A transformer intelligent adjustment system and method

Abstract

The present invention belongs to the technical field of power system control, and discloses a transformer intelligent regulation system and method; the method comprises: inputting real-time collected transformer operation data and weather data into a pre-trained load forecasting model to obtain predicted load data; inputting collected physical parameters and operation data, as well as pre-set calibration parameters, into a pre-built digital twin model to obtain a cooling capacity index; calculating the total heat source power of copper loss and iron loss based on the predicted load data and real-time collected electrical parameters of the transformer, and dynamically quantifying the balance state of the total heat source power and the cooling capacity index through a matching evaluation function in combination with the cooling capacity index to obtain a matching evaluation index; and systematically solving the problem of dynamic imbalance between the heat dissipation capacity and load demand of the transformer under extreme environments.

Description

Technical Field

[0001] The present invention relates to the technical field of power system control, and more particularly to a transformer intelligent regulation system and method. Background Art

[0002] With the deep integration of smart grids and new power systems, transformers face severe challenges in operating reliably in complex environments. For example, one power grid is characterized by high load density and a high frequency of extreme weather. Typhoons and persistent high temperatures affect over 2,000 transformers annually. Traditional transformer cooling systems rely on static models with fixed thresholds. This exposes core issues during extreme weather conditions (including high temperatures, high pressure, and typhoons, all of which exceed the corresponding preset thresholds): high temperatures reduce air density and increase the thermal resistance of heat sinks. Typhoon-force winds reduce cooling fan efficiency and accelerate the deposition of contaminants on heat sinks, directly reducing cooling system performance. Simultaneously, the sudden load fluctuations associated with extreme weather conditions cause cooling demand to surge in a short period of time, creating a dynamic imbalance between declining cooling capacity and surging load demand.

[0003] Existing technologies lack real-time coupled analysis of meteorological parameters, equipment status, and load fluctuations. Adjustment strategies lag behind actual needs, and a dynamic evaluation model for heat dissipation capacity attenuation has not been established. Adjustment based solely on manual experience can easily lead to insufficient or excessive cooling, resulting in a dynamic imbalance between the transformer's heat dissipation capacity and load demand in extreme environments.

[0004] In view of this, the present invention proposes a transformer intelligent regulation system and method to solve the above problems. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art and achieve the above-mentioned objectives, the present invention provides the following technical solutions: a transformer intelligent adjustment method, comprising:

[0006] Input the real-time collected transformer operation data and weather data into the pre-trained load forecasting model to obtain the predicted load data;

[0007] The collected physical parameters and operating data, as well as pre-set calibration parameters, are input into the pre-built digital twin model to obtain the cooling capacity index;

[0008] The total heat source power of copper loss and iron loss is calculated based on the predicted load data and the real-time collected electrical parameters of the transformer. Combined with the cooling capacity index, the matching evaluation function dynamically quantifies the balance between the total heat source power and the cooling capacity index to obtain the matching evaluation index.

[0009] If the matching evaluation index exceeds the set threshold, the real-time collected equipment status, operation data and weather data are input into the pre-trained cooling strategy model to obtain the cooling strategy;

[0010] After verifying that there are no abnormalities in the equipment status, adjustment instructions are generated and executed according to the cooling strategy. During the execution of the adjustment instructions, the calibration parameters of the digital twin model are dynamically corrected based on the real-time collected equipment status and operation data.

[0011] Furthermore, the training method of the load forecasting model includes:

[0012] Collect transformer operation data and weather data during historical extreme weather periods and mark them as training data;

[0013] The training data is normalized and preprocessed as model input samples, and the actual load data during the historical extreme weather period is used as the supervision signal;

[0014] A load forecasting model is constructed using a long short-term memory neural network; the long short-term memory neural network includes an input layer, multiple hidden layers, and a fully connected output layer; the dimensions of the input layer are consistent with the feature dimensions of the preprocessed model input samples; the fully connected output layer outputs predicted load data for a preset time period in the future and marks load surge points in the predicted load data;

[0015] With the goal of minimizing the mean square error between the predicted load and the supervisory signal, the load forecasting model parameters are iteratively optimized through the back propagation algorithm to obtain the final load forecasting model.

[0016] Furthermore, the method for determining whether a data point in the predicted load data is a load surge point includes:

[0017] Construct a multidimensional feature space including time characteristics, load historical trend characteristics and extreme weather characteristics; the time characteristics include periodic indicators of daily cycles, weekly cycles and monthly cycles; the load historical trend characteristics include statistical indicators of historical load mean, historical load standard deviation and historical load change rate; the extreme weather characteristics are weather data and its change rate;

[0018] Based on the multidimensional feature space, the characteristic value of each data point of the predicted load data is calculated in sequence to construct a characteristic value set; the characteristic value includes the load growth rate, the deviation from the historical load mean, and the correlation with the weather data;

[0019] Calculate the matching degree of each eigenvalue in the eigenvalue set with the preset surge point eigenvalue; if the matching degree exceeds the set matching degree threshold, the corresponding data point is determined to be a load surge point and the corresponding extreme scenario type is marked; if the matching degree does not exceed the set matching degree threshold, the corresponding data point is determined to be a normal data point; the extreme scenario types include high temperature, high pressure and strong wind;

[0020] For data points determined as load surge points in the forecast load data, verify whether the load surge persistence conditions are met.

[0021] Furthermore, the construction method of the digital twin model includes:

[0022] Construct a three-dimensional geometric model based on the collected three-dimensional geometric parameters of the transformer;

[0023] The collected transformer physical parameters and historical operating data are used as inputs to the digital twin model; the physical parameters include material thermal characteristics, oil viscosity-temperature relationship, and air density-temperature relationship;

[0024] Using computational fluid dynamics methods, the Navier-Stokes equations and the Fourier heat conduction equations are coupled to establish a multi-physics coupling model;

[0025] A particle swarm optimization algorithm was introduced to calibrate the calibration parameters in the multi-physics coupling model, including the convective heat transfer coefficient and the oil circulation resistance coefficient, with the optimization goal of minimizing the root mean square error between the oil temperature distribution predicted by the multi-physics coupling model and the oil temperature distribution measured in historical operating data.

[0026] The digital twin model is composed of a multi-physics field coupling model and a three-dimensional geometric model.

[0027] Furthermore, the method for calibrating the calibration parameters in the multi-physics field coupling model includes:

[0028] Constructing a particle swarm optimization space; each particle in the particle swarm optimization space is defined as a two-dimensional vector consisting of a convective heat transfer coefficient and an oil circulation resistance coefficient;

[0029] The parameter vectors of each particle are input into the multi-physics field coupling model, and numerical simulation is performed in combination with historical operating data to obtain the predicted oil temperature distribution. The predicted oil temperature distribution is compared with the historical measured oil temperature distribution during the same period, and the root mean square error is calculated as the particle fitness value.

[0030] Each particle in the particle swarm optimization space adjusts its speed and position according to the preset rules based on the minimum root mean square error parameter combination it has searched for and the minimum root mean square error parameter combination in the current population, and searches for a better solution area in the parameter space;

[0031] When the change in the global optimal fitness value after a preset positive integer number of iterations is less than the set fitness threshold, or the maximum number of iterations is reached, the optimization process is terminated;

[0032] The convective heat transfer coefficient and oil circulation resistance coefficient obtained by particle swarm optimization with minimized root mean square error are embedded in the multi-physics coupling model to complete the calibration of the calibration parameters in the multi-physics coupling model.

[0033] Furthermore, the method for obtaining the matching evaluation index includes:

[0034] Calculate the total heat source power of copper loss and iron loss based on the predicted load data and the electrical parameters of the transformer;

[0035] The cooling capacity index obtained from the digital twin model is calculated by comparing it with the total heat source power to form a basic matching index.

[0036] Monitor the temperature distribution of key parts of the transformer. If the temperature difference between the parts is greater than the set temperature difference, a penalty correction will be made to the basic matching index. The key parts include the winding, core and top layer of the oil tank.

[0037] The basic matching index is dynamically adjusted in combination with real-time weather data to obtain a dimensionless matching evaluation index.

[0038] Furthermore, the training method of the cooling strategy model includes:

[0039] Construct a multidimensional state space including transformer equipment status, operation data and weather data;

[0040] Construct an action space for discrete adjustment modes, including fan speed adjustment, standby equipment switching, and oil pump power adjustment;

[0041] A proximal policy optimization algorithm is used to construct a policy network; the policy network takes a multidimensional state space as input and an action probability distribution in an action space as output;

[0042] Design a compound reward function that includes safety reward, energy consumption reward, and lifespan reward;

[0043] Based on the composite reward function, the parameters of the policy network are iteratively optimized to maximize the cumulative reward value of the adjustment method selected by the policy network and obtain the cooling policy model.

