airflow required for stoichiometry, with a value range of [1.2, 2.5].

[0078] Heat recovery working fluid flow rate This refers to the mass flow rate of cooling water in a cascade heat recovery system, measured in kg / s, with a value range of [0.05, 0.3].

[0079] The modified digital twin model, acting as a predictor, predicts the future given the current state and candidate control sequences. within , , , .

[0080] The multi-objective optimization algorithm uses the aforementioned SQP solver to output global optimization settings. The optimized control layer runs as a software module on an industrial PC, sharing the same computing node as the digital twin model.

[0081] The feedforward compensation setting value output by the fast control layer is superimposed with the global optimization setting value output by the optimized control layer to generate the final integrated control command sent to each actuator.

[0082] The feedforward compensation setpoint output by the fast control layer includes the power increment of the desulfurization unit. and fuel valve opening increment The global optimization settings output by the optimized control layer include the fuel distribution ratio. Excess air coefficient and heat recovery working fluid flow rate The superposition rules are as follows: the final power setting value of the desulfurization unit is... ,in Depend on Indirectly determined.

[0083] The final opening setting value of the fuel supply valve for the fuel cell stack is The opening degree of the air regulating valve is determined by... The conversion yielded that the heat recovery pump speed was obtained from... The conversion is as follows. All integrated control commands are sent to the corresponding actuators via the Modbus TCP protocol, with an update cycle of 100ms (fast layer) or 60s (optimization layer), whichever is faster. If the feedforward compensation setpoint conflicts with the global optimization setpoint (e.g., ...), the command will be updated accordingly. If the limit is not met, the safety limit will be prioritized and the amplitude limit protection mechanism will be triggered.

[0084] S6. Convert the optimal operating setpoint into control commands to coordinate and regulate the fuel pretreatment process, the fuel cell stack operation process, and the waste heat recovery process in the combined heat and power system.

[0085] The optimal operating settings refer to a set of optimal operating settings obtained through multi-objective rolling optimization. and the feedforward compensation setting value of the fast control layer output ( , ).

[0086] The fuel pretreatment process includes temperature control and fuel gas purification in the desulfurization reactor; the stack operation process includes fuel supply and electrochemical reaction management in each stage; the waste heat recovery process includes working fluid flow regulation in the waste heat boiler or Organic Rankine Cycle (ORC) system; control commands are generated by the central controller in firmware form and sent to each actuator in real time via industrial Ethernet (IEEE 802.3). Coordinated regulation ensures a dynamic balance between fuel utilization, stack safety, and thermal efficiency across the three processes, with the regulation cycle consistent with the overall control command update cycle.

[0087] Specifically, the coordinated regulation of the fuel pretreatment process, the fuel stack operation process, and the waste heat recovery process includes: calculating and issuing power adjustment signals for controlling the temperature of the desulfurization reactor, valve control signals for adjusting the fuel flow rate of each section of the fuel stack, and frequency conversion control signals for adjusting the speed of the waste heat boiler feedwater pump or the organic working fluid circulation pump, based on the optimal operating setpoint.

[0088] Among them, the power control signal is used to control the temperature of the desulfurization reactor. Determined by the final power setpoint of the desulfurization unit, the expression is: The unit is kW, and the value range is [2.0, 10.0] kW. This range is set based on experiments on the heat capacity of the desulfurization reactor and H2S adsorption kinetics. The power adjustment signal is transmitted through... The analog output module sends signals to the thyristor power regulator, with a control accuracy of ±0.1kW. Valve control signals are used to adjust the fuel flow in different sections of the fuel cell stack. The final opening setting of the fuel supply valve at the fuel stack is determined by the percentage, ranging from 30% to 100%. A value below 30% will trigger carbon buildup risk protection. This signal is sent to the electric regulating valve actuator via the Modbus RTU protocol, with a response time ≤2s; it is also used for frequency conversion control of the waste heat boiler feedwater pump or organic working fluid circulation pump. The flow rate of the heat recovery working fluid It is derived from the conversion, and the conversion relationship is as follows:

[0089] ; in, , For the heat recovery working fluid flow boundary, , The frequency limit for the frequency converter; the frequency converter control signal is... The DC voltage output is sent to the frequency converter with a control accuracy of ±0.2Hz. The above three types of control signals are issued synchronously with a time deviation of no more than 100ms to achieve closed-loop coordinated regulation of the entire fuel, electricity, and heat process.

