Generator cooling efficiency evaluation method and system based on temperature rise mechanism

Abstract

This invention discloses a method and system for evaluating generator cooling efficiency based on temperature rise mechanisms, belonging to the field of generator cooling. The method includes: establishing a lumped-parameter thermal network model based on the target generator; real-time acquisition of three-phase current, real-time power, ambient temperature, and ambient relative humidity under unsteady-state operating conditions to obtain real-time operating data; inputting the data into a Gaussian mixture model for joint probability density estimation to obtain probability distribution characteristics, including expected value and variance space; sampling random parameters from the probability distribution characteristics, substituting random variables as dynamic boundary conditions into the model for numerical iterative solution to obtain the thermal field temperature rise probability distribution space within a preset future time period; evaluating the current cooling efficiency based on the distribution space to obtain the efficiency evaluation result. This application solves the technical problem that existing generators are difficult to warn of potential overheating risks in advance, which easily leads to accelerated aging of generator insulation materials, decreased insulation performance, and even breakdown.

Description

Technical Field

[0001] This invention relates to the field of generator cooling, and more specifically to a method and system for evaluating generator cooling performance based on temperature rise mechanisms. Background Technology

[0002] The generator cooling system is a core component ensuring the safe and stable operation of the generator set. Its efficiency directly determines the upper limit of the generator's output power and its service life. Especially for wind turbines, due to factors such as random fluctuations in wind speed, changes in grid dispatch commands, unit start-up and shutdown, and pitch regulation, they are in a long-term non-steady-state operating condition. The heat exchange capacity of the cooling system is significantly affected by multiple factors such as ambient temperature, relative humidity, and the degree of fouling on the heat dissipation surface, exhibiting significant dynamic changes. Under non-steady-state conditions, the convective heat transfer coefficient has extremely strong randomness and time-varying characteristics. The assumption of a fixed constant cannot truly reflect the actual heat exchange capacity of the cooling system, resulting in large errors in the calculated temperatures of key thermal nodes such as the generator stator windings and rotor. This makes it impossible to quantitatively assess the cooling efficiency and to provide early warning of potential overheating risks. It can easily lead to accelerated aging of generator insulation materials, decreased insulation performance, or even breakdown, seriously threatening the safe operation of wind turbine generator sets, and also increasing the operation and maintenance costs and equipment downtime losses of wind farms. Summary of the Invention

[0003] This application provides a generator cooling performance evaluation method and system based on the temperature rise mechanism, aiming to solve the technical problem that existing generators are unable to provide early warning of potential overheating risks, which can easily lead to accelerated aging of generator insulation materials, decreased insulation performance, or even breakdown. This achieves the technical effect of reducing the probability of equipment failure and improving the safety and reliability of generator operation.

[0004] In view of the above problems, this application provides a method and system for evaluating generator cooling efficiency based on temperature rise mechanism.

[0005] The first aspect disclosed in this application provides a method for evaluating the cooling efficiency of a generator based on a temperature rise mechanism, the method comprising: A lumped-parameter thermal network model is established based on the internal physical structure of the target generator. Real-time data on the three-phase current, real-time power, ambient temperature, and ambient relative humidity of the target generator under unsteady-state operating conditions are collected to obtain real-time operating data. This real-time operating data is input into a Gaussian mixture model for joint probability density estimation to obtain the probability distribution characteristics of the convective heat transfer coefficient, where the probability distribution characteristics include the expected value and variance space. Random parameters are sampled from the probability distribution characteristics of the convective heat transfer coefficient, and the sampled random variables are used as dynamic boundary conditions and substituted into the lumped-parameter thermal network model for numerical iterative solution to obtain the probability distribution space of the thermal field temperature rise of the target generator within a future preset time period. The current cooling efficiency of the target generator is evaluated based on the probability distribution space of the thermal field temperature rise to obtain the efficiency evaluation result.

[0006] Another aspect of this application discloses a generator cooling performance evaluation system based on a temperature rise mechanism, the system comprising: The model building module is used to establish a lumped-parameter thermal network model based on the internal physical structure of the target generator; the operating condition acquisition module is used to collect the three-phase current, real-time power, ambient temperature, and ambient relative humidity of the target generator under unsteady-state operating conditions in real time to obtain real-time operating condition data; the data input module is used to input the real-time operating condition data into a Gaussian mixture model for joint probability density estimation to obtain the probability distribution characteristics of the convective heat transfer coefficient, wherein the probability distribution characteristics include the expected value and variance space; the parameter sampling module is used to sample random parameters from the probability distribution characteristics of the convective heat transfer coefficient, and substitute the sampled random variables as dynamic boundary conditions into the lumped-parameter thermal network model for numerical iterative solution to obtain the thermal field temperature rise probability distribution space of the target generator in a future preset period; the thermal field evaluation module is used to evaluate the current cooling efficiency of the target generator based on the thermal field temperature rise probability distribution space to obtain the efficiency evaluation result.

[0007] One or more technical solutions provided in this application have at least the following technical effects or advantages: First, a coupled lumped parameter thermal network model, including stator winding nodes, rotor nodes, and cooling medium circuits, is established based on the internal physical structure of the generator. Real-time data on three-phase current, real-time power, ambient temperature, and ambient relative humidity under unsteady conditions are collected. After eliminating data coupling correlation through principal component analysis, the data is input into a Gaussian mixture model. The probability distribution characteristics of the convective heat transfer coefficient are solved using the expectation-maximization algorithm. Then, the dynamic boundary conditions are obtained using Latin hypercube sampling. The non-homogeneous state matrix is ​​numerically iterated and solved using the fourth-order Runge-Kutta method to obtain the probability distribution space of thermal field temperature rise within a preset future time period. Finally, the cooling efficiency is evaluated based on the probability of temperature over-temperature and graded early warning information is generated, which effectively reduces the probability of equipment failure and improves the safety and reliability of generator operation.

[0008] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0009] Figure 1 A flowchart illustrating a generator cooling performance evaluation method based on temperature rise mechanism is provided for embodiments of this application. Figure 2 A schematic diagram of the structure of a generator cooling efficiency evaluation system based on temperature rise mechanism is provided for the embodiments of this application.

[0010] Figure labeling: Model building module 11, operating condition acquisition module 12, data input module 13, parameter sampling module 14, thermal field evaluation module 15. Detailed Implementation

[0011] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0012] The overall concept of the technical solution provided in this application is as follows: This application provides a method and system for evaluating generator cooling efficiency based on temperature rise mechanism. By deeply integrating a deterministic physical mechanism model with a stochastic data-driven model, a coupled lumped parameter thermal network model including stator winding nodes, rotor nodes, and cooling medium circuits is first established based on the generator's internal physical structure. Real-time data on three-phase current, real-time power, ambient temperature, and ambient relative humidity under unsteady conditions are collected. After eliminating data coupling correlation through principal component analysis, the data is input into a Gaussian mixture model. The probability distribution characteristics of the convective heat transfer coefficient are solved using the expectation-maximization algorithm. Then, the dynamic boundary conditions are obtained using Latin hypercube sampling. The non-homogeneous state matrix is ​​numerically iterated and solved using the fourth-order Runge-Kutta method to obtain the probability distribution space of thermal field temperature rise within a preset future time period. Finally, the cooling efficiency is evaluated based on the probability of temperature overheating and a graded early warning information is generated. This enables early prediction of generator overheating failure risk, reduces wind farm operation and maintenance costs and unplanned downtime losses, and improves the safety and reliability of wind turbine operation.

[0013] After introducing the basic principles of this application, various non-limiting embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0014] Example 1, as Figure 1 As shown in the embodiments of this application, a method for evaluating the cooling efficiency of a generator based on a temperature rise mechanism is provided. This method includes: S100: Establish a lumped parameter thermal network model based on the internal physical structure of the target generator.

[0015] Specifically, a lumped parameter thermal network is established based on the internal physical structure of the target generator. The lumped parameter thermal network discretizes the continuously distributed three-dimensional thermal field inside the generator into a finite number of concentrated thermal nodes with uniform temperature characteristics. A simplified physical model is used to simulate the heat transfer and storage process between nodes through equivalent thermal elements. Specifically, the stator, rotor, cooling system and casing of the target wind turbine are first disassembled in three dimensions and their thermal characteristics are measured. The thermal nodes are divided according to the complete heat flow path from the loss generation part to the cooling medium inside the generator. The stator winding node corresponds to the winding conductor and the insulation layer area that generates copper loss on the stator side of the generator. The rotor node corresponds to the rotating iron core and the excitation winding area that generates iron loss and excitation loss on the rotor side of the generator. The cooling medium loop node corresponds to the air or coolant circulation channel that carries away heat. At the same time, key auxiliary thermal nodes such as the stator core and casing are retained to ensure the integrity of the heat flow transfer topology.

