A valve cooling system main waterway state prediction method based on deep learning

Abstract

The application discloses a valve cooling system main waterway state prediction method based on deep learning, and the method comprises the following steps: building an electric-flow-heat-mass transfer multi-physical field coupling simulation model of the valve cooling system main waterway; driving a deep learning algorithm with simulation data to obtain a deep learning valve cooling system main waterway data prediction model; using the obtained data prediction model to realize large-scale data acquisition, screening key state variables based on a variable control method; proposing a valve cooling system main waterway state evaluation method based on the deep learning model, and realizing full-automatic state division through an adaptive neural fuzzy system. The application can simulate and analyze the distribution of each key parameter under the condition of multi-physical field coupling of the valve cooling system main waterway, and reduce the dimension of the simulation model through the deep learning algorithm, realizes the high-speed of the valve cooling system main waterway state prediction, and judges the working state of the valve cooling system main waterway through the proposed state evaluation criterion.

Description

Technical Field

[0001] This invention relates to the field of high voltage direct current transmission engineering technology, and in particular to a method and system for predicting the main water circuit status of a valve cooling system based on deep learning. Background Technology

[0002] Ultra-high voltage direct current (UHVDC) transmission networks are the lifeline of the power system, and their safety and reliability are critical issues concerning people's livelihoods. However, a series of corrosion and scaling problems in the valve cooling system caused by leakage current seriously threaten system safety. Operational experience shows that the valve cooling system is a weak link in the converter valve, and the corrosion and scaling problems in the valve's internal cooling water system cannot be ignored. Domestic and foreign scholars have conducted research to varying degrees to address this problem. Research on the valve section of the converter valve cooling system has achieved certain results, resulting in a relatively reliable simulation model that comprehensively considers the effects of multi-physics field coupling. However, research on the main water circuit of the valve cooling system is severely insufficient, mostly limited to establishing models of the external electric field distribution of the valve tower, neglecting the study of electrode scaling problems in the main water circuit and failing to clarify the specific chemical reaction types of each electrode under multi-electrode pair conditions in the main water circuit. To ensure the safety and stability of the high-voltage power system and reduce the problems of water pipe burning and breakdown caused by voltage difference due to scaling of the main water circuit's equalizing electrodes, it is necessary to construct a multi-physics field coupling model that comprehensively considers the electrochemical field, turbulent field, temperature field, and mass transfer field, and to propose a method for predicting the state of the main water circuit. Summary of the Invention

[0003] The purpose of this invention is to provide a deep learning-based method for predicting the main water circuit status of a valve cooling system. This method can predict the voltage distribution and scaling status of the main water circuit after a period of operation, thereby providing guidance for the inspection and maintenance work of engineering technicians and ensuring the reliable operation of the high voltage DC system.

[0004] To achieve the objectives of this invention, the technical solution adopted is as follows:

[0005] A deep learning-based method for predicting the main water circuit state of a valve-cooled system includes the following steps:

[0006] S1. Build a multiphysics coupling simulation model of the main water circuit of the valve cooling system and define the material parameters;

[0007] S2. Train the simulation model to obtain a deep learning prediction model for the water circuit data of the valve cooling system;

[0008] S3. For the prediction model, use a variable control-based method to screen key state variables;

[0009] S4. Based on the obtained key state variables, a main waterway state assessment method is proposed, and an adaptive neural fuzzy system is used to achieve fully automatic state division.

[0010] Furthermore, the material parameters in S1 include the coefficient of thermal expansion, constant-pressure heat capacity, density, thermal conductivity, tangential coefficient of thermal expansion, thermal strain, electrolyte conductivity, dynamic viscosity, and the scale material defined in the chemical field. The scale material is defined in the chemical field as follows:

[0011] a. Define the electrochemical field potential of the main water channel and the reactions on the electrode surfaces involved to simulate the scale accumulation process:

[0012] At the high-potential scale-forming electrode, the potential distribution satisfies the following equation:

[0013]

[0014] In the above formula, For gradient, Let σ be the electrical potential, σ be the conductivity, ε be the dielectric constant, J be the current density, and t be the time.

[0015] The Tafel equation is used to describe the mathematical relationship between electrode reaction current density, overpotential, and the concentrations of reactants and products:

[0016]

[0017] In the above formula, i loc Let i be the local current density, i0 be the exchange current density, A be the Tafel slope representing the reciprocal of the voltage, and η be the overpotential.

