A method and system for predicting irreversible loss of a fuel cell centrifugal air compressor

Abstract

The application belongs to the technical field of fuel cells, and particularly relates to a fuel cell centrifugal air compressor irreversible loss prediction method and system, which comprises the following steps: collecting performance data based on a fuel cell centrifugal air compressor comprehensive performance test platform, establishing and verifying a multi-physical field coupling CFD simulation model; defining irreversible loss types and related structure parameters, constructing an integrated CFD simulation framework, generating a parameter combination table through orthogonal test and performing simulation calculation; proposing an RL-GHEM method, optimizing base model weights and combination strategies, dynamically switching linear and nonlinear combination modes, and constructing irreversible loss prediction models under different working conditions. The application proposes a hybrid integrated model based on reinforcement learning guidance, dynamically optimizes model combination strategies through a reward mechanism, optimally selects base models, improves the accuracy, stability and robustness of prediction, realizes accurate prediction of air compressor irreversible loss, and solves the problem of insufficient prediction accuracy of the prior art under variable working conditions.

Description

Technical Field

[0001] The present invention belongs to the technical field of fuel cells, and in particular relates to a method and system for predicting irreversible losses of a fuel cell centrifugal air compressor. Background Art

[0002] Fuel cell technology, as a clean and efficient energy conversion technology, plays a vital role in global energy transition and low-carbon economic development. Due to their high efficiency and environmental friendliness, fuel cells have been widely used in transportation, distributed power generation, and portable devices, and their advantages in achieving green energy utilization are particularly significant. However, to achieve optimal performance of fuel cell systems, the efficiency of the air supply system becomes one of the key technologies. In fuel cell systems, the efficiency and stability of the air supply system directly affect the performance of the electrochemical reaction and the overall efficiency of the system. The centrifugal air compressor is the core equipment of the fuel cell air supply system, and its performance determines the pressure, flow rate, and energy conversion efficiency of the air supply. However, due to the complex and changing working environment faced by centrifugal air compressors during operation, their internal flow, thermodynamic, and kinetic characteristics are complex and diverse, inevitably resulting in different types of irreversible losses. These losses significantly limit the performance of the compressor and become a major obstacle to further optimization of fuel cell systems.

[0003] The irreversible losses of centrifugal air compressors mainly include the following types: 1. Flow loss: Due to the vortex, flow separation and secondary flow phenomena of the gas in the impeller and diffuser, kinetic energy is converted into thermal energy, and the compression efficiency is reduced. 2. Thermodynamic loss: The entropy increase effect generated during the gas compression process causes part of the energy to be dissipated as heat, thereby reducing the energy conversion efficiency of the system. 3. Mechanical loss: The energy loss caused by friction between bearings, seals and rotating parts has a direct impact on the mechanical efficiency of the compressor. 4. Shock wave loss: Under super-design conditions, local shock waves may be generated in the impeller or diffuser, causing rapid dissipation of energy and decreased efficiency. The causes of the above losses involve complex fluid mechanics and thermodynamic phenomena. Current technologies still face many challenges in predicting and controlling these irreversible losses under variable conditions.

[0004] To reduce these losses, existing technologies primarily focus on the following areas: 1. Geometric design optimization: Improving the geometry of the impeller and diffuser to enhance compression efficiency. For example, optimizing the impeller curvature and diffuser flow path design can reduce vortexes and flow separation. However, such optimizations are typically designed for specific operating conditions (e.g., fixed speed, specific flow rate) and are not well adapted to changing requirements under dynamic operating conditions. 2. Advanced control methods: Such as model predictive control (MPC) and fuzzy control. These methods utilize system models or empirical rules to dynamically adjust the operating state of the air compressor. However, these control strategies rely on fixed mathematical models or manually designed rules, exhibiting limitations when dealing with highly nonlinear and rapidly changing operating conditions. 3. Data-driven optimization: With the application of artificial intelligence and machine learning, many studies have attempted to use algorithms to predict the operating state of the compressor and optimize its performance. Although these methods show potential in handling nonlinear problems, they still lack robustness, physical interpretability, and adaptability to external disturbances.

[0005] Taking patent CN110826270A as an example, this patent discloses a method for analyzing energy loss during the rotating stall of a centrifugal compressor. Its technical solution analyzes the characteristics of energy loss during the rotating stall of a centrifugal air compressor through a three-dimensional geometric model and a finite element model. Although this method has certain research value, it has the following shortcomings in practical applications: 1. Low computational efficiency: The finite element model is too complex and requires high computing resources, making it difficult to meet real-time optimization needs. 2. Single analysis dimension: It mainly focuses on the total amount of energy loss, and does not refine types such as flow loss and thermodynamic loss. 3. Limited applicability: The model relies on specific modeling assumptions, lacks adaptability across working conditions, and is difficult to promote and apply in complex dynamic environments.

[0006] Therefore, exploring efficient, accurate and cross-operating-condition adaptability prediction and control methods for the causes of irreversible losses of centrifugal air compressors under variable operating conditions is a key research direction for promoting the development of fuel cell air supply system technology. Summary of the Invention

[0007] In response to the shortcomings of the existing technology, the present invention proposes a method and system for predicting irreversible losses of a fuel cell centrifugal air compressor. This method combines reinforcement learning with a hybrid integration strategy and proposes a hybrid integration model method guided by reinforcement learning - RL-GHEM. The combination and selection process of the integrated model is dynamically guided by the reinforcement learning mechanism, and the reward mechanism in reinforcement learning is used to select and combine the base models with the best performance, thereby improving the accuracy, stability and robustness of the overall model prediction, and realizing accurate and reliable prediction of the irreversible losses of the air compressor.

[0008] In one aspect, the present invention provides a method for predicting irreversible losses of a fuel cell centrifugal air compressor, the method comprising:

[0009] Step 1: Based on the comprehensive performance test platform for fuel cell centrifugal air compressors, performance data of the air compressor is collected under different boundary parameters. Combined with the three-dimensional geometric model of the centrifugal air compressor, a multi-physics field coupled computational fluid dynamics (CFD) simulation model of the fuel cell centrifugal air compressor is established. The convergence and accuracy of the established CFD simulation model are verified.

[0010] Step 2: Define N types of irreversible loss for different air compressor components, and determine M structural parameters related to irreversible loss based on the structural characteristics of each component. Build an integrated CFD simulation framework based on the CFD simulation model of the fuel cell air compressor. Determine the number of levels and parameter combinations for the M structural parameters through orthogonal experiments. Generate a parameter combination table, import the parameter combination table into the CFD simulation framework, and use the CFD simulation framework to sequentially perform simulation calculations on all parameter combinations in the parameter combination table. Extract key data corresponding to each irreversible loss type from each set of simulation results.

[0011] Step 3: Propose the RL-GHEM method, construct multiple base models to form a model library, independently train and evaluate the performance of each base model, introduce a reinforcement learning controller, generate actions to adjust the base model weights or change the model combination strategy, optimize the weight distribution of the base model through the reward and punishment mechanism, dynamically switch between linear weighted combination and nonlinear combination according to real-time data characteristics, and use the exploration and utilization balance strategy to dynamically optimize the base model weights and combination strategies to construct an irreversible loss prediction model for air compressors under different working conditions.

[0012] Furthermore, step 2 includes:

[0013] Step 2.1: Based on the main sources of energy loss during air compressor operation, classify the irreversible loss types into flow loss, thermodynamic loss, mechanical loss, and shock wave loss. Based on the impact of different air compressor components on energy loss, determine the structural parameters that affect irreversible loss;

[0014] Step 2.2: Import the structural parameters of each air compressor component that affect irreversible losses, determined in Step 2.1, as input variables into the CFD simulation model. Parametric modeling is performed on the structural parameters to generate multiple CFD simulation models for different operating conditions. Based on the analysis requirements for irreversible loss types, the CFD simulation model is expanded and a multi-physics simulation framework based on the CFD simulation model is constructed. The calculation method for each type of irreversible loss is defined within the multi-physics simulation framework.

[0015] Step 2.3: Based on the M structural parameters and their levels that affect irreversible losses of each air compressor component, select the corresponding orthogonal table to generate a parameter combination table. Automatically import the parameter combination into the CFD simulation model constructed in step 2.2. Simulate all parameter combinations in turn. For each set of simulation results, extract the velocity field, temperature field, and other related data, and apply wavelet noise reduction processing to the data.

[0016] Furthermore, in step 2.1, the structural parameters affecting irreversible loss include but are not limited to: blade inlet installation angle, blade outlet installation angle, outlet blade thickness, volute cross-section ellipticity, volute outlet diameter, volute tongue angle, diffuser outlet radius, diffuser inlet diameter and diffuser outlet width.

[0017] Furthermore, in step 2.3, the flow loss in the multi-physics simulation framework is expressed by the entropy increase rate caused by viscous dissipation and turbulent dissipation. The entropy increase rate caused by viscous dissipation is determined by the following formula:

[0018]

[0019] Where: s v represents the entropy increase rate caused by viscous dissipation; μ represents the dynamic viscosity; T represents the thermodynamic temperature of the system; τ ij represents the viscous stress tensor; u i represents the velocity component in the i-th direction; x j represents the spatial coordinate of the jth direction; represents the velocity gradient;

[0020] The rate of entropy increase due to turbulent dissipation is determined by the following formula:

[0021]

[0022] Where: s v’ represents the entropy increase rate caused by turbulent dissipation; β represents the empirical coefficient; ρ m represents fluid density, reflecting the mass of the fluid per unit volume; k represents turbulent kinetic energy; w represents turbulent frequency; T represents temperature;

[0023] Thermodynamic losses are expressed by the rate of entropy increase due to the temperature gradient, which is determined by the following formula:

[0024]

[0025] Where: s t represents the entropy increase rate caused by temperature gradient; T represents temperature; k e represents the thermal conductivity; represents the temperature gradient; represents the square of the temperature gradient;

[0026] Mechanical losses are expressed by the entropy increase rate due to wall friction, which is determined by the following formula:

[0027]

[0028] Where: s w represents the entropy increase rate caused by wall friction; τ w represents the wall shear stress; v represents the first layer grid velocity near the wall; T represents the temperature;

[0029] The shock wave loss is expressed by the entropy increase rate caused by temperature fluctuations, which is determined by the following formula:

[0030]

[0031] Where: s t‘ represents the entropy increase rate caused by temperature pulsation; λ represents the thermal diffusivity; λ t represents the reference thermal diffusivity; s t represents the rate of entropy increase due to temperature gradient.