[0090] Specifically, a real-time optimization control method for an underground coalbed methane SOFC combined heat and power system further includes system adaptation and early warning steps: periodically using the latest collected system operation data and direct measurement data to incrementally update and train the network connection weight parameters in the machine learning algorithm.

[0091] The machine learning algorithm refers to a hybrid neural network model with a dual-channel input structure, whose network connection weight parameters include the weight matrix of the Long Short-Term Memory (LSTM) network layer. Weight matrices and bias vectors of fully connected neural network layers; latest collected system operating data includes total system current. Total fuel flow Oxygen concentration in anode tail gas Cathode air flow Average operating temperature of fuel cell stack Direct measurement data includes three-dimensional temperature field distribution data inside the fuel cell stack. Electrochemical impedance spectroscopy data of electrodes The above data is stored in a rolling cache queue with a length of 1000 samples after each valid S2 measurement. Incremental update training adopts an online learning strategy, triggering a fine-tuning every 200 new samples. The training algorithm is Stochastic Gradient Descent with Momentum (SGDM) with a momentum coefficient of 0.9 and a fixed learning rate. The loss function remains the multi-task mean squared error. ;

[0092] Training runs as a background task on an embedded AI acceleration chip, with each fine-tuning session taking no more than 30 seconds. If any item in the new data fails the data validity verification, the sample is removed and not included in the training. This incremental update mechanism ensures that the model continuously adapts to long-term dynamic changes such as electrode aging and sensor drift.

[0093] A long-term trend model for the electrode health status assessment indicators is established. When the indicator value predicted by the trend model will exceed the maintenance warning threshold within a preset time in the future, a predictive maintenance reminder is generated in advance. The electrode health status assessment indicators include a quantitative carbon buildup risk index. And quantitative sulfur poisoning risk index Both output data continuously at 10-second intervals and store it in a historical database; the long-term trend model adopts a first-order autoregressive moving average model. Each indicator is modeled independently, and the expression is: ; in, Indicates the current time The index value, These are the autoregressive coefficients. The moving average coefficient, Zero-mean Gaussian white noise (standard deviation) The parameters were determined using maximum likelihood estimation based on six months of field operation data. Future preset time. The maintenance warning threshold is set at 72 hours (259,200 seconds), a value determined based on typical maintenance cycles and fault development kinetics experiments of SOFC stacks. The threshold is set to quantify the carbon buildup risk index. The value was set at 0.85 to quantify the risk index of sulfur poisoning. The threshold is set to 0.80, based on the performance drop inflection point (voltage decay rate) in accelerated aging tests. This can be deduced by working backwards.

[0094] The prediction logic is a Boolean condition: if based on Model predictions or If the system detects an error, a predictive maintenance reminder will be generated immediately. This reminder will be pushed to the operator's monitoring terminal in the form of a text message and recorded in the system log. The predictive maintenance reminder includes suggested measures: check the activity of the desulfurizer or perform the anodizing regeneration procedure, which will be automatically matched according to the type of the out-of-limit indicator.

[0095] In a specific embodiment, a solid oxide fuel cell combined heat and power system with a rated power of 100kW installed in a coal mining area is used as the application object.

[0096] S1. Establish a fully physical field coupled digital twin model for simulating SOFC combined heat and power systems.

[0097] In the host computer of the control system, based on a multiphysics coupled simulation platform, a digital twin model covering the entire process of fuel purification, fuel cell stack reaction, and waste heat recovery is constructed according to the actual system's geometry, material properties, and process connections. The core of the model simulates a SOFC fuel cell stack module composed of 30 single-cell stacks. The model simultaneously solves the following control equations:

[0098] mass conservation equation: .in, The local density of the gas mixture. This is the gas velocity vector.

[0099] Momentum conservation equation: .in, For pressure, This refers to dynamic viscosity.

[0100] Energy conservation equation .in, For isobaric specific heat capacity, For temperature, For effective thermal conductivity, It serves as a volumetric heat source for electrochemical reactions.

[0101] Charge transport equation: .in, and These are ionic and electronic conductivity, respectively. and These are the ionic and electronic potentials, respectively.

[0102] Component transport equations: .in, For the first Mass fraction of the components Its effective diffusion coefficient, Its net generation rate.

[0103] The model was discretized using the finite element method, with a mesh count exceeding 50,000. The inlet boundary conditions were set as follows: methane mole fraction. hydrogen sulfide concentration It fluctuates between 50ppm and 150ppm. This model provides a high-fidelity mechanistic basis for subsequent state prediction and optimization.

