Gas-insulated transmission line temperature rise prediction method based on data-driven model

Abstract

The invention discloses a gas insulated transmission line temperature rise prediction method based on a data-driven model, and belongs to the field of temperature rise simulation, and the method comprises the steps: constructing a simulation model of a multi-field coupled gas insulated transmission line, carrying out the simulation of the simulation model, and obtaining sample data, the sample data comprises temperature data and temperature rise index characteristics in the simulation process; a fusion prediction model is constructed, the fusion prediction model adopts a BiTCN model, a BiGRU model and an XGBoost layer fusion model, and the fusion prediction model is optimized through an intelligent travel mouse algorithm according to the temperature sample data; and the temperature data and the temperature rise index characteristics of the gas insulated transmission line, which are acquired in real time, are predicted through the optimized fusion prediction model, so that the temperature data of the gas insulated transmission line can be obtained, and the temperature rise process of the GIL can be accurately predicted at low cost.

Description

Technical Field

[0001] This invention belongs to the field of temperature rise calculation technology for gas-insulated transmission lines, and particularly relates to a method for predicting the temperature rise of gas-insulated transmission lines based on a data-driven model. Background Technology

[0002] Gas-insulated transmission lines (GILs), as an alternative to overhead lines, are widely used in power systems due to their advantages such as high voltage capacity, large transmission capacity, low unit loss, high reliability, and environmental adaptability. GILs are an ideal solution for such problems and are also suitable for hydropower stations, nuclear power plants, and large substations requiring high-capacity outgoing lines. While GILs are often chosen for ultra-long, deeply buried tunnel transmission corridors, their completely enclosed nature, complex heat exchange processes, high risk of internal heat generation, and the potential for mechanical failures at high temperatures can lead to significant economic losses and safety accidents. Simultaneously, the increased temperature of the internal conductors significantly reduces the current-carrying capacity of the transmission line, further reducing the overall transmission capacity of the corridor and severely impacting power supply. The internal conductor temperature of GILs is often used as a key parameter for fault diagnosis; therefore, accurate prediction of GIL temperature is crucial for ensuring the safe and stable operation of the entire transmission system.

[0003] GILs (Gas Insulators) are primarily composed of coaxially arranged tubular aluminum alloy conductors. The internal conductors are under high voltage throughout operation, while the outer casing is grounded and encapsulated. To achieve insulation between the conductors and the casing, an insulating gas is filled between them. SF6 is commonly used, although some research uses environmentally friendly insulating gases instead. The pressure range of the internal insulating gas is typically 0.29 MPa to 0.51 MPa (20°C). In actual GIL operation, most temperature monitoring methods are applied to the casing; installing sensors internally may cause insulation problems. Numerical simulation techniques are often used as an economical and effective method to significantly reduce the need for necessary GIL temperature rise testing. Currently, commonly used numerical simulation methods include Computational Fluid Dynamics (CFD) and Thermal Network Models (TNM). CFD methods solve electromagnetic-fluid-thermal coupled partial differential equations (PDEs) using the finite element method (FEM), finite volume method (FVM), or finite difference time-domain method (FDTD).

[0004] CFD methods typically involve building two-dimensional or three-dimensional models to calculate the thermal field of a gas-liquid interconnect (GIL), coupling multiphysics calculations, and analyzing the temperature distribution of the GIL during operation. Mappings have been studied under different gas pressures, line currents, and ambient temperatures, corresponding to various practical operating conditions. However, the heat generation process of the GIL involves the coupling of electromagnetic, thermal, and flow fields, and these complex properties determine the significant computational cost. TNM methods, on the other hand, construct equivalent thermal circuit models, but rely on empirical parameters or assumptions for rapid optimization design. The inherent simplification assumptions of TNM methods and the challenge of accurately capturing multiphysics distributions further limit their versatility and accuracy. In summary, existing technologies lack a method for predicting the temperature rise characteristics of GILs that can reduce computational costs while improving versatility and accuracy. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention proposes a data-driven model-based method for predicting the temperature rise of gas-insulated transmission lines, thereby resolving the issues present in the prior art.

[0006] To achieve the above objectives, this invention provides a method for predicting the temperature rise of gas-insulated transmission lines based on a data-driven model, comprising:

[0007] A simulation model of a gas-insulated transmission line with multi-field coupling is constructed, and the simulation model is simulated to obtain sample data, including temperature data and temperature rise index characteristics during the simulation process.

[0008] A fusion prediction model is constructed, wherein the fusion prediction model adopts a model that combines BiTCN model, BiGRU model and XGBoost layer. Based on temperature sample data, the fusion prediction model is optimized by the intelligent lemming algorithm.

[0009] The optimized fusion prediction model is used to predict the temperature data and temperature rise characteristics of the gas-insulated transmission line in real time, and the temperature data of the gas-insulated transmission line is obtained.

[0010] Optionally, a simulation model can be constructed and simulated using computational fluid dynamics methods, wherein the simulation model takes into account the electromagnetic field-thermal field-flow field coupling.

[0011] Optional,

[0012] The electromagnetic field-thermal field coupling relationship in the simulation model includes:

[0013] ;

[0014] ;

[0015] in, Indicator element Ohmic loss, of which and These are the current density and the conjugate current density, respectively. It is the total number of finite element components for the conductive part. Indicates the area of ​​the component. The absolute temperature is The electrical conductivity of the material at that time; The electrical conductivity of the material at 20℃; The temperature coefficient of resistivity This refers to absolute temperature.

[0016] Optionally, the temperature field-flow field coupling relationship in the simulation model includes:

[0017] ;

[0018] ;

[0019] ;

[0020] ;

[0021] ;

[0022] in, They are respectively direction and Directional gas flow velocity; It is the acceleration due to gravity; This is due to the density difference caused by gas expansion; These represent the density, thermal conductivity, and dynamic viscosity of the gas, respectively. This refers to gas pressure. As a volumetric heat source; The blackbody radiation coefficient is... The outer diameter of the conductor; These are the conductor temperature and the casing temperature, respectively. For system blackness, The outer diameter of the outer casing; The blackness of the outer surface of the casing; Ambient temperature; The occlusion coefficient is... This represents the energy radiated from the surface of a conductor. This represents the radiant energy emanating from the shell surface. For time, For constant pressure heat capacity, T represents absolute temperature.