[0016] Next, the thermal connections and equivalent element parameters between each node are defined. Adjacent solid nodes are connected by conductive thermal resistance, which is used to simulate the heat conduction process inside the solid material. Its value is calculated from the geometric dimensions of the corresponding component and the thermal conductivity of the material. Solid nodes are connected to the cooling medium loop nodes by convective thermal resistance. Convective thermal resistance is inversely related to the convective heat transfer coefficient, i.e., convective thermal resistance... ,in The convective heat transfer coefficient is... The effective heat exchange area between the node and the cooling medium is defined. Each thermal node is configured with corresponding heat capacity parameters to simulate its heat storage capacity. The heat capacity value is calculated by multiplying the actual mass of the corresponding component of the node with the specific heat capacity of the material.

[0017] Subsequently, based on the law of conservation of energy, the heat balance equation for each thermal node is established. That is, the total heat flowing into the node at any time, i.e. the heat source generated by internal electromagnetic and mechanical losses, is equal to the heat storage of the node, i.e. the product of heat capacity and the node temperature change rate, plus the heat transferred from the node to all adjacent nodes, i.e. the temperature difference between adjacent nodes divided by the connection thermal resistance. The heat balance equations of all nodes are combined and arranged into a matrix form to obtain a non-homogeneous state matrix containing heat source terms, heat capacity matrix, and thermal resistance matrix, which constitutes the framework of the lumped parameter thermal network model.

[0018] Finally, the model is calibrated and verified, i.e., the model training process. Multiple sets of measured data such as stator winding temperature, rotor temperature, and inlet and outlet temperatures of the cooling medium are collected under steady-state rated operating conditions of the target generator. Based on the measured temperature, the empirical values ​​of the conductive thermal resistance and the initial convective heat transfer coefficient are finely adjusted through the parameter identification algorithm so that the absolute error between the calculated temperature and the measured temperature under steady-state operating conditions is controlled within the required temperature, thus completing the construction and verification of the basic physical model.

[0019] S200: Real-time acquisition of the three-phase current, real-time power, ambient temperature and ambient relative humidity of the target generator under non-steady-state operating conditions to obtain real-time operating data.

[0020] Specifically, the system collects real-time data on the three-phase current, real-time power, ambient temperature, and ambient relative humidity of the target generator under non-steady-state operating conditions, i.e., when the output power, speed, load, and other operating parameters of the wind turbine are continuously and dynamically changing due to factors such as random fluctuations in wind speed, changes in grid dispatch commands, unit start-up and shutdown, and pitch regulation. This real-time operating data is obtained by first deploying a multi-source heterogeneous sensor network tailored to the operating characteristics and data requirements of the target wind turbine; then, high-precision current transformers are installed at the stator output terminals of the generator to collect the three-phase current, i.e., the three-phase current in the generator stator windings. The instantaneous current value of the phase alternating current directly reflects the load level of the generator and is a parameter for calculating the copper loss of the stator winding. A power transmitter is installed at the generator outlet switchgear to collect real-time power, that is, the active power output by the generator at the current moment. An integrated temperature and humidity sensor is installed at the air inlet of the generator cooling system to collect ambient temperature and relative humidity, that is, the ratio of water vapor content in the air to saturated water vapor content, which affects the density and specific heat capacity of the cooling air, and thus changes the magnitude of the convective heat transfer coefficient. All sensors are selected with industrial-grade protection to adapt to the harsh outdoor operating environment of the wind farm.

[0021] Next, a unified synchronous acquisition system is configured, using a time protocol to provide nanosecond-level synchronous clocks for all sensors, ensuring that four types of data—three-phase current, real-time power, ambient temperature, and ambient relative humidity—are acquired synchronously at the same timestamp. The sampling frequency is set to 1Hz to capture the dynamic changes of parameters under unsteady conditions, avoiding unnecessary data storage and computational pressure caused by excessively high sampling frequencies.

[0022] Then, the raw data is preprocessed in real time. High-frequency noise generated by sensor electromagnetic interference is removed by the moving average filtering algorithm. The 3σ criterion is used to automatically identify and remove outliers caused by sensor faults or signal interruptions. Missing data is filled in by linear interpolation. At the same time, the raw data of different dimensions are converted into a standardized format.

[0023] Finally, the preprocessed real-time operating data is transmitted to the local edge computing node via industrial Ethernet for caching and simultaneously uploaded to the data analysis platform of the wind farm control center to provide real-time data input for subsequent model calculations.

[0024] S300: Input the real-time operating data into the Gaussian mixture model to perform joint probability density estimation and obtain the probability distribution characteristics of the convective heat transfer coefficient, wherein the probability distribution characteristics include the expected value and the variance space.

[0025] Specifically, the real-time operating data is input into a Gaussian mixture model. A Gaussian mixture model is a probability-based distribution fitting and clustering model that approximates arbitrarily complex continuous data distributions through a weighted linear combination of multiple independent Gaussian distributions. It can capture the multi-peak distribution characteristics and random fluctuations of the convective heat transfer coefficient under unsteady conditions, and perform joint probability density estimation to obtain the probability distribution characteristics of the convective heat transfer coefficient, including the expected value and variance space. The expected value reflects the most likely value of the convective heat transfer coefficient under the current operating conditions, while the variance space quantifies the uncertainty range affected by environmental and load fluctuations. Specifically, based on standardized real-time operating data, due to the three-phase current, real-time... There is a significant coupling correlation between power, ambient temperature, and ambient relative humidity. This includes a positive correlation between output power and three-phase current, and a seasonal correlation between ambient temperature and relative humidity. Directly inputting these into the model can lead to computational redundancy and overfitting risks. Therefore, principal component analysis is performed on the data first. Principal component analysis transforms a set of linearly correlated original variables into a set of linearly independent principal component variables through orthogonal transformation. Specifically, it calculates the covariance matrix of the standardized data, solves for its eigenvalues ​​and corresponding eigenvectors, and selects the first k principal components in descending order of eigenvalues, where k is the smallest integer with a cumulative contribution rate of 95%. This maps the original four-dimensional data into k-dimensional mutually independent feature matrices. Next, a Gaussian mixture model matching the dimension of the feature matrix is ​​constructed. The number of Gaussian components in the model is initialized, usually set to 3-5, corresponding to the typical low load, medium load, high load and extreme operating conditions of wind turbines, as well as the initial weights, mean vectors and covariance matrices of each component.

[0026] Then, the expectation-maximization algorithm, i.e., the iterative optimization algorithm, is used to gradually approach the optimal parameter solution by alternately executing the expectation step and the maximization step to train the model. In the E step, the posterior probability of each feature vector sample belonging to each Gaussian component is calculated based on the current model parameters. In the M step, the weight, mean vector and covariance matrix of each Gaussian component are re-estimated based on the posterior probability. The E step and the M step are repeated until the log-likelihood function of the model converges, that is, the difference between the log-likelihood functions of two adjacent iterations is less than the preset threshold 1e-6 or the maximum number of iterations of 100 is reached.

[0027] Finally, the joint probability density function of the convective heat transfer coefficient is calculated based on the converged Gaussian mixture model. The expected value is obtained by integrating the probability density function, which is the weighted average of the mean of all Gaussian components. The variance space is constructed by the weighted combination of the covariance matrix of each Gaussian component and the weight, thus completing the extraction of the probability distribution characteristics of the convective heat transfer coefficient.

[0028] S400: Random parameters are sampled from the probability distribution characteristics of the convective heat transfer coefficient, and the sampled random variables are substituted as dynamic boundary conditions into the lumped parameter thermal network model for numerical iterative solution to obtain the thermal field temperature rise probability distribution space of the target generator in the future preset time period.

[0029] Specifically, random parameter sampling is performed from the probability distribution characteristics of the convective heat transfer coefficient. This involves generating a large number of random samples that conform to the known probability distribution function. The Latin hypercube sampling method is used, which divides the probability space into multiple equal probability intervals and randomly samples within each interval. Compared with simple random sampling, this method can cover the entire probability distribution space more comprehensively with fewer samples. The sampled random variables are used as dynamic boundary conditions, i.e., model boundary parameters that are dynamically adjusted with time and operating conditions. These parameters are substituted into the lumped parameter thermal network model for numerical iteration to obtain the probability distribution space of the thermal field temperature rise of the target generator in the future preset time period. This means that at each future time node, the complete probability distribution set corresponding to the temperature values ​​of key thermal nodes such as the generator stator winding and rotor is obtained. Specifically, based on the extracted expected value and variance space of the convective heat transfer coefficient, the sampling sample size is set to 1000 sets. This sample size has been statistically verified to ensure the statistical significance of the probability distribution and control the computational complexity within an acceptable range for engineering. The Latin hypercube sampling method is used to generate 1000 random samples of the convective heat transfer coefficient that conform to the probability distribution.

