A new energy box transformer automatic temperature control method and system based on deep learning

Abstract

The application provides a new energy box transformer automatic temperature control method and system based on deep learning, which comprises the following steps: collecting box transformer load current, environmental temperature and humidity, winding and oil tank temperature multi-dimensional operation data, and constructing a sample set to be queried; setting feature sensitivity according to the output gradient of the benchmark temperature control prediction model, calculating the sample weighted state distance, and obtaining the local state concentration; inputting the sample into a model group composed of multiple heterogeneous temperature control prediction models, weighting the feature sensitivity of the gradient dispersion of each model output, combining the state change rate to obtain the composite uncertainty measure; fusing the sparsity index and the composite uncertainty measure after reverse weighting to obtain the query score, and screening the target query sample; solving the temperature control bias according to the composite uncertainty measure and the local state concentration of the real-time sample, correcting to obtain the real-time temperature control threshold, and triggering the temperature control instruction when the predicted temperature or temperature change slope meets the condition.

Description

Technical Field

[0001] This application belongs to the field of artificial intelligence, and in particular relates to an automatic temperature control method and system for new energy transformer substations based on deep learning. Background Technology

[0002] During operation, new energy prefabricated substations are subject to load current fluctuations and external environmental temperature and humidity, resulting in significant heat generation in the internal windings and oil tank. If this heat cannot be dissipated promptly, it will cause the equipment operating temperature to exceed limits, accelerating insulation material aging and even triggering electrical faults. Therefore, constructing a reliable automatic temperature control system is crucial for ensuring the safe and stable operation of new energy substations. Currently, prefabricated substation temperature control technology typically relies on data-driven algorithms. It collects operating status data such as load current, ambient temperature and humidity, and internal equipment temperature to train a temperature control prediction model, issuing start / stop commands for cooling equipment based on the predicted temperature trend. To maintain the model's prediction accuracy under different operating conditions, it is necessary to continuously collect and filter high-information operating data for iterative updates. However, during sample selection, it is impossible to accurately assess the local concentration and true information value of samples in the feature space. The selected data often contains a large amount of redundancy, increasing the computational load of data processing and failing to specifically improve the model's prediction performance under edge conditions. Automatic temperature control mechanisms generally use pre-set absolute temperature thresholds to trigger cooling operations. This control logic cannot be adjusted based on the sparsity of the data distribution and the reliability of the prediction results in the current real-time operating state, nor can it incorporate the slope of future temperature changes into the decision-making preconditions. When the equipment is in abnormal operating conditions or the model prediction is biased, the lack of a threshold for bias adjustment can easily cause lag in temperature control or frequent start-stop of the heat dissipation equipment, making it impossible to achieve the optimal balance between ensuring the thermal safety of the enclosure and reducing auxiliary power consumption. Summary of the Invention

[0003] According to one aspect of the present invention, a deep learning-based automatic temperature control method for new energy transformer substations is proposed to address the problem that existing methods cannot adjust based on the sparsity of data distribution and the reliability of prediction results in the current real-time operating state. The method includes the following steps:

[0004] The multi-dimensional operating status data of the transformer substation, including load current, ambient temperature and humidity, winding temperature, and oil tank temperature, are used to form a sample set to be queried. For each sample in the sample set to be queried, the feature sensitivity is set based on the gradient of the input of each dimension output by the benchmark temperature control prediction model. The weighted state distance between samples is calculated using the feature sensitivity to generate the local state concentration. The sample to be queried is input into a model group consisting of multiple temperature control prediction models with different structures. The gradient of the independent prediction results of each model with respect to each dimension of the input is extracted. When calculating the dispersion of the gradient of each model, the feature sensitivity is input for weighted summation. Combined with the rate of change of multidimensional operating status data, a composite uncertainty measure is generated. A final query score is generated by integrating a sparsity index that is inversely correlated with local state concentration and a composite uncertainty measure that is inversely weighted by local state concentration. The sample with the highest final score is selected as the target query sample. The current real-time operating state sample is obtained, and the temperature control bias is calculated based on the composite uncertainty measure and local state concentration of the sample. The temperature control bias is subtracted from the preset temperature target to obtain the real-time temperature control threshold. A temperature control command is triggered when the predicted temperature reaches the real-time temperature control threshold or the slope of future temperature change exceeds the preset benchmark.

[0005] Optionally, the multi-dimensional operating status data of the transformer substation, including load current, ambient temperature and humidity, winding and tank temperatures, constitutes a sample set to be queried, comprising: The communication node module is synchronously woken up by a time trigger at a preset duration, the communication port data is read, the actual ampere values ​​of all transmission and distribution circuits of the transformer are extracted, and the single current characteristic value is obtained. The ambient dry temperature and water mist relative humidity coefficient were extracted as real-time parameters of the surrounding environment. The temperature of the heating coil and the oil tank, which are read in real time from the internal pre-embedded and surface contact detection elements, are used as the true indicator characteristics of the internal temperature detection. By aligning different types of detection results from the same period side-by-side, a full-feature sequence state group is formed and placed into the corresponding variable database to obtain the feature matrix of the parameter vector feature group to be combined, which is used to construct the data extraction library query parameters.