[0044] Furthermore, based on the composite reward function, the parameters of the policy network are iteratively optimized to maximize the cumulative reward value of the adjustment method selected by the policy network, including:

[0045] Use digital twin models to simulate extreme weather scenarios and generate equipment status under different extreme weather scenarios;

[0046] The policy network selects an adjustment method from the action space based on the current policy state;

[0047] The digital twin model simulates the operation of the device according to the selected adjustment method, outputs the device status of the next unit time, and calculates the reward value of the corresponding composite reward function;

[0048] The device state of the next unit time and the reward value of the corresponding compound reward function are stored in the experience pool;

[0049] The relative value of each selected moderation approach was assessed using the generalized advantage estimation method;

[0050] Based on the data in the experience pool and the relative value of the adjustment method, the gradient descent algorithm is used to iteratively optimize the parameters of the policy network to maximize the cumulative reward value of the adjustment method selected by the policy network.

[0051] Furthermore, the method for dynamically correcting the calibration parameters of the digital twin model includes:

[0052] During the execution of the cooling strategy adjustment instructions, transformer operation data and equipment status are collected in real time as feedback data for dynamic correction;

[0053] Compare the real-time prediction results of the digital twin model with the feedback data, and calculate the deviation between the real-time prediction results and the feedback data;

[0054] If the deviation exceeds the set deviation, the calibration parameters that affect the prediction accuracy are identified, and the calibration parameters are iteratively adjusted using the gradient descent algorithm based on the direction and magnitude of the deviation;

[0055] After adjustment, the prediction accuracy of the corrected digital twin model in the subsequent set period is verified. If the deviation converges to the set deviation threshold in the subsequent set period, the parameter correction is confirmed to be effective and the digital twin model is updated; if the deviation does not converge to the set deviation threshold in the subsequent set period, the next round of correction is entered until the deviation converges to the set deviation threshold.

[0056] Furthermore, the transformer operation data includes the vibration amplitude of the cooling equipment, the status of the transformer body, electrical parameters, oil temperature distribution and load information; the weather data includes ambient temperature, humidity, wind speed, precipitation and weather warning type; and the electrical parameters include load current and winding resistance.

[0057] The present invention discloses a transformer intelligent regulation system, comprising:

[0058] The load forecasting module is used to input the real-time collected transformer operation data and weather data into the pre-trained load forecasting model to obtain the predicted load data;

[0059] The digital twin module is used to input the collected physical parameters and operating data, as well as pre-set calibration parameters, into the pre-built digital twin model to obtain the cooling capacity index;

[0060] The matching evaluation module calculates the total heat source power of copper loss and iron loss based on the predicted load data and the real-time collected electrical parameters of the transformer. Combined with the cooling capacity index, it dynamically quantifies the balance between the total heat source power and the cooling capacity index through the matching evaluation function to obtain the matching evaluation index.

[0061] The intelligent decision-making module inputs the real-time collected equipment status, operating data, and weather data into the pre-trained cooling strategy model to obtain a cooling strategy if the matching evaluation index exceeds the set threshold;

[0062] The closed-loop control module is used to verify that there are no abnormalities in the equipment status and then generate and execute adjustment instructions based on the cooling strategy. During the execution of the adjustment instructions, the calibration parameters of the digital twin model are dynamically corrected based on the real-time collected equipment status and operating data.

[0063] Compared with the prior art, the technical effects and advantages of the transformer intelligent regulation system and method of the present invention are as follows:

[0064] The system generates predicted load data by collecting transformer operating and weather data in real time. The system obtains the transformer's cooling capacity indicator through a constructed digital twin model. Combining the predicted load data and cooling capacity indicator, the system dynamically evaluates the balance between the transformer's current total heat source power and cooling capacity indicator, generating a matching evaluation indicator. Based on the matching evaluation indicator and the real-time collected equipment status, operating data, and weather data, the system selects the appropriate cooling strategy through the cooling strategy model. Adjustment instructions are generated and executed based on the cooling strategy. Equipment status is verified before execution, and digital twin model parameters are dynamically modified based on feedback data after execution. The entire system performs real-time coupled analysis of weather data, equipment status, and load fluctuations, establishing a dynamic assessment model for heat dissipation capacity degradation. This addresses the dynamic imbalance between transformer heat dissipation capacity and load demand in extreme environments, enabling accurate perception of equipment status, intelligent generation of cooling strategies, and continuous optimization of adjustment effects, significantly improving the safety, economy, and reliability of transformer operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 This is a schematic diagram of a transformer intelligent regulation system according to embodiment 1 of the present invention;

[0066] Figure 2 This is a flow chart of a transformer intelligent adjustment method according to embodiment 3 of the present invention;

[0067] Figure 3This is a flow chart of a method for determining whether a load surge point is present in Example 1 of the present invention;

[0068] Figure 4 This is a flow chart of a method for obtaining a matching evaluation index according to embodiment 1 of the present invention. DETAILED DESCRIPTION

[0069] The technical solutions in the embodiments of the present invention will be described in detail, clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention. It should be noted that the specific embodiments described below are only used to better illustrate and describe the technical solutions of the present invention, and are intended to enable those skilled in the art to better understand and implement the present invention, and should not be construed as limiting the scope of protection of the present invention. Without departing from the spirit and essence of the present invention, those skilled in the art may modify, adjust or make equivalent replacements based on the contents disclosed in the present invention, and these should all be regarded as the scope of protection of the present invention.

[0070] Example 1

[0071] See also Figure 1 As shown, this embodiment discloses a transformer intelligent regulation system, including a load forecasting module, a digital twin module, a matching evaluation module, an intelligent decision-making module and a closed-loop control module. Each module realizes data transmission through wired and / or wireless connections.

[0072] The load forecasting module is used to input the real-time collected transformer operation data and weather data into the pre-trained load forecasting model to obtain predicted load data.

[0073] The transformer operation data includes the vibration amplitude of the cooling equipment, the status of the transformer body, electrical parameters, oil temperature distribution and load information; the weather data includes ambient temperature, humidity, wind speed, precipitation and weather warning type.

[0074] The training methods for the load forecasting model include:

[0075] Collect transformer operation data and weather data during historical extreme weather periods and mark them as training data;

[0076] The training data is normalized and preprocessed as model input samples, and the actual load data during the historical extreme weather period is used as the supervision signal;

[0077] A load forecasting model is constructed using a long short-term memory neural network; the long short-term memory neural network includes an input layer, multiple hidden layers, and a fully connected output layer; the dimensions of the input layer are consistent with the feature dimensions of the preprocessed model input samples; the fully connected output layer outputs predicted load data for a preset time period in the future and marks load surge points in the predicted load data;

[0078] With the goal of minimizing the mean square error between the predicted load and the supervisory signal, the load forecasting model parameters are iteratively optimized through the back propagation algorithm to obtain the final load forecasting model.

[0079] Methods for collecting transformer operation data and weather data during historical extreme weather periods and marking them as training data include:

[0080] Transformer operating data and weather data from three consecutive extreme weather periods were collected and marked as training data. These extreme weather periods included days with orange high temperature warnings, typhoon landfalls, and cold wave passages. The training data was synchronously sampled at 15-minute intervals to form a raw data sequence with timestamps.

[0081] Methods for normalizing and preprocessing training data as model input samples and using actual load data during historical extreme weather periods as supervisory signals include:

[0082] Numerical features in transformer operating data and weather data were uniformly mapped to the [0, 1] range using the min-max scaling method. These numerical features included temperature, current, and wind speed. Categorical features, such as weather warning types, were converted to binary vectors using one-hot encoding. Each processed sample formed an input vector containing 20 features. Actual load data for the corresponding time period was extracted as a supervisory signal, resulting in a training dataset containing 100,000 samples.

[0083] The method of using a long short-term memory neural network (LSTM) to build a load forecasting model, outputting forecasted load data for a preset time period in the future, and marking load surge points in the forecasted load data includes:

[0084] The load forecasting model consists of an input layer, three LSTM hidden layers, and a fully connected output layer. The input layer has 20 neurons, matching the preprocessed feature dimensions, and receives the normalized input vector. Each LSTM hidden layer has 128 memory cells, with dropout technology used between layers to prevent overfitting. The tanh function is used as the activation function, with a dropout rate of 0.2. The fully connected output layer, with a total of 96 neurons, outputs forecasted load data for 15-minute intervals over the next 24 hours.

[0085] During model training, the mean squared error between the predicted load data and the supervisory signal is used as the primary loss function, while the auxiliary loss function uses binary cross-entropy loss. The total loss function is a weighted sum of the two. The weights of the primary and auxiliary loss functions are 0.8 and 0.2, respectively. The Adam algorithm is used for optimization, with an initial learning rate of 0.001, a batch size of 64, and 200 training cycles. During training, input samples are fed into the model in a time series. The LSTM hidden layer uses a gating mechanism to capture the temporal dependence of the load data and the abnormal fluctuations under the influence of extreme weather. The fully connected layer generates the predicted load series and the probability distribution of surge points based on the hidden layer output. After each training cycle, the loss value on the validation set is calculated. When the validation loss stops decreasing after 10 consecutive cycles, the early stopping mechanism is triggered to terminate training.

[0086] It should be noted that during the above-mentioned normalization preprocessing, the missing data is filled by the adjacent time mean interpolation method to ensure the integrity of the input samples; when constructing the LSTM hidden layer, the number of memory units in each layer can be adjusted according to computing resources and model complexity, with a typical range of 64 to 256; a dynamic learning rate adjustment strategy can be adopted during training. When the loss value decreases by less than 1%, the learning rate is attenuated by a factor of 0.9 to improve optimization efficiency.