[0030] Next, the random variable of convective heat transfer coefficient obtained from each sample is calculated according to the formula, and the thermal resistance is... ,in The convective heat transfer coefficient is... The effective heat exchange area between the nodes and the cooling medium is a fixed physical parameter, which is converted into the dynamic thermal resistance between each solid node and the cooling medium loop node. At the same time, the internal loss power of the stator winding node and the rotor node is calculated based on the three-phase current and real-time power in the real-time operating data, and it is used as the input heat source of the lumped parameter thermal network model.

[0031] Subsequently, the dynamic thermal resistance parameters are substituted into the pre-constructed non-homogeneous state matrix, which is a set of ordinary differential equations in matrix form obtained by combining the thermal balance equations of all thermal nodes. This matrix includes three core parts: the thermal capacity matrix, the thermal resistance matrix, and the heat source term. The fixed convection thermal resistance parameters in the original model are replaced to form a transient heat conduction solution model for the current sample.

[0032] Then, the future preset time period was set to 1 hour and the time step was 1 minute. The fourth-order Runge-Kutta method was adopted, that is, the function value of the next time step was calculated by weighted average of four different slopes. The transient heat conduction model corresponding to each sample was cyclically iterated for the time step to obtain 1000 sets of temperature response time series of stator winding nodes and rotor nodes in the next 60 time steps.

[0033] Finally, a time-step statistical analysis was performed on these 1000 sets of temperature response data to calculate the expected value, variance, and probability density function of the stator winding and rotor temperature at each time step. The probability distribution results of all time steps were integrated to form a thermal field temperature rise probability distribution space covering the entire future preset period.

[0034] S500: Evaluate the current cooling efficiency of the target generator based on the spatial distribution of the thermal field temperature rise probability, and obtain the efficiency evaluation result.

[0035] Specifically, based on the thermal field temperature rise probability distribution space, which refers to the complete probability distribution set of temperature values ​​of key thermal nodes such as the stator winding and rotor of the generator within a preset future time period, including the expected value, variance, and probability density function of the temperature at each time step, the current cooling efficiency of the target generator is evaluated to obtain the efficiency evaluation result. Specifically, from the solved thermal field temperature rise probability distribution space, the temperature probability density curves of the stator winding nodes and rotor nodes at each time step within the preset future time period are extracted. The temperature probability density curve is a continuous curve describing the probability density of the temperature random variable at different value points. The area under any interval of the curve represents the probability that the temperature falls within that interval.

[0036] Next, the heat resistance limit parameter of the insulation material of the target generator is obtained. The heat resistance limit of the insulation material refers to the highest temperature at which the insulation material used in the stator winding and rotor winding of the generator can operate safely for a long time. Exceeding this temperature will cause the insulation material to age faster, the insulation performance to decline, or even break down. It is the core threshold for safe operation of the generator and is usually determined by the insulation material grade.

[0037] Then, the numerical integration method is used to calculate the integral area in each temperature probability density curve where the temperature response value exceeds the heat resistance limit of the corresponding insulation material, and the ratio of this integral area to the total area of ​​the entire temperature probability density curve is calculated. This ratio is the probability that the generator will experience cooling failure at the corresponding time step.

[0038] Subsequently, a preset cooling efficiency risk threshold is established, which is set according to the wind farm's safety operation standards and maintenance strategies. It is generally set to 0.05, or 5%, indicating that when the cooling failure probability exceeds 5%, the current cooling system is considered to have a safety hazard. The calculated cooling failure probability at each time step is compared with the preset risk threshold point by point. If the cooling failure probability at all time steps is less than the preset risk threshold, the current cooling efficiency is determined to be sufficient, and the efficiency assessment result of the cooling system operating normally is output. If the cooling failure probability at any time step is greater than or equal to the preset risk threshold, the current cooling efficiency is determined to be insufficient. At the same time, the time point when the cooling failure probability first exceeds the threshold and the maximum failure probability value are recorded, and cooling failure early warning information containing specific risk time, risk location, and risk level is generated as the final efficiency assessment result.

[0039] Furthermore, in the method provided in the application embodiment, the lumped parameter thermal network model includes stator winding nodes, rotor nodes, and cooling medium loops, and the nodes are coupled to each other through convective heat transfer coefficients.

[0040] Specifically, the physical boundaries and thermal properties of the three thermal nodes are first defined. The stator winding node is a concentrated thermal unit corresponding to the copper winding conductor and its insulating layer on the stator side of the generator in the lumped parameter thermal network model. It is one of the main heat sources inside the generator, and its heat mainly comes from the copper losses generated when the three-phase current passes through the winding resistance. The magnitude of the copper losses is proportional to the square of the three-phase current. The rotor node is a concentrated thermal unit corresponding to the rotating iron core and excitation winding on the rotor side of the generator. Its heat mainly comes from the iron losses generated by the alternating magnetic field in the iron core and the excitation losses generated by the excitation current passing through the excitation winding. These two types of losses fluctuate dynamically with the changes in the generator's real-time power and speed. The cooling medium circuit is a concentrated thermal unit corresponding to the air or coolant circulation channel inside the generator. Its function is to carry away the heat generated by the stator winding node and rotor node through the flow of the medium. It is the final heat dissipation link in the generator's heat transfer path.

[0041] Next, basic thermal parameters are configured for each node, with a focus on assigning corresponding heat capacity parameters. Heat capacity is a physical quantity that characterizes the heat storage capacity of a node. Its value is equal to the product of the actual mass of the corresponding component and the specific heat capacity of the material. The heat capacity of the stator winding node is calculated by superimposing the heat capacities of the winding copper material and the insulating material. The heat capacity of the rotor node is calculated by superimposing the heat capacities of the iron core silicon steel sheet and the excitation winding. The heat capacity of the cooling medium circuit is calculated by multiplying the total mass of the circulating medium and the specific heat capacity of the medium.

[0042] Subsequently, the thermal coupling relationships between the nodes are established. Thermal coupling refers to the characteristic that different nodes influence each other through heat transfer and are interconnected by temperature changes. In this model, the three nodes are coupled to each other through the convective heat transfer coefficient. The convective heat transfer coefficient is a physical quantity that characterizes the heat transfer capacity between a solid surface and a flowing fluid; the larger the value, the more heat is transferred per unit area and per unit temperature difference. Specifically, the stator winding node is coupled to the cooling medium circuit through the stator-side convective heat transfer coefficient, the rotor node is coupled to the cooling medium circuit through the rotor-side convective heat transfer coefficient, and the stator winding node and the rotor node are indirectly coupled through the air gap convective heat transfer coefficient. Each coupling relationship corresponds to a convective thermal resistance, which is inversely related to the convective heat transfer coefficient. ,in The convective heat transfer coefficient is... The effective heat exchange area between the node and the cooling medium is represented by the convective thermal resistance, which constitutes the resistance to heat transfer between the nodes.

[0043] Finally, based on the law of conservation of energy, a set of coupled thermal balance equations is established. The thermal balance equations of each node take into account the temperature influence of adjacent coupled nodes. That is, the temperature change of the stator winding node depends not only on its own copper loss and heat capacity, but also on the temperature difference between it and the cooling medium loop and the rotor node, as well as the corresponding convective thermal resistance. The temperature change of the rotor node depends not only on its own iron loss and excitation loss, but also on the temperature difference between it and the cooling medium loop and the stator winding node, as well as the corresponding convective thermal resistance. The temperature change of the cooling medium loop depends on the total heat absorbed from the stator winding node and the rotor node, as well as the heat discharged to the external environment. Thus, a coupled thermal network system that is interconnected and mutually restrictive is formed.

[0044] Furthermore, in the method provided in the application embodiment, the real-time operating data is input into a Gaussian mixture model for joint probability density estimation to obtain the probability distribution characteristics of the convective heat transfer coefficient. This includes: performing principal component analysis on the real-time operating data to eliminate the coupling correlation between the three-phase current, real-time power, ambient temperature, and ambient relative humidity, and obtaining multiple independent feature vectors; combining the multiple feature vectors into a feature matrix, and inputting the feature matrix into the Gaussian mixture model; updating the model parameters through the expectation-maximization algorithm to obtain the probability density function of the convective heat transfer coefficient.