[0104] S2. Real-time acquisition of direct measurement data reflecting the internal state of the SOFC stack.

[0105] On the surface of the metal connector of the SOFC stack, micro-grooves are pre-machined along the airflow direction and embedded with distributed fiber optic temperature sensors to form a temperature measurement network. During system operation, this network continuously acquires data at a sampling frequency of 1Hz, generating a temperature measurement network with dimensions of... Three-dimensional temperature field distribution data It is used to monitor the internal thermal state of the fuel cell stack in real time.

[0106] Meanwhile, the system continuously calculates the total output current. With total fuel mass flow rate Average rate of change within the most recent 60-second time window and .when and At this point, the system is determined to have entered a quasi-steady-state window. Subsequently, the excitation source is controlled to inject a frequency consisting of 50 logarithmically spaced intervals (0.1 Hz to...). Small-amplitude alternating current disturbance signal composed of superimposed sinusoidal components (Hz) Each component amplitude is 0.1A. The voltage response across the fuel cell stack is simultaneously acquired at a frequency of 1MHz. Calculate each frequency point using Fast Fourier Transform. Complex impedance By calculating the coherence function at each frequency point (Threshold 0.95, based on the reliability requirements of harmonic measurement signals) Verify data validity. All frequency points measured in this study. The data is deemed valid.

[0107] S3. Based on the direct measurement data and the real-time operating parameters of the system, the system performs fusion calculations using machine learning algorithms to output an electrode health status assessment index.

[0108] A hybrid neural network model deployed in the system's edge computing unit (an embedded AI processor with at least 10 TOPS INT8 computing power) is invoked. This model has a dual-channel input structure:

[0109] 1. Time-series processing channel: The historical three-dimensional temperature field data sequence of the past 10 minutes (60 time steps in total) is input into the Long Short-Term Memory network layer to extract the time-series dynamic feature vectors characterizing the evolution trend of the temperature field. .

[0110] 2. Transient characteristic channel: The impedance values ​​of the three characteristic frequencies selected at the current moment ( , , ) and system real-time operating parameters , Mole fraction of carbon dioxide in anode tail gas Cathode air volume flow rate Average temperature of fuel cell stack The input vector is composed of these two elements, which are then mapped and fused through a two-layer fully connected neural network to output a transient feature vector. .

[0111] 3. Integration and Output: [This will involve...] and After being concatenated, the final output layer, through subsequent fully connected neural network calculations, provides two scalars: a quantified carbon deposition risk index. And quantitative sulfur poisoning risk index , as an indicator for assessing the health status of electrodes.

[0112] S4. Based on the electrode health status assessment index, dynamically correct the preset parameters in the all-physical field coupled digital twin model.

[0113] The control system is based on the evaluation indicators. and Based on the established mapping relationship, the key parameters of the corrected model are calculated in real time: anode exchange current density: .

[0114] Effective gas diffusion coefficient of the anode: .

[0115] Among them, nominal value , The weighting coefficients were determined based on Ni-YSZ anode performance benchmark data and parameter sensitivity analysis. The calculated parameters meet the following conditions: and Subsequently, it is updated in the material property library of the digital twin model so that it can more accurately reflect the current actual deterioration state of the electrode.

[0116] S5. Based on the modified digital twin model, perform multi-objective rolling optimization to solve for the optimal operating settings.

[0117] The system employs a hierarchical collaborative control architecture for rolling optimization.

[0118] 1. Optimize the control layer: with a fixed period Start. Using the modified digital twin model as the predictor (prediction time domain) Under the condition of meeting the stack temperature (850°C), excess air coefficient Under equal constraints, the sequential quadratic programming algorithm is used to solve the multi-objective optimization problem:

[0119] ; in, This is used as the control vector. This optimization selects a longevity-first approach, and sets weights accordingly. The solution yields a set of globally optimized settings: fuel distribution ratio. Excess air coefficient Mass flow rate of heat recovery working fluid .

[0120] 2. Fast Control Layer: Parallel operation of rule-based feedforward controllers. An online sulfur analyzer monitors the hydrogen sulfide concentration in the inlet fuel gas in real time. When its rate of change is detected At that time, the feedforward controller calculates and outputs the compensation amount within 100ms: the increment of the desulfurization unit power setpoint. Reduction in the setpoint of the fuel supply valve opening at the fuel stack front stage .