[0006] Optionally, for each sample in the query sample set, the feature sensitivity is set based on the gradient of the output of the benchmark temperature control prediction model with respect to the input of each dimension, including: Extract the input vector from the sample set to be queried, substitute it into the benchmark temperature control prediction model for forward calculation, and obtain the single temperature measurement result and the temperature prediction sequence within the future preset time period. Perform partial derivative calculations on the single-temperature measurement results to obtain the partial derivative values ​​corresponding to the direction of each input parameter as the gradient of each dimension input; The absolute values ​​of the partial derivatives of each gradient are calculated and converted into absolute gradient values, which are then summarized into a sample absolute gradient vector. The absolute gradient values ​​calculated from the collected samples are extracted, summed, and averaged to obtain the average absolute gradient of the class. The feature sensitivity is obtained by dividing each absolute gradient mean by the sum of all absolute gradient mean values.

[0007] Optionally, the step of calculating the weighted state distance between samples using feature sensitivity to generate local state concentration includes: For each sample to be compared, a single extractable feature parameter of the sample on each attribute data is obtained, and the feature parameter is dimensionless. Using a single extracted feature parameter after dimensionless processing, perform a difference operation between the two to obtain the absolute difference result; Perform a weighted product operation between the absolute difference result and the corresponding feature sensitivity; Sum all weighted product results to obtain the weighted state distance; The local state concentration degree, representing the proximity of samples, is calculated based on the reciprocal or negative exponential function of the weighted state distance.

[0008] Optionally, the step of inputting the sample to be queried into a model group consisting of multiple temperature control prediction models with different structures, extracting the gradient of the independent prediction results of each model with respect to each dimension of the input, and weighting and summing the input feature sensitivity when calculating the dispersion of the gradient of each model, and generating a composite uncertainty measure by combining the rate of change of multidimensional operating state data, includes: Extract the gradients of the independent prediction results of multiple structurally different temperature control prediction models with respect to each dimension of the input, and calculate the variance of the gradient of each model in each dimension as the basic dispersion. The basic dispersion of each dimension is multiplied by the corresponding feature sensitivity and then summed to obtain the prediction divergence metric on the amplified sensitive feature dimension. Extract the rate of change of multidimensional operating status data within a preset time window, and calculate the comprehensive scalar value of the rate of change of each dimension; The metric value and the comprehensive scalar value are normalized and weighted and merged to generate a composite uncertainty metric.

[0009] Optionally, the final query score is generated by fusing a sparsity index that is inversely correlated with the local state concentration and a composite uncertainty measure that is inversely weighted by the local state concentration, including: The sparsity index is obtained by dividing the numerical constant 1 as the numerator and the local state concentration as the denominator. Using the reciprocal of the local state concentration as the adjustment coefficient, multiplying it with the composite uncertainty measure, we obtain the composite uncertainty measure with inverse weighted adjustment; The final query score is generated by adding the sparsity index to the composite uncertainty measure with inverse weighting adjustment.

[0010] Optionally, the step of obtaining the current real-time operating status sample, calculating the temperature control bias based on the composite uncertainty measure and local state concentration of the sample, and subtracting the temperature control bias from the preset temperature target to obtain the real-time temperature control threshold includes: Obtain the composite uncertainty measure and local state concentration of the current real-time running state sample; Multiply the composite uncertainty measure by a preset positive scaling factor to obtain the uncertainty bias component; Divide the numerical constant 1 by the local state concentration to obtain the reciprocal of the concentration, and multiply the reciprocal of the concentration by the preset inverse scaling factor to obtain the concentration bias component. The temperature control bias is obtained by adding the uncertainty bias component and the concentration bias component.

[0011] Optionally, the step of triggering a temperature control command when the predicted temperature reaches a real-time temperature control threshold or the slope of future temperature changes exceeds a preset benchmark includes: Obtain the current real-time operating status sample, and obtain the predicted temperature of the current real-time operating status sample and the temperature prediction sequence within a preset time period from the benchmark temperature control prediction model. Determine whether the predicted temperature is greater than or equal to the real-time temperature control threshold; Calculate the slope of temperature change within a preset time period based on temperature prediction sequences; Determine if the slope of the temperature change is greater than a preset benchmark. If the predicted temperature is greater than or equal to the real-time temperature control threshold, or the temperature change slope is greater than the preset benchmark, then a temperature control command is generated and output to the temperature control actuator.

[0012] According to another aspect of the present invention, a new energy transformer substation automatic temperature control system based on deep learning is proposed, comprising the following modules: The generation module is used to acquire multi-dimensional operating status data of the transformer load current, ambient temperature and humidity, winding and oil tank temperature to form a sample set to be queried; for each sample in the sample set to be queried, the feature sensitivity is set based on the gradient of each dimension input output by the benchmark temperature control prediction model, and the weighted state distance between samples is calculated using the feature sensitivity to generate the local state concentration. The calculation module is used to input the sample to be queried into a model group consisting of multiple temperature control prediction models with different structures, extract the gradient of the independent prediction results of each model with respect to each dimension of input, input the feature sensitivity for weighted summation when calculating the dispersion of the gradient of each model, and generate a composite uncertainty measure by combining the rate of change of multidimensional operating status data. The calculation module is used to fuse a sparsity index that is inversely correlated with the local state concentration and a composite uncertainty measure adjusted by inverse weighting of the local state concentration to generate a final query score, and select the sample with the highest final score as the target query sample; obtain the current real-time operating state sample, calculate the temperature control bias based on the composite uncertainty measure and local state concentration of the sample, subtract the temperature control bias from the preset temperature target to obtain the real-time temperature control threshold; and trigger a temperature control command when the predicted temperature reaches the real-time temperature control threshold or the slope of future temperature change exceeds a preset benchmark.