[0023] Optionally, the electromagnetic field mathematical model in the simulation model is as follows:

[0024] ;

[0025] ;

[0026] in, It is a magnetic vector potential. It is an electric scalar potential. It is the permeability. Indicates electrical conductivity. It is angular frequency. It is the applied source current density within the conductive rod. This represents the Hamiltonian operator, where j is the imaginary unit.

[0027] Optionally, the fusion prediction model includes an input layer, a BiTCN layer, a BiGRU layer, an XGBoost layer, and an output layer connected in sequence.

[0028] Optionally, the process of optimizing the fusion prediction model using the intelligent lemming algorithm includes:

[0029] Initial candidate solutions are constructed based on the hyperparameters of the fusion prediction model. The initial candidate solutions include different individuals, each of which includes a set of hyperparameters of the fusion prediction model. The hyperparameters include the learning rate of the BiTCN model and the BiGRU model, the number of neurons, the number of filters, and the regularization factor of the BiGRU model, and the maximum number of iterations, depth, and learning rate of the XGBoost model.

[0030] Fitness is calculated for individuals in the initial candidate solutions. Based on the fitness calculation results, the current optimal solution is obtained. The energy coefficient is calculated based on the optimal solution. The energy coefficient is used to determine whether to proceed to the exploration phase or the development phase. The next candidate solution is generated based on the exploration phase and the development phase. Based on the next candidate solution, the process of fitness calculation, generation of the current optimal solution, and generation of the next candidate solution is repeated until the maximum number of iterations is reached. The optimal hyperparameters of the fusion prediction model are obtained to optimize the fusion prediction model. The fitness is calculated based on the accuracy of the fusion prediction model.

[0031] Optionally, the objective function of the intelligent lemming algorithm is calculated based on the mean absolute error, mean square error, root mean square error, and coefficient of determination of the fusion prediction model.

[0032] On the other hand, the present invention also provides a method for predicting the temperature rise of gas-insulated transmission lines based on a data-driven model, for performing the above-described method.

[0033] On the other hand, the present invention provides a computer-readable storage medium having stored thereon computer program instructions which, when executed by a processor, implement the steps of the above-described method.

[0034] Compared with the prior art, the present invention has the following advantages and technical effects:

[0035] To achieve low-cost and accurate prediction of the temperature rise process of GIL, this invention constructs a GIL model considering electromagnetic-thermal-fluid multiphysics coupling to provide training data, develops an ALA-BiTCN-BiGRU-XGBoost temperature rise prediction model, and proposes a temperature rise estimation strategy that combines CFD simulation data with a data-driven method. The following three main contributions distinguish this invention from existing technologies: (1) A data-driven method-based GIL temperature rise prediction model is proposed, which avoids the complex analysis of multiphysics coupling, achieves high prediction accuracy with only a small amount of training data and low cost, and is also universal. (2) The BiTCN-BiGRU feature extraction technique is used to improve the prediction performance of XGBoost, and the ALA algorithm is introduced to handle the hyperparameter adjustment problem to avoid manual parameter tuning. (3) Compared with the data in existing technologies, the newly developed GIL temperature rise prediction model has a lower error, while avoiding the problems existing in CFD and TNM methods when predicting GIL temperature rise. Attached Figure Description

[0036] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0037] Figure 1 This is a schematic diagram of the GIL heat transfer process according to an embodiment of the present invention;

[0038] Figure 2 This is a schematic diagram of the bidirectional dilated causal convolutional network architecture according to an embodiment of the present invention;

[0039] Figure 3 This is a flowchart of the BiGRU architecture according to an embodiment of the present invention;

[0040] Figure 4 This is a flowchart of the ALA-BiTCN-BiGRU-XGBoost temperature rise prediction model architecture according to an embodiment of the present invention;

[0041] Figure 5 This is a schematic diagram of the current density distribution under a current of 2600A according to an embodiment of the present invention;

[0042] Figure 6 This is a schematic diagram of the temperature and fluid distribution under a current of 2600A according to an embodiment of the present invention. Detailed Implementation

[0043] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0044] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0045] Compared to CFD and TNM methods, data-driven methods learn the temperature rise patterns of the GIL (Gas Inertial Ion) from historical data, avoiding the analytical requirements of complex multiphysics coupling equations. This invention addresses the GIL temperature rise prediction problem by proposing a hybrid model integrating a bidirectional temporal convolutional network (BiTCN), a bidirectional gated recurrent unit (BiGRU), and extreme gradient boosting (XGBoost), and incorporating the Intelligent Lemming Algorithm (ALA) for hyperparameter optimization. This architecture achieves high-precision temperature rise prediction based on multi-dimensional temporal features by integrating GIL temperature field monitoring data. In this research context, data-driven methods provide a new path for solving the GIL temperature rise problem.

[0046] As a primary carrier for long-distance, high-capacity power transmission in complex environments, the real-time and accurate prediction of conductor temperature in gas-insulated transmission lines (GILs) is crucial for ensuring their operational reliability and stability. This invention proposes a fusion-driven prediction model optimized using the Artificial Lemming Algorithm (ALA) to achieve high-precision prediction of the temperature rise effect of the inner conductor in GILs. The proposed model employs a BiTCN-BiGRU-XGBoost triple-layer composite architecture, where the bidirectional temporal convolutional network (BiTCN) and bidirectional gated recurrent unit (BiGRU) combine the powerful learning ability of XGBoost and the hyperparameter optimization characteristics of the ALA algorithm to enhance the adaptive extraction capability of multi-dimensional features. This improves upon the XGBoost model's reliance on manual feature extraction and insufficient generalization. Furthermore, computational fluid dynamics was used to simulate the electromagnetic-thermal-fluid coupling model of the GIL, obtaining the training data for this invention. The trained model exhibits good prediction accuracy and effectively avoids the repetitive modeling process under different operating conditions. Experimental results show that the maximum relative error between the fusion model and the measured data is 0.51%, and the minimum relative error is only 0.11%, which proves the high accuracy of the fusion model in predicting the temperature rise of GIL.