[0018] b. Define the fluid flow in the turbulent field of the main waterway to reveal the internal deionized water flow velocity and the pressure on the inner wall of the pipe: The turbulent field is defined by the turbulent k-ε(spf) function, and its main functional expressions are as follows:

[0019]

[0020] In the above formula, ρ is the liquid density, k is the turbulent energy, t is time, and v is the fluid velocity. Let μ be the gradient, μ be the dynamic viscosity, and μ be the viscosity gradient. T G is the turbulent viscosity coefficient. k σ is the turbulent kinetic energy generated by the average velocity gradient, ζ is the turbulent dissipation rate, and σ is the turbulent kinetic energy generated by the average velocity gradient. k C μ C 1s and C 2s For coefficients;

[0021] c. Define the temperature field distribution in the main waterway: The main functional expressions affecting the temperature field distribution in the main waterway are as follows.

[0022]

[0023] In the above formula, d zC is a constant, ρ is the fluid density, and C is the fluid density. ρ Let be the specific heat capacity per unit mass of the fluid, and v be the fluid velocity. Let q be the temperature gradient, and q be the heat flux. Let Q be the divergence of heat flux, Q be the energy source term (mainly the heat of chemical reaction here), q0 be the total heat flux, and k be the energy source term. h is the thermal conductivity.

[0024] d. Define the migration conditions of each ion in the main waterway mass transfer field: Define the mass transfer field as the dilute mass transfer (tds) condition, and add five substances "OH", "H", "Al3", "AlOH4", and "AlOH3" as the dependent variables for migration in the main waterway; the main functional expressions included in the mass transfer field are as follows:

[0025]

[0026] In the above formula, N i D represents the flux of scale-forming ions. i The diffusion coefficient is... For the gradient, c i Z represents the ion concentration. i The number of scale-forming ions by charge; u m,i The constant is the electrotransfer number, and F is the Faraday constant. ν represents the electrolyte potential, v represents the fluid velocity, and the subscript i represents the ion name.

[0027] Furthermore, the specific steps for training the simulation model in S2 are as follows:

[0028] 4) Obtain and organize the scaling and voltage distribution data of the main water channel equalizing electrode obtained from the multiphysics coupling model;

[0029] 5) The acquired data is divided into an experimental group dataset and a test group training set. The experimental group dataset is then further divided into a training set and a validation set. To improve the training efficiency of the neural network prediction model and accelerate network convergence, Z-score is used to normalize the simulation data. The formula is as follows:

[0030]

[0031] Where, N i D represents the normalized feature quantity. i σ represents the feature data before normalization; μ represents the mean of the feature data; σ represents the standard deviation of the feature data.

[0032] 6) Training is performed using a Deep Operator Network (DeepONet).

[0033] Furthermore, the training process for the deep operator network in step 3) is as follows:

[0034] 3.1) Determine the hidden layers and the number of neurons within the network;

[0035] 3.2) Determine the basic information of the input layer and output layer, and initialize the weights inside the neurons;

[0036] 3.3) Input the dataset into the branch network and trunk network for forward propagation, and calculate the value of each neuron. Obtain the output of DeepONet.

[0037] 3.4) Calculate the loss function, perform backpropagation, and find the error term for each neuron.

[0038] 3.5) Update the weights within the network according to the error term and the weight update formula.

[0039] 3.6) Iterate through steps three to five until the maximum number of iterations is reached.

[0040] Furthermore, the main evaluation value of the water circuit status in S4 is voltage, and the voltage distortion rate ΔUi is defined to characterize the changing trend of the main water circuit voltage distribution after electrode scaling:

[0041]

[0042] Among them, U i-初始 U represents the potential on each electrode when the initial scale-free layer is formed. i-结垢 This represents the potential of each electrode after scaling has occurred during a period of operation.

[0043] Voltage distortion at the electrode also means that a voltage difference appears between the inside and outside of the water pipe wall around the electrode. Combining the breakdown characteristics data of the main water pipe and the initial potential at the corresponding electrode, the voltage distortion rate U when the pressure equalization capability is completely unbalanced is obtained. i-F The failure rate of the voltage equalization capability at the corresponding electrode, f i Defined as:

[0044]

[0045] Based on this, the overall pressure equalization capacity failure rate f of the main waterway is defined as follows:

[0046]

[0047] Based on the failure rate f of the main water circuit pressure equalization capacity, the health status classification standard of the valve cooling system's main water circuit is set, dividing the main water circuit status into four levels, among which:

[0048] When 0 ≤ f ≤ 0.2, it is in a healthy state;

[0049] When 0.2 < f ≤ 0.6, it is a moderate state;

[0050] When 0.6 < f ≤ 1, it is a severe condition;

[0051] When f > 1, it is a fault state.

[0052] Furthermore, the system architecture of the adaptive neural fuzzy system includes:

[0053] First layer: Fuzzification layer, input variables x1 and x2 are fuzzified through membership functions;

[0054] The second layer is the rule reasoning layer, which performs a product operation on the input signals using ∏ and receives the fuzzification layer neurons to calculate the intensity of the rule excitation represented by the fuzzification layer neurons.