[0121] 3. Command Synthesis: The central controller superimposes two layers of commands. Among them, the desulfurization power reference value used for superposition... The output value is taken from the previous optimization cycle (a preset safety value is used when the system starts). After synthesis, the final integrated control command is obtained.

[0122] S6. Convert the optimal operating setpoint into control commands to coordinate and regulate the combined heat and power system.

[0123] Based on the integrated control commands, the central controller synchronously generates and sends out three types of control signals: 1. Desulfurization reactor power adjustment signal: ,pass Analog output module is distributed.

[0124] 2. Control signals for the fuel supply valve at the fuel stack front stage: The data is transmitted to the electric control valve via the Modbus RTU protocol.

[0125] 3. Variable frequency control signal for waste heat recovery working fluid circulation pump: based on And the linear conversion relationship between flow rate and frequency, calculated to obtain ,pass The analog output module sends the signal to the frequency converter.

[0126] The above signals are issued synchronously within a 100ms time window, enabling coordinated closed-loop regulation of the three sub-processes: fuel pretreatment, stack operation, and waste heat recovery.

[0127] S7, System Adaptation and Early Warning.

[0128] The system executes the following adaptive and alerting steps in the background: 1. Model Incremental Update: Every 200 sets of valid new running data samples, the system automatically calls an online learning algorithm to fine-tune the weight parameters of the hybrid neural network model. The optimization algorithm used is stochastic gradient descent with momentum (momentum coefficient 0.9, learning rate...). Each training session takes no more than 30 seconds, enabling the evaluation model to continuously adapt to the long-term aging drift of the system.

[0129] 2. Predictive maintenance early warning: The system continuously records electrode health status assessment indicators. and Historical values. Based on the operating data of the past 6 months, a first-order autoregressive moving average model is independently established for each indicator. Its parameters , The goodness of fit is determined by the maximum likelihood estimation method. When the model predicts the next 72 hours... When the preset maintenance warning threshold of 0.80 is exceeded, the system immediately generates a predictive maintenance reminder, which predicts that the risk of sulfur poisoning will soon exceed the limit, suggests checking the activity of the desulfurizer, and pushes the reminder to the operator monitoring interface.

[0130] 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.

Claims

1. A real-time optimization control method for an underground coalbed methane SOFC combined heat and power system, characterized in that, Includes the following steps: Establish a fully physical field coupled digital twin model for simulating SOFC combined heat and power systems; Real-time acquisition of direct measurement data reflecting the internal state of SOFC stacks, including at least three-dimensional temperature field distribution data and electrode electrochemical impedance spectroscopy data; Based on the direct measurement data and real-time system operating parameters, a machine learning algorithm is used to perform fusion calculations and output an evaluation index characterizing the current health status of the stack electrodes. Based on the electrode health status assessment index, the preset parameters in the all-physics field coupled digital twin model are dynamically corrected; Based on the modified digital twin model, multi-objective rolling optimization is performed with at least one of the following as optimization objectives: overall system efficiency, internal temperature uniformity of the fuel cell stack, and electrode health status assessment index, to obtain a set of optimal operating settings. The optimal operating setpoints are converted into control commands to coordinate and regulate the fuel pretreatment process, the fuel cell stack operation process, and the waste heat recovery process in the combined heat and power system.

2. The real-time optimization control method for an underground coalbed methane SOFC combined heat and power system according to claim 1, characterized in that, The real-time acquisition of three-dimensional temperature field distribution data inside the fuel cell stack includes: A distributed fiber optic temperature sensing network deployed on the surface of the SOFC stack metal connector is used to continuously collect temperature profile data along the airflow direction and stacking direction of the stack at a preset spatial resolution. The sensing optical fiber is fixed in a pre-processed micro-groove on the surface of the metal connector, or is attached to the surface of the metal connector by a high-temperature resistant thermally conductive adhesive layer to ensure thermal contact.

3. The real-time optimization control method for an underground coalbed methane SOFC combined heat and power system according to claim 2, characterized in that, After acquiring the temperature profile data, a data preprocessing step is also included: The raw temperature data is filtered using a moving average to suppress measurement noise; Based on the filtered temperature data, spatial temperature gradient distribution data reflecting the degree of drastic change in the temperature field are calculated.