[0087] With the goal of minimizing the mean square error between the predicted load and the supervisory signal, the load forecasting model parameters are iteratively optimized through the back propagation algorithm to obtain the final load forecasting model. The method includes:

[0088] The model parameters are iteratively optimized through the back-propagation algorithm; the load forecasting model parameters include the weight matrix and bias vector of the LSTM layer and the connection parameters of the fully connected layer; when the total loss function converges to below 0.05, the final load forecasting model is obtained.

[0089] See Figure 3 As shown, the method for determining whether a data point in the predicted load data is a load surge point includes:

[0090] Construct a multidimensional feature space including time characteristics, load historical trend characteristics and extreme weather characteristics; the time characteristics include periodic indicators of daily cycles, weekly cycles and monthly cycles; the load historical trend characteristics include statistical indicators of historical load mean, historical load standard deviation and historical load change rate; the extreme weather characteristics are weather data and its change rate;

[0091] Based on the multidimensional feature space, the characteristic value of each data point of the predicted load data is calculated in sequence to construct a characteristic value set; the characteristic value includes the load growth rate, the deviation from the historical mean, and the correlation with the weather data;

[0092] The matching degree of each eigenvalue in the eigenvalue set is calculated with the preset surge point eigenvalue; if the matching degree exceeds the set matching degree threshold, the corresponding data point is determined to be a load surge point and the corresponding extreme scenario type is marked; if the matching degree does not exceed the set matching degree threshold, the corresponding data point is determined to be a normal data point; the extreme scenario types include high temperature, high pressure, high humidity and strong wind;

[0093] For the data points determined as load surge points in the predicted load data, further verification is conducted to determine whether the load surge continuity conditions are met.

[0094] The methods for constructing a multidimensional feature space including time characteristics, load historical trend characteristics, and extreme weather characteristics include:

[0095] Based on the time series characteristics of the predicted load data, three types of features are extracted to construct the feature space: the time feature uses the sliding window technology to calculate the load mean within the daily cycle (24 hours), weekly cycle (7 days), and monthly cycle (30 days) as well as the periodic indicators of the peak occurrence time, which are used to identify the time pattern of load fluctuations; the load historical trend feature is based on the load data of the past three years, and calculates the historical load mean, standard deviation and 15-minute sliding window change rate corresponding to each time point. The historical mean reflects the normal load level, the standard deviation represents the load fluctuation range, and the change rate is used to measure the load change speed; the extreme meteorological feature collects real-time ambient temperature, wind speed, precipitation and their hourly change rate. For example, the temperature change rate is defined as the temperature difference between the current moment and the previous hour, and the wind speed change rate is defined as the wind speed difference between the current moment and the previous hour, which are used to quantify the impact of extreme weather on load.

[0096] Based on the multidimensional feature space, the eigenvalue of each data point of the predicted load data is calculated in sequence. The method of constructing the eigenvalue set includes:

[0097] For time characteristics, calculate the deviation between the load mean value and the historical daily mean value in the same time period of the daily cycle corresponding to the current data point. The calculation method is:

[0098] ;

[0099] Where, The deviation between the load mean value and the historical daily mean value in the same time period of the daily cycle corresponding to the current data point; is the current load value; The load average value of the same day and time period in history; It is the standard deviation of load during the same day and time period in history.

[0100] The calculation method for the load growth rate in the load history trend feature is:

[0101] ;

[0102] Where, is the load growth rate; is the load value of the previous sampling period.

[0103] The deviation between the load historical trend characteristics and the historical mean is calculated as:

[0104] ;

[0105] Where, is the deviation between the historical trend characteristics of the load and the historical mean; It is the historical load average for the entire period.

[0106] The correlation of extreme meteorological characteristics is calculated using the Pearson correlation coefficient. For example, the temperature correlation is calculated as:

[0107] ;

[0108] Where, is the temperature dependence; is the historical temperature data; is the corresponding load data; is the mean of the temperature data; is the mean value of load data; is the standard deviation of the temperature data; is the standard deviation of the load data; i is the index of the data, representing the i-th sample, for example, the first temperature and load data is and n is the number of data, i.e., the number of temperature and load data. A five-dimensional feature value set is formed, including time deviation, load growth rate, historical deviation, temperature correlation, and wind speed correlation.

[0109] The method of calculating the matching degree between each eigenvalue in the eigenvalue set and the preset surge point eigenvalue, and determining that the corresponding data point is a load surge point if the matching degree exceeds the set matching degree threshold, and marking the corresponding extreme scenario type includes:

[0110] The preset characteristic values ​​are derived based on statistics of historical extreme scenario data: for example, the load growth rate threshold in high temperature scenarios is 20%, the historical deviation threshold is 1.5, and the temperature correlation threshold is 0.8; the load growth rate threshold in high pressure scenarios is 15%, the historical deviation threshold is 1.3, and the wind speed correlation threshold is 0.6; the load growth rate threshold in strong wind scenarios is 18%, the historical deviation threshold is 1.4, and the wind speed correlation threshold is 0.7.

[0111] The matching degree is calculated using a weighted summation method, with the following weights: load growth rate 0.4, historical deviation 0.3, temperature correlation 0.15, and wind speed correlation 0.15. The load growth rate directly reflects the magnitude of the surge; the historical deviation reflects the degree of anomaly; the temperature correlation is a sensitivity factor for high-temperature scenarios; and the wind speed correlation is a sensitivity factor for strong-wind scenarios. When the matching degree is greater than or equal to 0.7, the data point is identified as a candidate load surge point, and the corresponding extreme scenario type is labeled based on the maximum correlation feature.

[0112] For data points determined as load surge points in the forecast load data, methods for further verifying whether the load surge persistence conditions are met include:

[0113] Check whether the load growth rate at the data point and for the next two consecutive sampling periods (a total of three periods) is greater than 15% for each 45-minute period, and whether the historical deviation is greater than 1.2. This condition is set to eliminate short-term noise interference. For example, a transient pulse load may cause a single data point to be abnormal, but if the duration is insufficient, it does not constitute a true surge scenario. If the persistence condition is met, the point is ultimately determined to be a load surge point; if not, the candidate point is eliminated.

[0114] It should be noted that in the construction of the above-mentioned multidimensional feature space, the length of the time window can be adjusted according to the actual operating cycle of the transformer. The typical daily cycle window is 24 hours, the weekly cycle is 7 days, and the monthly cycle is 30 days, so as to cover the diurnal fluctuations of the load, weekend effects and seasonal changes; when calculating the characteristic values, the sample size of historical data must be no less than 3 years to ensure the stability of the statistical parameters; the matching weight distribution is determined by machine learning training of historical surge events, and the logistic regression model is used to optimize the weights of each feature to make the misjudgment rate less than 5%; the number of sampling cycles for the continuity condition is set to 3, corresponding to a duration of 45 minutes, which meets the minimum response cycle requirements for power system load regulation, which can both avoid misjudgment and ensure the timeliness of the warning.

[0115] It should be noted that the above method can systematically identify load surge points from three dimensions: time patterns, load fluctuation characteristics, and extreme weather impacts. Combined with a continuous verification mechanism, it can improve the accuracy of judgment and provide a reliable decision-making basis for the pre-adjustment of transformer cooling strategies.

[0116] The digital twin module is used to input the collected physical parameters and operating data, as well as pre-set calibration parameters, into a pre-built digital twin model to obtain cooling capacity indicators.

[0117] Methods for building digital twin models include:

[0118] Construct a three-dimensional geometric model based on the collected three-dimensional geometric parameters of the transformer;

[0119] The collected transformer physical parameters and historical operating data are used as inputs to the digital twin model; the physical parameters include material thermal characteristics, oil viscosity-temperature relationship, and air density-temperature relationship;

[0120] Using computational fluid dynamics methods, the Navier-Stokes equations and the Fourier heat conduction equations are coupled to establish a multi-physics coupling model;

[0121] A particle swarm optimization algorithm was introduced to calibrate the calibration parameters in the multi-physics coupling model, including the convective heat transfer coefficient and the oil circulation resistance coefficient, with the optimization goal of minimizing the root mean square error between the oil temperature distribution predicted by the multi-physics coupling model and the actual oil temperature distribution measured in historical operating data.

[0122] The digital twin model is composed of a multi-physics field coupling model and a three-dimensional geometric model.

[0123] Based on the collected three-dimensional geometric parameters of the transformer, the method of constructing a three-dimensional geometric model includes:

[0124] Laser scanning technology was used to capture the three-dimensional geometric parameters of the transformer's internal winding arrangement, oil channel distribution, and radiator structure. Scanning accuracy was controlled to the millimeter level to ensure the model accurately reflects the transformer's internal spatial structure. Computer-aided design software was used to construct the 3D geometric model. The acquired point cloud data was gridded to generate a 3D solid model containing at least 100,000 grid cells. The grid density of key heat-generating components, including the high-voltage and low-voltage windings, was maintained at no less than 20 cells per square centimeter to ensure the accuracy of subsequent thermal analysis.