[0045] Specifically, the real-time operating data is input into a Gaussian mixture model for joint probability density estimation. Joint probability density estimation refers to simultaneously estimating the probability distribution of a target variable under the combined influence of multiple random variables. Specifically, it estimates the joint probability distribution between four input variables (three-phase current, real-time power, ambient temperature, and ambient relative humidity) and the target variable (convective heat transfer coefficient), obtaining the probability distribution characteristics of the convective heat transfer coefficient. Specifically, firstly, principal component analysis is performed on the preprocessed standardized real-time operating data. Principal component analysis transforms a set of linearly coupled variables (where coupling correlation refers to the linear dependence between multiple variables, including a positive correlation between generator output power and three-phase current, and a seasonal negative correlation between ambient temperature and relative humidity) into a set of linearly independent principal component variables. Specifically, the standardized data matrix is ​​calculated. ; in, Let N be the dataset matrix, and N be the number of samples. Let be the set of real numbers, the set of all real numbers; 4 is the dimension of the original variables, corresponding to three-phase current, real-time power, ambient temperature, and ambient relative humidity; and the covariance matrix of the standardized data matrix is ​​as follows: ; in, The sample covariance matrix, The number of samples is N, which is consistent with the number of samples in the dataset matrix. The dataset matrix is ​​centered, with one sample per row and one feature per column. Let X be the transpose of matrix X. The coefficients are unbiased estimates. N-1 is used to obtain an unbiased estimate of the population covariance. The eigenvalues ​​of the covariance matrix C are then calculated. ; and the corresponding 4 orthogonal unit eigenvectors Based on the principle that the cumulative contribution rate of eigenvalues ​​reaches 95%, the first k principal components are selected, where k is usually 2 or 3. The original data matrix X is then projected onto a projection matrix composed of the first k eigenvectors. ; We obtain mutually independent k-dimensional feature vectors: ; in, Let i be the feature vector after dimensionality reduction of the i-th monitoring sample. Let i be the vector of the i-th monitoring sample in the original data. The eigenvector projection transformation matrix, For sample number, To monitor the total number of samples, all feature vectors are combined row-wise to form a feature matrix: ; in The feature matrix after dimensionality reduction. The vectors are dimensionality-reduced sample vectors, each vector corresponding to one set of dimensionality-reduced monitoring data. T is the transpose, which converts column vectors into row vectors. N×k is the matrix dimension. Y is an N-row, k-column matrix. The feature matrix is ​​a matrix composed of independent feature vectors after dimensionality reduction and decorrelation. Its column vectors are linearly independent principal components, and its row vectors are the feature representations corresponding to each sample. It is the standard input format of the Gaussian mixture model.

[0046] Next, a Gaussian mixture model matching the dimension of the feature matrix is ​​constructed. The probability density function of its convective heat transfer coefficient h is given. The probability density function describes the probability of a continuous random variable occurring within a unit length around a given value. The area under the curve of any interval represents the probability that the random variable falls within that interval. The expression is: ; in, For input features, Here, K represents the number of Gaussian mixture components, corresponding to four typical operating modes of the wind turbine: low load, medium load, high load, and extreme conditions. Let be the weighting coefficient of the k-th Gaussian component, satisfying and , representing the prior probability that the sample comes from the k-th Gaussian component. Let be the probability density function of the k-th one-dimensional Gaussian distribution. Let be the mean of the k-th Gaussian component. Let be the variance of the k-th Gaussian component.

[0047] Then, the expectation-maximization algorithm is used to solve the parameter estimation problem of a probabilistic model with latent variables. Latent variables are variables that cannot be directly observed but affect the distribution of observed data. By alternately executing the expectation step, E-step, and maximization step, M-step, the global optimal parameter solution is gradually approximated, and the model parameters are updated. Specifically, the first step is to initialize the model parameters by randomly setting the initial weights of K Gaussian components. Initial mean and initial variance And set the iteration termination threshold. The maximum number of iterations is T=100; the second step is to execute step E, which calculates y for each sample based on the model parameters of the current t-th iteration. i Posterior probability of belonging to the k-th Gaussian component The calculation formula is: ; in, For latent variables, a value of 1 indicates that sample y is a latent variable. i It comes from the k-th Gaussian component; a value of 0 indicates that it is not. Let be the k-dimensional covariance matrix of the k-th Gaussian component at the t-th iteration; the third step executes the M-step, updating the parameters of all Gaussian components based on the posterior probabilities calculated in the E-step, with the weight update formula as follows: ; in, The weight coefficient of the k-th Gaussian component after the (t+1)-th iteration update. This represents the total number of all monitored samples. Let t be the posterior probability that the i-th sample belongs to the k-th state obtained in the t-th iteration; the mean update formula is: ; in, Let y be the mean vector of the k-th Gaussian component after the (t+1)-th iteration update. i Let be the feature vector of the i-th monitoring sample after dimensionality reduction; the covariance matrix update formula is: ; in, Let be the covariance matrix of the (t+1)th iteration, and let represent the dispersion of the feature distribution after the update of the k-th health state. Let T be the mean vector of the (t+1)th iteration, i.e., the mean of the feature of the kth state that has just been updated, and let T be the vector transpose. The fourth step is to calculate the log-likelihood function of the current model: ; in, Let be the log-likelihood function value of the model at the t-th iteration, representing the degree of fit of the model to the sample data. This is the set of all model parameters for the t-th iteration, including weights, mean, and covariance matrix. Given the model parameter set for the t-th iteration, determine the difference in log-likelihood between two adjacent iterations. Is it less than the iteration termination threshold? If the maximum number of iterations T has been reached, the iteration stops; otherwise, it returns to the second step. Finally, based on the converged Gaussian mixture model parameters, the probability distribution of the feature space is transformed into the probability density function of the convective heat transfer coefficient h through regression mapping, and its expected value is calculated. ; in, Let h be the expected value of the health status variable, representing the overall comprehensive health assessment value. The weight coefficients are for the k-th health state. Let be the feature mean corresponding to the k-th health state, where k is the total number of health state categories, and be the variance space. ; in, Let h be the population variance of the health status variable, representing the degree of dispersion and fluctuation in health status. The variance within a single component reflects data fluctuations under similar conditions, enabling the extraction of the probability distribution characteristics of the convective heat transfer coefficient.

[0048] Furthermore, in the method provided in the application embodiment, the multiple feature vectors are combined into a feature matrix, and the feature matrix is ​​input into a Gaussian mixture model. The model parameters are updated using the expectation-maximization algorithm to obtain the probability density function of the convective heat transfer coefficient. This includes: using the expectation-maximization algorithm to iteratively update the weight parameters, mean vector, and covariance matrix in the Gaussian mixture model until the log-likelihood function of the Gaussian mixture model converges; based on the converged Gaussian mixture model, the feature matrix is ​​analyzed to obtain the probability density function of the convective heat transfer coefficient, and the probability density function is used as the probability distribution feature.

[0049] Specifically, the multiple eigenvectors are combined into a feature matrix, which is a two-dimensional matrix formed by stacking independent eigenvectors obtained from principal component analysis in rows. Each row corresponds to a standard chemical condition sample at a time step, and each column corresponds to a linearly independent principal component feature. The feature matrix is ​​then input into a Gaussian mixture model. An expectation-maximization algorithm, which is an iterative optimization algorithm for solving the parameters of a probability model containing latent variables (where latent variables refer to the Gaussian component numbers to which each sample belongs), is used to iteratively update the weight parameters, mean vector, and covariance matrix in the Gaussian mixture model until the log-likelihood function of the Gaussian mixture model converges.

[0050] Subsequently, based on the converged Gaussian mixture model, the feature matrix is ​​analyzed to obtain the probability density function of the convective heat transfer coefficient, and the probability density function is used as the probability distribution feature. Specifically, the feature matrix is ​​first standardized by arranging the k-dimensional feature vectors corresponding to each sample output by principal component analysis, where k is the number of principal components with a cumulative contribution rate of 95% (usually 2 or 3), in chronological order to form a feature matrix of dimension N×k, where N is the total number of valid operating condition samples participating in model training. This ensures that each row in the matrix corresponds to a complete and independent unsteady-state condition, and that all feature values ​​have been standardized to zero mean. The core parameters of the Gaussian mixture model are initialized based on typical wind turbine operating modes (low load, medium load, high load, and extreme conditions). The number of Gaussian components, K, is set to 4. The weight parameters of each Gaussian component are randomly initialized, representing its proportion in the mixture model and reflecting the prior probability of the corresponding operating condition. All weights are strictly summed to 1. The mean vector represents the center position of each Gaussian component, corresponding to the average value of the convective heat transfer coefficient under a typical operating condition. The covariance matrix describes the distribution shape and dispersion of each Gaussian component, reflecting the fluctuation range of the convective heat transfer coefficient and the correlation between variables under the corresponding operating condition. An iteration termination threshold is also set. The maximum number of iterations, T=100, is set to prevent the algorithm from getting stuck in local optima or infinite loops.