[0055] The third layer: the normalization layer, which uses the neurons in the rule layer to normalize the activation intensity for a given rule;

[0056] The fourth layer is the inverse fuzzification layer. This layer connects to the normalization layer neurons, receives input variables x1 and x2, and calculates the weighted adaptive value of the given rule.

[0057] Fifth layer: Output layer, which calculates the sum of the adaptive output values ​​of the neurons in the upper layer, and is a single-output fixed node.

[0058] Compared with the prior art, the advantages of the present invention are as follows:

[0059] This invention can simulate the distribution of key parameters under multi-physics coupling conditions in the main water circuit of a valve cooling system, and reduce the dimensionality of the simulation model through deep learning algorithms to achieve high-speed prediction of the main water circuit state of the valve cooling system. The invention also determines the working state of the main water circuit of the valve cooling system through the proposed state evaluation criteria. Attached Figure Description

[0060] Figure 1 This is a schematic diagram of the basic process of the method in an embodiment of the present invention;

[0061] Figure 2 This is a model diagram of the main water inlet of the valve cooling system in this invention;

[0062] Figure 3 This is a diagram showing the distribution of the turbulent flow field obtained from the simulation in this invention;

[0063] Figure 4 This is a basic architecture diagram of the DeepOnet network in this invention;

[0064] Figure 5 This is a flowchart of the DeepOnet neural network training process in this invention;

[0065] Figure 6This is a diagram illustrating the deep learning prediction results described in this invention.

[0066] Figure 7 This is a structural diagram of the main waterway status assessment model based on ANFIS in this invention;

[0067] Figure 8 This is a diagram showing the change in clothing thickness obtained from simulation in this invention. Detailed Implementation

[0068] To further illustrate the technical means and effects adopted by the present invention to achieve the intended purpose, the following detailed description is provided in conjunction with the accompanying drawings and preferred embodiments:

[0069] like Figure 1 As shown, the present invention proposes a deep learning-based method for predicting the main water circuit state of a valve-cooled system, comprising the following steps:

[0070] S1. Construct a two-dimensional multiphysics coupled simulation model of the main water circuit of the valve cooling system, including:

[0071] A multiphysics coupled model was built based on the main water channel of the valve cooling system, and material parameters were defined. Given the large scale of the main water channel structure and the complexity of the electrochemical reactions, the main water channel model was simplified as follows to reduce computation time and workload while maintaining computational accuracy:

[0072] ① The three-dimensional pipe entity is simplified into a two-dimensional waterway, and the scaling situation on the cross-section is mainly studied;

[0073] ②The main waterway has a symmetrical structure; half of it is used to construct the model.

[0074] ③ Ignore the external equalizing cover and terminals of the equalizing electrode;

[0075] ④ Ignore the water pipe wall and the equalizing electrode.

[0076] After model simplification, it is necessary to define the material parameters required for deionized water, specifically including: ① coefficient of thermal expansion, ② constant pressure heat capacity, ③ density, ④ thermal conductivity, ⑤ tangential coefficient of thermal expansion, ⑥ thermal strain, ⑦ electrolyte conductivity, and ⑧ dynamic viscosity. Additionally, the scale material is defined within the electrochemical field.

[0077] The electrochemical field potential of the main water channel and the reactions on the electrode surfaces involved are defined to simulate the scale accumulation process. The electrochemical field is defined using a secondary current distribution (cd), and each electrode is defined as an "electrode surface." The corresponding electrode potential and equilibrium potential are determined, and the electrode kinetic expression revealing the chemical reaction process is completed. At high-potential scale-forming electrodes, the molar mass, molar density, and film resistance of the deposits also need to be refined. For the potential distribution in the cooling system, Ohm's law, the current continuity equation, and Gauss's law need to be considered, and the potential distribution satisfies the following equation:

[0078]

[0079] In the above formula, For gradient, Let σ be the electrical potential, ε be the conductivity, J be the current density, and t be the time. On the other hand, the Tafel equation is used to describe the mathematical relationship between the electrode reaction current density and the overpotential, reactant and product concentrations:

[0080]

[0081] In the above formula, i loc Let i be the local current density, i0 be the exchange current density, A be the Tafel slope representing the reciprocal of the voltage, and η be the overpotential.

[0082] The fluid flow in the turbulent field of the main waterway is defined to reveal the internal deionized water flow velocity and the pressure on the pipe wall. The turbulent field is defined using the k-ε(spf) metric, treating each inlet of the main waterway as a velocity inlet (i.e., a constant inlet velocity) and each outlet as a pressure outlet, thus avoiding contradictions in the flow field definition. The turbulent field distribution is obtained through simulation. Gravity conditions are added to ensure the reliability of the flow field simulation. The main functional expressions included in the turbulent field are:

[0083]

[0084] In the above formula, ρ is the liquid density, k is the turbulent energy, t is time, and v is the fluid velocity. Let μ be the gradient, μ be the dynamic viscosity, and μ be the viscosity gradient. T G is the turbulent viscosity coefficient. k σ is the turbulent kinetic energy generated by the average velocity gradient, ζ is the turbulent dissipation rate, and σ is the turbulent kinetic energy generated by the average velocity gradient. k C μ C 1s and C 2s is a coefficient.