[0013] Preferably, the multi-dimensional operating status data of the transformer substation, including load current, ambient temperature and humidity, winding and oil tank temperatures, constitute the sample set to be queried, comprising: The communication node module is synchronously woken up by a time trigger at a preset duration, the communication port data is read, the actual ampere values ​​of all transmission and distribution circuits of the transformer are extracted, and the single current characteristic value is obtained. The ambient dry temperature and water mist relative humidity coefficient were extracted as real-time parameters of the surrounding environment. The temperature of the heating coil and the oil tank, which are read in real time from the internal pre-embedded and surface contact detection elements, are used as the true indicator characteristics of the internal temperature detection. By aligning different types of detection results from the same period side-by-side, a full-feature sequence state group is formed and placed into the corresponding variable database to obtain the feature matrix of the parameter vector feature group to be combined, which is used to construct the data extraction library query parameters.

[0014] Preferably, for each sample in the query sample set, the feature sensitivity is set based on the gradient of the input dimension output by the benchmark temperature control prediction model, including: Extract the input vector from the sample set to be queried, substitute it into the benchmark temperature control prediction model for forward calculation, and obtain the single temperature measurement result and the temperature prediction sequence within the future preset time period. Perform partial derivative calculations on the single-temperature measurement results to obtain the partial derivative values ​​corresponding to the direction of each input parameter as the gradient of each dimension input; The absolute values ​​of the partial derivatives of each gradient are calculated and converted into absolute gradient values, which are then summarized into a sample absolute gradient vector. The absolute gradient values ​​calculated from the collected samples are extracted, summed, and averaged to obtain the average absolute gradient of the class. The feature sensitivity is obtained by dividing each absolute gradient mean by the sum of all absolute gradient mean values.

[0015] Preferably, the step of generating local state concentration by calculating the weighted state distance between samples using feature sensitivity includes: For each sample to be compared, a single extractable feature parameter of the sample on each attribute data is obtained, and the feature parameter is dimensionless. Using a single extracted feature parameter after dimensionless processing, perform a difference operation between the two to obtain the absolute difference result; Perform a weighted product operation between the absolute difference result and the corresponding feature sensitivity; Sum all weighted product results to obtain the weighted state distance; The local state concentration degree, representing the proximity of samples, is calculated based on the reciprocal or negative exponential function of the weighted state distance.

[0016] Preferably, the step of inputting the sample to be queried into a model group consisting of multiple temperature control prediction models with different structures, extracting the gradient of the independent prediction results of each model with respect to each dimension of the input, and weighting and summing the input feature sensitivity when calculating the dispersion of the gradient of each model, and generating a composite uncertainty measure by combining the rate of change of multidimensional operating state data, includes: Extract the gradients of the independent prediction results of multiple structurally different temperature control prediction models with respect to each dimension of the input, and calculate the variance of the gradient of each model in each dimension as the basic dispersion. The basic dispersion of each dimension is multiplied by the corresponding feature sensitivity and then summed to obtain the prediction divergence metric on the amplified sensitive feature dimension. Extract the rate of change of multidimensional operating status data within a preset time window, and calculate the comprehensive scalar value of the rate of change of each dimension; The metric value and the comprehensive scalar value are normalized and weighted and merged to generate a composite uncertainty metric.

[0017] Preferably, the final query score is generated by fusing a sparsity index that is inversely correlated with the local state concentration and a composite uncertainty measure adjusted by inverse weighting of the local state concentration, including: The sparsity index is obtained by dividing the numerical constant 1 as the numerator and the local state concentration as the denominator. Using the reciprocal of the local state concentration as the adjustment coefficient, multiplying it with the composite uncertainty measure, we obtain the composite uncertainty measure with inverse weighted adjustment; The final query score is generated by adding the sparsity index to the composite uncertainty measure with inverse weighting adjustment.

[0018] Preferably, the step of acquiring the current real-time operating status sample, calculating the temperature control bias based on the composite uncertainty metric and local state concentration of the sample, and subtracting the temperature control bias from the preset temperature target to obtain the real-time temperature control threshold includes: Obtain the composite uncertainty measure and local state concentration of the current real-time running state sample; Multiply the composite uncertainty measure by a preset positive scaling factor to obtain the uncertainty bias component; Divide the numerical constant 1 by the local state concentration to obtain the reciprocal of the concentration, and multiply the reciprocal of the concentration by the preset inverse scaling factor to obtain the concentration bias component. The temperature control bias is obtained by adding the uncertainty bias component and the concentration bias component.

[0019] Preferably, the step of triggering the temperature control command when the predicted temperature reaches the real-time temperature control threshold or the slope of future temperature change exceeds a preset benchmark includes: Obtain the current real-time operating status sample, and obtain the predicted temperature of the current real-time operating status sample and the temperature prediction sequence within a preset time period from the benchmark temperature control prediction model. Determine whether the predicted temperature is greater than or equal to the real-time temperature control threshold; Calculate the slope of temperature change within a preset time period based on temperature prediction sequences; Determine if the slope of the temperature change is greater than a preset benchmark. If the predicted temperature is greater than or equal to the real-time temperature control threshold, or the temperature change slope is greater than the preset benchmark, then a temperature control command is generated and output to the temperature control actuator.