[0051] The algorithm then proceeds to the iterative update process of Expectation-Maximization (EM). The first step, the Expectation Step (E-step), calculates the posterior probability of each sample in the feature matrix belonging to each Gaussian component based on the model parameters of the current t-th iteration. This probability is calculated by multiplying the probability density of each Gaussian component corresponding to the sample by its weight, and then dividing by the weighted sum of the probability densities of all Gaussian components corresponding to the sample. The second step, the Maximization Step (M-step), re-estimates the parameters of each Gaussian component based on the posterior probabilities of all samples calculated in the E-step: weights. The parameters are updated to the average of the posterior probabilities of all samples belonging to that component; the mean vector is updated to the weighted average of all samples by their posterior probabilities; and the covariance matrix is ​​updated to the weighted covariance of all samples by their posterior probabilities. The third step is to calculate the log-likelihood function of the current model, a function that measures the degree of fit between the model parameters and the observed feature matrix. A larger value indicates a better fit to the actual working conditions data; this is calculated by summing the natural logarithms of the mixture probability densities of all samples. The fourth step is to determine the convergence condition: if the absolute difference between the log-likelihood function values ​​of two adjacent iterations is less than a preset threshold... If the iteration count reaches the maximum limit T, the iteration stops; otherwise, it returns to step E and continues execution. After the algorithm converges, the optimal Gaussian mixture model parameter set is obtained. At this time, based on the converged model, a probability analysis is performed on the feature matrix. Through the pre-established linear regression mapping relationship between the principal component feature space and the physical quantity of the convective heat transfer coefficient, the multidimensional Gaussian mixture distribution of the feature space is transformed into a one-dimensional probability density function of the convective heat transfer coefficient. That is, the probability density function of the continuous random variable of the convective heat transfer coefficient near different value points. The area under any interval of the curve represents the probability that the convective heat transfer coefficient falls within that interval. This function completely describes all possible values ​​of the convective heat transfer coefficient and their corresponding probabilities under the current unsteady condition, and uses this as the core probability distribution feature for subsequent random parameter sampling.

[0052] Furthermore, in the method provided in the application embodiment, random parameters are sampled from the probability distribution characteristics of the convective heat transfer coefficient, and the sampled random variables are substituted as dynamic boundary conditions into the lumped parameter thermal network model for numerical iterative solution to obtain the thermal field temperature rise probability distribution space of the target generator in a future preset time period. This includes: calculating the internal loss power of the stator winding nodes and rotor nodes based on the real-time power and three-phase current in the real-time operating data, and using the internal loss power as the input heat source; converting the convective heat transfer coefficient sampled from the expected value and variance space into the dynamic thermal resistance between each node, and combining the input heat source to solve the transient heat conduction process between the stator winding nodes, rotor nodes and cooling medium circuit to obtain the thermal field temperature rise probability distribution space.

[0053] Specifically, the process of sampling random parameters from the probability distribution characteristics of the convective heat transfer coefficient, and substituting the sampled random variables as dynamic boundary conditions into the lumped parameter thermal network model for numerical iteration to obtain the probability distribution space of the thermal field temperature rise of the target generator in a future preset period is as follows: First, the internal loss power of the stator winding nodes and rotor nodes is calculated based on the real-time power and three-phase current in the real-time operating data. This internal loss power is the part of electromagnetic energy and mechanical energy that is irreversibly converted into heat energy during generator operation. It includes three types: stator winding copper loss, rotor core iron loss, and rotor excitation winding excitation loss. This loss is used as the input heat source, i.e., the only source of heat in the lumped parameter thermal network model. Its value directly determines the temperature rise level of each node of the generator. The specific calculation method is as follows: the stator winding copper loss is calculated according to the effective value of the three-phase current I and the DC resistance R of the stator winding using the formula: ; in, This refers to the copper loss power of the three-phase motor windings. This represents the effective value of the single-phase winding current of the motor. 3 represents the equivalent winding resistance of a single phase of the motor, and 3 represents the phase number correction coefficient for a three-phase motor. All of these are used as input heat sources for the stator winding nodes. Rotor iron loss is positively correlated with the generator's real-time power P and speed n, and is determined by a pre-calibrated iron loss coefficient. Calculate using the formula: ; in, This refers to the power loss of the motor core. This is the core loss correction factor. This is a reference value for core loss under baseline operating conditions. This refers to the actual operating speed of the motor. The rated speed of the generator; rotor excitation loss is based on the excitation current. and excitation winding resistance Calculate using the formula: ; in, The power loss in the motor excitation circuit. This is the effective value of the operating current of the excitation winding. The equivalent resistance of the excitation winding is used, and the sum of rotor iron loss and excitation loss is taken as the input heat source of the rotor node. Then, based on the probability distribution characteristics of the convective heat transfer coefficient, including the expected value and variance space, the expected value reflects the most likely value of the convective heat transfer coefficient under the current operating conditions, and the variance space quantifies the uncertainty range affected by factors such as wind speed fluctuations, changes in ambient temperature and humidity, and fouling on the heat dissipation surface, random parameter sampling is performed. This scheme adopts Latin hypercube sampling, which divides the probability space evenly into multiple equal probability intervals and randomly selects a sample in each interval. According to engineering verification, the sampling sample size is set to 1000 groups, which can ensure the statistical significance of the probability distribution and control the computational complexity within the range that the wind farm edge computing node can bear.

[0054] Then, the random variable h of the convective heat transfer coefficient obtained from each sample is calculated using the formula: ; This is converted into the dynamic thermal resistance between each solid node and the cooling medium loop node, which is a thermal parameter that changes in real time with the convective heat transfer coefficient. For surface convection heat transfer thermal resistance, Let A be the convective heat transfer coefficient of the fluid surface, and let A be the effective heat transfer area between the node and the cooling medium. A is a fixed parameter determined by the physical structure of the generator. It constitutes a dynamic boundary condition that can truly reflect the real-time fluctuation characteristics of the heat transfer capacity of the cooling system under unsteady conditions.

[0055] Subsequently, the dynamic thermal resistance is substituted into the non-homogeneous state matrix of the pre-constructed lumped parameter thermal network model, which is obtained by simultaneously solving the thermal balance equations of all thermal nodes. It includes three core parts: the thermal capacity matrix, the thermal resistance matrix, and the heat source term. The fixed convective thermal resistance parameter in the original model is replaced, and the previously calculated input heat source is loaded to form a transient heat conduction solution model for each sample. The future preset time period is set to 1 hour and the time step is 1 minute. The fourth-order Runge-Kutta method is used to calculate the function value of the next time step by weighted averaging four different slopes. The model corresponding to each sample is cyclically iterated for the time step. Through parallel calculation of 1000 samples, 1000 sets of temperature response time series of stator winding nodes and rotor nodes under the next 60 time steps are obtained.

[0056] Finally, a time-step statistical analysis was performed on these 1000 sets of temperature response data to calculate the expected value, variance, and probability density function of the stator winding and rotor temperature at each time step. The probability distribution results of all time steps were integrated to form a thermal field temperature rise probability distribution space covering the entire future preset period, that is, a complete probability distribution set of the generator's key thermal node temperature values ​​at each future time node.

[0057] Furthermore, in the method provided in the application embodiment, the convective heat transfer coefficient sampled from the expected value and variance space is converted into the dynamic thermal resistance between each node. Combined with the input heat source, the transient heat conduction process between the stator winding node, rotor node, and cooling medium circuit is solved to obtain the thermal field temperature rise probability distribution space. This includes: establishing a heat flow transfer topology based on the relative geometric positions between the stator winding node, rotor node, and cooling medium circuit; based on the heat flow transfer topology, the input heat source is loaded as a heat flow injection term onto the corresponding stator winding node and rotor node, and the dynamic thermal resistance and the internal heat capacity of the stator winding node and rotor node are combined into a matrix according to the law of energy conservation to obtain a non-homogeneous state matrix; the non-homogeneous state matrix is ​​solved iteratively using a numerical iteration algorithm to obtain the thermal field temperature rise probability distribution space of the target generator in a future preset time period.

[0058] Specifically, the convective heat transfer coefficient sampled from the expected value and variance space is converted into the dynamic thermal resistance between each node. Combined with the input heat source, the transient heat conduction process between the stator winding nodes, rotor nodes, and cooling medium loop is solved to obtain the thermal field temperature rise probability distribution space. Specifically, firstly, based on the relative geometric positions between the stator winding nodes, rotor nodes, and cooling medium loop, a heat flow transfer topology is established. This topology is a directed graph structure formed by connecting each thermal node with equivalent thermal resistance, based on the physical paths of heat generation, transfer, and dissipation within the generator. The flow direction and connection relationships of heat from the heat source to the final dissipation in the cooling medium are the foundation for constructing a mathematical model of the thermal network. Specifically, based on the physical structure of the generator stator and rotor being coaxially arranged and the cooling medium flowing around the stator and rotor, three core heat flow transfer paths are determined: the radial convective heat flow path between the stator winding nodes and the cooling medium loop; the axial and radial combined convective heat flow path between the rotor nodes and the cooling medium loop; and the indirect convective heat flow path formed between the stator winding nodes and the rotor nodes through the air gap. The dynamic thermal resistance type corresponding to each path is also labeled, forming a complete and realistic model. The heat transfer topology is based on the laws of physics. Then, based on this topology, the previously calculated stator winding copper loss and rotor total loss are used as heat injection terms, i.e., the amount of heat injected into the corresponding thermal nodes per unit time. This serves as the input excitation for the thermal network model, and its value directly determines the temperature rise rate and final steady-state temperature of the nodes. This heat is applied to the corresponding stator winding nodes and rotor nodes, respectively. The internal heat capacity of each node is also introduced, i.e., the physical quantity representing the node's heat storage capacity. Its value is equal to the product of the actual mass of the corresponding component and the specific heat capacity of the material, determining the node's temperature response rate to heat changes. The larger the capacity, the slower the temperature change, and the dynamic thermal resistance derived from the convective heat transfer coefficient obtained from sampling, i.e., the heat transfer resistance that changes in real time with the convective heat transfer coefficient, follows the law of conservation of energy, i.e., the total heat flowing into the node at any time is equal to the heat storage of the node plus the total heat flowing out of the node. A heat balance equation is established for each core node. The heat balance equations of the three nodes are combined and arranged into a matrix form to obtain the non-homogeneous state matrix, which is the matrix representation of a set of ordinary differential equations composed of the heat capacity matrix, the thermal resistance matrix, and the heat source vector. This is the core mathematical model for numerically solving the transient heat conduction process, and its general form is: ; Where C is the diagonal heat capacity matrix, and the diagonal elements are the heat capacity values ​​of each node; K is the thermal resistance matrix, which is composed of the reciprocals of the dynamic thermal resistance between each node; T is the node temperature vector; and P is the heat source vector, which includes the heat flow injection terms of each node.