[0085] The temperature field distribution in the main water channel is defined to ensure accurate simulation. Temperature affects chemical reaction rates and flow field conditions. The temperature field is defined using fluid heat transfer (ht). Based on experimental data, the inlet and outlet temperatures are treated as fixed values ​​to constrain the temperature field distribution. The "inflow" attribute at each inlet and the "outflow" attribute at each outlet are supplemented. The main functional expression affecting the temperature field distribution in the main water channel is:

[0086]

[0087] In the above formula, d z C is a constant, ρ is the fluid density, and C is the fluid density. ρLet be the specific heat capacity per unit mass of the fluid, and v be the fluid velocity. Let q be the temperature gradient, and q be the heat flux. Let Q be the divergence of heat flux, Q be the energy source term (mainly the heat of chemical reaction here), q0 be the total heat flux, and k be the energy source term. h is the thermal conductivity.

[0088] The migration conditions of each ion in the main water channel mass transfer field are defined to achieve multi-physics coupling. The mass transfer field is defined as the dilute mass transfer (tds) condition, and five substances, "OH", "H", "Al3", "AlOH4", and "AlOH3", are added as dependent variables for migration in the main water channel. The charge number and diffusion coefficient of each substance are supplemented to improve the reaction coefficient of the electrochemical reaction inside the main water channel. The corresponding local current density in the electrochemical field is selected, and the surface equilibrium reaction caused by scaling at the high-pressure scaling electrode is additionally added. The inlet and outlet of each part of the main water channel are regarded as open boundaries, and the external substance concentration of each open boundary is defined based on the experimental data. The internal and external ion exchange is dominated by the electrochemical field and the turbulent field. The main functional expressions included in the mass transfer field are:

[0089]

[0090] In the above formula, N i D represents the flux of scale-forming ions. i The diffusion coefficient is... For the gradient, c i Z represents the ion concentration. i The number of scale-forming ions by charge; u m,i The constant is the electrotransfer number, and F is the Faraday constant. ν represents the electrolyte potential, v represents the fluid velocity, and the subscript i represents the ion name.

[0091] The main waterway exhibits a coupling relationship between multiple physical fields, specifically manifested in the following ways:

[0092] In the temperature field expression (4), the energy source term is mainly the heat of chemical reaction. There is heat generation and consumption in the electrochemical field, which affects the temperature field distribution. The mass transfer field expression (5) includes the electrolyte potential. The migration process of ions is influenced by electric field force, which is then incorporated into mass transfer field calculations.

[0093] The temperature field expression (4) and the mass transfer field expression (5) contain the fluid velocity v. The turbulent field affects the temperature distribution and ion migration process of the main water channel through the fluid flow, thus coupling the temperature field and the mass transfer field.

[0094] In Tafel equation (2), the exchange current density i0, the deionized water density ρ and dynamic viscosity μ in the turbulent field expression (3), and the diffusion coefficient D in the mass transfer field expression (5) are all functions of temperature. The temperature field thus implicitly achieves coupling with the electric, fluid, and mass transfer fields.

[0095] Under the influence of the mass transfer field, the scale layer gradually accumulates and thickens on the surface of the equalizing electrode, which will lead to changes in the characteristics of the water flow channel and the electrode surface, thereby affecting the distribution of the turbulent field and the electrochemical field reaction during the transient process.

[0096] S2. A deep learning-based prediction model for the main water circuit of the valve-cooling system is obtained by training a deep learning algorithm using simulation data. The model is endowed with associative memory and reasoning capabilities through deep machine learning. This includes the following key steps:

[0097] Acquire and organize the scaling and voltage distribution data of the main water circuit equalizing electrode obtained from the multiphysics coupling model, as well as the corresponding operating conditions of the converter valve, including but not limited to: voltage level, electrolyte conductivity, cooling water flow rate, temperature of each inlet and outlet of the main water circuit, and external ion concentration of each inlet and outlet of the main water circuit.

[0098] The simulation data was divided into experimental and test datasets. The experimental dataset was then further divided into training and validation sets according to a specific ratio. To improve the training efficiency of the neural network prediction model and accelerate network convergence, the simulation data was normalized using the Z-score, as shown in the following formula:

[0099]

[0100] Where, N i D represents the normalized feature quantity. i σ represents the feature data before normalization; μ represents the mean of the feature data; σ represents the standard deviation of the feature data.

[0101] Training is performed using a Deep Operator Network (DeepONet), whose core idea lies in operator regression rather than function regression. The DeepONet network structure contains two sub-networks: the trunk network and the branch network. The branch network takes the system's input function as input, while the trunk network takes the output location information as input. The inputs of the two networks are multiplied internally to fuse the outputs of the branch and trunk networks, thus obtaining the output of the DeepONet network.