[0020] This invention generates local state concentration by extracting gradient feature sensitivity from multidimensional operating state data, identifying the actual impact of various factors on transformer temperature and improving the accuracy of multi-condition assessment. It utilizes structurally dissimilar model sets to calculate gradient dispersion and combines it with data change rates to generate a composite uncertainty metric, overcoming the prediction bias of a single model and covering multi-dimensional unknown risks. By fusing sparsity and uncertainty metrics, it selects target samples with both high information content and high representativeness, reducing interference from extreme isolated noise points. Uncertainty and concentration are converted into temperature control bias to set real-time temperature control thresholds, and combined with temperature prediction values ​​and future change slopes as dual triggering conditions, a more stringent safety margin is established in advance under high-risk conditions, enabling predictive temperature intervention and ensuring the operational safety and service life of new energy transformer equipment. Attached Figure Description

[0021] Figure 1 A flowchart of the first embodiment; Figure 2 The overall algorithm flowchart for new energy transformer substations; Figure 3 This is a logic diagram for sample scoring fusion screening and threshold temperature control. Detailed Implementation

[0022] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0023] In the first embodiment, the present invention proposes an automatic temperature control method for new energy transformer substations based on deep learning, such as... Figure 1 It includes the following steps: S1. Obtain multi-dimensional operating status data of transformer load current, ambient temperature and humidity, winding and oil tank temperature to form a sample set to be queried; for each sample in the sample set to be queried, set the feature sensitivity based on the gradient of each dimension input output by the benchmark temperature control prediction model, and use the feature sensitivity to calculate the weighted state distance between samples to generate local state concentration.

[0024] The load current of the transformer substation is collected by current transformers installed on the low-voltage and high-voltage sides. The external ambient temperature and humidity are collected by temperature and humidity sensors. The transformer winding temperature and oil tank temperature are collected by fiber optic temperature probes and oil temperature sensors, respectively. The collected data from multiple consecutive time periods are processed using max-min normalization and concatenated into a multi-dimensional operating state vector containing multiple dimensions. Combining these multi-dimensional operating state vectors from multiple time periods yields the query sample set. The benchmark temperature control prediction model employs a deep neural network containing an input layer, multiple fully connected hidden layers, and an output layer. A single sample from the query sample set is input into the benchmark temperature control prediction model, and the predicted temperature value is calculated through forward propagation. The partial derivative of the predicted temperature value with respect to each dimension of the input feature in the sample is calculated using the backpropagation algorithm, and the absolute value of the partial derivative is taken as the feature sensitivity of the corresponding dimension. When calculating the distance between any two samples in the query sample set, the difference between the two samples in each dimension is calculated. The absolute value of this difference is multiplied by the feature sensitivity of the corresponding dimension, and then the products calculated for all dimensions are summed to obtain the weighted state distance. For the current sample in the query sample set, the weighted state distances of this sample and all other samples are calculated and sorted in ascending order. The top ten nearest neighbor samples are selected, and the weighted state distances of these ten nearest neighbor samples and the current sample are summed. The reciprocal of the sum is taken as the local state concentration of the current sample. In an alternative embodiment, the top ten nearest neighbor samples are selected, the average of the weighted state distances of these ten nearest neighbor samples and the current sample is calculated, and this average value is substituted into the Gaussian radial basis negative exponential function to obtain the local state concentration of the current sample.

[0025] In an optional embodiment, the acquisition of multi-dimensional operating status data such as transformer load current, ambient temperature and humidity, winding and tank temperatures constitutes a query sample set, including: The communication node module is synchronously woken up by a time trigger at a preset duration, the communication port data is read, the actual ampere values ​​of all transmission and distribution circuits of the transformer are extracted, and the single current characteristic value is obtained. The ambient dry temperature and water mist relative humidity coefficient were extracted as real-time parameters of the surrounding environment. The temperature of the heating coil and the oil tank, which are read in real time from the internal pre-embedded and surface contact detection elements, are used as the true indicator characteristics of the internal temperature detection. By aligning different types of detection results from the same period side-by-side, a full-feature sequence state group is formed and placed into the corresponding variable database to obtain the feature matrix of the parameter vector feature group to be combined, which is used to construct the data extraction library query parameters.

[0026] The hardware time triggers inside the core control unit, such as the RTC real-time clock chip, are set to every... A synchronous hardware interrupt signal is sent at a fixed interval of 5 minutes. When the trigger reaches the 5-minute interval, it immediately wakes up the data acquisition modules deployed at each monitoring node of the transformer substation via the RS485 industrial bus. The communication node module parses the Modbus RTU protocol and reads the data in the register, extracts the actual current values ​​of phases A, B, and C of the three transmission and distribution branches on the low-voltage side of the transformer, and calculates the arithmetic mean or effective value of the three-phase current of each branch as the actual current value of that branch. The current of branch 1 is 450.5A, the current of branch 2 is 320.2A, and the current of branch 3 is 510.8A. These values ​​are then integrated into a single-dimensional single-current characteristic value through vector superposition or averaging, and denoted as [value missing]. By deploying louvered box-type meteorological sensors outside the transformer substation, the ambient temperature is captured. =35.2℃ and relative humidity coefficient of water mist =65.5%RH, serving as a real-time parameter representing the external heat dissipation environment of the transformer substation. The data processing module obtains the temperature of the heating coil through a PT100 platinum resistance probe embedded in the transformer winding layers. =85.6℃, and the oil tank temperature is obtained using a patch or immersion temperature sensor. =68.4℃, both together constitute the true internal temperature detection index characteristic representing the transformer's heat accumulation. To ensure consistency in the time dimension, a timestamp comparison algorithm is used to horizontally align single current characteristic values, surrounding real-time parameters, and the true internal temperature detection index characteristic with the same timestamp, such as 2023-10-15 14:05:00. Specifically, the integrated single current characteristic value... , , , , The scalar data from the five dimensions are sequentially arranged into one-dimensional row vectors of 427.1, 35.2, 0.655, 85.6, and 68.4, forming a state group for the entire feature sequence. After 24 hours of continuous data collection, the 288 row vectors obtained are appended to the time-series database and concatenated to form a 288×5 feature matrix of the parameter vector feature group to be combined. This matrix serves as the basic query parameter set for model querying and scheduling.