[0125] It should be noted that during the construction of the aforementioned three-dimensional geometric model, the accuracy of the meshing affects the accuracy and efficiency of the calculation results. A finer mesh provides more accurate results, but increases the computational effort and time; a coarser mesh provides faster computation, but may compromise accuracy. Therefore, it is necessary to make a trade-off based on the actual situation and select an appropriate meshing accuracy. In this embodiment, after multiple experiments and verifications, a meshing scheme was determined that ensures both accuracy and efficiency.

[0126] Methods for using collected transformer physical parameters and historical operating data as input to the digital twin model include:

[0127] The system collects thermal parameters of transformer materials, including the thermal conductivity of the winding copper material and the specific heat capacity of the core silicon steel sheet. The typical thermal conductivity of the winding copper material is 385W / (m・K), and the typical specific heat capacity of the core silicon steel sheet is 460J / (kg・K). The system also obtains the viscosity-temperature curve of the transformer oil, which is experimentally measured over a temperature range of -20°C to 120°C and is used to describe the impact of oil temperature changes on oil fluidity. The system also collects data on the air density-temperature relationship, which conforms to the ideal gas state equation and is used to calculate the heat exchange characteristics of the air surrounding the radiator. Furthermore, the system collects historical operating data from the past three years, including oil temperature distribution and cooling equipment power under different load conditions. The data is sampled every 15 minutes, resulting in a historical dataset containing at least 100,000 records.

[0128] It's important to note that when collecting physical parameters and historical operating data, data accuracy and completeness are crucial to model performance. To ensure data accuracy, high-precision measuring instruments and equipment are used for data collection, and the collected data undergoes multiple verification and corrections. Furthermore, to ensure data integrity, as much historical operating data as possible is collected, including data from different seasons and load conditions, so that the model can better adapt to various actual operating conditions.

[0129] Methods for establishing a multi-physics coupling model by coupling the Navier-Stokes equations and the Fourier heat conduction equation using computational fluid dynamics include:

[0130] Computational fluid dynamics (CFD) methods were used to couple the Navier-Stokes and Fourier heat conduction equations. The Navier-Stokes equations describe the flow characteristics of transformer oil, taking into account the temperature-dependent variations in oil viscosity and density. The Fourier heat conduction equations are used to calculate the heat transfer process within the transformer, including heat conduction through solid components and convection through the fluid. At least five monitoring points were set up in the model, located at key locations such as the winding hotspots, the center of the core, and the oil conservator, for subsequent comparison with measured data.

[0131] It should be noted that in the multi-physics field coupling model, the solution of the Navier-Stokes equations and the Fourier heat conduction equation requires the use of appropriate numerical methods and algorithms. In this embodiment, the two equations are discretized using the finite volume method, and the discretized equations are solved using an iterative solution algorithm. The finite volume method is a commonly used numerical calculation method with the advantages of high computational accuracy and good stability. The iterative solution algorithm can effectively process complex nonlinear equations, and by continuously iteratively updating the numerical value of the solution, it eventually converges to a solution that meets the accuracy requirements.

[0132] A particle swarm optimization algorithm is introduced, with the optimization goal of minimizing the root mean square error between the oil temperature distribution predicted by the multi-physics coupling model and the oil temperature distribution measured in historical operating data. The calibration parameters in the multi-physics coupling model are calibrated using the following methods:

[0133] Constructing a particle swarm optimization space; each particle in the particle swarm optimization space is defined as a two-dimensional vector consisting of a convective heat transfer coefficient and an oil circulation resistance coefficient;

[0134] The parameter vectors of each particle are input into the multi-physics field coupling model, and numerical simulation is performed in combination with historical operating data to obtain the predicted oil temperature distribution. The predicted oil temperature distribution is compared with the historical measured oil temperature distribution during the same period, and the root mean square error is calculated as the particle fitness value.

[0135] Each particle in the particle swarm optimization space adjusts its speed and position according to the preset rules based on the minimum root mean square error parameter combination it has searched for and the minimum root mean square error parameter combination in the current population, and searches for a better solution area in the parameter space;

[0136] When the change in the global optimal fitness value after a preset positive integer number of iterations is less than the set fitness threshold, or the maximum number of iterations is reached, the optimization process is terminated;

[0137] The convective heat transfer coefficient and oil circulation resistance coefficient obtained by particle swarm optimization with minimized root mean square error are embedded in the multi-physics coupling model to complete the calibration of the calibration parameters in the multi-physics coupling model.

[0138] Methods for constructing particle swarm optimization space include:

[0139] The key parameters to be optimized are defined as a two-dimensional parameter vector ,in, The convective heat transfer coefficient of the radiator surface is used to characterize the heat exchange efficiency between the air and the radiator. Its initial value range is set to 50 to 200, which covers the changes in the convective heat transfer capacity under normal operation and extreme environments of the transformer. Represents the transformer oil circulation resistance coefficient, which describes the resistance characteristics of oil flow in the circulation path. Its initial value range is set to 0.1 to 0.3, based on the transformer oil channel structure and fluid mechanics. Each particle initially generates a random initial position and initial velocity within the above parameter range. The initial velocity range is set to 20% of the position range, i.e., -30 to 30 and -0.04 to 0.04, respectively, to ensure sufficient initial exploration of the particle within the search space.

[0140] The parameter vector of each particle is input into the multi-physics field coupling model, and numerical simulation is performed in combination with historical operating data to obtain the predicted oil temperature distribution. The predicted oil temperature distribution is then compared with the historical measured oil temperature distribution during the same period. The root mean square error is calculated as the particle fitness value. The method includes:

[0141] The parameter vector of each particle is input into the established multi-physics coupling model, and numerical simulation is performed in combination with historical operating data. The historical operating data includes historical data of at least 10 typical operating conditions, such as high temperature, low temperature, rated load, and overload. Each operating condition contains no less than 500 sets of sampling data, and the sampling interval is 10 minutes. During the simulation process, the multi-physics coupling model calculates the temperature distribution of the transformer winding, core, and oil based on the input convective heat transfer coefficient and oil circulation resistance coefficient. The predicted oil temperature distribution is compared with the measured oil temperature data of the same period, and the root mean square error formula is used to calculate the fitness value. The measured oil temperature data of the same period is collected by the transformer's built-in temperature sensor, with an accuracy of ±0.5°C, and is arranged at five key locations, such as the winding hotspot, the middle of the core, and the oil surface.

[0142] In the particle swarm optimization space, each particle adjusts its speed and position according to the preset rules based on the minimum error parameter combination it has searched for and the minimum error parameter combination in the current population. The method of searching for a better solution region in the parameter space includes:

[0143] Each particle tracks two optimal positions during the search process: one is the minimum error parameter combination searched by the particle itself, that is, the individual optimal position; the other is the minimum error parameter combination currently searched by the entire particle swarm, that is, the global optimal position. The particle's speed and position are updated according to the following rules:

[0144] ;

[0145] ;

[0146] Where, is the current velocity of the particle; is the current position of the particle; is the updated velocity of the particle; The updated position of the particle; is the optimal position of the individual; is the global optimal position; is the inertia weight, which is initially set to 0.9 and linearly decreases to 0.4 with the number of iterations to balance the global search and local development capabilities; and is the learning factor, which is set to 2.0 and is used to adjust the learning intensity of particles toward the individual optimal and global optimal positions; 、 is a random number between 0 and 1, used to increase the randomness of the search process. Through the above update rules, particles continuously move to the region of better solutions in the parameter space, achieving the coordinated optimization of individual experience and group information.

[0147] When the change in the global optimal fitness value after a preset positive integer number of iterations is less than the set fitness threshold, or the maximum number of iterations is reached, the method for terminating the optimization process includes:

[0148] The maximum number of iterations was set to 200 to avoid excessive consumption of computing resources. Furthermore, when the change in the global optimal fitness value after 10 consecutive iterations was less than 0.1, the algorithm was considered to have converged and the optimization process was terminated early. This termination condition ensured sufficient search time for the algorithm while avoiding ineffective recalculation, balancing calibration accuracy and efficiency.

[0149] The method of embedding the minimum root mean square error convective heat transfer coefficient and the oil circulation resistance coefficient obtained by particle swarm optimization into the multi-physics coupling model to complete the calibration of the calibration parameters in the multi-physics coupling model includes:

[0150] After the optimization process terminates, the parameter combination that minimizes the root mean square error (RMS) found by the particle swarm search—that is, the optimal convective heat transfer coefficient and oil circulation resistance coefficient—is embedded in the multiphysics coupling model. After embedding, the model's accuracy is verified: under normal operating conditions, the RMS error between the model's predicted and measured oil temperatures should not exceed 1.5; under extreme operating conditions, such as a sudden 10°C rise in ambient temperature and a 30% increase in load, the error should not exceed 2.0. If verification passes, calibration is complete. If not, the initial parameter range can be adjusted or the algorithm parameters can be optimized before recalibration.

[0151] It should be noted that the parameter setting of the above-mentioned particle swarm optimization algorithm is determined based on a balance between computational complexity and search accuracy. Numerous simulation experiments have verified that this number can both ensure population diversity and avoid excessive computational time. The selection of historical operating data must cover typical operating conditions throughout the transformer's life cycle to ensure the wide applicability of the calibrated model. The placement of monitoring points must conform to the transformer's thermal distribution characteristics, prioritizing areas with large temperature gradients to improve the accuracy of fitness calculations. This method effectively optimizes the key parameters of the multi-physics coupling model, improves the digital twin model's prediction accuracy for the transformer's thermal state, and provides reliable physical model support for the intelligent generation of subsequent cooling strategies.