[0102] Although the input u of the DeepONet branch network represents the input in function form, the data still needs to be discretized throughout the computer's training process. Here, we sample random points of the input function to form a discretized sequence and input it into the branch network. This discretization process represents certain specific functions. The branch network has n inputs u. (1) u (2) , ......, u (n)Both are represented as discrete points of a finite number of representative input functions. A single input function u is of the form X = {x1, x2, ..., x...}. m}, where m represents the number of discrete points sampled from this input function. For different input functions u of the branch network... (1) u (2) , ......, u (n) It is necessary to ensure that the location information of the selected discrete points on each input function is the same, that is, the sampling method of discrete points on the trajectories of different representative input functions should be the same.

[0103] In DeepONet, the input to the trunk network is the position information output by the system, in the form: Y = {y1, y2, ..., y3}. p} represents the output G(u)(y) obtained after inputting u into the real system, respectively, within the range {y1, y2, ..., y}. p The corresponding true output G(u)(y) is obtained at the position}. The position information can be randomly selected on the system output function, and it does not need to be the same as the selection of m, so as to meet the requirement of the same point sampling method.

[0104] The network output is a scalar, denoted as G. θ (u)(y), where θ=(W,β) represents the parameters trained inside the DeepONet network, and W and β represent the weight matrix and bias coefficient inside the neuron, respectively.

[0105] The forward propagation of the DeepONet network is a standard feedforward neural network propagation. Its unique feature is that the output nodes of the neurons in the trunk and branch networks, after being forward propagated separately, are multiplied again to obtain the overall network output. The internal output of the branch networks is as follows:

[0106]

[0107] Where σ l br β represents the nonlinear activation function of the l-th layer of the neural network. l This represents the bias coefficient for each neuron at this point. The output of the branch network is in the form [b1, b2, ..., bb]. q ] T .

[0108] The corresponding trunk network output is as follows:

[0109]

[0110] Similarly, σ l br β represents the nonlinear activation function of the l-th layer of the neural network.l This represents the bias coefficient for each neuron at this point. The output of the trunk network is represented as [t1, t2, ..., tt]. q ] T .

[0111] The outputs of the branch network and trunk network after forward propagation are multiplied to obtain the output of the deeoponet network, which is G. θ (u (i) )(y):

[0112]

[0113] The backpropagation process of the DeepONet network employs the traditional BP algorithm. Due to the use of two sub-networks, the loss function needs to be backpropagated simultaneously to both the trunk and branch networks to update the weights within both networks. The loss function selected for backpropagation in this network is as follows:

[0114]

[0115] The weight update formula is as follows:

[0116]

[0117] in and Let these represent the weight matrix and bias coefficients within the l-th layer of the branch network, respectively. and represents the weight matrix and bias coefficient within the l-th layer of the trunk network, respectively. μ represents the learning rate.

[0118] In summary, the DeepONet network training process is as follows:

[0119] 1. Determine the hidden layers and number of neurons within the network.

[0120] 2. Determine the basic information of the input layer and output layer, and initialize the weights inside the neurons.

[0121] 3. Input the dataset into the branch network and trunk network for forward propagation, and calculate the value of each neuron. Obtain the output of DeepONet.

[0122] Fourth, calculate the loss function, perform backpropagation, and find the error term for each neuron.

[0123] 5. Update the weights within the network according to the error term and the weight update formula.

[0124] 6. Iterate through steps three through five until the maximum number of iterations is reached.

[0125] The operating conditions of the converter valves at each time point are used as input information, and the corresponding state information is used as output information. The deep learning network model is trained using training and validation sets. The training loss function and validation loss function of the network are used as validation metrics. The candidate structures are determined by empirical formulas, and the model accuracy is gradually improved by traversing and validating the results through a grid search method.

[0126] The mean squared error is selected as the loss function. Based on the simple cross-validation method, the optimal network structure and the optimal number of training iterations are determined by the trend of the graph changes of training loss and validation loss. In this way, a deep learning model for predicting the main water circuit state of the valve cooling system is formed by training finite element simulation data.

[0127] The processed test dataset was then fed into a data-driven deep learning data prediction model to verify the accuracy and effectiveness of the network model.

[0128] This deep learning data prediction model can quickly obtain the feature vector values ​​of scale thickness and voltage at the main water channel electrodes based on the input conditions. By inverse processing these values, the changes in scale thickness and voltage at each equalizing electrode over time can be obtained. This enables the deep learning model to replace the finite element simulation for rapid calculation of multi-field information of the main water channel.