[0032] Furthermore, step 3 includes:

[0033] Step 3.1: Select multiple types of base models and build a base model library. Divide the irreversible loss data into training and test subsets. Train each base model independently on its corresponding training subset. Evaluate the performance of each base model on the test subset and calculate the corresponding mean squared error. Based on the evaluation results, assign initial weights through a reinforcement learning mechanism to complete the initialization of the base model library.

[0034] Step 3.2: Initialize the reinforcement learning controller, set the controller parameters including state information, action set, and reward function, receive the current state information of the base model, and generate actions based on the state information. The actions may include adjusting the weights of the base model or changing the model combination strategy. The reward signal is calculated based on the overall model prediction performance, and the controller strategy is optimized and adjusted using the reinforcement learning algorithm.

[0035] Step 3.3: Monitor the performance of the base models in real time, recording mean squared error, computational complexity, and generalization ability. Adjust the weights of the base models based on the reward mechanism, optimizing overall model performance by increasing the weights of high-performing base models or decreasing the weights of low-performing base models. Dynamically select model combinations based on real-time data characteristics, including linear weighted average and nonlinear weighted combination. Apply the updated weight distribution and combination strategy to the hybrid ensemble model to complete dynamic adjustments.

[0036] Step 3.4: Set a trade-off strategy between exploration and exploitation behavior to determine whether the current state is to perform exploration or exploitation behavior. In the exploration phase, expand the search space by randomly perturbing the base model weights or introducing new model combination strategies. In the exploitation phase, optimize the overall prediction performance by selecting the current best-performing model combination strategy. Dynamically adjust the ratio of exploration and exploitation based on historical reward feedback to achieve continuous optimization of the base model configuration and performance improvement.

[0037] Furthermore, the base model in step 3.1 includes but is not limited to a deep neural network, a support vector machine, a classification regression decision tree, and a long short-term memory network.

[0038] Furthermore, in step 3.2, the state information of the controller input includes the predicted mean square error, weight distribution and historical performance of the base model.

[0039] Furthermore, in step 3.2, the actions generated by the controller include adjusting the weights of the base models and changing the combination strategy of the base models, where the combination strategy is switching between linear weighted or nonlinear weighted combination methods.

[0040] Furthermore, step 1 includes:

[0041] Step 1.1: Build a comprehensive performance test platform for fuel cell centrifugal air compressors to test the performance parameters, operating parameters, and control parameters of the centrifugal air compressors under all operating conditions, and obtain test data for the air compressors under different operating conditions;

[0042] Step 1.2: Use 3D modeling software to construct a 3D geometric model of the centrifugal air compressor;

[0043] Step 1.3: Based on the test data measured in step 1.1 and the three-dimensional geometric model constructed in step 1.2, a multi-physics field coupling calculation CFD simulation model for the fuel cell centrifugal air compressor is established. Multiple iterative calculations of key performance parameters are performed through the CFD simulation model to complete the convergence verification of the simulation results. The simulation calculation results are compared and analyzed with the test data obtained in step 1.1 to verify the convergence and accuracy of the CFD simulation model.

[0044] On the other hand, the present invention also provides a fuel cell centrifugal air compressor irreversible loss prediction system, comprising:

[0045] The data acquisition and model building module is used to collect air compressor performance data under different boundary parameters based on the comprehensive performance test platform of the fuel cell centrifugal air compressor, and to establish a multi-physics field coupling calculation CFD simulation model of the fuel cell centrifugal air compressor in combination with the three-dimensional geometric model of the centrifugal air compressor, and to verify the convergence and accuracy of the CFD simulation model;

[0046] The parameter optimization and key data extraction module is used to define N types of irreversible loss for different components of the air compressor, and determine M structural parameters related to the irreversible loss based on the structural characteristics of each component. Based on the CFD simulation model of the fuel cell air compressor, an integrated CFD simulation framework is constructed. By designing orthogonal experiments, the number of levels and parameter combinations of the M structural parameters are determined, and a parameter combination table is generated. The parameter combination table is imported into the CFD simulation framework. Using the CFD simulation framework, simulation calculations are sequentially performed on all parameter combinations in the parameter combination table, and key data corresponding to each irreversible loss type are extracted from each set of simulation results.

[0047] The prediction model construction module is used to propose the RL-GHEM method, construct multiple base models to form a model library, independently train and evaluate the performance of each base model, introduce a reinforcement learning controller, generate actions to adjust the base model weights or change the model combination strategy, optimize the weight distribution of the base model through the reward and punishment mechanism, dynamically switch between linear weighted combination and nonlinear combination methods according to real-time data characteristics, and use the exploration and utilization balance strategy to dynamically optimize the base model weights and combination strategies to construct an irreversible loss prediction model for air compressors under different working conditions.

[0048] The beneficial effects of the present invention are:

[0049] By constructing a test bench for a fuel cell centrifugal air compressor and a multi-physics field coupled CFD simulation model, this invention enables precise performance parameter collection and analysis under diverse boundary conditions, ensuring high model fit and reliability under complex operating conditions. Combining experimental validation with CFD simulation, this invention comprehensively reflects the multidimensional characteristics and dynamic changes in air compressor operation, improving prediction accuracy and model versatility, laying a solid foundation for refined energy loss management.

[0050] By selecting key structural parameters and designing orthogonal experiments, this method implements batch simulation and data preprocessing of multi-dimensional parameter combinations within a CFD integrated simulation framework. Compared to traditional methods, this systematic and efficient loss extraction approach more accurately identifies and characterizes various sources of irreversible losses, providing detailed data support and optimization paths for subsequent optimization, thereby improving the overall energy efficiency and operational performance of the air compressor.

[0051] The proposed RL-GHEM method combines reinforcement learning with a hybrid integrated model, significantly improving the model's prediction accuracy and adaptability for irreversible loss data under different operating conditions. It possesses strong learning and self-optimization capabilities. Under complex and changing operating conditions, the RL-GHEM method's intelligent control and combination strategies enable continuous optimization of prediction performance, providing more powerful data support and optimization tools for air compressor energy management and performance improvement. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 is a flow chart of a method for predicting irreversible losses of a fuel cell centrifugal air compressor according to the present invention;

[0053] Figure 2 2. It is a schematic diagram of the layout of a monitoring system for a comprehensive performance test platform for a fuel cell air compressor according to an embodiment of the present invention;

[0054] Figure 3 is a schematic structural diagram of a three-dimensional model of a fuel cell air compressor according to an embodiment of the present invention;

[0055] Figure 4 1 is a schematic diagram of a CFD integrated simulation framework of a fuel cell air compressor according to an embodiment of the present invention;

[0056] Figure 5 4 is a flow chart of the MMPS algorithm according to an embodiment of the present invention. DETAILED DESCRIPTION

[0057] In order to enable those skilled in the art to better understand the technical solution of the present application, the present application will be further described in detail below with reference to the accompanying drawings and embodiments.

[0058] In the description of this application, the terms "first", "second", and "third" are used for descriptive purposes only and should not be understood as indicating or implying relative importance; the term "plurality" refers to two or more, unless otherwise expressly defined. Terms such as "installed", "connected", "connected", and "fixed" should be understood in a broad sense. For example, "connected" can mean a fixed connection, a detachable connection, or an integral connection; "connected" can mean a direct connection or an indirect connection through an intermediate medium. For those of ordinary skill in the art, the specific meanings of the above terms in this application can be understood according to the specific circumstances.

[0059] In the description of this application, it should be understood that the terms "up", "down", "left", "right", "front", "back", etc., indicating directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing this application and simplifying the description, rather than indicating or implying that the device or unit referred to must have a specific direction, be constructed and operated in a specific direction. Therefore, they should not be understood as limitations on this application.

[0060] Throughout this specification, terms such as "one embodiment / method," "some embodiments / methods," and "specific embodiments / methods" mean that the specific features, structures, materials, or characteristics described in conjunction with that embodiment / method or example are included in at least one embodiment / method or example of the present application. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiment / method or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments / methods or examples.

[0061] As the instruction manual Figure 1 The purpose of the present invention is to provide a method for predicting irreversible losses of a fuel cell centrifugal air compressor, comprising:

[0062] Step 1: Based on the comprehensive performance test platform for fuel cell centrifugal air compressors, the performance data of the air compressor is collected under different boundary parameters. Combined with the three-dimensional geometric model of the centrifugal air compressor, a multi-physics field coupled computational fluid dynamics (CFD) simulation model of the fuel cell centrifugal air compressor is established; the convergence and accuracy of the established CFD simulation model are verified.

[0063] The purpose of Step 1 is to construct a high-precision and high-reliability simulation model using the performance data and three-dimensional geometric models under multiple boundary parameters collected by the test platform. This aims to provide accurate theoretical support and optimization basis for the flow characteristics analysis, thermal management optimization, and performance improvement of the centrifugal air compressor, thereby reducing test costs, improving design efficiency, and ensuring the efficient and stable operation of the centrifugal air compressor in the fuel cell system.