[0047] The technical solution is explained in light of the above:

[0048] The relevant technical process of this invention is as follows:

[0049] S1. Using CFD methods, a multi-field coupled GIL simulation model is built. During the construction process, electromagnetic-thermal-flow field coupling is considered. Under the condition of electromagnetic-thermal-flow field coupling, the GIL simulation model is simulated, and the temperature rise index characteristics of the GIL simulation model are determined to obtain the time-series temperature at different points of the GIL simulation model, thereby constructing a dataset.

[0050] S2. A BiTCN-BiGRU-XGBoost model is constructed, using temperature rise index characteristics and temperature time-series data as input. Specifically, the input includes real-time monitoring data of the shell's top, bottom, left, and right vertex temperatures, and the shell's average temperature from a GIL simulation model built using CFD methods. The temperature at the corresponding point on the conductor is the final output. The aforementioned BiTCN-BiGRU-XGBoost model comprises an input layer, a BiTCN layer, a BiGRU layer, an XGBoost layer, and an output layer connected sequentially. The model is optimized using the intelligent lemming algorithm, where relevant parameters are used as the solution content, and accuracy is used as the target value. The optimized model is then used as the solution tool, with the required temperature rise index characteristics and temperature time-series data acquired in real-time as input to obtain real-time predicted model temperature data.

[0051] S1. Building a GIL simulation model considering electromagnetic-thermal-flow field coupling:

[0052] 1.1. GIL Simulation Model Setup:

[0053] To obtain the training and test set data required for the prediction model, this invention constructs three types of GIL simulation models, including two models applicable to voltage conditions above 1000kV and one model for 550kV. Specific parameters for the three types of GILs are shown in Table 1, which contains information on material property parameters.

[0054] Table 1

[0055]

[0056] In this table, A represents GIL applicable to 550kV models, while B and C represent two other types of GIL applicable to voltage levels above 1000kV. The parameters shown in Table 1 are material properties at 20 degrees Celsius.

[0057] 1.2. GIL Multiphysics Coupling:

[0058] 1.2.1 Mathematical Model of Electromagnetic Field:

[0059] The reason why GILs generate heat is that when a GIL carries a large current, based on the inherent resistive characteristics of the conductor material, ohmic losses occur between the conductor and the metal casing. The resulting heat energy diffuses to the surrounding environment in three ways: heat conduction, heat radiation, and heat convection. Figure 1 The distribution characteristics of this heat transfer path are presented intuitively.

[0060] To simplify calculations and ensure the scientific validity and effectiveness of the electromagnetic field calculation model, this invention makes the following assumptions:

[0061] 1) The outer casing provides good shielding, so the proximity effect of the GIL pipe can be disregarded.

[0062] 2) For low-frequency alternating current, the effects of space charge and displacement current can be ignored.

[0063] 3) The conductor and shell materials are isotropic linear media with constant relative permeability and electrical conductivity that is only a function of temperature.

[0064] The ohmic loss of the conductive rod current and the induced current within the casing can be quantified by solving the eddy current field equations. The eddy current field control equations established by GIL in the frequency domain are as follows:

[0065] (1)

[0066] (2)

[0067] in, It is the magnetic vector potential (MVP). It is an electric scalar potential (ESP). It is the permeability. Indicates electrical conductivity. It is angular frequency. It is the applied source current density within the conductive rod. This represents the Hamiltonian operator, where j is the imaginary unit.

[0068] 1.2.2 Electromagnetic field-temperature field coupling:

[0069] The conductor and the outer shell are finely divided using the finite element method. Through the finite element division, the loss per unit length in the conductor can be obtained by solving the eddy current field as shown in equation (3).

[0070] (3)

[0071] in Indicator element Ohmic loss, of which and These are the current density and the conjugate current density, respectively. It is the total number of finite element components for the conductive part. The absolute temperature is The electrical conductivity of the material at that time The area represents the component. Since the conductivity of the GIL conductor and the shell material changes with temperature, the power loss during actual operation is also affected by temperature. In order to accurately calculate the power loss, it is necessary to perform bidirectional coupling of the electromagnetic field and the temperature field. The bidirectional coupling of the electromagnetic field and the temperature field is essentially an iterative calculation of conductivity. Under normal circumstances, the resistivity of metallic materials changes linearly with temperature, and the relationship between its conductivity and temperature can be expressed as shown in equation (4).

[0072] (4)

[0073] In the formula: The absolute temperature is The electrical conductivity of the material at that time; The electrical conductivity of the material at 20℃, where T represents the absolute temperature; is the temperature coefficient of resistivity, with a typical value of 0.0042.

[0074] 1.2.3 Temperature field-flow field coupling:

[0075] Heat transfer in the GIL includes three modes: conduction, convection, and radiation. To accelerate model convergence, the GIL temperature field calculation in this invention is simplified as follows:

[0076] 1) Since the gas in the GIL is thin and its heat reflection capability is very weak, the radiative heat transfer of the gas in the model is ignored.

[0077] 2) Assume the GIL is indoors, so the effects of wind speed and solar radiation on temperature rise are not considered.

[0078] 3) The air boundary is far enough away from the GIL that its temperature is not affected by the heat generated by the GIL.

[0079] 4) The density, viscosity, and thermal conductivity of SF6 gas and air are all temperature-dependent, but the specific heat of the gas is considered constant.