[0129] S3. Utilize the obtained data to predict the model and achieve large-scale data acquisition, screening key state variables based on variable control methods. Under default conditions, a 50kV voltage is applied between the high and low voltage electrodes in the main water channel of the electrochemical field, the electrolyte conductivity is 8μS / m, the normal phase inflow velocity at the velocity inlet in the turbulent field is 0.3m / s, the default outlet temperature in the temperature field is 303.15K, and the default external concentration at each open boundary in the mass transfer field is 10⁻⁵ mol / m³. Specific data acquisition and research include:

[0130] Different substations have different voltage levels, and the cutoff voltage will vary accordingly during rectification and inversion. Different voltage levels result in different leakage currents, which in turn lead to different reaction rates. Different potential differences also cause variations in electric field strength, resulting in different ion transport rates. The effects of different voltage levels (10kV, 30kV, 50kV, 70kV, and 90kV) on scaling and voltage distribution changes in the main water circuit's equalizing electrode were analyzed.

[0131] Considering that operating the internal cooling water at high conductivity may lead to increased leakage current and losses, and could even cause electrical leakage accidents, a quantitative analysis of the impact of electrolyte conductivity was conducted. Conductivities of 2 μS / m, 6 μS / m, 8 μS / m, 11 μS / m, and 14 μS / m were set to analyze the effects of different conductivity levels on scaling on the main water circuit's equalizing electrode and voltage distribution changes.

[0132] The flow rate of the cooling water in the main water channel affects both the temperature distribution and ion migration. The inflow velocities at each inlet were combined and distributed, with values ​​of 0.1 m / s, 0.2 m / s, 0.3 m / s, 0.4 m / s, and 0.5 m / s respectively. The effects of the flow rate changes at each inlet on the scaling of the equalizing electrode and the voltage distribution in the main water channel were analyzed.

[0133] Temperature affects the reaction rate and flow field of electrochemical reactions. The model mainly relies on the temperature of each inlet and outlet of the main water channel to simulate the overall temperature field. The inlet and outlet temperatures are taken from 283.15K, 293.15K, 303.15K, 313.15K and 323.15K respectively. The influence of different temperature conditions at each inlet and outlet on the scaling of the equalizing electrode and the voltage distribution change of the main water channel is analyzed.

[0134] The scale layer on the equalizing electrode is mainly composed of aluminum, primarily due to corrosion from the aluminum radiator in the valve section. Therefore, the ion concentration in the valve section is a key factor determining the ion concentration in the main water circuit, and thus a significant factor influencing scale formation on the equalizing electrode. External concentrations (mol / m³) at each inlet and outlet were set to 10⁻⁶, 5*10⁻⁶, 10⁻⁵, 10⁻⁴, and 10⁻³, respectively, to analyze the impact of ion concentration changes at each inlet and outlet on scale formation and voltage distribution on the equalizing electrode in the main water circuit.

[0135] Based on variable control and using sensitivity analysis, the obtained data results were observed and organized to identify factors that have a significant impact on the scaling and voltage distribution of the main water circuit equalizing electrode. Then, the effective range and law of each key factor affecting the main water circuit equalizing capability were grasped, and the key state quantities were identified to support the optimization planning guidance for the long-term operation of the valve cooling system.

[0136] S4. A method for assessing the main water circuit state of a valve-cooling system is proposed based on a deep learning model, and a fully automatic state division is achieved through an adaptive neurofuzzy system, including:

[0137] The main water circuit is relatively wide, so there is no need to worry about scale buildup clogging the circuit and causing overheating. However, because the conductivity of the scale is several orders of magnitude lower than that of the cooling water, the voltage difference between the inside and outside of the pipe wall will gradually increase as the scale thickness increases. This may lead to the pipe wall burning and damage, or even cause the pipe wall thickness to be unable to withstand the radial pressure inside the cooling water circuit, resulting in leakage. To address this major issue in the main water circuit, the condition of the main water circuit is assessed by voltage conditions, and the voltage distortion rate ΔUi is defined to characterize the trend of voltage distribution change in the main water circuit after electrode scaling.

[0138]

[0139] U i-初始U represents the potential on each electrode when the initial scale-free layer is formed. i-结垢 This represents the potential of each electrode after scaling has occurred during a period of operation.

[0140] Voltage distortion at the electrode also means that a voltage difference appears between the inside and outside of the water pipe wall around the electrode. Combining the breakdown characteristics data of the main water pipe and the initial potential at the corresponding electrode, the voltage distortion rate U when the pressure equalization capability is completely unbalanced is obtained. i-F The failure rate of the voltage equalization capability at the corresponding electrode, f i Defined as:

[0141]

[0142] Based on this, the overall pressure equalization capacity failure rate f of the main waterway is defined as:

[0143]

[0144] Based on the failure rate f of the main water circuit pressure equalization capacity, the health status of the main water circuit of the valve cooling system is classified into four levels, as shown in Table 1: healthy, moderate, severe, and faulty.