[0059] Subsequently, the fourth-order Runge-Kutta method is used to update the function value of the next time step by calculating the slopes at four different points and taking a weighted average. The non-homogeneous state matrix is ​​solved iteratively. First, the future preset time period is set to 1 hour and the time step size is 1 minute. The entire solution process is divided into 60 consecutive time steps. For each sampled dynamic thermal resistance sample, starting from the measured initial temperature state at the current moment, the temperature values ​​of the stator winding, rotor and cooling medium at each time step are calculated in sequence to obtain the temperature response time sequence of the key thermal node corresponding to the sample in the next 60 time steps. The above solution process is repeated for all 1000 independent dynamic thermal resistance samples to obtain 1000 sets of mutually independent temperature response time sequences.

[0060] Finally, a time-step statistical analysis was performed on these 1000 sets of temperature response data. The expected value, variance, and probability density function of the stator winding and rotor temperature were calculated at each time step. The probability distribution results of all time steps were integrated to form a thermal field temperature rise probability distribution space covering the entire future preset period. That is, at each future time node, it is a complete set of all possible temperature values ​​and their corresponding probabilities of the generator's key thermal nodes, rather than a single deterministic temperature value. This can comprehensively quantify the impact of environmental and load uncertainties on the generator's thermal field distribution under unsteady operating conditions.

[0061] Furthermore, in the method provided in the application embodiment, the non-homogeneous state matrix is ​​solved iteratively using a numerical iterative algorithm to obtain the thermal field temperature rise probability distribution space of the target generator within a future preset time period, including: setting the total iteration duration and time step; using the fourth-order Runge-Kutta method to perform iterative calculation of the non-homogeneous state matrix at the time step until the total iteration duration is met, obtaining multiple temperature response values ​​of the stator winding nodes and rotor nodes at each time step; and analyzing the multiple temperature response values ​​to obtain the thermal field temperature rise probability distribution space.

[0062] Specifically, the standard expression for the non-homogeneous state matrix, i.e., the matrix form of the ordinary differential equations obtained by simultaneously solving the heat balance equations of the stator winding nodes, rotor nodes, and cooling medium circuit, through a numerical iterative algorithm, is as follows: ; Where C is the diagonal heat capacity matrix, with diagonal elements representing the heat capacity values ​​of each node; K is the thermal conduction matrix containing dynamic thermal resistance, with elements composed of the reciprocals of the thermal resistance between nodes; T is the node temperature vector; and P is the heat source vector. This is the core mathematical description of the transient heat conduction process. A cyclic solution is performed to obtain the probability distribution space of the thermal field temperature rise of the target generator within a preset future time period. Specifically: First, the total iteration duration and time step are set. The total iteration duration refers to the future time range covered by the thermal field prediction. This scheme sets it to 1 hour based on the actual response requirements of wind farm operation and maintenance early warning, allowing sufficient time for equipment inspection and maintenance personnel. The time step refers to the time interval between two adjacent temperature calculations during the numerical iteration process, set to 1 minute. This value ensures that the rapid temperature fluctuation characteristics under unsteady conditions are captured while keeping the computational load of a single sample within the tolerance range of the wind farm edge calculation nodes.

[0063] Next, using the fourth-order Runge-Kutta method, the function value for the next time step is updated by calculating the slopes at four different points and taking a weighted average. The non-homogeneous state matrix is ​​then iteratively calculated using time steps. Specifically, the measured temperatures of the generator stator winding, rotor, and cooling medium at the current moment are used as the initial temperature vector T0. For each independent dynamic thermal resistance sample obtained by Latin hypercube sampling, 60 time-step iterations are performed sequentially starting from the initial moment. In each iteration, the slopes at four different stages are calculated first. , , , ,in The slope at the initial moment. Based on the midpoint of the time step The predicted slope, Based on the midpoint of the time step The corrected slope, At the end of the time step, based on The predicted slope is then calculated using the weighted average formula: ; Obtain the temperature vector at the next time step. , Let be the instantaneous temperature of the device at the nth time step, where Using the time step as the time step, this process is repeated until 60 iterations are completed to obtain the temperature response values ​​of the stator winding node and rotor node corresponding to the sample at 60 time steps. That is, under a single dynamic thermal resistance sample, the temperature calculation results of the key thermal nodes of the generator at each time step reflect the transient change process of the thermal field under specific heat transfer conditions. The above iterative calculation process is repeated for all 1000 independent dynamic thermal resistance sampling samples to obtain 1000 sets of mutually independent temperature response time series, that is, 1000 stator winding temperature values ​​and 1000 rotor temperature values ​​corresponding to each time step.

[0064] Finally, based on the multiple temperature response values, statistical analysis is performed to obtain the thermal field temperature rise probability distribution space, which is the complete set of all possible temperature values ​​and their corresponding probabilities of the generator's key thermal nodes at each future time node. Specifically, statistical analysis is performed on 1000 temperature response values ​​at each time step, and their arithmetic mean is calculated as the expected temperature value for that time step. Their sample variance is calculated as the temperature fluctuation range for that time step. Then, the temperature probability density function for that time step is obtained by fitting it using the kernel density estimation method. The expected temperature values, variances, and probability density functions of all 60 time steps are integrated in chronological order to form a continuous thermal field temperature rise probability distribution space covering the entire preset period of the next 1 hour.

[0065] Furthermore, in the method provided in the application embodiment, the thermal field temperature rise probability distribution space is obtained by analyzing the plurality of temperature response values, including: performing expectation analysis and variance analysis on the plurality of temperature response sample values ​​to obtain the expectation value and variance; and constructing the thermal field temperature rise probability distribution space according to the expectation and variance.

[0066] Specifically, based on the multiple temperature response sample values, i.e., after solving 1000 independent random samples of convective heat transfer coefficients using the fourth-order Runge-Kutta method, the stator winding nodes and rotor nodes are obtained at each future time step, with each sample value corresponding to a transient response of the thermal field under a specific heat transfer condition. Analysis is then performed to obtain the probability distribution space of the thermal field temperature rise. Specifically: firstly, expectation analysis and variance analysis are performed on the temperature response sample values ​​at each time step. Expectation analysis is a statistical method that determines the most likely value of a random variable by calculating the sample arithmetic mean, while variance analysis is a statistical method that quantifies the fluctuation range of a random variable by calculating the sample dispersion. The specific calculation method is as follows: for the t-th time step within a future preset period, t=1, 2, ..., 60, corresponding to a 1-hour prediction duration and a 1-minute time step, the stator winding temperature values ​​of all 1000 samples at that time step are extracted. ; Rotor temperature values: ; Calculate the expected values ​​of stator winding temperature separately: ; in, Here are the temperature data for the i-th sampling point on the stator and the expected rotor temperature: ; in, The temperature value at the i-th sampling time of the rotor measuring point is the expected value, which reflects the most likely temperature value of the generator's key thermal node at that time step under the current unsteady-state condition. Simultaneously, the temperature variance of the stator windings is calculated. ; and rotor temperature variance ; Variance quantifies the degree of dispersion of temperature values ​​around the expected value. A larger variance indicates a greater range of temperature fluctuations due to uncertainties in operating conditions, resulting in poorer operational stability of the cooling system. Next, based on the expected value and variance of each time step, a temperature probability distribution for that time step is constructed. Due to the central limit theorem, the mean distribution of a large number of independent random samples approximately follows a normal distribution. Therefore, this scheme uses a normal distribution to fit the temperature probability density function for each time step. The stator winding temperature probability density function is: ; in, Let be the normal distribution probability density function of the stator temperature at time t. Here, represents the stator temperature sample mean, and represents the center value of the distribution. The standard deviation of the stator temperature sample represents the magnitude of data dispersion. Let temperature be a random variable. For pi and natural constant, the rotor temperature probability density function is: ; in, Let be the normal distribution probability density function of the rotor temperature at time t. The mean of the rotor temperature samples is located at the center of the normal distribution. The standard deviation of rotor temperature samples reflects the degree of data dispersion. Let temperature be a random variable. Pi is the natural constant.