[0080] (1) Natural convection heat transfer:

[0081] The gas flow inside the GIL is induced by conductor heating, which is a natural convection phenomenon. In this model, the movement of air and SF6 is a low-speed (Mach number Ma < 0.3) Newtonian fluid flow, so a laminar flow model is used for calculation. For the two-dimensional temperature field of this model, the natural convection of the gas is described by the Navier-Stokes equations. This process follows the conservation of mass, momentum, and energy, as shown in equations (5)-(7), respectively.

[0082] (5)

[0083] (6)

[0084] (7)

[0085] In the formula: They are respectively direction and Directional gas flow velocity; It is the acceleration due to gravity; This is due to the density difference caused by gas expansion; These represent the density, thermal conductivity, and dynamic viscosity of the gas, respectively. This refers to gas pressure. The heat source is a volumetric heat source, where x and y represent the horizontal and vertical coordinates, respectively. This represents the constant-pressure specific heat capacity of the corresponding gas. For time, For constant pressure heat capacity, T represents absolute temperature.

[0086] (2) Radiative heat transfer;

[0087] In the heat transfer analysis of gas-insulated transmission lines (GILs), the radiative heat transfer mechanism mainly involves two key regions: one is the enclosed space between the outer surface of the conductor rod and the inner surface of the shell, and the other is the open area between the outer surface of the shell and the surrounding environment. Since the reflection effect of the gas medium is usually ignored in the system design, the path of radiative energy exchange is simplified to occur directly between the two pairs of surfaces mentioned above. For the enclosed annular space formed by the conductor and the shell, its radiative heat transfer characteristics conform to the heat transfer model between long concentric cylinders. Therefore, the radiative heat transfer between the radiating surface and the absorbing surface can be expressed as shown in Equation (8):

[0088] (8)

[0089] In the formula: This represents the energy radiated from the surface of a conductor. The blackbody radiation coefficient is taken as 5.67 W / (m·K); Let the outer diameter of the conductor be m; These are the conductor temperature and the casing temperature, respectively, in K; The system's emissivity is calculated using the following formula:

[0090] (9)

[0091] In the formula: These represent the blackness of the conductor's outer surface and the blackness of the shell's inner surface, respectively. Let the outer diameter of the conductor be m. Let be the inner diameter of the casing, in meters (m). The radiation from the GIL casing to the surrounding air is equivalent to the radiative heat dissipation of a long cylinder. For a horizontally installed GIL, the radiative heat dissipation per unit length of a single-phase casing is:

[0092] (10)

[0093] In the formula: The blackbody radiation coefficient is... This represents the radiant energy emanating from the shell surface. Where is the outer diameter of the outer shell, in meters (m); The blackness of the outer surface of the casing; The ambient temperature, in K; Let K be the conductor temperature. This is the occlusion coefficient, which is 0 when there is no occlusion.

[0094] S2. The proposed BiTCN-BiGRU-XGBoost prediction strategy:

[0095] 2.1. Sub-components of the ALA-BiTCN-BiGRU-XGBoost fusion prediction model:

[0096] 2.1.1 Bidirectional Temporal Convolutional Networks:

[0097] Temporal Convolutional Networks (TCNs) combine the advantages of CNNs and RNNs, possessing causal convolutions, dilated convolutions, and residual connections. This helps alleviate gradient vanishing, and compared to RNNs / LSTMs, TCNs can process the entire sequence in parallel. Traditional TCN methods only perform forward convolution calculations on the input sequence, neglecting the information potentially hidden in the backward features. Therefore, this invention employs a BiTCN structure to capture both forward and backward features in GIL temperature field data, thereby improving information acquisition capabilities and ultimately enhancing overall accuracy. The bidirectional dilated causal convolution structure is as follows: Figure 2 As shown.

[0098] The core objective of causal convolution in TCNs is to ensure that the model relies only on historical information during prediction, thereby preserving the causality of the time series. In the input sequence... The output is then obtained through the network. In the BiTCN network, the output at each time t is... It only depends on the input before the current time t. Causal convolution requires deeper network layers to capture historical information dependencies over a longer time range, while dilated convolution, by adjusting the dilation factor, allows TCN to flexibly expand the receptive field without pooling operations while maintaining the integrity of the information. For one-dimensional input... and convolution kernel ,but The dilated convolution at the point is shown in equation (11). Wherein, Represents the kernel size. Represents the past direction. It is the expansion rate. This represents the weights of the convolution kernel.

[0099] (11)

[0100] The dilation factor controls the insertion spacing of zero vectors between convolutional kernels, achieving an exponential expansion of the receptive field. In BiTCN, layer-by-layer dilated convolutions enable the network to quickly capture long-range dependencies, but an excessively large receptive field can lead to gradient vanishing and a decrease in convergence speed. Introducing a residual block structure can effectively maintain gradient stability while improving the efficiency of sequence feature extraction. Residual blocks are as follows: Figure 2 As shown in (b) of the diagram.

[0101] 2.1.2 Bidirectional Gated Loop Unit:

[0102] Bidirectional Gated Recurrent Unit (BiGRU) is a variant of recurrent neural network (RNN) that combines a bidirectional structure with GRU units. It addresses the vanishing / exploding gradient problem of traditional RNNs through a gating mechanism. Compared to LSTM, which has three gates, GRU reduces model complexity while retaining the core ideas of LSTM, including update and reset gates. It requires fewer parameters and lower training resource requirements. The basic structure of GRU is as follows: Figure 3 As shown in (a), it includes an update gate. and reset door The network propagation process of GRU can be described as follows: Equation (12)-Equation (15):

[0103] (12)

[0104] (13)

[0105] (14)

[0106] (15)

[0107] in, This indicates that the door is being reset. Indicates an update to the door. This represents the Sigmoid function. This represents the weight matrix of the reset gate input vector. This represents the input vector at time t. The weight matrix represents the weights used to reset the hidden state of the door. This represents the hidden state at time t-1. This represents the bias matrix of the reset gate. This represents the weight matrix of the update gate input vector. The weight matrix represents the weights used to update the hidden state of the gate. This represents the bias matrix of the update gate. The weight matrix representing the candidate hidden state. The weight matrix represents the input vector of the candidate hidden state. The bias matrix represents the candidate hidden state. Represents the hyperbolic tangent function. It is the Hadamard product. Indicates the candidate hidden state. This is the final output. Similar to BiTCN, BiGRU consists of two independent GRUs and can handle forward and reverse sequences. Its structure is as follows: Figure 3 As shown.