[0152] The matching evaluation module calculates the total heat source power of copper loss and iron loss based on the predicted load data and the real-time collected electrical parameters of the transformer, and combines it with the cooling capacity index to dynamically quantify the balance state of the total heat source power and the cooling capacity index through the matching evaluation function to obtain the matching evaluation index.

[0153] See Figure 4 As shown, the method for obtaining the matching evaluation index includes:

[0154] Calculate the total heat source power of winding copper loss and core iron loss based on the predicted load data and transformer electrical parameters;

[0155] The cooling capacity index obtained from the digital twin model is calculated by comparing it with the total heat source power to form a basic matching index.

[0156] Monitor the temperature distribution of key parts of the transformer. If the temperature difference between the parts is greater than the set temperature difference, a penalty correction will be made to the basic matching index. The key parts include the winding, core and top layer of the oil tank.

[0157] The basic matching index is dynamically adjusted in combination with real-time weather data to obtain a dimensionless matching evaluation index.

[0158] Based on the predicted load data and the electrical parameters of the transformer, the method for calculating the total heat source power of the winding copper loss and the core iron loss includes:

[0159] The predicted load data includes real-time monitoring data of load current and voltage for the next 15 minutes to 1 hour. The electrical parameters of the transformer include winding resistance, core material characteristics, and rated frequency. The winding copper loss is calculated using Joule's law; the core iron loss is calculated based on the hysteresis and eddy current loss characteristics of the core material. The calculation method is:

[0160] ;

[0161] Where, is the hysteresis loss coefficient; is the eddy current loss coefficient; the hysteresis loss coefficient and the eddy current loss coefficient are determined by the magnetization curve test of the core material; is the grid frequency, which is 50Hz; is the maximum magnetic flux density of the core, which can be calculated from the transformer's rated voltage and core cross-sectional area. The total heat source power is the sum of the winding copper losses and the core iron losses, and is used to characterize the transformer's internal heat generation rate under the predicted load.

[0162] The method of calculating the ratio of the cooling capacity index obtained from the digital twin model to the total heat source power to form a basic matching index includes:

[0163] The cooling capacity index is extracted from the digital twin model in real time; the cooling capacity index includes the oil circulation flow rate, the effective heat dissipation power of the fan, and the radiator surface convection heat transfer coefficient. The cooling capacity index is substituted into the heat exchange model to calculate the actual heat dissipation power; the heat exchange model comprehensively considers oil convection, fan forced heat dissipation, and radiator natural convection. The calculation method is:

[0164] ;

[0165] Where, It is a heat exchange model; It is the oil circulation flow, measured by the oil flow relay; ; h is the convective heat transfer coefficient of the radiator surface, which is obtained by simulating the ambient wind speed and temperature through the digital twin model; is the specific heat capacity of transformer oil, which is taken as 1900 J / (kg・K); is the temperature difference between the oil inlet and outlet; is the heat dissipation area of ​​the radiator; The difference between the radiator surface and the ambient temperature. The basic matching index is calculated as the ratio of heat dissipation power to total heat source power. This index directly reflects the matching relationship between cooling capacity and heat dissipation requirements. A value greater than 1 indicates sufficient heat dissipation capacity, while a value less than 1 indicates insufficient heat dissipation capacity.

[0166] Monitor the temperature distribution of key parts of the transformer. If the temperature difference between each part is greater than the set temperature difference, the method of punitive correction of the basic matching index includes:

[0167] Temperature sensors placed at the transformer winding hotspot, the core center column, and the top oil level of the tank monitor the temperature of key locations in real time, calculating the difference between the temperature of each location and the winding hotspot temperature. If the temperature difference between any two locations exceeds the set temperature difference threshold, such as the temperature difference between the winding and the core is greater than 15°C, or the temperature difference between the winding and the top oil level of the tank is greater than 20°C, it is determined that the temperature distribution is uneven and there is a risk of local overheating. At this time, a punitive correction is made to the basic matching index. The correction formula is:

[0168] ;

[0169] Where, The basic matching index after punitive correction; As the basic matching indicator; is the correction factor, with a value of 0.02°C⁻², determined through thermal stability experiments; is the variance of the temperature of the key parts, which is used to quantify the unevenness of the temperature distribution. The larger the variance, the greater the correction amplitude.

[0170] The method of dynamically adjusting the basic matching index by combining real-time weather data to obtain a dimensionless matching evaluation index includes:

[0171] Adopt corresponding adjustment strategies for different weather types: When the ambient temperature is greater than 35℃, the air density decreases, resulting in a decrease in the fan cooling efficiency. According to the formula Calculate the temperature correction factor; where is the real-time ambient temperature; is the temperature correction factor; when the wind speed is greater than 10m / s, strong wind enhances the convective heat transfer on the radiator surface. Calculate the wind speed correction factor; where, is the real-time wind speed; is the wind speed correction factor; when the relative humidity is greater than 85%, the thermal conductivity of the insulation material changes and affects the internal heat conduction. Calculate the humidity correction factor; where, is the relative humidity; is the humidity correction factor. Comprehensive correction factor The final matching evaluation index is the ratio of the basic matching index after penalty correction to the comprehensive correction factor. This index quantifies the heat dissipation matching, temperature uniformity and environmental impact in a unified manner through dimensionless processing.

[0172] It should be noted that the placement of temperature sensors at these key locations is determined in accordance with transformer thermal design specifications to ensure the capture of the highest temperature points and typical heat dissipation nodes. The temperature difference threshold is determined based on the GB / T 1094.2-2013 standard, "Power Transformers - Part 2: Temperature Rise of Liquid-Immersed Transformers," in combination with thermal stability parameters provided by the equipment manufacturer. Weather data collection is synchronized with forecasted load data (once per minute) to ensure real-time dynamic adjustments. The coefficients in each correction formula are validated through historical operating data fitting and simulation experiments. For example, the temperature correction factor, trained on 1,000 sets of high-temperature operating data, achieved a correlation of over 92% between the corrected matching index and actual overheating faults. This matching evaluation index accurately reflects the real-time status of the transformer's heat dissipation system, providing a quantitative basis for intelligent adjustment of cooling strategies.

[0173] In the intelligent decision-making module, if the matching evaluation index exceeds the set threshold, the real-time collected equipment status, operation data and weather data are input into the pre-trained cooling strategy model to obtain the cooling strategy.

[0174] The training method of the cooling strategy model includes:

[0175] Construct a multidimensional state space including transformer equipment status, operation data and weather data;

[0176] Construct an action space for discrete adjustment modes, including fan speed adjustment, standby equipment switching, and oil pump power adjustment;

[0177] A proximal policy optimization algorithm is used to construct a policy network; the policy network takes a multidimensional state space as input and an action probability distribution in an action space as output;

[0178] Design a compound reward function that includes safety reward, energy consumption reward, and lifespan reward;

[0179] Based on the composite reward function, the parameters of the policy network are iteratively optimized to maximize the cumulative reward value of the adjustment method selected by the policy network.

[0180] The method for constructing a multidimensional state space including transformer equipment status, operation data, and weather data includes:

[0181] The state space encompasses three characteristic dimensions: transformer status, operating data, and weather data. Transformer status includes winding hotspot temperature, core center column temperature, tank top oil temperature, and cooling system operating status. Operating data includes load current, winding voltage, oil circulation flow rate, and radiator inlet and outlet pressure differential. Weather data includes ambient temperature, wind speed, relative humidity, and precipitation. The data collection frequency is synchronized with the cooling strategy adjustment cycle, set at 1 minute per session, to ensure real-time status information.

[0182] The method for constructing the action space of discrete regulation modes including fan speed regulation, standby equipment switching, and oil pump power regulation includes:

[0183] The action space is defined as a finite set of discrete adjustment methods, specifically including: ① fan speed adjustment, with three adjustment modes set: low speed 500r / min, medium speed 1000r / min, and high speed 1500r / min; ② standby equipment switching, including standby fans and standby oil pumps; there are two groups of standby fans, with a status of "0-off" or "1-on"; there is one group of standby oil pumps, with a status of "0-off" or "1-on"; ③ oil pump power adjustment, divided into five levels according to the output frequency of the oil pump variable frequency controller: 20Hz, 30Hz, 40Hz, 50Hz, and 60Hz, corresponding to power of 10kW, 15kW, 20kW, 25kW, and 30kW. The above discretization process facilitates the precise execution of control equipment in engineering practice and covers the main adjustment methods of the transformer cooling system.

[0184] Methods for constructing policy networks using proximal policy optimization algorithms include:

[0185] The policy network is a three-layer fully connected neural network. The number of neurons in the input layer matches the dimensionality of the state space, totaling 15 dimensions. The hidden layer consists of two layers with 128 neurons each. The ReLU activation function introduces nonlinear mapping capabilities. The number of neurons in the output layer matches the dimensionality of the action space, totaling 10 discrete actions. The probability distribution of each action is output via the Softmax function. The network parameters are initialized using the Xavier initialization method to ensure that the variance of the activation values ​​in each layer remains consistent, avoiding vanishing or exploding gradients.