[0064] Step 1 includes:

[0065] Step 1.1: Build a comprehensive performance test platform for fuel cell centrifugal air compressors to test the performance parameters, operating parameters, and control parameters of the centrifugal air compressors under all operating conditions, and obtain test data for the air compressors under different operating conditions.

[0066] It should be noted that the full operating condition range includes but is not limited to operating conditions with different intake flow rates, intake pressures, outlet pressures, air temperatures, motor speeds, and dynamic load changes; by adjusting the boundary conditions of the test platform, the key performance data of the air compressor under rated conditions, partial load conditions, overload conditions, and dynamic response conditions are collected. The collected data include but are not limited to intake pressure, intake temperature, outlet pressure, outlet temperature, air flow, motor speed, and energy consumption, etc., providing comprehensive and reliable basic data support for subsequent model establishment and optimization.

[0067] As the instruction manual Figure 2The comprehensive performance test platform for fuel cell centrifugal air compressor includes the air compressor and its motor, air circuit system, water cooling system, power supply and control system.

[0068] The air flow system is the core component of the test platform, guiding air through the air compressor and measuring relevant performance parameters. It includes an air filter, intake valve, intercooler, and exhaust valve. The air filter removes particulate matter from the air, protecting the compressor. The intake valve controls the flow of air entering the compressor. The intercooler reduces the temperature of the compressed air. The exhaust valve controls the output of compressed air. The air filter is connected to the compressor's air inlet via the intake valve and piping. A pressure sensor (PT) and a temperature sensor (THA) are installed at the compressor's air inlet to monitor the pressure and temperature of the incoming air. Driven by a motor, the air compressor compresses air, and its outlet is connected to the intercooler's air inlet via piping. A pressure sensor (PT) and a temperature sensor (THA) are installed at the compressor's air outlet to monitor the compressed air's outlet parameters. The intercooler's outlet is connected to the exhaust pipe via the exhaust valve and piping. A temperature sensor (THA) and a flow sensor (FS) are installed at the compressor's air outlet to monitor the cooling effect.

[0069] The water cooling system is used to maintain the thermal management performance of the air compressor and intercooler, dissipating heat through circulating water, and includes a water tank, a water pump, a cooling water flow path, and multiple sensors. The cooling water is stored in the water tank, and the water pump is connected to the water tank to provide circulating water flow. The water pump is connected to the intercooler and the air compressor cooling circuit inlet through a pipe. The cooling water enters the intercooler through the pipe, taking away the heat in the compressed air. The water flowing out of the intercooler flows back to the water tank through the pipe, forming a closed loop. The multiple sensors in the water system include a flow sensor (FS) for measuring the cooling water flow, a pressure sensor (PT) for monitoring the water pressure, and a temperature sensor for monitoring the temperature of the cooling water.

[0070] The power supply and control system, which drives the air compressor and collects data, includes a DC power supply module, controller, and sensors. The DC power supply module (high-voltage power supply: Q-100V adjustable power supply and low-voltage power supply: 12V power supply) is connected to the air compressor motor via control circuits. The Q-100V adjustable power supply drives the compressor's main motor, while the 12V power supply powers auxiliary equipment such as sensors and controllers. The controller is directly connected to all sensors and actuators, monitoring inlet and outlet pressure, flow, and temperature data, and adjusting motor speed, intake and exhaust valve openings in real time.

[0071] The fuel cell centrifugal air compressor comprehensive performance test platform is used to test and record the operating performance of the air compressor under different boundary parameter conditions. During the test, by adjusting multiple variables within the platform (such as gas flow, inlet and outlet pressures, temperature, and motor speed), a comprehensive evaluation of the air compressor performance is achieved. The specific test process is as follows:

[0072] Test Preparation: First, activate the low-voltage power supply (12V) to power the controller and sensors, and check the sensor connections and readings. Activate the high-voltage power supply (Q-100V) to power the air compressor's main motor and check its operating status. Next, based on the test objectives, set the boundary conditions within the test range. These include air flow, inlet and outlet pressures, and air temperature. Air flow is controlled by adjusting the opening of the intake valve, inlet and outlet pressures by adjusting the opening of the intake and exhaust valves, and air temperature by adjusting the test environment or the intercooler's cooling capacity.

[0073] Test steps: Test the compressor's performance parameters in its initial state, including inlet and outlet pressures, temperatures, and air flow, and record the relevant data. The initial state involves operating the compressor under specified initial parameters (such as standard atmospheric pressure, room temperature, and rated speed). Flow rate variation test: Adjust the intake valve opening, set multiple flow rates (such as 20%, 50%, and 80% of rated flow), and operate the compressor. Record the inlet pressure (PT), temperature (THA), outlet pressure (PT), outlet temperature (THA), and flow rate (FS) at each flow rate. Pressure variation test: Adjust the exhaust valve opening to vary the outlet pressure and test performance under different pressure ratios. Record the inlet and outlet pressure (PT), temperature (THA), flow rate (FS), and intercooler cooling efficiency under these conditions. Temperature variation test: Adjust the ambient temperature or intercooler, set the intake air temperature (such as -10°C, 25°C, and 50°C), and operate the test. Record the compressed air outlet temperature (THA), pressure (PT), and flow rate (FS) at different intake air temperatures. Speed ​​Variation Test: Adjust the compressor's main motor speed (e.g., 50% low speed, 100% rated speed, 120% high speed) and operate the compressor at various speeds. Record the inlet and outlet pressures (PT), temperatures (THA), and flow rates (FS) at the corresponding speeds. Boundary Condition Test: Operate the compressor under exceeding design operating conditions, such as high flow, extremely high / low inlet temperatures, or exceeding design pressure ratios. Record operating data under these extreme conditions, noting performance changes and equipment operational stability.

[0074] The controller collects and records real-time data including inlet pressure (PT), temperature (THA), and flow rate (FS); outlet pressure (PT), temperature (THA), and flow rate (FS); motor speed and energy consumption data; and intercooler inlet and outlet cooling water temperature (THA) and flow rate (FS). Performance curves, such as flow-pressure characteristic curves and flow-temperature rise curves, are plotted based on this collected data to analyze compressor efficiency, cooling effectiveness, and dynamic response capabilities under different operating conditions.

[0075] Step 1.2: Use 3D modeling software to construct a 3D geometric model of the centrifugal air compressor.

[0076] It should be noted that the geometric model of the centrifugal air compressor created in step 1.2 is the same as the geometric structure of the centrifugal air compressor used in the fuel cell air compressor comprehensive performance test platform test in step 1.1.

[0077] As the instruction manual Figure 3 , Figure 3 This is a component of the 3D model of a fuel cell air compressor. Specifically, the core design data for the centrifugal air compressor is first obtained, including parameters such as impeller diameter, number of blades, blade curvature, diffuser structure, and volute geometry. The dimensions and relative positions of boundary conditions such as the compressor's air inlet, outlet, and cooling structure are then determined.

[0078] Then select 3D modeling software suitable for modeling complex fluid equipment (such as CATIA, SolidWorks or Creo), initialize the modeling environment, set the unit system, coordinate system and default precision to ensure the accuracy and consistency of the model.

[0079] Model the impeller, diffuser, and other components of a centrifugal air compressor. Create an impeller model based on design parameters, including parameterized design of the blades' spatial distribution, curvature, thickness, and blade spacing. Generate an axial channel at the center of the impeller, and ensure that the coupling between the impeller and the main shaft meets design requirements. Construct the diffuser's flow path geometry, including the inlet and outlet cross-sectional dimensions, channel curvature, and diffusion angle. Seamlessly connect the diffuser and impeller outlet to ensure smooth gas flow. Build a geometric model of the volute. The volute design must accurately reflect the boundary conditions of the inlet, outlet, and intercooler connection to ensure structural integrity. Assemble the constructed impeller, diffuser, volute, and other components, and use constraints to calibrate the positions and angles of each component. Refine the mesh in critical areas (such as blade edges and diffuser flow paths) to ensure accurate model details for subsequent simulation analysis. Finally, export the 3D geometry to a file format that supports CFD analysis (such as STEP, IGES, or Parasolid).

[0080] Step 1.3: Based on the test data measured in step 1.1 and the three-dimensional geometric model constructed in step 1.2, a multi-physics field coupling calculation CFD simulation model for the fuel cell centrifugal air compressor is established. Multiple iterative calculations of key performance parameters are performed through the CFD simulation model to complete the convergence verification of the simulation results. The simulation calculation results are compared and analyzed with the test data obtained in step 1.1 to verify the convergence and accuracy of the CFD simulation model.

[0081] It should be noted that the simulation model comprehensively considers gas flow, heat conduction, and dynamic interactions to accurately simulate the flow characteristics, temperature distribution, and performance parameters of centrifugal air compressors under different operating conditions. Key performance parameters include flow, pressure, temperature, and efficiency. This validates the accuracy and reliability of the CFD simulation model, ensuring it meets the needs of compressor performance optimization and design improvements.

[0082] Specifically, CFD simulation software (such as ANSYS Fluent and OpenFOAM) is first used to import the 3D geometry model and set the computational domain and boundary conditions. Based on experimental data acquired from the test platform (including inlet and outlet pressures, temperatures, and flow rates), the boundary parameters of the simulation model are defined. Taking into account the interaction of multiple physical fields, the gas flow equation (NS equation), the energy conservation equation, and the rotating machinery dynamics equation are coupled to establish a coupled fluid dynamics model.

[0083] The 3D model of the centrifugal air compressor was then meshed, using a combination of structured and unstructured meshes to refine the mesh in key areas (such as the impeller, diffuser flow path, and boundary layer). Mesh independence was verified and an appropriate mesh size was selected to ensure a balance between computational accuracy and efficiency.