[0108] 2.1.3 Limiting Gradient Boosting:

[0109] Extreme Gradient Boosting (XGBoost) is based on the boosting ensemble learning framework. It iteratively generates a sequence of weak learners, where each new base learner is specifically fitted to the prediction residuals of the previous model. This algorithm uses the gradient boosting framework to optimize the decision tree base model, accelerates convergence through second-order derivative information, and introduces a regularization term to control model complexity. It has significant advantages in capturing complex nonlinear relationships and high-order feature interactions in the data. Its mathematical essence is to construct an ensemble model with strong generalization ability through the mapping relationship between the feature matrix and the target variable. The objective function minimized by XGBoost is defined in formula (16):

[0110] (16)

[0111] in It's a real label. It is a predicted label. It is a loss function. This is the regularization term, where n represents the number of samples. This represents the i-th real label. This represents the i-th predicted label. Represents the prediction function. Regularization term. Used to control the complexity of the model. The definition is as follows:

[0112] (17)

[0113] The prediction function is The number of leaf nodes in the tree is The characteristic number is The regularization parameter is The L2 regularization parameter is , No. The weights of the leaf nodes are .

[0114] 2.1.4 Lemming Optimization Algorithm:

[0115] The Artificial-Lemming-Algorithm (ALA) is based on the characteristics of lemmings, namely, their long-distance migration, burrowing, foraging, and predator avoidance behaviors. Long-distance migration and burrowing represent the exploratory behavior of ALA, while foraging and predator avoidance reflect its exploitative behavior. This algorithm mathematically models these four behaviors to optimize the solution of the problem, better balancing exploration and exploitation while maintaining computational efficiency. It effectively addresses challenges such as premature convergence, insufficient exploration, and lack of robustness in high-dimensional, non-convex search spaces. This invention optimizes the learning rate, number of BiGRU neurons, number of filters, regularization factor, and maximum number of iterations, depth, and learning rate of XGBoost using the ALA algorithm, thereby improving prediction accuracy. The ALA optimization of BiTCN-BiGRU-XGBoost consists of the following five main steps.

[0116] (1) Initialization phase:

[0117] Before performing iterative optimization in the ALA algorithm, a population initialization phase must be completed. The initial candidate solution set is a matrix, determined by the population size. Number of rows and dimensions ( It consists of columns.

[0118] (18)

[0119] (19)

[0120] In the formula, This represents the candidate solution set. This represents the element corresponding to the i-th individual in the j-th dimension. It is a random value in the range [0-1]. express The lower bound of the dimension yes The upper limit of dimensions.

[0121] (2) Exploration phase:

[0122] This stage can be divided into two exploration methods: long-distance migration and burrowing. The choice between these two methods is randomized; 30% of individuals are randomly assigned to use long-distance migration, while the remaining individuals use burrowing. During long-distance migration, lemmings explore the search space based on their current location and the locations of random individuals in the population, aiming to find habitats with abundant food resources and thus obtain better living conditions and resources. The mathematical model is as follows:

[0123] (20)

[0124] (twenty one)

[0125] (twenty two)

[0126] In the formula, For the current energy factor, Indicates the first During the second search The position of the next iteration. This represents the current optimal solution. The ALA algorithm uses the root mean square error (RMSE) as the objective function, and the optimal solution for each iteration is determined by this. Used as a flag to change the search direction, it helps avoid local optima. This represents a vector that is characterized by Brownian motion. It is a size of The vector. Indicates the first Location during search, Denotes a randomly selected individual in the population for the search, where , This represents the maximum number of iterations, and t represents the current time. It is a random value within the range of [0-1].

[0127] Meanwhile, when lemmings explore by digging, they will randomly dig new burrows based on the current location of their existing burrow and the location of random individuals within the population. This behavior provides them with shelter and serves as a food storage method, and its mathematical expression is shown in formulas (23-24), where... It is a random number related to the current iteration number. This refers to a search individual randomly selected from the population. .

[0128] (twenty three)

[0129] (twenty four)

[0130] (3) Development phase:

[0131] Lemmings move extensively and freely within their burrows, relying on their keen sense of smell and hearing to locate food sources. The mathematical model for this is as follows:

[0132] (25)

[0133] (26)

[0134] (27)

[0135] in, This represents a spiral shape indicating random searching during foraging. Indicates the radius of the foraging range. It is a random value in the range [0-1]. This represents the best element in the j-th dimension at time t. This represents the corresponding element of the i-th individual in the j-th dimension at time t.

[0136] Ultimately, ALA focused on the avoidance and protective behaviors of lemmings when faced with danger, employing deceptive maneuvers to evade predators. Its mathematical model is as follows:

[0137] (28)

[0138] (29)

[0139] In the formula, This indicates the lemming's ability to escape. This represents the maximum number of iterations. This is the Levy flight function, used to simulate the deceptive maneuvers of a rabbit during its escape.

[0140] 2.2. ALA-BiTCN-BiGRU-XGBoost fusion prediction model:

[0141] This invention proposes a GIL (Gas Intake System) temperature rise prediction method based on a BiTCN-BiGRU-XGBoost fusion approach. It extracts temperature rise index features from a CFD simulation model. The input includes real-time monitoring data of the shell's top, bottom, left, and right vertex temperatures, as well as the average shell temperature, from a CFD-built GIL simulation model. This data is then passed to the XGBoost model to learn the temperature rise index sequence. The developed fusion data-driven model comprises four parts: an input layer, a BiTCN-BiGRU layer, an XGBoost layer, and an output layer. BiTCN captures the hidden features of the temperature sequences at different points on the GIL shell in both the forward and reverse directions. The output of BiTCN is further processed by BiGRU. BiGRU excels at capturing the forward and reverse dependencies of the temperature sequences at different points on the GIL shell and can simultaneously consider past and future information. This hybrid feature extraction architecture fully leverages the advantages of both methods to handle complex temporal dependencies. The basic flow of the fusion prediction model is as follows: Figure 4 As shown.