[0186] Methods for designing a compound reward function that includes safety rewards, energy consumption rewards, and life rewards include:

[0187] The composite reward function is specifically defined as:

[0188] ;

[0189] Where, is the compound reward function; Reward for safety; Rewards for energy consumption; Reward for life; 、 and is the weight coefficient; the safety reward is designed based on the matching evaluation index and satisfies the following relationship:

[0190] ;

[0191] It should be noted that when the matching evaluation index When it is less than 0.8, it indicates that the heat dissipation capacity is insufficient; when the matching evaluation index is greater than 1.2, it indicates that the heat dissipation capacity is redundant; this design prompts the strategy network to prioritize the safe operation of the transformer. The energy consumption reward is calculated based on the real-time power consumption of the cooling equipment, which is the product of the real-time power consumption of the cooling equipment and -0.01, guiding the strategy network to choose a lower energy consumption adjustment method. The life reward takes into account the number of equipment starts and stops and the operating time. A single start-stop penalty of -5 is set for the switching action of the standby equipment, and a penalty of -2 per hour is set for the high-speed operation state of the oil pump and fan to delay equipment aging. Weight coefficient Take 0.6, Take 0.3, The value is set to 0.1, determined through expert experience and historical failure data statistics, to ensure that safety goals take priority over energy consumption and life goals.

[0192] It should be noted that the layout of each temperature sensor in the state space is determined according to the JB / T 8450-2016 standard "Transformer Winding Temperature Controller" to ensure accurate monitoring of the temperature of key parts; the adjustment accuracy of the action space matches the actual control accuracy of the cooling equipment to ensure the executable nature of the policy output; the thresholds and coefficients in the composite reward function are optimized through 1000 hours of hardware-in-the-loop simulation testing, which reduces energy consumption of the trained policy by more than 15% under safety constraints and extends equipment life by 20%.

[0193] Based on the composite reward function, the method of iteratively optimizing the parameters of the policy network to maximize the cumulative reward value of the adjustment method selected by the policy network includes:

[0194] Use digital twin models to simulate extreme weather scenarios and generate equipment status under different extreme weather scenarios;

[0195] The policy network selects an adjustment method from the action space based on the current policy state;

[0196] The digital twin model simulates the operation of the device according to the selected adjustment method, outputs the device status of the next unit time, and calculates the reward value of the corresponding composite reward function;

[0197] The device state of the next unit time and the reward value of the corresponding compound reward function are stored in the experience pool;

[0198] The relative value of each selected moderation approach was assessed using the generalized advantage estimation method;

[0199] Based on the data in the experience pool and the relative value of the adjustment method, the gradient descent algorithm is used to iteratively optimize the parameters of the policy network to maximize the cumulative reward value of the adjustment method selected by the policy network.

[0200] Methods for using digital twin models to simulate extreme weather scenarios and generate equipment states under different extreme weather scenarios include:

[0201] By inputting extreme weather parameters such as an ambient temperature of 40°C, a wind speed of 15m / s, and a relative humidity of 90%, as well as initial equipment states such as a load current of 1000A, a winding hotspot temperature of 85°C, and a fan speed of 1000r / min, the digital twin model solves the Navier-Stokes equations and the Fourier heat conduction equation to calculate the oil flow distribution inside the transformer, temperature field changes, and cooling equipment power consumption. It then outputs a device state vector containing 15-dimensional state parameters. The simulation time step is set to 1 minute, consistent with the cooling strategy adjustment period.

[0202] The policy network selects the adjustment method from the action space based on the current policy state, including:

[0203] The policy network uses an ε-greedy strategy to select specific adjustment actions: the action with the highest probability is selected with a 90% probability, and other actions are randomly selected with a 10% probability. This ensures both exploratory and developmental aspects of the strategy. For example, if the probability of selecting a high-speed fan (1500 rpm) is calculated to be 60%, the strategy may directly select that action or randomly select another action.

[0204] The method of simulating the operation of the device according to the selected adjustment method, outputting the device state of the next unit time, and calculating the reward value of the corresponding composite reward function includes:

[0205] Selected adjustment actions, such as starting a backup fan and adjusting the oil pump power to 40Hz, are input into the digital twin model as boundary conditions. The cooling equipment operating parameters are updated, and the multiphysics coupling equations are re-solved to obtain the equipment state vector for the next unit time. Simultaneously, the composite reward function value is calculated based on real-time equipment power consumption, matching evaluation indicators, and the number of equipment starts and stops.

[0206] Methods for evaluating the relative value of each alternative moderation approach using generalized odds estimation include:

[0207] The next state and reward are stored in the experience pool. This is a circular buffer with a capacity of 100,000 entries. Each data record contains the current state, the selected action, the reward value, and the next state. A prioritized experience replay strategy is used during storage. Samples with larger absolute reward values, such as negative rewards due to insufficient heat dissipation, are given higher priority. This increases the training weight of important data and prevents gradient updates from being dominated by common samples.

[0208] The relative value of the adjustment methods is evaluated using the generalized advantage estimation method. The advantage function is calculated to measure the goodness of the current action compared to the average policy, and the formula is:

[0209] ;

[0210] in, is the advantage function; s is the current time step, which indicates the current moment of evaluating the advantage; m is the future step offset, which indicates the offset of the future steps, traversing each step from s+1 to f; f is the cutoff future step, which indicates the farthest future moment considered when calculating the advantage; is the discount factor, which is set to 0.99, reflecting the decay of the present value of future rewards; is the advantage estimation parameter, which takes a value of 0.95 and is used to balance bias and variance; In state The state-value function under can be estimated by a separately trained value network. is the immediate reward, that is, the instantaneous feedback obtained in the next time step s+m+1 after the action is performed at time step s+m; In state This method effectively reduces the variance of reward estimation and improves the stability of policy evaluation by accumulating future rewards and deducting the current state value.

[0211] Based on the data in the experience pool and the relative value of the adjustment methods, the gradient descent algorithm is used to iteratively optimize the parameters of the policy network to maximize the cumulative reward value of the adjustment method selected by the policy network. The following methods are used:

[0212] Use gradient descent algorithm to minimize the objective function :

[0213] ;

[0214] in, is the current policy network; For the old policy network, it is updated every 10 iterations; For the clipping parameter, we set it to 0.2 to prevent the policy update from being too large and causing performance degradation; As the clipping function, the probability ratio of the new and old policy actions is limited to [ ] within the scope; is the adjustment method in the action space; is the state feature in the state space; is the expectation operator, which represents the state-action pair sampled from the experience pool D Calculate the expected value. The optimizer uses the Adam algorithm with an initial learning rate of 3×10⁻ 4 The reward decays by a factor of 0.95 every 50 training cycles to ensure rapid convergence in the early stages and fine-tuning in the later stages. Termination conditions are when the cumulative reward value fluctuation is less than 1% for 100 consecutive cycles or when the maximum number of training cycles is 2000, balancing training accuracy and computational efficiency.

[0215] It should be noted that the simulation accuracy of the digital twin model is calibrated using historical operating data, and the oil temperature prediction error in extreme scenarios does not exceed ±2°C, ensuring the engineering reliability of the generated equipment state. The exploration rate of the ε-greedy strategy is set at 10% at the beginning of training and linearly decays to 5% with increasing iterations, gradually transitioning from the exploration phase to the development phase. The discount factor and advantage parameter in the generalized advantage estimation are determined through 1000 hours of hardware-in-the-loop testing, keeping the reward evaluation error within 5%. The priority calculation of the prioritized experience replay uses TD error absolute value sorting, increasing the training frequency of key samples by more than 30%. Through the above method, the policy network can gradually learn a cooling adjustment strategy with lower energy consumption and longer equipment life under safety constraints. The cumulative reward value of the trained model in typical extreme scenarios is significantly improved compared to the initial random strategy, achieving optimal control of the cooling system.

[0216] The closed-loop control module is used to verify that there are no abnormalities in the equipment status and then generate and execute adjustment instructions based on the cooling strategy. During the execution of the adjustment instructions, the calibration parameters of the digital twin model are dynamically corrected based on the real-time collected equipment status and operating data.

[0217] Methods for dynamically correcting the calibration parameters of the digital twin model include:

[0218] During the execution of the cooling strategy adjustment instructions, transformer operation data and equipment status are collected in real time as feedback data for dynamic correction;

[0219] Compare the real-time prediction results of the digital twin model with the feedback data, and calculate the deviation between the real-time prediction results and the feedback data;

[0220] If the deviation exceeds the set deviation, the calibration parameters that affect the prediction accuracy are identified, and the calibration parameters are iteratively adjusted using the gradient descent algorithm based on the direction and magnitude of the deviation;

[0221] After adjustment, the prediction accuracy of the corrected digital twin model in the subsequent set period is verified. If the deviation converges to the set deviation threshold in the subsequent set period, the parameter correction is confirmed to be effective and the digital twin model is updated; if the deviation does not converge to the set deviation threshold in the subsequent set period, the next round of correction is entered until the deviation converges to the set deviation threshold.