[0084] Verify convergence: Run CFD simulations with different initial conditions and iteration steps, observing how residuals (such as flow, pressure, and temperature residuals) change with the number of iterations. Determine whether the simulation model has met convergence criteria based on the following indicators: all residuals decrease by at least three orders of magnitude (e.g., from 10^-3 to 10^-6); global physical quantities (such as inlet and outlet flow, pressure, and temperature) are stable and their fluctuations are below a set threshold (e.g., less than 0.1%). Compare simulation results with different grid densities and time steps to verify the stability and consistency of the results.

[0085] Verify accuracy: Run CFD simulations under several typical operating conditions (e.g., rated speed, partial load, overload), and record key performance parameters, including flow, pressure, temperature, and efficiency. Compare and analyze the simulation results with experimental data obtained from the testbed, and calculate error metrics (e.g., absolute error, relative error, or root mean square error). If the error range for key parameters is within the allowable range (e.g., within 5%), the simulation model is considered highly accurate. Analyze the source of the error to determine if adjustments to model assumptions, boundary conditions, or mesh optimization are necessary.

[0086] Finally, for working conditions where there are deviations between the test data and the simulation results, adjust the model parameters (such as turbulence model, boundary conditions) or increase the simulation complexity (such as considering the unsteady turbulence model). Re-verify the convergence and accuracy of the model until the simulation results fully conform to the test data. Through the above process, the convergence and accuracy verification of the CFD simulation model is completed, ensuring that it can accurately reflect the flow characteristics, thermodynamic properties and efficiency characteristics of the centrifugal air compressor under various boundary conditions. The final verified simulation model provides a reliable theoretical basis and technical support for the performance optimization and design improvement of centrifugal air compressors.

[0087] Step 2: Define N types of irreversible losses for different components of the air compressor, and determine M structural parameters related to irreversible losses based on the structural characteristics of each component. Build an integrated CFD simulation framework based on the CFD simulation model of the fuel cell air compressor. Determine the number of levels and parameter combinations of the M structural parameters by designing orthogonal experiments, generate a parameter combination table, and import the parameter combination table into the CFD simulation framework. Using the CFD simulation framework, perform simulation calculations on all parameter combinations in the parameter combination table in sequence, and extract key data corresponding to each type of irreversible loss from each set of simulation results.

[0088] The purpose of Step 2 is to systematically extract various types of irreversible losses (such as flow losses and thermodynamic losses) in different compressor components under multiple structural parameter combinations by building an integrated CFD simulation framework for fuel cell air compressors, combining orthogonal experimental design with automated batch simulation methods, quantifying their influencing characteristics and establishing a correlation between losses and structural parameters. Through efficient data generation and preprocessing, a comprehensive and reliable data foundation is provided for the development of subsequent performance optimization models, thereby supporting the optimized design and performance improvement of air compressors.

[0089] Step 2 includes:

[0090] Step 2.1: Based on the main sources of energy loss during air compressor operation, classify the types of irreversible losses into flow loss, thermodynamic loss, mechanical loss, and shock wave loss. Based on the impact of different air compressor components on energy loss, determine the structural parameters that affect irreversible loss.

[0091] Specifically, flow losses refer to the dissipation of kinetic energy caused by eddies, flow separation, and secondary flows. Thermodynamic losses refer to the dissipation of heat energy due to the increase in entropy during compression. Mechanical losses refer to friction losses between bearings, seals, and rotating components. Shock wave losses refer to the rapid dissipation of energy caused by shock waves generated in high-speed flows.

[0092] The key components of an air compressor include impellers, diffusers, and volutes. Based on the key components of the air compressor, the structural parameters that affect irreversible losses are determined. The structural parameters of the blades that affect irreversible losses include but are not limited to: blade inlet installation angle, blade outlet installation angle, and outlet blade thickness. Among them, the blade inlet installation angle affects flow separation and vortex generation, the blade outlet installation angle affects fluid momentum loss, and the blade outlet thickness affects turbulence intensity and flow loss. The structural parameters of the volute that affect irreversible losses include but are not limited to: volute cross-sectional ellipticity, volute outlet diameter, and volute tongue angle. Among them, the volute cross-sectional ellipticity affects secondary flow and flow distribution, the volute outlet diameter affects the outlet velocity and shock wave loss, and the volute tongue angle affects flow stability. The structural parameters of the diffuser that affect irreversible losses include but are not limited to: diffuser outlet radius, diffuser inlet diameter, and diffuser outlet width. Among them, the diffuser outlet radius affects the fluid pressure recovery efficiency, the diffuser inlet diameter affects the fluid contraction effect, and the diffuser outlet width affects the flow field uniformity.

[0093] Step 2.2: Import the structural parameters of each air compressor component that affects irreversible losses determined in step 2.1 as input variables into the CFD simulation model, perform parametric modeling on the structural parameters, and generate CFD simulation models for multiple different working conditions. Based on the analysis requirements of irreversible loss types, expand the CFD simulation model, build a multi-physics field simulation framework based on the CFD simulation model, and define the calculation method for each type of irreversible loss in the multi-physics field simulation framework.

[0094] The multi-physics simulation framework based on CFD simulation model is as shown in the attached manual. Figure 4 As shown, the multi-physics simulation framework demonstrates a complete closed-loop process from geometric modeling to performance analysis. The framework uses CFturbo (turbomachinery design software) to create a parametric geometric model of the compressor's key components, which is then imported into ANSYS Mesh or TurboGrid for high-quality meshing. Subsequently, boundary conditions and physical models are set using CFX to complete CFD simulation solutions for flow, thermodynamics, and mechanical properties. Finally, CFD-Post extracts velocity, temperature, and pressure field data for post-processing analysis. The simulation results are then used to optimize design parameters, forming an iterative design cycle that improves compressor performance and quantifies irreversible losses.

[0095] After the introduction of irreversible loss types, the multi-physics field simulation framework includes fluid field analysis modules, thermodynamic field analysis modules, mechanical field analysis modules and shock wave field analysis modules, which are used to analyze the corresponding irreversible losses.

[0096] Flow loss is calculated by analyzing the vortex intensity and pressure loss in the flow separation region. Flow loss is expressed by the entropy increase rate due to viscous dissipation and turbulent dissipation. The entropy increase rate due to viscous dissipation describes the energy dissipated by the fluid due to viscous friction (such as shear force), which mainly manifests as significant shear dissipation near the boundary layer. The entropy increase rate due to viscous dissipation is determined by the following formula:

[0097]

[0098] Where: s v represents the entropy increase rate caused by viscous dissipation; μ represents the dynamic viscosity (reflecting the internal friction characteristics of the fluid); T represents the thermodynamic temperature of the system; τ ij represents the viscous stress tensor (describing the stress component inside the fluid); u i represents the velocity component in the i-th direction; x j represents the spatial coordinate of the jth direction; Represents the velocity gradient (reflecting the rate of change of velocity in the spatial coordinate direction).

[0099] The entropy increase rate due to turbulent dissipation describes the inevitable energy dissipation of turbulent eddies during energy fractionation, and the kinetic energy loss caused by vortices or vortex streets in turbulent flow fields. The entropy increase rate due to turbulent dissipation is determined by the following formula:

[0100]

[0101] Where: s v’ represents the entropy increase rate caused by turbulent dissipation; β represents the empirical coefficient used to characterize the effect of eddy dissipation on entropy increase; ρ m represents the fluid density, reflecting the mass of the fluid per unit volume; k represents the turbulent kinetic energy, representing the kinetic energy per unit mass in turbulence; w represents the turbulent frequency, reflecting the rate of energy dissipation in turbulence; T represents the temperature (absolute temperature scale).

[0102] Thermodynamic loss is calculated by quantitatively analyzing heat dissipation based on the entropy increase formula. Thermodynamic loss is expressed as the entropy increase rate caused by the temperature gradient. The temperature gradient reflects the irreversible thermodynamic loss caused by temperature differences during heat conduction. A typical scenario involves heat transfer from a high-temperature area to a low-temperature area during compression. The entropy increase rate caused by the temperature gradient is determined by the following formula:

[0103]

[0104] Where: s t represents the rate of entropy increase due to temperature gradient; T represents temperature (Kelvin); k e Indicates the thermal conductivity coefficient (reflects the thermal conductivity of the fluid); represents the temperature gradient (the rate of change of temperature in space); Represents the square of the temperature gradient (reflects the intensity of temperature change).

[0105] Mechanical losses are calculated by combining bearing friction torque and speed to calculate friction power. Mechanical losses are expressed as the entropy increase rate due to wall friction. Wall friction reflects the losses caused by friction and heat transfer when the fluid contacts the wall. Typical scenarios include friction losses in the inner flow channel of an air compressor or on the diffuser wall. The entropy increase rate due to wall friction is determined by the following formula:

[0106]

[0107] Where: s w represents the entropy increase rate caused by wall friction; τ w represents the wall shear stress, which indicates the shear force of the fluid on the wall; v represents the first layer grid velocity near the wall, which reflects the flow velocity near the wall; T represents temperature (Kelvin).

[0108] Shock wave losses are calculated by locating the shock wave region through flow field analysis and calculating the losses caused by the shock wave intensity. Shock wave losses are expressed as the entropy increase rate due to temperature fluctuations. Temperature fluctuations describe the pressure energy loss caused by shock waves in high-speed flows, such as shock wave losses in impeller flow paths or diffusers. The entropy increase rate due to temperature fluctuations is determined by the following formula:

[0109]

[0110] Where: s t‘ represents the entropy increase rate caused by temperature pulsation; λ represents thermal diffusivity, which describes the ability of heat diffusion; λ t Represents the reference thermal diffusivity, used for normalization calculation; s t represents the entropy increase rate caused by the temperature gradient (same as formula (3)).