[0142] This invention employs a BiTCN-BiGRU hybrid neural network architecture to extract multi-level features from input temperature data. It captures local temporal features through a bidirectional temporal convolutional network and models long-range dependencies using bidirectional gated recurrent units, generating a high-dimensional feature representation. After feature compression, the compressed features are input into an XGBoost ensemble learning model for training, while an ALA optimization algorithm is introduced to dynamically adjust hyperparameters.

[0143] To evaluate the accuracy of this method in predicting GIL temperature rise, the mean absolute error (MAE), mean square error (MSE), root mean square error (RMSE), and coefficient of determination R were used. 2 To verify the accuracy of the model proposed in this invention. The calculation process of the performance evaluation indicators MAE, MSE and RMSE is mathematically expressed as Equation (30)-Equation (32):

[0144] (30)

[0145] (31)

[0146] (32)

[0147] In the formula, This represents the predicted temperature value. This represents the measured temperature value, and N represents the total sample size.

[0148] Based on relevant data, the experimental results for the above technical solution are described below:

[0149] 3. The experimental design of this invention is as follows: 1) For Type A GILs, CFD simulation experiments were conducted at three provided temperatures. The experimental environment temperature for current carrying capacity in the range of 6200A-5000A was 297K, starting at 6200A, with subsequent experiments conducted every 300A, for a total of 5 sets of data. The experimental environment temperature for current carrying capacity in the range of 4700A-3500A was 289K, and the experimental environment temperature for current carrying capacity in the range of 3200A-2000A was 299K, starting at 4700A and 3200A respectively, with subsequent experiments conducted every 300A, collecting 5 sets of experimental data for each. 2) For Type B and Type C GILs suitable for 1000kV voltage levels, CFD simulation experiments were conducted at a temperature of 293.15K. The current carrying capacity was designed to be between 8500A and 4000A, with subsequent experiments conducted every 500A, for each type, collecting 10 sets of experimental data. The experimental data mentioned above will provide a sufficient data foundation for verifying the superiority of the ALA-BiTCN-BiGRU-XGBoost method.

[0150] 3.1. Simulation model verification:

[0151] To systematically verify the accuracy of the GIL model constructed in this invention, the assumptions of the GIL in the simulation are consistent with the experimental conditions outlined in the literature. Here, the conductor temperature rise characteristics of type A GIL under rated current loads of 2600A, 4400A, and 5600A are simulated and verified. Taking the experimental results of 2600A as an example, the numerical results of the two-dimensional eddy current field distribution of the GIL system under power frequency current conditions show that the peak magnetic induction intensity generated on the conductor surface due to the skin effect reaches... The outer shell, through the electromagnetic barrier constructed by the eddy current antimagnetic effect, effectively reduces the magnetic field energy radiated outward by the system. This provides theoretical support for the safe deployment of GIL in sensitive electromagnetic environments.

[0152] In this invention, the emissivity of the outer wall of the conductive rod and the outer wall of the outer shell are simultaneously set to 0.8 and 0.9, respectively. The steady-state electromagnetic-thermal-fluid coupled field is solved at a current of 2600A, where... Figure 5 The distribution of current density is shown. Due to the electromagnetic induction generated by the current flow in the conductive rod, currents of equal magnitude but opposite direction will be induced in the enclosed space. Due to the skin effect, the current in the conductive rod is concentrated on the outer surface, while the induced current generated in the GIL shell is concentrated on the inner surface. Meanwhile, the steady-state temperature field and internal flow field of the GIL are obtained as follows: Figure 6 As shown in (a) and (b):

[0153] The small cross-sectional area of ​​the conductive rod results in a high loss density, causing the maximum temperature of the conductive rod to reach 315.6K. Figure 6It can be seen that the temperature at the top and bottom of the conductive rod is the same. Meanwhile, under the natural convection of the insulating gas, the temperature at the top apex of the shell is about 1.2K higher than that at the bottom apex, which verifies the conclusions of experimental data in existing technologies. The final experimental statistical results are shown in Table 2, revealing the temperature rise characteristics of the GIL system under different current-carrying conditions. For three typical operating currents of 2600A, 4400A, and 5600A, the numerical simulation and experimental data show good agreement. The steady-state temperature rise of the conductive rod in CFD is 16.6K, 42.1K, and 59.5K, respectively. The maximum deviation from the experimental value is 1.6K. Furthermore, the temperature rise at the top of the shell in CFD is 5.8K, 18.2K, and 26.8K, respectively. The temperature gradients between the bottom and top of the shell are 1.2K, 1.4K, and 5.6K, respectively, verifying the thermal convection effect of SF6 gas. Table 2 compares the results obtained from CFD simulation and experimental data in existing literature.

[0154] Table 2

[0155]

[0156] As shown in Table 2, with increasing load current, both existing experimental and CFD simulations demonstrate a gradual increase in the temperature difference between the top and bottom of the GIL casing. Furthermore, the data reveals that as the current increases, the difference between the temperature calculated by CFD at the top of the casing and the actual value widens. This is because the experiment was conducted in a closed indoor space, where the actual ambient temperature undergoes dynamic changes. In the simulation model constructed in this invention, the air domain is finite, and a constant temperature boundary condition is applied at the air boundary.

[0157] The above analysis verifies the effectiveness of the CFD model constructed in this invention for calculating the temperature field of GILs. This invention will use three GIL models (A, B, and C) as verification objects, and, combined with data obtained from the CFD model, employ the ALA-BiTCN-BiGRU-XGBoost method to accurately predict the transient temperature rise process of GILs.