[0027] In an optional embodiment, the feature sensitivity of each sample in the query sample set is set based on the gradient of the baseline temperature control prediction model output with respect to the input of each dimension, including: Extract the input vector from the sample set to be queried, substitute it into the benchmark temperature control prediction model for forward calculation, and obtain the single temperature measurement result and the temperature prediction sequence within the future preset time period. Perform partial derivative calculations on the single-temperature measurement results to obtain the partial derivative values ​​corresponding to the direction of each input parameter as the gradient of each dimension input; The absolute values ​​of the partial derivatives of each gradient are calculated and converted into absolute gradient values, which are then summarized into a sample absolute gradient vector. The absolute gradient values ​​calculated from the collected samples are extracted, summed, and averaged to obtain the average absolute gradient of the class. The feature sensitivity is obtained by dividing each absolute gradient mean by the sum of all absolute gradient mean values.

[0028] Randomly or sequentially extract the i-th test input vector from the feature matrix. The input includes the load current. =450A, ambient temperature =35℃, humidity =65%, winding temperature =85℃, fuel tank temperature =68℃, substituted into the pre-trained benchmark prediction measurement network. The benchmark prediction measurement network is a deep neural network containing fully connected layers and long short-term memory network units. The structure includes an input layer, multiple hidden layers, and an output layer. The hidden layers contain recurrent computation units and linear mapping feedforward layers. The nonlinear feature transformation between the network layers is completed through a linear rectification function. The input of the benchmark prediction measurement network is the extracted input vector and a time series matrix composed of multiple time-time data within a historical preset time window. This vector contains the feature dimension values ​​of load current, ambient temperature, humidity, winding temperature, and oil tank temperature. The output of the benchmark prediction measurement network is the calculated single-temperature measurement result at the current moment and the continuous temperature prediction sequence of multiple nodes within a preset time period in the future. Through multi-layer matrix multiplication and forward propagation operation of the activation function, the single-temperature measurement result Y at the current moment is output, the predicted winding hot spot temperature Y=86.5℃ is given, and the temperature prediction sequence of 12 points with a step size of 5 minutes within the next 60 minutes is given. Using an automatic differentiation engine, the output scalar Y is compared with the input vector. Perform backpropagation differentiation, i.e., calculate the partial derivatives. Where j = 1, 2, 3, 4, 5. The calculated partial derivative values ​​for each dimension are 1.2 for load current gradient, 0.8 for ambient temperature gradient, -0.1 for humidity gradient, 0.5 for winding temperature gradient, and 0.3 for oil tank temperature gradient, which are the characteristic numbers of the partial variables.

[0029] To remove the interference of the gradient direction's sign on sensitivity evaluation, the absolute value of each partial derivative is calculated, i.e., |-0.1|=0.1, which is converted into the absolute value of the gradient. The absolute values ​​of the gradients for all dimensions of the same sample are then integrated into a vector, i.e., 1.2, 0.8, 0.1, 0.5, 0.3, to obtain the sample's absolute gradient vector. This process is performed on N samples (N=1000) in the entire query sample set, for a specific dimension, such as load current. The arithmetic mean of the absolute gradient values ​​of the load current class is obtained by summing all 1000 absolute gradient values ​​calculated from 1000 samples and dividing by N. =1.15, and similarly, the absolute average characteristics of the other four dimensions are calculated to be 0.75, 0.15, 0.45, and 0.25. The absolute average characteristics of all dimensions are summed to obtain a total S = 1.15 + 0.75 + 0.15 + 0.45 + 0.25 = 2.75. Each average characteristic is then divided by the total S to perform L1 norm normalization, yielding a single-feature calculation constant for each dimension as the final feature sensitivity. The resulting load current sensitivity is 1.15 / 2.75 ≈ 0.418, and the ambient temperature sensitivity is 0.273. This set of weight vectors measures the dominant influence of each state parameter on the temperature change result.

[0030] In an optional embodiment, the step of calculating the weighted state distance between samples using feature sensitivity to generate local state concentration includes: For each sample to be compared, a single extractable feature parameter of the sample on each attribute data is obtained, and the feature parameter is dimensionless. Using a single extracted feature parameter after dimensionless processing, perform a difference operation between the two to obtain the absolute difference result; Perform a weighted product operation between the absolute difference result and the corresponding feature sensitivity; Sum all weighted product results to obtain the weighted state distance; The local state concentration degree, representing the proximity of samples, is calculated based on the reciprocal or negative exponential function of the weighted state distance.