[0145] Table 1. Criteria for Classifying the Main Water Circuit Status of Valve Cooling Systems

[0146]

[0147] Based on an Adaptive Neural Fuzzy System (ANFIS), a state assessment model for the main water circuit of a valve cooling system is constructed. Voltage and scaling levels are used as input parameters, and the pressure equalization failure rate f is selected as the output. The system structure includes:

[0148] The first layer is the fuzzification layer. The input variables x1 and x2 are fuzzified through the membership function. That is, the degree to which each neuron belongs to a certain fuzzy rule is represented by the fuzzy set A (A1, A2, B1, B2), which is the membership degree.

[0149] O 1,i =u Ai (x1), i = 1, 2O 1,i =u B(i-2) (x2), i = 3, 4, (15)

[0150] This paper uses the most commonly used Gaussian function. {d i σ i} is the membership function u Ai The set of parameters of (x1).

[0151] The second layer is the rule reasoning layer, which performs a product operation on the input signals using ∏ and receives the fuzzification layer neurons to calculate the rule excitation intensity represented by the fuzzification layer neurons.

[0152] O 2,i =w i =u Ai (x1)u B(i) (x2), i = 1, 2, (16)

[0153] The third layer: the normalization layer, which uses the neurons in the rule layer to normalize the activation intensity according to a given rule.

[0154]

[0155] The fourth layer is the inverse fuzzification layer. This layer connects to the normalization layer neurons, receives input variables x1 and x2, and calculates the weighted adaptive value of the given rule.

[0156]

[0157] Fifth layer: Output layer, which calculates the sum of the adaptive output values ​​of the neurons in the upper layer, and is a single-output fixed node.

[0158]

[0159] Data normalization is performed to eliminate network errors and their impact caused by differences in units and orders of magnitude, accelerate convergence, and improve efficiency when processing massive amounts of data. Input feature parameters are selected based on the potential at various points. The scaling degree index dg is the number of network input nodes n=2; the pressure equalization capacity failure rate f is selected as the model output quantity, which is the number of network output nodes m=1.

[0160] This valve-cooling system main water circuit status assessment model can receive the output data from the DeepOnet model and output the health status of the main water circuit pressure equalization capability accordingly, realizing automatic and efficient assessment of the valve-cooling system pressure equalization capability status. Combining the two deep learning models yields the complete valve-cooling system main water circuit status assessment model.

[0161] The operating conditions and other parameters of the main water circuit of the valve cooling system to be evaluated are obtained and input into the main water circuit condition assessment model to obtain the scaling, voltage time series, and condition assessment results of the main water circuit. The results will provide guidance for the inspection and maintenance work of engineering technicians, ensuring the reliable operation of the high-voltage direct current system.

[0162] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0163] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0164] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications, equivalent changes, and alterations made by those skilled in the art to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for predicting the main water circuit state of a valve-cooled system based on deep learning, characterized in that: Includes the following steps: S1. Build a multiphysics coupling simulation model of the main water circuit of the valve cooling system and define the material parameters; S2. Train the simulation model to obtain a deep learning prediction model for the water circuit data of the valve cooling system; S3. For the prediction model, use a variable control-based method to screen key state variables; S4. Based on the obtained key state variables, a main waterway state assessment method is proposed, and an adaptive neural fuzzy system is used to achieve fully automatic state division. The main evaluation value of the water circuit status in S4 is voltage. The voltage distortion rate ΔUi is defined to characterize the changing trend of the main water circuit voltage distribution after electrode scaling. in, U i-初始 This represents the potential on each electrode when the initial scale-free layer is formed. U i-结垢 This represents the potential of each electrode after scaling has occurred during a period of operation. Voltage distortion at the electrode also means that a voltage difference exists between the inside and outside of the water pipe wall around the electrode. Combining the breakdown characteristics data of the main water pipe and the initial potential at the corresponding electrode, the voltage distortion rate value when the pressure equalization capacity is completely unbalanced can be obtained. U i-F The failure rate of voltage equalization capability at the corresponding electrode f i Defined as: Based on this, the failure rate of the overall pressure equalization capacity of the main waterway is defined. f： Based on the failure rate of the main waterway pressure equalization capacity f The standard for classifying the health status of the main water circuit in a valve-cooled system divides the main water circuit status into four levels, among which: When 0 ≤f A value ≤0.2 indicates a healthy state; When 0.2 < f When the value is ≤0.6, it is considered a moderate state; When 0.6 < f When the value is ≤1, it is considered a severe condition; when f When the value is greater than 1, it indicates a fault state.