[0158] 3.2. Dataset:

[0159] In actual GIL operation, the internal conductor temperature is affected by multi-physics coupling effects, with key influencing factors including load current, ambient temperature, wind speed, and the density and isobaric heat capacity of the insulating gas. To balance the measurability of engineering data with the generalization requirements of different insulating gas inversion models, this invention considers the variations in the top, bottom, and bilateral temperatures of the casing, as well as the average casing temperature. A sample set is constructed using CFD simulation data to establish a high-precision temperature rise prediction model for the GIL's internal conductors, integrating physics modeling and data-driven approaches. This solution ensures the measurability of engineering data while also addressing the model's generalization requirements under different GIL operating conditions.

[0160] This invention uses CFD numerical simulation to compare the thermal field distribution characteristics of three types of gas-insulated liquids (GILs) (A, B, and C) under different current-carrying conditions. Based on the time-temperature rise curves of the conductor top, shell top, shell bottom, shell left side, shell right side, and shell average temperature, the temperature field evolution of the GIL is characterized by a typical exponential growth pattern in the time-varying temperature curves at all monitoring points, gradually converging to a steady-state value within a 200-400 minute time domain. Due to the relatively low thermal conductivity of SF6 gas, increasing the current-carrying capacity exacerbates the temperature difference between the conductor and the shell.

[0161] Based on experimental data, this invention divides the Type A GIL into training and testing sets. Specifically, for the Type A GIL, four current conditions (5600A, 5300A, 4400A, and 2600A) are selected as the testing set, while all other current conditions are included in the training set for model training.

[0162] 3.3. Estimation Results:

[0163] To better demonstrate the advantages of the proposed method, this invention introduces four other prediction models for comparison. For ease of presentation later, M1, M2, M3, M4, and M5 are described in abbreviation form as LSTM, TCN, XGBoost, ALA-XGBoost, and ALA-BiTCN-BiGRU-XGBoost, respectively. This invention mainly focuses on the transient temperature rise process of GILs. Although the prediction accuracy of different data-driven methods varies, they can all effectively achieve dynamic tracking of the temperature at the top of the conductive rod inside the GIL. Observation shows that the system exhibits temperature stability characteristics at 22,000 seconds, at which point the temperature curve tends to flatten and no longer increases with time. By 36,000 seconds, the relative error between the temperature monitoring data and the theoretical steady-state value is less than 0.1%.

[0164] Experimental data show that the temperature rise characteristics of type A GIL under different thermodynamic boundary conditions are as follows: all data-driven methods have the ability to predict temperature rise, but they show significant performance differences: in the initial stage (t<5000s), there is a large prediction deviation. After parameter optimization, the prediction accuracy is improved to a certain extent. After structure-algorithm co-optimization, high-precision tracking prediction throughout the entire cycle can be achieved (mean absolute error 0.5%).

[0165] The four key evaluation metrics for the Type A GIL temperature rise prediction model are: MAE, MSE, RMSE, and the coefficient of determination R². A higher R² value indicates a better model fit, and lower error metrics (MAE / MSE / RMSE) indicate better prediction accuracy. Experimental data show that the M5 model exhibits optimal prediction performance under different flow rates. The maximum MAE, MSE, and RMSE errors all do not exceed 0.262 K under all experimental conditions. When the flow rate is 2600 A, all evaluation metrics of the M5 method do not exceed 0.032 K, demonstrating the high accuracy of the prediction method proposed in this invention. The R² value of the M5 method under all operating conditions for Type A GIL... 2 Both are 0.999, and the R-value of the M1-M4 method is also 0.999. 2 The value of at least 0.991 further demonstrates the feasibility of the data-driven method for predicting the temperature rise of the GIL. Table 3 presents the analysis of the steady-state temperature rise prediction results for the GIL using multiple models.

[0166] Table 3

[0167]

[0168] Table 3 presents the steady-state error indices of the various GIL temperature rise prediction methods of this invention. Experimental data shows that the maximum steady-state errors of each model are M1 (1.49K), M2 (0.96K), M3 (1.44K), and M4 (0.89K), with the maximum steady-state error of M5 not exceeding 0.26K and the maximum relative error not exceeding 0.08%. The temperature rise prediction errors of all methods are below the 1.5K threshold, fully verifying the effectiveness of the data-driven method. Analysis of the comparison between data M3 and M4 in the table shows that the ALA method can effectively reduce model errors, while M5 exhibits the best accuracy, verifying its high-resolution modeling capability for the GIL temperature rise process. Therefore, high-fidelity CFD simulation provides a sufficient and reliable training data foundation for the data-driven model. Taking type A GIL as an example, the comparison results with the data obtained from previous researchers' experiments are shown in Table 4: Table 4 compares the experimental results with the experimental data in the literature.

[0169] Table 4

[0170]

[0171] As shown in the comparative analysis in Table 4, the M5 method, trained with high-precision CFD simulation data, achieved high-resolution modeling. The relative error between the model prediction and the actual measurement was controlled within 0.5%, verifying the accuracy of the method in practical engineering thermodynamic modeling.

[0172] This invention proposes a novel method for calculating the temperature rise of gas-cooled induction generators (GILs) by integrating electromagnetic-thermal-fluid multiphysics coupled CFD calculation data with the ALA-BiTCN-BiGRU-XGBoost method. Experimental comparisons were conducted on three types of GILs under two voltage levels and different load currents, demonstrating the feasibility and high accuracy of the proposed method. The main conclusions of this invention are as follows:

[0173] This invention uses CFD numerical simulation to build a simulation model of GIL, and the simulation accuracy reaches within 1.58%, providing sufficient data sources for training the data-driven model proposed in this invention.