[0031] For the A-th sample to be evaluated and the B-th sample in the reference set, their attribute data in each dimension are read. To remove calculation biases caused by different dimensions such as current (0-1000A), temperature (-20-150℃), and humidity (0-100%), the minimum-maximum normalization method is used to map the single extracted feature parameters of each dimension to a closed interval of 0 to 1. The historical minimum value of the transformer load current is set to 100A and the maximum value to 900A. If the current of sample A is 500A, the dimensionless value is (500-100) / (900-100) = 0.50; if the current of sample B is 420A, the normalized value is (420-100) / 800 = 0.40. The dimensionless parameter values ​​of samples A and B in the same dimension are extracted, and the difference is calculated by subtracting them and taking the absolute value, resulting in an absolute difference of |0.50-0.40| = 0.10. Similarly, the absolute difference for ambient temperature is 0.15, and the absolute difference for humidity is 0.05. The absolute difference for each dimension is multiplied by the corresponding feature sensitivity. The feature sensitivity for current is 0.418, and for ambient temperature it is 0.273. Therefore, the weighted product for current is 0.0418, and the weighted product for ambient temperature is 0.04095. The weighted products for all five dimensions are summed to obtain 0.1255, which is the weighted state distance between the two samples. The weighted state distances between sample A and its K neighboring samples (K=10) are calculated, and the average distance is obtained. =0.150. To convert this value into a concentration index, a negative exponential function of the Gaussian radial basis function type is applied, with the attenuation bandwidth parameter σ=0.2, and substituted into the formula. The value is approximately 0.755. This value falls within the range greater than 0 and less than or equal to 1. The value 0.755 represents the density of sample A within the current multidimensional operational feature space, i.e., the local state concentration. The process of calculating feature sensitivity and local state concentration based on multidimensional operational state data, and the overall algorithm flow are as follows: Figure 2 As shown in the figure, in one embodiment, GRU, LSTM, and 1D-CNN are used as the baseline temperature control prediction model.

[0032] S2: Input the sample to be queried into a model group consisting of multiple temperature control prediction models with different structures, extract the gradient of the independent prediction results of each model with respect to each dimension of the input, input the feature sensitivity for weighted summation when calculating the dispersion of the gradient of each model, and generate a composite uncertainty measure by combining the rate of change of multidimensional operating status data.

[0033] A model set comprising three different network structures—Long Short-Term Memory (LSTM), gated recurrent units (GRU), and one-dimensional convolutional neural networks (CNNs)—is constructed. The normalized query samples are input into these three structurally distinct temperature prediction models, yielding independent temperature predictions from each model. For each model, the partial derivatives of its predictions with respect to each dimension of the multidimensional operational state data are calculated using backpropagation. These partial derivatives serve as the gradients of each model's output with respect to each dimension's input. For each dimension of the multidimensional operational state data, the variance of the gradients for each of the three models is calculated, representing the dispersion of each model's gradient. The gradient variance for each dimension is multiplied by the corresponding dimension's feature sensitivity obtained from the baseline temperature prediction model in step one. The products of all dimensions are then summed to obtain the weighted gradient dispersion. The rate of change for each dimension is obtained by subtracting the multidimensional operational state data collected at the previous moment from the current moment's data and dividing by the time interval between the two moments. The sum of the squares of the rates of change for each dimension using the Euclidean norm is then taken as the square root to obtain the overall rate of change, i.e., the comprehensive scalar value. After normalizing the weighted gradient dispersion (predicted divergence metric) and the overall rate of change, the composite uncertainty metric of the sample is generated by linearly combining and weighting the results using preset fusion weight coefficients.

[0034] In an optional embodiment, the step of inputting the sample to be queried into a model group consisting of multiple temperature control prediction models with different structures, extracting the gradient of the independent prediction results of each model with respect to each dimension of the input, and weighting and summing the input feature sensitivity when calculating the dispersion of the gradient of each model, and generating a composite uncertainty measure by combining the rate of change of multidimensional operating state data, includes: Extract the gradients of the independent prediction results of multiple structurally different temperature control prediction models with respect to each dimension of the input, and calculate the variance of the gradient of each model in each dimension as the basic dispersion. The basic dispersion of each dimension is multiplied by the corresponding feature sensitivity and then summed to obtain the prediction divergence metric on the amplified sensitive feature dimension. Extract the rate of change of multidimensional operating status data within a preset time window, and calculate the comprehensive scalar value of the rate of change of each dimension; The metric value and the comprehensive scalar value are normalized and weighted and merged to generate a composite uncertainty metric.

[0035] M three pre-set prediction models with distinct structures (M=3) are used. For the same input state samples, the partial derivative gradient values ​​of the predicted temperature output by these three models with respect to a certain dimension, such as load current, are extracted, and are 1.25, 1.35, and 1.15, respectively. The variance of these three gradient values ​​along the load current dimension is calculated to be approximately 0.0067. This variance serves as the basic dispersion for evaluating the consistency of the sensitivity of multiple models to the load current dimension. Similarly, the basic dispersion of ambient temperature is calculated to be 0.0042. The basic dispersion of each dimension is multiplied by the pre-calculated feature sensitivity: the product for load current is 0.418 × 0.0067 = 0.0028, and the product for ambient temperature is 0.273 × 0.0042 = 0.0011. The products of these five dimensions are summed to obtain the prediction divergence metric that highlights and amplifies the key temperature influence features. =0.0055. Rewind to the previous preset time window. The multi-dimensional operating status data over a 15-minute period is optionally preprocessed using methods such as normalization to calculate the instantaneous rate of change for each characteristic dimension. The rate of change for the load current dimension is 0.02 / min, for the ambient temperature dimension it is 0.01 / min, and for the winding temperature dimension it is 0.05 / min. The Euclidean norm is then used to sum the squares of multiple rate-of-change vectors and then take the square root to calculate the comprehensive scalar value. ≈0.055. This is the aforementioned measure of prediction discrepancy. Combined scalar value of the rate of change The range standardization algorithm is used to map the data to the interval between 0 and 1, and the normalized data is then... =0.45, =0.60. Fixed fusion weight coefficients α=0.7 and β=0.3 are set, and the result is obtained through a linear combination formula. The composite uncertainty measure U = 0.7 × 0.45 + 0.3 × 0.60 = 0.495 was calculated. This value represents the endogenous uncertainty of the model prediction itself and the exogenous uncertainty caused by the change in the current operating conditions of the equipment.