[0111] Furthermore, formulas (1)-(5) are the rate of increase of entropy generated by irreversible processes (viscous dissipation, heat conduction, eddy current dissipation, etc.) in a specific area per unit time. By calculating the local entropy increase rate s v 、s v’ 、s t 、s w 、s t‘ Integrate over the entire space to get the corresponding total entropy increase rate S' v 、S' v’ 、S' t 、S' w 、S' t‘ The corresponding total entropy increase rate can be expressed by the following formula

[0112] S'=∫sdV(6)

[0113] The total entropy increase rate reflects the comprehensive effect of irreversible energy transfer in the system (such as heat conduction, viscosity, turbulent dissipation, etc.). By analyzing each component, it can be determined which physical mechanism contributes the most to the entropy increase, thereby optimizing the system design or control method to reduce irreversible losses.

[0114] Step 2.3: Based on the M structural parameters and their levels that affect irreversible losses in each air compressor component, select the corresponding orthogonal table to generate a parameter combination table. These parameter combinations are automatically imported into the CFD simulation model constructed in Step 2.2, and simulation calculations are performed on all parameter combinations in sequence. For each set of simulation results, the velocity field, temperature field, and other relevant data are extracted and subjected to wavelet noise reduction.

[0115] Specifically, first determine the test parameters and levels, M structural parameters, and set several levels for each parameter (such as 3 levels: small value, medium value, and large value). For example: blade inlet angle [10°, 15°, 20°], volute outlet diameter [100mm, 120mm, 140mm]. According to the number of parameters M and the number of levels L for each parameter, select a suitable orthogonal table. In the implementation method of this application, there are 9 structural parameters of each component of the air compressor that affect the irreversible loss (blade inlet installation angle, blade outlet installation angle, outlet blade thickness, volute section ellipticity, volute outlet diameter, volute tongue angle, diffuser outlet radius, diffuser inlet diameter, and diffuser outlet width). Therefore, for 9 structural parameters and 3 levels for each parameter, an L27 orthogonal table (27 groups of combinations) can be selected to ensure that all parameter combinations cover the effects of all factor interactions as much as possible. Use tools (such as Excel, Python scripts, or professional software) to generate all parameter combinations. Then, use an automated script (such as Python, MATLAB, or Shell script) to import the parameter combinations of the orthogonal table. Import the geometric models of the different working condition models generated in step 2.2 into the CFD software, set the simulation conditions (such as mesh division and boundary conditions), run the script, and simulate each set of parameter combinations in turn using the calculation method of each irreversible loss in step 2.2. Extract the fluid dynamics field, thermodynamic field, mechanical loss, and shock wave loss from each simulation result. In the fluid dynamics field, extract the flow velocity distribution for analyzing flow loss; extract the pressure gradient for calculating shock wave loss. In the thermodynamic field, extract the temperature distribution for calculating thermodynamic loss; and extract entropy increase data through the energy equation. In the mechanical loss, extract the relevant data of wall friction loss. In the shock wave loss, extract the shock wave position and intensity data for calculating energy dissipation. Store the extracted data in an easy-to-process format (such as CSV, SQL database), with each set of parameter combinations corresponding to a set of data, to facilitate subsequent correlation analysis. Finally, each set of simulation data (such as velocity field and temperature field) is decomposed by wavelet, the threshold is set, and the high-frequency components are processed by soft / hard threshold. The denoised data is obtained through wavelet reconstruction, and all simulation data are normalized to a unified range (such as [0,1]).

[0116] Step 3: Propose the RL-GHEM method, construct multiple base models to form a model library, independently train and evaluate the performance of each base model, introduce a reinforcement learning controller, generate actions to adjust the base model weights or change the model combination strategy, optimize the weight distribution of the base model through the reward and punishment mechanism, dynamically switch between linear weighted combination and nonlinear combination according to real-time data characteristics, and use the exploration and utilization balance strategy to dynamically optimize the base model weights and combination strategies to construct an irreversible loss prediction model for air compressors under different working conditions.

[0117] The purpose of step 3 is to propose the RL-GHEM method (i.e., an optimization method based on reinforcement learning (RL) and a hybrid ensemble model (GHEM)) to select and independently train multiple base models to comprehensively capture the multidimensional characteristics of irreversible losses, provide a reliable data foundation for subsequent dynamic optimization, and adjust the weight distribution and combination strategy of the base models in real time through the reinforcement learning controller, so that the model can adapt to real-time changing working conditions and data characteristics, thereby improving the flexibility and accuracy of the prediction.

[0118] As the instruction manual Figure 5 , showing the cyclic process of reinforcement learning in the RL-GHEM method, combining the interactive relationship between "environment" and "agent". The environment represents a base model library with different weights, which includes various trained models (such as DNN, SVM, LSTM, etc.). The agent represents the reinforcement learning controller and is the core optimization module of the entire system. The state describes the characteristic information of the current environment, including the mean square error calculated based on the current model configuration and the current weights of each base model. The reward is used to measure the effect of the action generated by the agent. It is the feedback mechanism of reinforcement learning. The reward is usually directly related to the mean square error (MSE) and reflects the quality of the model combination strategy. The action is the operation generated by the agent based on the current state, which is used to optimize the weights and combination strategies of the base models. The action directly acts on the environment and updates the base model configuration.

[0119] Step 3 specifically includes:

[0120] Step 3.1: Select multiple types of base models and build a base model library. Divide the irreversible loss data into training subsets and test subsets. Perform independent training on each base model on its corresponding training subset. Evaluate the performance of each base model on the test subset and calculate the corresponding mean square error. Based on the evaluation results, assign initial weights through a reinforcement learning mechanism to complete the initialization of the base model library.

[0121] Specifically, first, multiple types of base models are selected to build a model library. The base models include but are not limited to deep neural networks (DNN), support vector machines (SVM), classification and regression decision trees (CART), long short-term memory networks (LSTM), etc. Deep neural networks (DNN) are used to process complex nonlinear features and high-dimensional data. Support vector machines (SVM) are good at modeling small sample data and capturing boundary features in high-dimensional space. Classification and regression decision trees (CART) are suitable for processing nonlinear and discrete features. Long short-term memory networks (LSTM) focus on capturing the temporal characteristics and dynamic changes of data. According to the characteristics of irreversible loss data (such as time series, nonlinear trends, etc.), appropriate data feature processing methods are selected. According to the characteristics of different base models, specific hyperparameters are preset (such as the number of hidden layers of deep neural networks, the kernel function type of support vector machines, and the time step of LSTM).

[0122] Independent training and testing environments are then set up for each base model to ensure that each base model in the model library performs well in specific tasks and covers different modeling advantages. Irreversible loss data includes irreversible loss data (such as flow losses and thermodynamic losses) for fuel cell air compressors under different operating conditions. The irreversible loss data is divided into training and testing subsets according to a certain ratio (e.g., 80% training set, 20% testing set), ensuring that the data distribution of the training and testing subsets is consistent to avoid model performance being affected by differences in data distribution. Each base model is independently trained on its own training subset to avoid interference between models. Cross-validation is used to optimize the model's hyperparameters and ensure the generalization ability of each base model. The performance of each base model is evaluated on the test subset, and its mean squared error (MSE) is calculated as the primary metric. For each base model, its predictive ability is tested under different operating conditions, and its performance is compared for various types of irreversible losses.

[0123] Finally, we assign initial weights based on the mean squared error (MSE) calculated on the test subset. Using reinforcement learning, we assign initial weights to each base model, providing a foundation for subsequent dynamic optimization. The weight assignment principle is: higher initial weights are assigned to models with better performance (lower MSE), while lower weights are assigned to models with poorer performance.

[0124] Step 3.2: Initialize the reinforcement learning controller, set the controller parameters including state information, action set and reward function, receive the state information of the current base model, and generate actions based on the state information. The actions include adjusting the weights of the base model or changing the model combination strategy; calculate the reward signal based on the overall model prediction performance, and optimize and adjust the controller strategy through the reinforcement learning algorithm.

[0125] Specifically, deep reinforcement learning (such as Q-learning or policy gradient methods) is first adopted, and the controller is designed as an intelligent agent that can learn and optimize in the state-action-reward cycle. The input state information of the controller includes the prediction mean square error (MSE), weight distribution and historical performance of the base model. The prediction mean square error (MSE) of the base model is used to evaluate the current prediction performance of the base model, the weight distribution is used to evaluate the current importance of the base model in the ensemble model, and the historical performance is used to record the performance trend of the base model in multiple rounds of training and testing. The actions generated by the controller include adjusting the weights of the base model and changing the combination strategy of the base model (such as switching between linear weighted and nonlinear weighted combination methods).

[0126] The controller then sets a set of actions it can take, including adjusting base model weights, changing the combination strategy, retraining specific base models, and introducing new base models. Adjusting base model weights refers to increasing or decreasing the weight of a base model. Changing the combination strategy involves switching between linear weighted averaging and nonlinear weighted combination. Retraining specific base models involves retraining a base model when performance is low. Introducing new base models involves introducing a new base model when the existing model library fails to meet performance requirements.

[0127] Furthermore, a reward function is defined to evaluate the effectiveness of the controller's actions and provide feedback for the next decision. The reward value is designed to improve overall prediction accuracy and system stability. The reward function includes overall prediction error, weight distribution efficiency, and action complexity penalty.

[0128] The overall prediction error can be expressed as:

[0129] R1=-MSE total (7)

[0130] Where: MSE total Represents the overall prediction mean square error of the hybrid ensemble model.