4. The real-time optimization control method for an underground coalbed methane SOFC combined heat and power system according to claim 1, characterized in that, The electrode electrochemical impedance spectroscopy data includes: Continuously calculate the rate of change of total system current and total fuel flow within the most recent time window; When the rate of change is lower than its corresponding first steady-state determination threshold, the determination system enters a quasi-steady-state window that can be used for measurement. Within the quasi-steady-state window, the excitation source is controlled to inject a small-amplitude multi-frequency AC current disturbance signal of a preset waveform into the fuel cell stack; The voltage response signals at both ends of the fuel cell stack are acquired synchronously at high speed, and the complex impedance spectrum corresponding to the multi-frequency current disturbance is obtained by frequency domain analysis.

5. The real-time optimization control method for an underground coalbed methane SOFC combined heat and power system according to claim 4, characterized in that, The frequency domain analysis method is either Fast Fourier Transform or Lock-in Amplification. After obtaining the complex impedance spectrum, the method further includes a data validity verification step: calculating the coherence function or total harmonic distortion of the impedance spectrum. If the value is lower than the second steady-state determination threshold, the measurement data is determined to be valid and adopted.

6. The real-time optimization control method for an underground coalbed methane SOFC combined heat and power system according to claim 1, characterized in that, The fusion calculation using machine learning algorithms includes: Construct a hybrid neural network model with a dual-channel input structure; The first channel is a time-series processing channel, configured to input the historical sequence of the three-dimensional temperature field distribution data inside the fuel cell stack into the long short-term memory network layer to extract the implicit time-series dynamic features. The second channel is a transient feature channel, configured to input the characteristic frequency impedance value in the current electrode electrochemical impedance spectroscopy data and the real-time operating parameters of the system into a fully connected neural network layer for feature mapping and fusion; The time-series dynamic features are concatenated with the fused features output from the second channel and input into a subsequent fully connected neural network layer for calculation. The final output is the electrode health status assessment index, which includes a quantified carbon deposition risk index and a quantified sulfur poisoning risk index.

7. A real-time optimization control method for an underground coalbed methane SOFC combined heat and power system according to claim 6, characterized in that, The dynamic correction of preset parameters in the digital twin model includes: Establish a mapping relationship between the electrode health status assessment index and the anode kinetic parameters and mass transfer parameters in the digital twin model; Based on the mapping relationship, the values ​​of the anode exchange current density and the anode effective gas diffusion coefficient in the digital twin model are adjusted in real time using the quantitative carbon deposition risk index and the quantitative sulfur poisoning risk index at the current moment.

8. The real-time optimization control method for an underground coalbed methane SOFC combined heat and power system according to claim 1, characterized in that, The multi-objective rolling optimization is implemented using a hierarchical collaborative control architecture. A fast control layer operates a rule-based feedforward controller in parallel. The feedforward controller is configured to: when the rate of change of hydrogen sulfide concentration in the inlet fuel gas exceeds the mutation threshold, immediately adjust the power setting value of the desulfurization unit and the opening setting value of the fuel supply valve in front of the stack according to a preset ratio. The control layer is optimized by calling a multi-objective optimization algorithm at a fixed period. The multi-objective optimization algorithm uses the modified digital twin model as a predictor to perform collaborative optimization of fuel allocation ratio, air excess coefficient and heat recovery working fluid flow rate, and generate global optimization setpoints. The feedforward compensation setting value output by the fast control layer is superimposed with the global optimization setting value output by the optimized control layer to generate the final integrated control command sent to each actuator.

9. A real-time optimization control method for an underground coalbed methane SOFC combined heat and power system according to claim 8, characterized in that, It also includes system adaptation and early warning steps: The network connection weight parameters in the machine learning algorithm are incrementally updated and trained using the latest collected system operation data and direct measurement data periodically. A long-term trend model for the electrode health status assessment indicators is established. When the indicator value predicted by the trend model will exceed the maintenance warning threshold within a preset time in the future, a predictive maintenance reminder is generated in advance.

10. A real-time optimization control method for an underground coalbed methane SOFC combined heat and power system according to claim 1, characterized in that, The coordinated regulation of the fuel pretreatment process, the stack operation process, and the waste heat recovery process is described below: Based on the optimal operating settings, power adjustment signals are calculated and issued for controlling the temperature of the desulfurization reactor, valve control signals for adjusting the fuel flow rate of each section of the fuel cell stack, and frequency conversion control signals for adjusting the speed of the waste heat boiler feedwater pump or the organic working fluid circulation pump.

Images