[0067] Finally, the expected temperature values, variances, and probability density functions of all 60 time steps are integrated in chronological order to form a three-dimensional data structure that changes continuously over time. The x-axis represents time, the y-axis represents temperature, and the z-axis represents probability density. This data structure is the probability distribution space of the thermal field temperature rise, which refers to the complete set of all possible temperature values ​​and their corresponding probabilities of key thermal nodes such as the generator stator winding and rotor within a preset future time period. It includes the probability envelope of expected temperature values, fluctuation ranges, and occurrence probabilities.

[0068] Furthermore, in the method provided in the application embodiment, the current cooling efficiency of the target generator is evaluated based on the thermal field temperature rise probability distribution space to obtain an efficiency evaluation result, including: extracting the temperature probability density curve in the thermal field temperature rise probability distribution space; calculating the integral area in the temperature probability density curve where the temperature response value exceeds the heat resistance limit of the insulation material of the target generator, and obtaining the ratio of the integral area to the total area of ​​the temperature probability density curve; determining whether the ratio is greater than a preset risk threshold, and if so, determining that the current cooling efficiency is insufficient, and obtaining cooling failure warning information as the efficiency evaluation result.

[0069] Specifically, based on the thermal field temperature rise probability distribution space, which is a complete three-dimensional data set of all possible temperature values ​​and their corresponding probabilities of key thermal nodes such as the generator stator winding and rotor within a preset future time period, including the expected value, variance, and probability density function of the temperature at each time step, the current cooling efficiency of the target generator is evaluated to obtain the efficiency evaluation result. Specifically, the temperature probability density curves corresponding to the stator winding nodes and rotor nodes in the thermal field temperature rise probability distribution space are extracted step by step. The temperature probability density curve is a continuous smooth curve describing the probability density of the occurrence of the temperature random variable within a unit length near different value points. The integral area of ​​any interval under the curve represents the probability that the temperature falls within that interval, and its total area is always equal to 1. For each 1-minute time step in the next 1 hour, the temperature probability density curve data of the stator winding and rotor are extracted respectively. Each curve contains 100 uniformly distributed temperature sampling points and their corresponding probability density values.

[0070] Next, the heat resistance limit parameter of the insulation material of the target generator is obtained. The heat resistance limit of the insulation material refers to the highest temperature at which the insulation material used in the generator winding can operate safely for a long time without irreversible performance degradation. It is the threshold for safe operation of the generator and is determined by the insulation material grade. It is used as the critical value for over-temperature judgment.

[0071] Then, the trapezoidal numerical integration method is used to calculate the integral area of ​​the temperature response value exceeding the heat resistance limit of the insulation material in each temperature probability density curve. That is, the definite integral is calculated from the heat resistance limit temperature to the maximum temperature value of the curve. This integral area is the probability of cooling failure of the generator at the corresponding time step. The cooling failure probability is the likelihood that the temperature of the generator's key thermal node exceeds the heat resistance limit of the insulation material, leading to accelerated aging or even breakdown of the insulation. It is the core indicator for quantitatively assessing the operating risk of the cooling system. Since the total area of ​​the probability density curve is always equal to 1, this integral area is directly equal to the cooling failure probability, and there is no need to calculate the proportion separately.

[0072] Subsequently, a preset cooling efficiency risk threshold is established, typically set based on the wind farm's safety operation standards, equipment importance, and maintenance strategies. It is generally set to 0.05 (5%), indicating that when the cooling failure probability exceeds 5%, the current cooling system is considered to have an unacceptable safety hazard, requiring maintenance measures. The cooling failure probability of the stator winding and rotor at each time step is compared point-by-point with the preset risk threshold. If the cooling failure probability at all time steps is less than the preset risk threshold, the current cooling efficiency is deemed sufficient, and the cooling system is output as operating normally, with no overheating risk in the next hour. If the cooling failure probability at any time step is greater than or equal to the preset risk threshold, the current cooling efficiency is deemed insufficient. Simultaneously, the time point when the cooling failure probability first exceeds the threshold, the maximum failure probability value, and the corresponding risk location (stator winding or rotor) are automatically recorded. A cooling failure warning message containing "insufficient cooling efficiency warning," the specific risk time, and the risk level (divided into low, medium, and high levels based on the maximum failure probability and the risk location) is generated as the final efficiency assessment result.

[0073] In summary, the generator cooling efficiency evaluation method based on temperature rise mechanism provided in this application has the following technical effects: First, a coupled lumped-parameter thermal network model, including stator winding nodes, rotor nodes, and cooling medium circuits, is established based on the generator's internal physical structure. Real-time data on three-phase current, real-time power, ambient temperature, and relative humidity under unsteady-state conditions are collected. After principal component analysis to eliminate data coupling correlations, the data is input into a Gaussian mixture model. The probability distribution characteristics of the convective heat transfer coefficient are solved using the expectation-maximization algorithm. Then, Latin hypercube sampling is used to obtain dynamic boundary conditions. The non-homogeneous state matrix is ​​numerically iterated using the fourth-order Runge-Kutta method to obtain the probability distribution space of thermal field temperature rise within a preset future time period. Finally, the cooling efficiency is quantitatively evaluated based on the temperature over-temperature probability, and graded early warning information is generated, effectively reducing the probability of equipment failure and improving the safety and reliability of generator operation. Example 2, based on the same inventive concept as the generator cooling efficiency evaluation method based on the temperature rise mechanism in the previous examples, such as... Figure 2 As shown in the embodiment of this application, a generator cooling efficiency evaluation system based on temperature rise mechanism is provided. The system includes: The model building module 11 is used to establish a lumped parameter thermal network model based on the internal physical structure of the target generator; the operating condition acquisition module 12 is used to collect the three-phase current, real-time power, ambient temperature, and ambient relative humidity of the target generator under unsteady operating conditions in real time to obtain real-time operating condition data; the data input module 13 is used to input the real-time operating condition data into a Gaussian mixture model for joint probability density estimation to obtain the probability distribution characteristics of the convective heat transfer coefficient, wherein the probability distribution characteristics include the expected value and variance space; the parameter sampling module 14 is used to sample random parameters from the probability distribution characteristics of the convective heat transfer coefficient, and substitute the sampled random variables as dynamic boundary conditions into the lumped parameter thermal network model for numerical iterative solution to obtain the thermal field temperature rise probability distribution space of the target generator in a future preset period; the thermal field evaluation module 15 is used to evaluate the current cooling efficiency of the target generator based on the thermal field temperature rise probability distribution space to obtain the efficiency evaluation result.

[0074] Furthermore, the model building module 11 is also used to perform the following steps: the lumped parameter thermal network model includes stator winding nodes, rotor nodes and cooling medium loops, and each node is coupled to the other through convective heat transfer coefficients.

[0075] Furthermore, the data input module 13 is also used to perform the following steps: perform principal component analysis on the real-time operating data to eliminate the coupling correlation between the three-phase current, real-time power, ambient temperature and ambient relative humidity, and obtain multiple independent feature vectors; combine the multiple feature vectors into a feature matrix, and input the feature matrix into a Gaussian mixture model, update the model parameters through the expectation-maximization algorithm, and obtain the probability density function of the convective heat transfer coefficient.

[0076] Furthermore, the data input module 13 is also used to perform the following steps: using the expectation-maximization algorithm to iteratively update the weight parameters, mean vector and covariance matrix in the Gaussian mixture model until the log-likelihood function of the Gaussian mixture model converges; based on the converged Gaussian mixture model, analyzing the feature matrix to obtain the probability density function of the convective heat transfer coefficient, and using the probability density function as the probability distribution feature.

[0077] Furthermore, the parameter sampling module 14 is also used to perform the following steps: calculate the internal loss power of the stator winding nodes and rotor nodes based on the real-time power and three-phase current in the real-time operating data, and use the internal loss power as the input heat source; convert the convective heat transfer coefficient sampled from the expected value and variance space into the dynamic thermal resistance between each node, and combine the input heat source to solve the transient heat conduction process between the stator winding nodes, rotor nodes and cooling medium circuit to obtain the thermal field temperature rise probability distribution space.