[0131] The weight distribution efficiency can be expressed as:

[0132]

[0133] The action complexity penalty can be expressed as:

[0134]

[0135] Where: C a Indicates the complexity of the actions taken by the controller (such as the penalty value for frequently adjusting weights).

[0136] According to formulas (7)-(9), the reward function is obtained:

[0137] R=R1+αR2+βP(10)

[0138] Where: α and β represent adjustment coefficients, balancing prediction accuracy and computational complexity.

[0139] The above is the process of initializing the reinforcement learning controller. Then, the current base model state information and the state information of the aggregated ensemble model are collected and passed to the controller using a vectorized representation method. The current base model state information includes the latest prediction error (MSE) of each base model, the current weight distribution of the extracted base model, the historical performance of the extracted base model, and the analysis of its performance trend. The state information of the aggregated ensemble model includes the total prediction error of the aggregated model (such as the overall MSE) and the recording of the current combination strategy (such as linear weighting or nonlinear weighting). The state information is passed to the controller using a vectorized representation method:

[0140] S=[MSE1,MSE2,……,MSE N ,w1,w2,……,w N , strategy](11)

[0141] Where: N represents the number of base models; w i represents the weight of the i-th base model.

[0142] After the state information is input into the controller, the controller selects the optimal action A based on the current state information S through the reinforcement learning algorithm. Action A includes adjusting the weight of a certain model, changing the combination strategy, etc. In the initial stage, more action combinations are explored through random sampling, and the optimal action is selected through strategy optimization in the later stage. After selecting the optimal action, the weight of the base model or the combination strategy is adjusted. Adjust the updated weight, and the original weight w i Based on the implementation of w i +Δw, Δw represents the weight adjustment value calculated by the controller. Switching combination strategy, for example: when the action is A1, the switching combination strategy is linear weighting; when the action is A2, the switching combination strategy is nonlinear weighting.

[0143] Furthermore, since the controller adjusts the parameters (such as weights or strategies) of the hybrid ensemble model after generating an action, which is reflected in the prediction performance of the hybrid ensemble model, the reward signal is updated according to the prediction performance of the hybrid ensemble model. First, the reward signal is calculated according to the overall prediction performance of the adjusted model according to formula (10), and the reward value is generated by comparing the total prediction error and the rationality of the weight distribution before and after the adjustment. Then, the reward signal is used to optimize the controller's strategy and update the Q value in Q-learning. η represents the learning rate and γ represents the discount factor. Optimize the policy parameters in the policy gradient method: Among them, J(θ) represents the policy objective function.

[0144] By initializing the reinforcement learning controller, setting state information, action sets and reward functions, dynamically inputting the current model state, generating actions for weight adjustment or strategy change, and using reward signals to optimize the controller strategy, the predictive performance and adaptability of the hybrid integrated model can be gradually improved.

[0145] Step 3.3: Monitor the performance of the base models in real time, record the mean square error, computational complexity, and generalization ability, adjust the weights of the base models based on the reward mechanism, optimize the overall model performance by increasing the weights of the base models with good performance or reducing the weights of the base models with poor performance, and dynamically select the model combination method based on the real-time data characteristics. The model combination methods include linear weighted average and nonlinear weighted combination. Apply the updated weight distribution and combination strategy to the hybrid ensemble model to complete the dynamic adjustment.

[0146] Specifically, the prediction mean square error (MSE), computational complexity, generalization ability, and stability of each base model are first monitored. After each prediction, the above indicators are recorded and updated to the state information of the reinforcement learning controller. The prediction mean square error (MSE) is used to evaluate the error of the base model on the test set, reflecting its prediction accuracy. Computational complexity measures the running time and resource consumption of the base model to prevent high-complexity models from occupying too much weight. Generalization ability is to evaluate the ability of the base model to cope with new data through cross-validation or new data testing. Stability is to observe the performance fluctuations of the base model under different working conditions and real-time data characteristics.

[0147] Then increase the weight of the base model with good performance (low MSE, strong generalization ability, and high stability). The specific formula is:

[0148]

[0149] Where: α represents the learning rate; Reward i Represents the reward value.

[0150] Reduce the weight of the base model with poor performance (high MSE, high complexity, poor stability), formula:

[0151]

[0152] Where: β represents the learning rate; Penalty i Indicates the penalty value.

[0153] Make sure the adjusted weights meet the normalization conditions:

[0154]

[0155] Furthermore, the optimal model combination strategy is dynamically selected based on the real-time data characteristics. When the data characteristics are simple or the working conditions are stable, a linear weighted combination is used:

[0156]

[0157] Where: w i represents the basis model weight; y i Represents the base model output.

[0158] When the data features are complex or the working conditions are changeable, a nonlinear combination strategy (such as power averaging) is used:

[0159]

[0160] Where: p represents the adjustment index, which determines the degree of nonlinearity.

[0161] After each prediction, the adjusted weights and combined strategies are updated to the hybrid ensemble model based on the actions of the reinforcement learning controller (weight adjustments and strategy switching). The new prediction performance is fed back to the reinforcement learning controller to update its state information to optimize the next round of decision-making.

[0162] Step 3.4: Set a trade-off strategy between exploration and exploitation behavior to determine whether the current state is to perform exploration or exploitation behavior. In the exploration phase, expand the search space by randomly perturbing the base model weights or introducing new model combination strategies. In the exploitation phase, optimize the overall prediction performance by selecting the current best-performing model combination strategy. Dynamically adjust the ratio of exploration and exploitation based on historical reward feedback to achieve continuous optimization of the base model configuration and performance improvement.

[0163] Specifically, the exploration and utilization mechanisms are defined. The exploration mechanism is that the controller randomly adjusts the weights of the base models, tries new model combinations, or introduces new base models to expand the search space. The utilization mechanism is that the controller selects the best performing model combination and weight distribution based on the current historical feedback to optimize the current prediction performance. In the initial stage, exploration is prioritized and a higher exploration probability is set to obtain a comprehensive understanding of the characteristics of the irreversible loss model. As the optimization process progresses, the exploration probability is gradually reduced, the utilization ratio is increased, and the stability of system performance is improved. A dynamic adjustment formula based on time steps is adopted:

[0164]

[0165] Where: P 探索 represents the probability of exploration; t represents the current iteration number; t c It represents the equilibrium point, which determines the speed of switching from exploration to exploitation; τ represents the adjustment rate, which controls the slope of exploration decline.

[0166] Furthermore, in the exploration phase, the search space is expanded and random perturbations are introduced to the existing base model weights:

[0167] w' i =wi +η·rand(-1,1) (18)

[0168] Where: w' i represents the new weight after perturbation; w i represents the current weight; η represents the disturbance amplitude, set a small value to avoid violent fluctuations; rand(-1,1) represents the random number generator.

[0169] Then, unused base models are randomly selected to be added to the current combination, and new weighting strategies (such as nonlinear weighted combination and power-weighted average) are adopted for the existing combination. The performance of each explored model configuration is evaluated (mean square error, computational complexity, etc.), and the results are stored in the experience pool for subsequent reference.

[0170] In the utilization phase, the current model performance is optimized, and the optimal model combination strategy is extracted from the experience pool based on the historical reward function or the current model configuration performance. The weighted update rule is used:

[0171]

[0172] Where: represents the updated weight; w i Indicates the current weight; represents the optimal weight; α represents the smoothing coefficient, which controls the update speed.

[0173] Based on the current optimal configuration, small-scale perturbations are made to fine-tune the weights to ensure the local optimality of the model combination strategy.

[0174] Furthermore, based on historical rewards, the ratio of exploration and utilization is dynamically adjusted. By analyzing the trend of the reward function over time, the ratio of exploration and utilization is dynamically adjusted: when the reward value grows slowly or stagnates, the exploration ratio is increased to escape the local optimum; when the reward value increases rapidly, the utilization ratio is increased to consolidate the current optimal strategy. The reward function form is:

[0175] R=-MSE+λ·S(20)

[0176] Where: MSE represents the mean square error of model prediction; S represents the model stability index (such as the degree of fluctuation of weight changes); λ represents the stability weight factor.

[0177] The present invention achieves accurate prediction and dynamic optimization of multidimensional parameters by constructing a fuel cell centrifugal air compressor test bench and CFD simulation model, and combining reinforcement learning with the RL-GHEM method of hybrid integrated model, providing efficient and intelligent data support and optimization path for energy loss management and performance improvement of the air compressor.

[0178] Example

[0179] The method of the present invention is based on which the performance of a fuel cell centrifugal air compressor is optimized.

[0180] Step 1: Build a comprehensive performance test platform for fuel cell centrifugal air compressors for multi-operating performance data collection. The platform includes: Air compressor: impeller diameter 150mm, maximum speed 60,000rpm. Air system: includes air filter, intake valve, intercooler, and exhaust valve. Water cooling system: used for thermal management, equipped with a circulating water pump and cooling sensor. Power supply and control system: includes a Q-100V adjustable DC power supply to drive the air compressor main motor, and a 12V power supply to power the controller and sensors. A pressure sensor (PT) and temperature sensor (THA) are installed at the air inlet, a pressure sensor (PT), temperature sensor (THA), and flow sensor (FS) are installed at the air outlet, and a flow sensor (FS) and temperature sensor are installed in the cooling water circuit.

[0181] Set up the following test conditions:

[0182] Rated operating conditions: inlet pressure 100kPa, flow rate 50g / s, outlet pressure 200kPa.

[0183] Partial load: inlet pressure 90kPa, flow rate 30g / s, outlet pressure 150kPa.

[0184] High load: inlet pressure 110kPa, flow rate 70g / s, outlet pressure 250kPa.

[0185] Dynamic conditions: simulate rapid pressure fluctuations and flow changes.