[0222] Methods for comparing the real-time prediction results of the digital twin model with the feedback data and calculating the deviation between the real-time prediction results and the feedback data include:

[0223] The digital twin model outputs real-time predicted oil temperature distribution, including three key locations: the winding, the core, and the oil level, as well as oil circulation flow and cooling device power consumption. Deviations are calculated using the root mean square error method. Deviation thresholds are set at: oil temperature prediction error exceeding 2°C, oil circulation flow error exceeding 15%, and cooling device power error exceeding 10%. These thresholds are determined based on the transformer thermal protection setting and the equipment's operating accuracy requirements, and are used to determine whether the model prediction deviates from actual operating conditions.

[0224] If the deviation exceeds the set deviation, identify the calibration parameters that affect the prediction accuracy, and use the gradient descent algorithm to iteratively adjust the calibration parameters according to the direction and magnitude of the deviation. The following methods are used:

[0225] The calibration parameters are the convection heat transfer coefficient and oil circulation resistance coefficient, which are directly related to heat exchange efficiency in the digital twin model. These two factors influence the oil temperature distribution and oil flow state prediction through the multi-physics field coupling model. A gradient descent algorithm is used to adjust the parameters, with the objective function being the current cycle deviation value. The parameter update formula is:

[0226] ;

[0227] Where, is the value of the parameter to be adjusted at iteration step t; for the value of iteration step 1, that is, is the convective heat transfer coefficient; the value of iterative step 2, that is, is the oil circulation resistance coefficient; is the learning rate, which is set to 0.01 and can be determined through simulation experiments to avoid parameter oscillation; is the gradient of the objective function with respect to the parameter. The direction of adjustment is determined by the sign of the deviation: if the model-predicted oil temperature is higher than the measured value, meaning the heat dissipation capacity is underestimated, the convective heat transfer coefficient is increased or the oil circulation resistance coefficient is decreased. If the model-predicted oil temperature is lower than or equal to the measured value, the convective heat transfer coefficient is decreased or the oil circulation resistance coefficient is increased. The adjustment amplitude is positively correlated with the absolute value of the deviation, and a single adjustment should not exceed 10% of the current parameter value to ensure stability during the adjustment process.

[0228] After adjustment, the prediction accuracy of the revised digital twin model is verified within the subsequent set period. The verification period is set to three consecutive control cycles. If the deviations of each monitored parameter converge to the set threshold within three cycles, the parameter adjustment is confirmed to be effective, and the optimal parameter combination is updated to the digital twin model. If the target is not met, the original parameters are retained and the next round of correction is entered until the deviations converge. This verification mechanism ensures the effectiveness of parameter adjustments and avoids incorrect corrections caused by short-term fluctuations.

[0229] It should be noted that before calculating the deviation, the measured data must be subjected to a sliding average filter to eliminate high-frequency noise. The gradient descent algorithm uses a batch gradient descent method, using all monitoring data from the current cycle for each iteration to ensure the accuracy of the gradient calculation. Limits on parameter adjustment are set to prevent oscillations in simulation results caused by drastic changes in model parameters, in line with the principle of prioritizing stability in engineering control. This method can correct parameter drift in the digital twin model caused by equipment aging and environmental changes in real time, ensuring that the model's prediction accuracy consistently meets engineering application requirements and providing reliable model support for precise adjustment of cooling strategies.

[0230] Example 2

[0231] This embodiment provides a transformer intelligent regulation system, further comprising:

[0232] The pre-response and adaptive module is used to start equipment health detection and redundancy strategy adjustment in advance in combination with meteorological warnings. At the same time, it regularly updates the model's time-varying parameters and optimizes the algorithm strategy library to dynamically balance heat dissipation capacity and equipment life.

[0233] The method of starting equipment health detection and redundancy strategy adjustment in advance in combination with meteorological warnings includes:

[0234] Obtain weather forecast data for the next 72 hours through the weather station's weather warning interface, including extreme weather type, estimated arrival time, ambient temperature peak, and wind speed peak;

[0235] When an extreme weather warning is detected, the equipment health check process is initiated; if it is determined that the cooling equipment has health risks, a redundancy strategy trigger signal is generated;

[0236] Based on the redundant strategy trigger signal and the meteorological warning level, the redundant strategy library is called to generate pre-adjustment instructions;

[0237] During the pre-response execution process, device response data is collected in real time and the pre-adjustment effect index is calculated; if the pre-adjustment effect index is less than the preset value, the redundancy strategy is automatically upgraded.

[0238] The device health detection process includes:

[0239] For example, the historical data of the vibration amplitude of the cooling equipment every minute for the past 7 days is called to calculate the mean and standard deviation of the vibration amplitude. If the sum of the mean and twice the standard deviation of the vibration amplitude is greater than the bearing wear warning threshold, it is determined that there is a health hazard in the cooling equipment and a redundant strategy trigger signal is generated.

[0240] The method for generating the pre-adjustment instruction is:

[0241] For example, if the weather warning is a high temperature warning and the equipment health status is abnormal, the backup fan rotation strategy is started 3 hours in advance to transfer the main fan load to the backup equipment; if the weather warning is a typhoon warning, the fan blade angle is locked to 0 2 hours in advance, and the oil pump power is increased to bipolar operation to reduce the physical impact of strong winds on the cooling system.

[0242] The calculation of the pre-adjustment effect index The method is:

[0243] ;

[0244] Where, To regulate the front oil temperature; is the peak ambient temperature; To pre-adjust the oil temperature; is the current ambient temperature.

[0245] It should be noted that this indicator reflects the ability of pre-response to suppress temperature rise in extreme weather, and its value range is 0 to 1. The closer it is to 1, the better the pre-regulation effect.

[0246] The method of regularly updating the model's time-varying parameters and optimizing the algorithm strategy library to dynamically balance heat dissipation capacity and equipment life includes:

[0247] Obtain the average ambient temperature and load factor for the day, and update the air density and transformer oil viscosity in the digital twin model;

[0248] Based on the adjustment records of the past preset time, an online learning algorithm is used to update the composite reward function weight of the policy network;

[0249] Through cluster analysis, historical extreme scenarios are divided into a preset number of typical operating condition clusters. The policy effect is evaluated on the adjustment data within each operating condition cluster, and the optimal action sequence is extracted and updated to the policy library.

[0250] The output includes an adaptive adjustment report of the pre-response strategy execution results, model time-varying parameter update records, and strategy library optimization logs.

[0251] The method for updating the composite reward function weight of the policy network includes:

[0252] For example, based on the adjustment records of the past 24 hours, if the emergency strategy is frequently triggered in a high temperature scenario, the safety reward weight is automatically increased to 0.7 to strengthen the safety limited logic.

[0253] The method of dividing historical extreme scenarios into a preset number of typical operating condition clusters through cluster analysis, evaluating the strategy effect of the adjustment data in each operating condition cluster, extracting the optimal action sequence and updating it to the strategy library includes:

[0254] For example, the algorithm strategy library optimization is performed once a month. Through cluster analysis, historical extreme scenarios, such as "high temperature + high load" and "typhoon + sudden load", are divided into 5 typical working condition clusters to extract the optimal action sequence; for the "typhoon + sudden load" working condition, if historical data shows that the strategy of "starting the backup fan first and then gradually increasing the main fan speed" has lower energy consumption than "turning on all equipment" and has a similar oil temperature control effect, then this sequence will be set as the default strategy for this working condition.

[0255] It should be noted that the pre-response and adaptive module transforms passive response into active defense through the full-chain mechanism of "meteorological warning-health detection-redundancy adjustment-parameter adaptation-strategy evolution". While improving the adaptability of equipment to extreme environments, it continuously enhances the long-term operating efficiency of the system through periodic parameter updates and strategy optimization.

[0256] Example 3

[0257] See also Figure 2 As shown, this embodiment provides a transformer intelligent adjustment method, including:

[0258] Input the real-time collected transformer operation data and weather data into the pre-trained load forecasting model to obtain the predicted load data;

[0259] The collected physical parameters and operating data, as well as pre-set calibration parameters, are input into the pre-built digital twin model to obtain the cooling capacity index;

[0260] The total heat source power of copper loss and iron loss is calculated based on the predicted load data and the real-time collected electrical parameters of the transformer. Combined with the cooling capacity index, the matching evaluation function dynamically quantifies the balance between the total heat source power and the cooling capacity index to obtain the matching evaluation index.

[0261] If the matching evaluation index exceeds the set threshold, the real-time collected equipment status, operation data and weather data are input into the pre-trained cooling strategy model to obtain the cooling strategy;

[0262] After verifying that there are no abnormalities in the equipment status, adjustment instructions are generated and executed according to the cooling strategy. During the execution of the adjustment instructions, the calibration parameters of the digital twin model are dynamically corrected based on the real-time collected equipment status and operation data.