[0174] This invention uses three types of GILs to build simulation models and obtains a total of 35 different sets of data through CFD calculations, including 15 sets of type A GILs, 10 sets of type B GILs, and 10 sets of type C GILs. The training and test sets of type A GILs are divided into 11 sets for training and 4 sets for testing. The division method for types B and C is the same, with 6 sets for training and 4 sets for testing.

[0175] This invention proposes using the ALA method to optimize the BiTCN-BiGRU-XGBoost fusion model, where BiTCN-BiGRU is responsible for feature extraction from the training data, and XGBoost is responsible for learning and predicting the final GIL temperature change. Compared with the other four proposed data-driven models, this model has the best accuracy, with a maximum relative error of no more than 0.08% compared with CFD calculation results.

[0176] The ALA-BiTCN-BiGRU-XGBoost fusion model proposed in this invention has a maximum relative error of 0.5% when compared with the measured data of type A GIL, which can meet engineering requirements. At the same time, the data-driven method proposed in this invention provides a new approach for predicting the temperature rise of GIL.

[0177] With the deepening application of data-driven methods in energy systems, this invention proposes a novel GIL (Gas Intake System) temperature rise prediction framework that integrates data-driven methods and CFD (Computational Fluid Dynamics) calculation data. Research results show that this hybrid modeling method effectively characterizes the GIL temperature rise pattern by coupling multiphysics calculation data with equipment simulation characteristics, providing new methodological support for intelligent monitoring of power transmission equipment. However, the boundary conditions of the CFD calculation results and GIL simulation model used in this invention still deviate from actual operating conditions, resulting in certain limitations of the method. Future research will focus on constructing a learning framework that integrates multiphysics calculation results and measured data, improving the boundary condition constraint mechanism and increasing the dimensionality of the training dataset to enhance the model's generalization ability to different operating conditions.

[0178] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for predicting temperature rise in gas-insulated transmission lines based on a data-driven model, characterized in that, include: A simulation model of a gas-insulated transmission line with multi-field coupling is constructed, and the simulation model is simulated to obtain sample data, including temperature data and temperature rise index characteristics during the simulation process. A fusion prediction model is constructed, wherein the fusion prediction model adopts a model that combines BiTCN model, BiGRU model and XGBoost layer. Based on temperature sample data, the fusion prediction model is optimized by the intelligent lemming algorithm. The optimized fusion prediction model is used to predict the temperature data and temperature rise characteristics of the gas-insulated transmission line in real time, and the temperature data of the gas-insulated transmission line is obtained.

2. The method according to claim 1, characterized in that, A simulation model is constructed and simulated using computational fluid dynamics methods, wherein the simulation model takes into account the coupling of electromagnetic field, thermal field and flow field.

3. The method according to claim 1, characterized in that, The electromagnetic field-thermal field coupling relationship in the simulation model includes: ； ； in, Indicator element Ohmic loss, of which and These are the current density and the conjugate current density, respectively. It is the total number of finite element components for the conductive part. Indicates the area of ​​the component. The absolute temperature is The electrical conductivity of the material at that time; The electrical conductivity of the material at 20℃; The temperature coefficient of resistivity This refers to absolute temperature.

4. The method according to claim 1, characterized in that, The temperature field-flow field coupling relationship in the simulation model includes: ； ； ； ； ； in, They are respectively direction and Directional gas flow velocity; It is the acceleration due to gravity; This is due to the density difference caused by gas expansion; These represent the density, thermal conductivity, and dynamic viscosity of the gas, respectively. This refers to gas pressure. As a volumetric heat source; The blackbody radiation coefficient is... The outer diameter of the conductor; These are the conductor temperature and the casing temperature, respectively. For system blackness, The outer diameter of the outer casing; The blackness of the outer surface of the casing; Ambient temperature; The occlusion coefficient is... This represents the energy radiated from the surface of a conductor. This represents the radiant energy emanating from the shell surface. For time, For constant pressure heat capacity, T represents absolute temperature.

5. The method according to claim 1, characterized in that, The electromagnetic field mathematical model in the simulation model is as follows: ； ； in, It is a magnetic vector potential. It is an electric scalar potential. It is the permeability. Indicates electrical conductivity. It is angular frequency. It is the applied source current density within the conductive rod. This represents the Hamiltonian operator, where j is the imaginary unit.

6. The method according to claim 1, characterized in that, The fusion prediction model includes an input layer, a BiTCN layer, a BiGRU layer, an XGBoost layer, and an output layer connected in sequence.

7. The method according to claim 1, characterized in that, The process of optimizing the fusion prediction model using the intelligent lemming algorithm includes: Initial candidate solutions are constructed based on the hyperparameters of the fusion prediction model. The initial candidate solutions include different individuals, each of which includes a set of hyperparameters of the fusion prediction model. The hyperparameters include the learning rate of the BiTCN model and the BiGRU model, the number of neurons, the number of filters, and the regularization factor of the BiGRU model, and the maximum number of iterations, depth, and learning rate of the XGBoost model. Fitness is calculated for individuals in the initial candidate solutions. Based on the fitness calculation results, the current optimal solution is obtained. The energy coefficient is calculated based on the optimal solution. The energy coefficient is used to determine whether to proceed to the exploration phase or the development phase. The next candidate solution is generated based on the exploration phase and the development phase. Based on the next candidate solution, the process of fitness calculation, generation of the current optimal solution, and generation of the next candidate solution is repeated until the maximum number of iterations is reached. The optimal hyperparameters of the fusion prediction model are obtained to optimize the fusion prediction model. The fitness is calculated based on the accuracy of the fusion prediction model.

8. The method according to claim 1, characterized in that, The objective function of the intelligent lemming algorithm is calculated based on the mean absolute error, mean square error, root mean square error, and coefficient of determination of the fusion prediction model.

9. A temperature rise prediction system for gas-insulated transmission lines based on a data-driven model, characterized in that, Used to perform the method described in any one of claims 1-8.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When executed by a processor, the computer program instructions implement the steps of the method according to any one of claims 1-8.

Images