[0078] Furthermore, the parameter sampling module 14 is also used to perform the following steps: establishing a heat flow transfer topology based on the relative geometric positions between the stator winding nodes, rotor nodes, and cooling medium circuits; based on the heat flow transfer topology, loading the input heat source as a heat flow injection term onto the corresponding stator winding nodes and rotor nodes respectively, and performing a matrix combination of the dynamic thermal resistance and the internal heat capacity of the stator winding nodes and rotor nodes according to the law of conservation of energy to obtain a non-homogeneous state matrix; and using a numerical iterative algorithm to iteratively solve the non-homogeneous state matrix to obtain the thermal field temperature rise probability distribution space of the target generator in a future preset time period.

[0079] Furthermore, the parameter sampling module 14 is also used to perform the following steps: setting the total iteration duration and time step; using the fourth-order Runge-Kutta method to perform time step iterative calculation on the non-homogeneous state matrix until the total iteration duration is met, to obtain multiple temperature response values ​​of the stator winding node and rotor node at each time step; and analyzing the multiple temperature response values ​​to obtain the thermal field temperature rise probability distribution space.

[0080] Furthermore, the parameter sampling module 14 is also used to perform the following steps: perform expectation analysis and variance analysis on the multiple temperature response sample values ​​to obtain the expectation value and variance; and construct the thermal field temperature rise probability distribution space based on the expectation and variance.

[0081] Furthermore, the thermal field assessment module 15 is also used to perform the following steps: extract the temperature probability density curve in the thermal field temperature rise probability distribution space; calculate the integral area in the temperature probability density curve where the temperature response value exceeds the heat resistance limit of the insulation material of the target generator, and obtain the ratio of the integral area to the total area of ​​the temperature probability density curve; determine whether the ratio is greater than a preset risk threshold, and if so, determine that the current cooling efficiency is insufficient, and obtain cooling failure warning information as the efficiency assessment result.

[0082] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made 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 generator cooling efficiency evaluation method based on temperature rise mechanism, characterized in that, The method includes: Establish a lumped parameter thermal network model based on the internal physical structure of the target generator; Real-time data acquisition is performed on the three-phase current, real-time power, ambient temperature and relative humidity of the target generator under non-steady-state operating conditions to obtain real-time operating data. The real-time operating data is input into a Gaussian mixture model for joint probability density estimation to obtain the probability distribution characteristics of the convective heat transfer coefficient, wherein the probability distribution characteristics include the expected value and the variance space. Random parameters are sampled from the probability distribution characteristics of the convective heat transfer coefficient, and the sampled random variables are substituted into the lumped parameter thermal network model as dynamic boundary conditions for numerical iterative solution to obtain the thermal field temperature rise probability distribution space of the target generator in the future preset time period. The current cooling efficiency of the target generator is evaluated based on the spatial distribution of the thermal field temperature rise probability, and the efficiency evaluation result is obtained.

2. The generator cooling efficiency evaluation method based on temperature rise mechanism as described in claim 1, characterized in that, The lumped parameter thermal network model includes stator winding nodes, rotor nodes, and cooling medium loops, and the nodes are coupled to each other through convective heat transfer coefficients.

3. The generator cooling efficiency evaluation method based on temperature rise mechanism as described in claim 1, characterized in that, The real-time operating data is input into a Gaussian mixture model for joint probability density estimation to obtain the probability distribution characteristics of the convective heat transfer coefficient, including: Principal component analysis is performed on the real-time operating data to eliminate the coupling correlation between the three-phase current, real-time power, ambient temperature and ambient relative humidity, and to obtain multiple independent feature vectors. The multiple feature vectors are combined into a feature matrix, and the feature matrix is ​​input into a Gaussian mixture model. The model parameters are updated by the expectation-maximization algorithm to obtain the probability density function of the convective heat transfer coefficient.

4. The generator cooling efficiency evaluation method based on temperature rise mechanism as described in claim 3, characterized in that, The multiple feature vectors are combined into a feature matrix, which is then input into a Gaussian mixture model. The model parameters are updated using an expectation-maximization algorithm to obtain the probability density function of the convective heat transfer coefficient, including: The expectation-maximization algorithm is used to iteratively update the weight parameters, mean vector, and covariance matrix in the Gaussian mixture model until the log-likelihood function of the Gaussian mixture model converges. Based on the converged Gaussian mixture model, the feature matrix is ​​analyzed to obtain the probability density function of the convective heat transfer coefficient, and the probability density function is used as the probability distribution feature.

5. The generator cooling efficiency evaluation method based on temperature rise mechanism as described in claim 1, characterized in that, Random parameters are sampled from the probability distribution characteristics of the convective heat transfer coefficient. These sampled random variables are then substituted as dynamic boundary conditions into the lumped-parameter thermal network model for numerical iterative solution. This yields the probability distribution space of the thermal field temperature rise of the target generator within a predetermined future time period, including: Based on the real-time power and three-phase current in the real-time operating data, the internal loss power of the stator winding nodes and rotor nodes is calculated, and the internal loss power is used as the input heat source. The convective heat transfer coefficient sampled from the expected value and variance space is converted into the dynamic thermal resistance between each node. Combined with the input heat source, the transient heat conduction process between the stator winding node, rotor node and cooling medium circuit is solved to obtain the thermal field temperature rise probability distribution space.

6. The generator cooling efficiency evaluation method based on temperature rise mechanism as described in claim 5, characterized in that, The convective heat transfer coefficient sampled from the expected value and variance space is converted into the dynamic thermal resistance between each node. Combined with the input heat source, the transient heat conduction process between the stator winding nodes, rotor nodes, and cooling medium loop is solved to obtain the thermal field temperature rise probability distribution space, including: A heat flow transfer topology is established based on the relative geometric positions of the stator winding nodes, rotor nodes, and cooling medium loops. Based on the heat transfer topology, the input heat source is loaded as a heat injection term onto the corresponding stator winding node and rotor node, and the dynamic thermal resistance and the internal heat capacity of the stator winding node and rotor node are combined into a matrix according to the law of energy conservation to obtain a non-homogeneous state matrix. The non-homogeneous state matrix is ​​solved iteratively using a numerical iterative algorithm to obtain the thermal field temperature rise probability distribution space of the target generator in a future preset time period.

7. The generator cooling efficiency evaluation method based on temperature rise mechanism as described in claim 6, characterized in that, The non-homogeneous state matrix is ​​solved iteratively using a numerical iterative algorithm to obtain the thermal field temperature rise probability distribution space of the target generator within a preset future time period, including: Set the total iteration duration and time step; The non-homogeneous state matrix is ​​calculated iteratively over time step by time step using the fourth-order Runge-Kutta method until the total iteration time is satisfied, thereby obtaining multiple temperature response values ​​of the stator winding node and rotor node at each time step. Based on the analysis of the multiple temperature response values, the probability distribution space of the temperature rise in the thermal field is obtained.

8. The generator cooling efficiency evaluation method based on temperature rise mechanism as described in claim 7, characterized in that, Based on the analysis of the multiple temperature response values, the probability distribution space of the thermal field temperature rise is obtained, including: Expectation analysis and variance analysis are performed on the multiple temperature response sample values ​​to obtain the expected value and variance. Based on the expected value and variance, the probability distribution space of the thermal field temperature rise is constructed.

9. The generator cooling efficiency evaluation method based on temperature rise mechanism as described in claim 1, characterized in that, The current cooling efficiency of the target generator is evaluated based on the spatial distribution of the thermal field temperature rise probability, and the efficiency evaluation results are obtained, including: Extract the temperature probability density curve from the thermal field temperature rise probability distribution space; Calculate the integral area in the temperature probability density curve where the temperature response value exceeds the heat resistance limit of the insulation material of the target generator, and obtain the ratio of the integral area to the total area of ​​the temperature probability density curve. Determine whether the ratio is greater than a preset risk threshold. If so, determine that the current cooling performance is insufficient and obtain cooling failure warning information as the performance evaluation result.

10. A generator cooling efficiency evaluation system based on temperature rise mechanism, characterized in that, The system is used to implement the generator cooling efficiency evaluation method based on the temperature rise mechanism according to any one of claims 1 to 9, the system comprising: The model building module is used to build a lumped parameter thermal network model based on the internal physical structure of the target generator. The operating condition acquisition module is used to collect the three-phase current, real-time power, ambient temperature and ambient relative humidity of the target generator under non-steady-state operating conditions in real time, and obtain real-time operating condition data. The data input module is used to input the real-time operating data into the Gaussian mixture model for joint probability density estimation to obtain the probability distribution characteristics of the convective heat transfer coefficient, wherein the probability distribution characteristics include the expected value and the variance space. The parameter sampling module is used to sample random parameters from the probability distribution characteristics of the convective heat transfer coefficient, and substitute the sampled random variables as dynamic boundary conditions into the lumped parameter thermal network model for numerical iterative solution to obtain the thermal field temperature rise probability distribution space of the target generator in the future preset time period. The thermal field assessment module is used to assess the current cooling efficiency of the target generator based on the thermal field temperature rise probability distribution space, and obtain the efficiency assessment result.

Images