[0186] The test platform was then activated to record data under each operating condition, including inlet and outlet pressure (PT), temperature (THA), flow rate (FS), as well as motor speed and energy consumption. By adjusting the openings of the inlet and exhaust valves, the pressure ratio and flow rate were dynamically adjusted, and performance curve data such as flow-pressure curves and temperature rise curves were collected.

[0187] Based on the actual structural parameters of the air compressor on the test platform, the following parameters were used: impeller: 10 blades, 30° blade outlet angle, 150mm diameter; diffuser: 50mm flow path curvature radius, 10mm outlet width; volute: 40mm inlet diameter, 60mm outlet diameter. A 3D geometric model was generated using SolidWorks software. The impeller, diffuser, and volute were parametrically designed to ensure geometric consistency with the test platform. The model was then exported in STEP format for CFD analysis.

[0188] Import the 3D geometry into ANSYS Fluent. Set the computational domain and boundary conditions: inlet and outlet pressures and temperatures are based on test data, the fluid is air, and a turbulence model (k-ε model) is used. A combination of structured and unstructured meshing is used, with a high-density mesh in the impeller region and a refined mesh of 800,000 elements. The diffuser and volute regions are moderately refined, totaling 2 million elements. Mesh independence is verified in critical areas (such as blade edges), and an appropriate mesh size is selected.

[0189] The simulation was then run for 1000 iterations, monitoring the convergence residuals: pressure and temperature residuals < 10^-6. The simulation results were compared with test data: rated operating conditions: simulated flow rate 50.5 g / s (1% error), outlet pressure 198 kPa (1% error). High-load operating conditions: simulated flow rate 71 g / s (1.5% error), outlet pressure 245 kPa (2% error). These results demonstrate good simulation model accuracy and convergence.

[0190] Key performance data extracted from the simulation results: Flow field: Observing the vortex distribution at the impeller outlet reveals significant flow separation in the high-speed region. Temperature field: A temperature rise region appears at the diffuser flow channel outlet, indicating significant thermodynamic losses. Pressure field: The uneven pressure distribution at the volute outlet suggests the possibility of shock wave losses.

[0191] Step 2: Based on the three-dimensional CFD simulation model established in step 1 and the experimental data, classify the irreversible losses in the operation of the air compressor and determine the structural parameters related to the irreversible losses. Impeller: The blade inlet angle (10°~20°) affects the flow separation, the blade outlet angle (30°~50°) affects the fluid momentum loss, and the blade thickness (2mm~5mm) affects the turbulence intensity. Diffuser: The outlet radius (50mm~70mm) affects the pressure recovery efficiency, the inlet diameter (40mm~60mm) affects the fluid contraction effect, and the outlet width (10mm~20mm) affects the flow field uniformity. Volute: The cross-sectional ellipticity (0.8~1.2) affects the secondary flow, the outlet diameter (100mm~140mm) affects the shock wave loss, and the volute tongue angle (10°~30°) affects the flow stability.

[0192] CFturbo software was used to generate parametric geometric models for the impeller, diffuser, and volute. Different levels of value were set for each parameter, such as the blade inlet angle (10°, 15°, 20°) and the volute cross-sectional ellipticity (0.8, 1.0, 1.2). The exported geometry was imported into ANSYS Mesh and meshed using a high-quality structured mesh of 1 million cells for the impeller region and 1.5 million cells for the diffuser and volute regions, with mesh refinement applied to the boundary layer.

[0193] A multi-physics simulation framework was constructed using the fluid field analysis module, the thermodynamic field analysis module, the mechanical field analysis module, and the shock wave field analysis module. The simulation results for the flow, temperature, and pressure fields under different parameter combinations were compared with the experimental data from Step 1. The results showed that the flow rate, temperature, and pressure errors of the simulation model were all within 3%, verifying the reliability of the model.

[0194] Then, based on the 9 structural parameters and the 3 levels of each parameter, the L27 orthogonal table (27 groups of parameter combinations) was selected, and the parameter combination table was generated using a Python script and automatically imported into the CFD simulation framework. For each group of parameter combinations, a CFD simulation was run. Each simulation calculation took about 3 hours to complete all 27 groups of simulations. The 27 groups of simulation results include fluid mechanics field data (such as flow velocity distribution and pressure gradient), thermodynamic field data (such as temperature distribution and entropy increase rate), mechanical loss-related data (such as wall friction shear force), and shock wave loss-related data (such as shock wave intensity and range). Wavelet denoising was used to denoise the simulation data (such as flow velocity and temperature) and extract key features. All simulation data were normalized to [0,1] and stored as CSV files for subsequent analysis.

[0195] Determine the structural parameters and optimization directions that have the greatest impact on losses: a blade inlet angle of 15° reduces flow separation and increases flow by 5%. A diffuser outlet width of 15mm reduces entropy increase by 15%. A volute outlet diameter of 120mm reduces shock wave losses by 10%.

[0196] Step 3: Select a variety of models suitable for irreversible loss data characteristics: deep neural networks, support vector machines, classification and regression decision trees, and long-term memory networks to build a base model library. Based on the experimental data collected in step 1 and the orthogonal experimental simulation data in step 2, construct a training set and a test set (80% training set, 20% test set). Data features include: flow loss related features (eddy current intensity, flow velocity distribution), thermodynamic loss related features (temperature gradient, entropy increase rate), mechanical loss related features (wall friction coefficient, shear force), and shock wave loss related features (pressure gradient and intensity in the shock wave region).

[0197] Each base model was trained independently. For the deep neural network, the number of hidden layers was set to 4, and the activation function was ReLU. For the S-support vector machine, a radial basis kernel function was used, and the C and γ parameters were optimized. For the classification and regression decision tree, the maximum depth was limited to 10 to prevent overfitting. For the long short-term memory network, the time step was set to 10 and the number of hidden units was set to 64. The mean squared error (MSE) of each base model evaluated on the test set was 0.008 for the deep neural network, 0.015 for the support vector machine, 0.02 for the classification and regression decision tree, and 0.007 for the long short-term memory network. Based on the test set MSE results, initial weights were assigned to the base models through a reinforcement learning mechanism: 0.3 for the deep neural network, 0.2 for the support vector machine, 0.1 for the classification and regression decision tree, and 0.4 for the long short-term memory network. This resulted in the trained deep neural network, support vector machine, classification and regression decision tree, and long short-term memory network models.

[0198] A reinforcement learning controller is designed, using deep Q-learning (DQN) to optimize the controller policy. In the initial stage, various weight assignment combinations are tried through random exploration (ε-greedy strategy). Later, the weight assignment and combination strategy are optimized based on the policy gradient method.

[0199] After each prediction, the MSE, computational complexity, and generalization ability of the base model are recorded. Increase the weight of the base model with excellent performance (low MSE and strong generalization ability), and reduce the weight of the base model with poor performance (high MSE and high computational complexity). When the real-time data characteristics change little, a linear weighted combination strategy is adopted. When the real-time data characteristics are complex or change drastically, switch to a nonlinear combination strategy. Then randomly adjust the weights of the base model or try a new combination strategy. Based on historical rewards, select the current optimal combination strategy and weight distribution. In the early stages, set a higher exploration probability (ε=0.5) to expand the search space. As training progresses, gradually reduce the exploration probability (ε=0.1), increase utilization behavior, and ensure system stability.

[0200] Under dynamic conditions (flow fluctuation of 20% and pressure change of 15%), the optimized hybrid ensemble model reduced the total prediction error (MSE) by 30% and improved computational efficiency by 25%. Compared with the fixed-weight model, the RL-GHEM model significantly improved prediction accuracy and robustness under different operating conditions.

[0201] Another object of the present invention is to provide a fuel cell centrifugal air compressor irreversible loss prediction system, comprising:

[0202] The data acquisition and model building module is used to collect the performance data of the air compressor under different boundary parameters based on the comprehensive performance test platform of the fuel cell centrifugal air compressor, and combine the three-dimensional geometric model of the centrifugal air compressor to establish a multi-physics field coupling calculation CFD simulation model of the fuel cell centrifugal air compressor, and verify the convergence and accuracy of the CFD simulation model.

[0203] The parameter optimization and key data extraction module is used to define N types of irreversible losses for different components of the air compressor, and determine M structural parameters related to the irreversible losses based on the structural characteristics of each component. Based on the CFD simulation model of the fuel cell air compressor, an integrated CFD simulation framework is constructed. By designing orthogonal experiments, the number of levels and parameter combinations of the M structural parameters are determined, a parameter combination table is generated, and the parameter combination table is imported into the CFD simulation framework. Using the CFD simulation framework, simulation calculations are performed on all parameter combinations in the parameter combination table in sequence, and key data corresponding to each irreversible loss type are extracted from each set of simulation results.

[0204] The prediction model construction module is used to propose the RL-GHEM method, construct multiple base models to form a model library, independently train and evaluate the performance of each base model, introduce a reinforcement learning controller, generate actions to adjust the base model weights or change the model combination strategy, optimize the weight distribution of the base model through the reward and punishment mechanism, dynamically switch between linear weighted combination and nonlinear combination methods according to real-time data characteristics, and use the exploration and utilization balance strategy to dynamically optimize the base model weights and combination strategies to construct an irreversible loss prediction model for air compressors under different working conditions.

[0205] The above is only an embodiment of the present invention, and common sense such as the specific structure and characteristics of the scheme are not described in detail here. For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and the present invention can be implemented in other specific forms without departing from the spirit or basic characteristics of the present invention. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-restrictive. The scope of the present invention is defined by the appended claims rather than the above description, and it is intended that all changes that fall within the meaning and scope of the equivalent elements of the claims are included in the present invention. Any figure mark in the claims should not be regarded as limiting the claim involved.