[0263] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

[0264] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A transformer intelligent adjustment method, characterized in that: include: Input the real-time collected transformer operation data and weather data into the pre-trained load forecasting model to obtain the predicted load data; The collected physical parameters and operating data, as well as pre-set calibration parameters, are input into the pre-built digital twin model to obtain the cooling capacity index; The total heat source power of copper loss and iron loss is calculated based on the predicted load data and the real-time collected electrical parameters of the transformer. Combined with the cooling capacity index, the matching evaluation function dynamically quantifies the balance between the total heat source power and the cooling capacity index to obtain the matching evaluation index. If the matching evaluation index exceeds the set threshold, the real-time collected equipment status, operation data and weather data are input into the pre-trained cooling strategy model to obtain the cooling strategy; After verifying that the equipment status is normal, generate and execute adjustment instructions based on the cooling strategy; During the execution of the adjustment instructions, the calibration parameters of the digital twin model are dynamically corrected based on the real-time collected equipment status and operation data.

2. A transformer intelligent adjustment method according to claim 1, characterized in that: The training methods for the load forecasting model include: Collect transformer operation data and weather data during historical extreme weather periods and mark them as training data; The training data is normalized and preprocessed as model input samples, and the actual load data during the historical extreme weather period is used as the supervision signal; A load forecasting model is constructed using a long short-term memory neural network; the long short-term memory neural network includes an input layer, multiple hidden layers, and a fully connected output layer; the dimensions of the input layer are consistent with the feature dimensions of the preprocessed model input samples; the fully connected output layer outputs predicted load data for a preset time period in the future and marks load surge points in the predicted load data; With the goal of minimizing the mean square error between the predicted load and the supervisory signal, the load forecasting model parameters are iteratively optimized through the back propagation algorithm to obtain the final load forecasting model.

3. A transformer intelligent adjustment method according to claim 2, characterized in that: Methods for determining whether a data point in the forecast load data is a load surge point include: Construct a multidimensional feature space including time characteristics, load historical trend characteristics and extreme weather characteristics; the time characteristics include periodic indicators of daily cycles, weekly cycles and monthly cycles; the load historical trend characteristics include statistical indicators of historical load mean, historical load standard deviation and historical load change rate; the extreme weather characteristics are weather data and its change rate; Based on the multidimensional feature space, the characteristic value of each data point of the predicted load data is calculated in sequence to construct a characteristic value set; the characteristic value includes the load growth rate, the deviation from the historical load mean, and the correlation with the weather data; Calculate the matching degree of each eigenvalue in the eigenvalue set with the preset surge point eigenvalue; if the matching degree exceeds the set matching degree threshold, the corresponding data point is determined to be a load surge point and the corresponding extreme scenario type is marked; if the matching degree does not exceed the set matching degree threshold, the corresponding data point is determined to be a normal data point; the extreme scenario types include high temperature, high pressure and strong wind; For data points determined as load surge points in the forecast load data, verify whether the load surge persistence conditions are met.

4. A transformer intelligent adjustment method according to claim 1, characterized in that: Methods for building digital twin models include: Construct a three-dimensional geometric model based on the collected three-dimensional geometric parameters of the transformer; The collected transformer physical parameters and historical operating data are used as inputs to the digital twin model; the physical parameters include material thermal characteristics, oil viscosity-temperature relationship, and air density-temperature relationship; Using computational fluid dynamics methods, the Navier-Stokes equations and the Fourier heat conduction equations are coupled to establish a multi-physics coupling model; A particle swarm optimization algorithm was introduced to calibrate the calibration parameters in the multi-physics coupling model, including the convective heat transfer coefficient and the oil circulation resistance coefficient, with the optimization goal of minimizing the root mean square error between the oil temperature distribution predicted by the multi-physics coupling model and the oil temperature distribution measured in historical operating data. The digital twin model is composed of a multi-physics field coupling model and a three-dimensional geometric model.

5. A transformer intelligent adjustment method according to claim 4, characterized in that: Methods for calibrating calibration parameters in multiphysics coupled models include: Constructing a particle swarm optimization space; each particle in the particle swarm optimization space is defined as a two-dimensional vector consisting of a convective heat transfer coefficient and an oil circulation resistance coefficient; The parameter vectors of each particle are input into the multi-physics field coupling model, and numerical simulation is performed in combination with historical operating data to obtain the predicted oil temperature distribution. The predicted oil temperature distribution is compared with the historical measured oil temperature distribution during the same period, and the root mean square error is calculated as the particle fitness value. Each particle in the particle swarm optimization space adjusts its speed and position according to the preset rules based on the minimum root mean square error parameter combination it has searched for and the minimum root mean square error parameter combination in the current population, and searches for a better solution area in the parameter space; When the change in the global optimal fitness value after a preset positive integer number of iterations is less than the set fitness threshold, or the maximum number of iterations is reached, the optimization process is terminated; The convective heat transfer coefficient and oil circulation resistance coefficient obtained by particle swarm optimization with minimized root mean square error are embedded in the multi-physics coupling model to complete the calibration of the calibration parameters in the multi-physics coupling model.

6. The transformer intelligent adjustment method according to claim 1, characterized in that: Methods for obtaining matching evaluation indicators include: Calculate the total heat source power of copper loss and iron loss based on the predicted load data and the electrical parameters of the transformer; The cooling capacity index obtained from the digital twin model is calculated by comparing it with the total heat source power to form a basic matching index. Monitor the temperature distribution of key parts of the transformer. If the temperature difference between the parts is greater than the set temperature difference, a penalty correction will be made to the basic matching index. The key parts include the winding, core and top layer of the oil tank. The basic matching index is dynamically adjusted in combination with real-time weather data to obtain a dimensionless matching evaluation index.

7. The transformer intelligent adjustment method according to claim 1, characterized in that: The training method of the cooling strategy model includes: Construct a multidimensional state space including transformer equipment status, operation data and weather data; Construct an action space for discrete adjustment modes, including fan speed adjustment, standby equipment switching, and oil pump power adjustment; A proximal policy optimization algorithm is used to construct a policy network; the policy network takes a multidimensional state space as input and an action probability distribution in an action space as output; Design a compound reward function that includes safety reward, energy consumption reward, and lifespan reward; Based on the composite reward function, the parameters of the policy network are iteratively optimized to maximize the cumulative reward value of the adjustment method selected by the policy network and obtain the cooling policy model.

8. The transformer intelligent adjustment method according to claim 7, characterized in that: Based on the composite reward function, the method of iteratively optimizing the parameters of the policy network to maximize the cumulative reward value of the adjustment method selected by the policy network includes: Use digital twin models to simulate extreme weather scenarios and generate equipment status under different extreme weather scenarios; The policy network selects an adjustment method from the action space based on the current policy state; The digital twin model simulates the operation of the device according to the selected adjustment method, outputs the device status of the next unit time, and calculates the reward value of the corresponding composite reward function; The device state of the next unit time and the reward value of the corresponding compound reward function are stored in the experience pool; The relative value of each selected moderation approach was assessed using the generalized advantage estimation method; Based on the data in the experience pool and the relative value of the adjustment method, the gradient descent algorithm is used to iteratively optimize the parameters of the policy network to maximize the cumulative reward value of the adjustment method selected by the policy network.

9. The transformer intelligent adjustment method according to claim 1, characterized in that: Methods for dynamically correcting the calibration parameters of the digital twin model include: During the execution of the cooling strategy adjustment instructions, transformer operation data and equipment status are collected in real time as feedback data for dynamic correction; Compare the real-time prediction results of the digital twin model with the feedback data, and calculate the deviation between the real-time prediction results and the feedback data; If the deviation exceeds the set deviation, the calibration parameters that affect the prediction accuracy are identified, and the calibration parameters are iteratively adjusted using the gradient descent algorithm based on the direction and magnitude of the deviation; After adjustment, the prediction accuracy of the corrected digital twin model in the subsequent set period is verified. If the deviation converges to the set deviation threshold in the subsequent set period, the parameter correction is confirmed to be effective and the digital twin model is updated; if the deviation does not converge to the set deviation threshold in the subsequent set period, the next round of correction is entered until the deviation converges to the set deviation threshold.

10. The transformer intelligent adjustment method according to claim 1, characterized in that: The transformer operation data includes the vibration amplitude of the cooling equipment, the status of the transformer body, electrical parameters, oil temperature distribution and load information; the weather data includes ambient temperature, humidity, wind speed, precipitation and weather warning type; the electrical parameters include load current and winding resistance.

11. A transformer intelligent regulation system, configured to implement a transformer intelligent regulation method according to any one of claims 1 to 10, characterized in that: include: The load forecasting module is used to input the real-time collected transformer operation data and weather data into the pre-trained load forecasting model to obtain the predicted load data; The digital twin module is used to input the collected physical parameters and operating data, as well as pre-set calibration parameters, into the pre-built digital twin model to obtain the cooling capacity index; The matching evaluation module calculates the total heat source power of copper loss and iron loss based on the predicted load data and the real-time collected electrical parameters of the transformer. Combined with the cooling capacity index, it dynamically quantifies the balance between the total heat source power and the cooling capacity index through the matching evaluation function to obtain the matching evaluation index. The intelligent decision-making module inputs the real-time collected equipment status, operating data, and weather data into the pre-trained cooling strategy model to obtain a cooling strategy if the matching evaluation index exceeds the set threshold; The closed-loop control module verifies that the equipment is in good condition and then generates and executes adjustment instructions based on the cooling strategy. During the execution of the adjustment instructions, the calibration parameters of the digital twin model are dynamically corrected based on the real-time collected equipment status and operation data.

Images