[0036] S3. A final query score is generated by integrating a sparsity index that is inversely correlated with the local state concentration and a composite uncertainty measure that is inversely weighted by the local state concentration. The sample with the highest final score is selected as the target query sample. The current real-time operating state sample is obtained. The temperature control bias is calculated based on the composite uncertainty measure and local state concentration of the sample. The temperature control bias is subtracted from the preset temperature target to obtain the real-time temperature control threshold. A temperature control command is triggered when the predicted temperature reaches the real-time temperature control threshold or the slope of future temperature change exceeds the preset benchmark.

[0037] The reciprocal of the local state concentration obtained in step one is used as the sparsity index of the sample. The composite uncertainty measure obtained in step two is divided by the local state concentration obtained in step one, thereby using the local state concentration to perform inverse weighting adjustment of the composite uncertainty measure. The calculated sparsity index is added to the adjusted composite uncertainty measure, and the sum is used as the final query score for the corresponding query sample. The final query scores of all samples in the query sample set are traversed, and the query sample corresponding to the highest score is selected as the target query sample and sent to the human expert system for labeling. The real temperature operation label corresponding to the target query sample is obtained; the target query sample with the real label is added to the historical training dataset, and the baseline temperature control prediction model and various different temperature control prediction models in the model group are fine-tuned and trained using backpropagation and gradient descent algorithms with the goal of minimizing the error between the predicted temperature and the real label.

[0038] During the real-time control phase, the latest multi-dimensional operating status data of the new energy transformer substation is acquired through various sensors as the current real-time operating status sample. Following the same calculation process as steps one and two, the local state concentration and composite uncertainty measure of the current real-time operating status sample are obtained. The composite uncertainty measure of the current real-time operating status sample is multiplied by a preset positive scaling factor to obtain the uncertainty bias component. The constant 1 is divided by the local state concentration of the current real-time operating status sample and multiplied by a preset negative scaling factor to obtain the concentration bias component. The uncertainty bias component and the concentration bias component are added together, and the result is used as the temperature control bias. The rated safe maximum operating temperature of the transformer substation is used as the preset temperature target. The temperature control bias is then subtracted from the preset temperature target to generate the real-time temperature control threshold. The system uses a baseline temperature control prediction model to predict the temperature of the current real-time operating status sample, and calculates the difference between the predicted temperature and the previous predicted temperature, divided by the time interval, as the future temperature change slope. Alternatively, the system uses the same model to predict the temperature of the current real-time operating status sample, obtaining the current predicted temperature and a temperature prediction sequence for a preset future time period. The difference between the temperature at the future target time and the current predicted temperature is extracted from the temperature prediction sequence and divided by the corresponding time interval to obtain the future temperature change slope. When the predicted temperature is determined to be greater than or equal to the real-time temperature control threshold, or the future temperature change slope is greater than the preset temperature rise rate safety benchmark value, the control system sends a temperature control command to start the operation of the transformer substation cooling fan and air conditioning equipment.

[0039] In an optional embodiment, the final query score is generated by fusing a sparsity index that is inversely correlated with local state concentration and a composite uncertainty measure that is inversely weighted by local state concentration, including: The sparsity index is obtained by dividing the numerical constant 1 as the numerator and the local state concentration as the denominator. Using the reciprocal of the local state concentration as the adjustment coefficient, multiplying it with the composite uncertainty measure, we obtain the composite uncertainty measure with inverse weighted adjustment; The final query score is generated by adding the sparsity index to the composite uncertainty measure with inverse weighting adjustment.

[0040] For any candidate query sample, obtain the local state concentration value previously calculated for that sample. =0.755 and the composite uncertainty measure U=0.495. The processor executes an arithmetic division instruction, using the fixed constant 1.0 as the numerator, and divides the local state concentration by... Using 0.755 as the denominator, a division operation is performed to calculate the sparsity index, which represents the degree of isolation of the samples in the feature space. =1.0 / 0.755≈1.3245, ensuring the inverse inverse correlation property that lower concentration leads to sparser samples. The inverse of the local state concentration, 1.3245, is reused as an adjustment coefficient. This adjustment coefficient is multiplied by the composite uncertainty measure U, and the result is substituted into the numerical value to obtain the inversely weighted composite uncertainty measure. =1.3245×0.495≈0.6556, this mechanism makes the model tend to nonlinearly amplify and penalize the inherent uncertainty of samples in sparse marginal regions of the data. The adder accumulator is then called to apply the basic sparsity index... With penalized amplified inverse weighted adjustment uncertainty Perform the addition operation to generate the final query score: Score = 1.3245 + 0.6556 = 1.9801. Calculate and iterate through all the query scores, sort them in descending order, and extract the sample with the highest score, which combines the characteristics of highest sparsity and least reliable prediction, as the target query sample.

[0041] In an optional embodiment, the step of obtaining a sample of the current real-time operating status, calculating a temperature control bias based on the composite uncertainty metric and local state concentration of the sample, and subtracting the temperature control bias from the preset temperature target to obtain the real-time temperature control threshold includes: Obtain the composite uncertainty measure and local state concentration of the current real-time running state sample; Multiply the composite uncertainty measure by a preset positive scaling factor to obtain the uncertainty bias component; Divide the numerical constant 1 by the local state concentration to obtain the reciprocal of the concentration, and multiply the reciprocal of the concentration by the preset inverse scaling factor to obtain the concentration bias component. The temperature control bias is obtained by adding the uncertainty bias component and the concentration bias component.