Claims

1. A method for predicting irreversible loss of a fuel cell centrifugal air compressor, characterized in that: Methods include: Step 1: Based on the comprehensive performance test platform for fuel cell centrifugal air compressors, collect air compressor performance data under different boundary parameters. Combined with the three-dimensional geometric model of the centrifugal air compressor, establish a multi-physics field coupling calculation CFD simulation model for the fuel cell centrifugal air compressor. Verify the convergence and accuracy of the established CFD simulation model. Step 2: Define N types of irreversible loss for different air compressor components, and determine M structural parameters related to irreversible loss based on the structural characteristics of each component. Build an integrated CFD simulation framework based on the CFD simulation model of the fuel cell air compressor. Determine the number of levels and parameter combinations for the M structural parameters through orthogonal experiments. Generate a parameter combination table, import the parameter combination table into the CFD simulation framework, and use the CFD simulation framework to sequentially perform simulation calculations on all parameter combinations in the parameter combination table. Extract key data corresponding to each irreversible loss type from each set of simulation results. Step 3: Propose the RL-GHEM method, construct multiple base models to form a model library, independently train and evaluate the performance of each base model, introduce a reinforcement learning controller, generate actions to adjust the base model weights or change the model combination strategy, optimize the weight distribution of the base model through the reward and punishment mechanism, dynamically switch between linear weighted combination and nonlinear combination according to real-time data characteristics, and use the exploration and utilization balance strategy to dynamically optimize the base model weights and combination strategies to construct an irreversible loss prediction model for air compressors under different working conditions.

2. The method for predicting irreversible loss of a fuel cell centrifugal air compressor according to claim 1, characterized in that: Step 2 includes: Step 2.1: Based on the main sources of energy loss during air compressor operation, classify the irreversible loss types into flow loss, thermodynamic loss, mechanical loss, and shock wave loss. Based on the impact of different air compressor components on energy loss, determine the structural parameters that affect irreversible loss; Step 2.2: Import the structural parameters of each air compressor component that affect irreversible losses, determined in Step 2.1, as input variables into the CFD simulation model. Parametric modeling is performed on the structural parameters to generate multiple CFD simulation models for different operating conditions. Based on the analysis requirements for irreversible loss types, the CFD simulation model is expanded and a multi-physics simulation framework based on the CFD simulation model is constructed. The calculation method for each type of irreversible loss is defined within the multi-physics simulation framework. Step 2.3: Based on the M structural parameters and their levels that affect irreversible losses of each air compressor component, select the corresponding orthogonal table to generate a parameter combination table. Automatically import the parameter combination into the CFD simulation model constructed in step 2.

2. Simulate all parameter combinations in turn. For each set of simulation results, extract the velocity field, temperature field, and other related data, and apply wavelet noise reduction processing to the data.

3. The method for predicting irreversible loss of a fuel cell centrifugal air compressor according to claim 2, characterized in that: In step 2.1, the structural parameters that affect irreversible losses include but are not limited to: blade inlet installation angle, blade outlet installation angle, outlet blade thickness, volute cross-section ellipticity, volute outlet diameter, volute tongue angle, diffuser outlet radius, diffuser inlet diameter, and diffuser outlet width.

4. The method for predicting irreversible loss of a fuel cell centrifugal air compressor according to claim 2, wherein: In step 2.3, the flow loss in the multiphysics simulation framework is expressed by the entropy increase rate caused by viscous dissipation and turbulent dissipation. The entropy increase rate caused by viscous dissipation is determined by the following formula: Where: s v represents the entropy increase rate caused by viscous dissipation; μ represents the dynamic viscosity; T represents the thermodynamic temperature of the system; τ ij represents the viscous stress tensor; u i represents the velocity component in the i-th direction; x j represents the spatial coordinate of the jth direction; represents the velocity gradient; The rate of entropy increase due to turbulent dissipation is determined by the following formula: Where: s v’ represents the entropy increase rate caused by turbulent dissipation; β represents the empirical coefficient; ρ m represents fluid density, reflecting the mass of the fluid per unit volume; k represents turbulent kinetic energy; w represents turbulent frequency; T represents temperature; Thermodynamic losses are expressed by the rate of entropy increase due to the temperature gradient, which is determined by the following formula: Where: s t represents the entropy increase rate caused by temperature gradient; T represents temperature; k e represents the thermal conductivity; represents the temperature gradient; represents the square of the temperature gradient; Mechanical losses are expressed by the entropy increase rate due to wall friction, which is determined by the following formula: Where: s w represents the entropy increase rate caused by wall friction; τ w represents the wall shear stress; v represents the first layer grid velocity near the wall; T represents the temperature; The shock wave loss is expressed by the entropy increase rate caused by temperature fluctuations, which is determined by the following formula: Where: s t‘ represents the entropy increase rate caused by temperature pulsation; λ represents the thermal diffusivity; λ t represents the reference thermal diffusivity; s t represents the rate of entropy increase due to temperature gradient.

5. The method for predicting irreversible loss of a fuel cell centrifugal air compressor according to claim 1, characterized in that: Step 3 includes: Step 3.1: Select multiple types of base models and build a base model library. Divide the irreversible loss data into training and test subsets. Train each base model independently on its corresponding training subset. Evaluate the performance of each base model on the test subset and calculate the corresponding mean squared error. Based on the evaluation results, assign initial weights through a reinforcement learning mechanism to complete the initialization of the base model library. Step 3.2: Initialize the reinforcement learning controller, set the controller parameters including state information, action set, and reward function, receive the current state information of the base model, and generate actions based on the state information. The actions may include adjusting the weights of the base model or changing the model combination strategy. The reward signal is calculated based on the overall model prediction performance, and the controller strategy is optimized and adjusted using the reinforcement learning algorithm. Step 3.3: Monitor the performance of the base models in real time, recording mean squared error, computational complexity, and generalization ability. Adjust the weights of the base models based on the reward mechanism, optimizing overall model performance by increasing the weights of high-performing base models or decreasing the weights of low-performing base models. Dynamically select model combinations based on real-time data characteristics, including linear weighted average and nonlinear weighted combination. Apply the updated weight distribution and combination strategy to the hybrid ensemble model to complete dynamic adjustments. Step 3.4: Set a trade-off strategy between exploration and exploitation behavior to determine whether the current state is to perform exploration or exploitation behavior. In the exploration phase, expand the search space by randomly perturbing the base model weights or introducing new model combination strategies. In the exploitation phase, optimize the overall prediction performance by selecting the current best-performing model combination strategy. Dynamically adjust the ratio of exploration and exploitation based on historical reward feedback to achieve continuous optimization of the base model configuration and performance improvement.

6. The method for predicting irreversible loss of a fuel cell centrifugal air compressor according to claim 5, characterized in that: In step 3.1, the base model includes but is not limited to deep neural networks, support vector machines, classification and regression decision trees, and long short-term memory networks.

7. The method for predicting irreversible loss of a fuel cell centrifugal air compressor according to claim 5, characterized in that: In step 3.2, the input controller state information includes the predicted mean square error, weight distribution and historical performance of the base model.

8. The method for predicting irreversible loss of a fuel cell centrifugal air compressor according to claim 5, characterized in that: In step 3.2, the actions generated by the controller include adjusting the weights of the base models and changing the combination strategy of the base models. The combination strategy is to switch between linear weighted and nonlinear weighted combination methods.

9. The method for predicting irreversible loss of a fuel cell centrifugal air compressor according to claim 1, characterized in that: Step 1 includes: Step 1.1: Build a comprehensive performance test platform for fuel cell centrifugal air compressors to test the performance parameters, operating parameters, and control parameters of the centrifugal air compressors under all operating conditions, and obtain test data for the air compressors under different operating conditions; Step 1.2: Use 3D modeling software to construct a 3D geometric model of the centrifugal air compressor; Step 1.3: Based on the test data measured in step 1.1 and the three-dimensional geometric model constructed in step 1.2, a multi-physics field coupling calculation CFD simulation model for the fuel cell centrifugal air compressor is established. Multiple iterative calculations of key performance parameters are performed through the CFD simulation model to complete the convergence verification of the simulation results. The simulation calculation results are compared and analyzed with the test data obtained in step 1.1 to verify the convergence and accuracy of the CFD simulation model.

10. A fuel cell centrifugal air compressor irreversible loss prediction system, characterized in that: include: The data acquisition and model building module is used to collect air compressor performance data under different boundary parameters based on the comprehensive performance test platform of the fuel cell centrifugal air compressor, and to establish a multi-physics field coupling calculation CFD simulation model of the fuel cell centrifugal air compressor in combination with the three-dimensional geometric model of the centrifugal air compressor, and to verify the convergence and accuracy of the CFD simulation model; The parameter optimization and key data extraction module is used to define N types of irreversible loss for different components of the air compressor, and determine M structural parameters related to the irreversible loss based on the structural characteristics of each component. Based on the CFD simulation model of the fuel cell air compressor, an integrated CFD simulation framework is constructed. By designing orthogonal experiments, the number of levels and parameter combinations of the M structural parameters are determined, and a parameter combination table is generated. The parameter combination table is imported into the CFD simulation framework. Using the CFD simulation framework, simulation calculations are sequentially performed on all parameter combinations in the parameter combination table, and key data corresponding to each irreversible loss type are extracted from each set of simulation results. The prediction model construction module is used to propose the RL-GHEM method, construct multiple base models to form a model library, independently train and evaluate the performance of each base model, introduce a reinforcement learning controller, generate actions to adjust the base model weights or change the model combination strategy, optimize the weight distribution of the base model through the reward and punishment mechanism, dynamically switch between linear weighted combination and nonlinear combination methods according to real-time data characteristics, and use the exploration and utilization balance strategy to dynamically optimize the base model weights and combination strategies to construct an irreversible loss prediction model for air compressors under different working conditions.

Images