2. The method for predicting the main water circuit state of a valve-cooling system based on deep learning as described in claim 1, characterized in that: The material parameters in S1 include the coefficient of thermal expansion, constant-pressure heat capacity, density, thermal conductivity, tangential coefficient of thermal expansion, thermal strain, electrolyte conductivity, dynamic viscosity, and the scale material defined in the chemical field. The scale material is defined in the chemical field as follows: a. Define the electrochemical field potential of the main water channel and the reactions on the electrode surfaces involved to simulate the scale accumulation process: At the high-potential scale-forming electrode, the potential distribution satisfies the following equation: In the above formula, ∇ is the gradient, φ is the potential, σ is the conductivity, ε is the dielectric constant, J is the current density, and t is the time. The Tafel equation is used to describe the mathematical relationship between electrode reaction current density, overpotential, and the concentrations of reactants and products: In the above formula, i loc For local current density, i 0 represents the exchange current density. A The Tafel slope represents the reciprocal of the voltage. η It is the overpotential; b. Define the fluid flow in the turbulent field of the main waterway to reveal the internal deionized water flow velocity and the pressure on the inner wall of the pipe: The turbulent field is defined using k-ε turbulence, and its main functional expressions are as follows: In the above formula, ρ For the density of the liquid, k For turbulent energy, t Let v be time, v be the fluid velocity, and ∇ be the gradient. μ For dynamic viscosity, μ T The viscosity coefficient is the turbulent viscosity coefficient. G k σ is the turbulent kinetic energy generated by the average velocity gradient, ζ is the turbulent dissipation rate, and σ is the turbulent kinetic energy generated by the average velocity gradient. k , C μ C 1s and C 2s For coefficients; c. Define the temperature field distribution in the main waterway: The main functional expressions affecting the temperature field distribution in the main waterway are as follows. In the above formula, d z It is a constant. ρ For fluid density, C ρ ρ is the specific heat capacity per unit mass of fluid, v is the fluid velocity, and ∇ T Let be the temperature gradient, q be the heat flux, and ∇·q be the divergence of the heat flux. Q This is the energy source term, primarily the heat of chemical reaction. q 0 represents the total heat flux. k h Thermal conductivity; d. Define the migration conditions of each ion in the main waterway mass transfer field: Define the mass transfer field as the transport conditions of dilute substances, and add five substances "OH", "H", "Al3", "AlOH4", and "AlOH3" as the dependent variables for migration in the main waterway; the main functional expressions included in the mass transfer field are as follows: In the above formula, N i This refers to the flux of scale-forming ions. D i Let ∇ be the diffusion coefficient and ∇ be the gradient. c i This refers to the ion concentration. z i The number of charge ions that form scale; u m,i The constant is the electrotransfer number, and F is the Faraday constant. φ l Here, is the electrolyte potential, and v is the fluid velocity; subscript i This is the name of the ion.

3. The method for predicting the main water circuit state of a valve-cooling system based on deep learning as described in claim 1, characterized in that: The specific steps for training the simulation model in S2 are as follows: 1) Obtain and organize the scaling and voltage distribution data of the main waterway pressure equalization electrode obtained from the multiphysics coupling model; 2) The acquired data was divided into an experimental group dataset and a test group training set. The experimental group dataset was then further divided into a training set and a validation set. To improve the training efficiency of the neural network prediction model and accelerate network convergence, Z-score was used to normalize the simulation data. The formula is as follows: in, N i Represents the normalized feature quantity. D i σ represents the feature data before normalization; μ represents the mean of the feature data; σ represents the standard deviation of the feature data. 3) Training is performed using deep operator networks.

4. The method for predicting the main water circuit state of a valve-cooling system based on deep learning as described in claim 3, characterized in that: The training process for the deep operator network in step 3) is as follows: 3.1) Determine the hidden layers and the number of neurons within the network; 3.2) Determine the basic information of the input layer and output layer, and initialize the weights inside the neurons; 3.3) Input the dataset into the branch network and trunk network for forward propagation, calculate the value of each neuron; obtain the output of the deep operator network; 3.4) Calculate the loss function, perform backpropagation, and find the error term for each neuron; 3.5) Update the weights within the network according to the error term and the weight update formula; 3.6) Iterate through steps three to five until the maximum number of iterations is reached.

5. The method for predicting the main water circuit state of a valve-cooling system based on deep learning as described in claim 1, characterized in that: The system architecture of the adaptive neurofuzzy system includes: First layer: Fuzzification layer, input variables x1 and x2 are fuzzified through membership functions; The second layer is the rule reasoning layer, which performs a product operation on the input signals using ∏ and receives the fuzzification layer neurons to calculate the intensity of the rule excitation represented by the fuzzification layer neurons. The third layer: the normalization layer, which uses the neurons in the rule layer to normalize the activation intensity for a given rule; The fourth layer is the inverse fuzzification layer. This layer connects to the normalization layer neurons, receives input variables x1 and x2, and calculates the weighted adaptive value of the given rule. Fifth layer: Output layer, which calculates the sum of the adaptive output values ​​of the neurons in the upper layer, and is a single-output fixed node.

Images