[0042] In each control cycle, the latest real-time operating conditions are collected, and the composite uncertainty measure of the current real-time operating state sample is obtained according to the same calculation process in steps S1 and S2 above. =0.65 and local state concentration =0.80. A positive scaling factor is pre-set in memory. Set to 5.0℃ per unit measure to measure the currently extracted composite uncertainty. =0.65 and the positive proportionality coefficient Perform multiplication to calculate the downward adjustment value of the safety margin based on uncertainty, i.e., the uncertainty bias component. =0.65 × 5.0 = 3.25℃. The control algorithm performs a reciprocal operation, dividing the constant 1 by the local state concentration. =0.80, resulting in the reciprocal of the concentration, 1 / 0.80 = 1.25. Extract the inverse scaling factor preset in the configuration file. Setting the reciprocal to 2.0℃ per unit, multiplying the calculated reciprocal of the concentration ratio by this inverse proportionality coefficient yields a conservative compensation value representing the missing data confidence level, i.e., the concentration bias component. =1.25 × 2.0 = 2.50℃. Combining the two conservative adjustment values ​​using arithmetic addition, the overall temperature control offset is calculated to be 3.25℃ + 2.50℃ = 5.75℃. If the original factory preset target for the air-cooling start-up temperature is 90.0℃, the system under current operating conditions will conservatively tighten and correct the real-time temperature control threshold to 90.0℃ - 5.75℃ = 84.25℃, achieving proactive heat dissipation to prevent potential problems.

[0043] In an optional embodiment, triggering the temperature control command when the predicted temperature reaches the real-time temperature control threshold or the future temperature change slope exceeds a preset benchmark includes: Obtain the current real-time operating status sample, and obtain the predicted temperature of the current real-time operating status sample and the temperature prediction sequence within a preset time period from the benchmark temperature control prediction model. Determine whether the predicted temperature is greater than or equal to the real-time temperature control threshold; Calculate the slope of temperature change within a preset time period based on temperature prediction sequences; Determine if the slope of the temperature change is greater than a preset benchmark. If the predicted temperature is greater than or equal to the real-time temperature control threshold, or the temperature change slope is greater than the preset benchmark, then a temperature control command is generated and output to the temperature control actuator.

[0044] The control system extracts the latest real-time operating status sample and feeds it into a pre-deployed baseline temperature control prediction model for forward inference. It then obtains the current predicted temperature Y=85.3℃ and a future temperature prediction sequence for a preset time period of 30 minutes with a 5-minute step size: 86.1℃, 87.5℃, 89.2℃, 91.4℃, 93.8℃, 96.5℃. The logic judgment module compares the current predicted temperature Y=85.3℃ with the real-time temperature control threshold calculated in the previous steps. A numerical comparison is performed at 84.25℃ to check if the condition 85.3℃ ≥ 84.25℃ is met. In parallel, the first and last data of the temperature prediction sequence are extracted, namely the current predicted temperature Y = 85.3℃ and the future time window. For example, the predicted temperature at the end of 30 minutes. The temperature is 96.5℃, using the difference formula. The average temperature change slope within the next 30-minute time window is calculated as Slope = (96.5℃ - 85.3℃) / 30min ≈ 0.373℃ / min. The logic judgment module extracts the preset baseline temperature rise rate stored in the register, with a warning threshold of 0.30℃ / min, and judges whether the calculated slope satisfies 0.373℃ / min > 0.30℃ / min. The above two judgment conditions are comprehensively evaluated through a logic OR gate. Since 85.3℃ ≥ 84.25℃ and 0.373 > 0.30 are both true, the trigger condition is met. The main control module immediately encapsulates a standard communication message containing the control code and outputs the temperature control command to the intelligent fan or oil pump of the transformer substation through the hardware interface to dissipate heat and avoid the risk of thermal runaway. The processing logic of sample sparsity, composite uncertainty metric fusion scoring and threshold temperature control is as follows. Figure 3 As shown.

[0045] In the second embodiment, the present invention also proposes a new energy transformer substation automatic temperature control system based on deep learning, comprising the following modules: The generation module is used to acquire multi-dimensional operating status data of the transformer load current, ambient temperature and humidity, winding and oil tank temperature to form a sample set to be queried; for each sample in the sample set to be queried, the feature sensitivity is set based on the gradient of each dimension input output by the benchmark temperature control prediction model, and the weighted state distance between samples is calculated using the feature sensitivity to generate the local state concentration. The calculation module is used to input the sample to be queried into a model group consisting of multiple temperature control prediction models with different structures, extract the gradient of the independent prediction results of each model with respect to each dimension of input, input the feature sensitivity for weighted summation when calculating the dispersion of the gradient of each model, and generate a composite uncertainty measure by combining the rate of change of multidimensional operating status data. The calculation module is used to fuse a sparsity index that is inversely correlated with the local state concentration and a composite uncertainty measure adjusted by inverse weighting of the local state concentration to generate a final query score, and select the sample with the highest final score as the target query sample; obtain the current real-time operating state sample, calculate the temperature control bias based on the composite uncertainty measure and local state concentration of the sample, subtract the temperature control bias from the preset temperature target to obtain the real-time temperature control threshold; and trigger a temperature control command when the predicted temperature reaches the real-time temperature control threshold or the slope of future temperature change exceeds a preset benchmark.

[0046] The above description represents the preferred embodiments of the present invention. It should be noted that, for those skilled in the art, various improvements and modifications can be made without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for automatic temperature control of new energy transformer substations based on deep learning, characterized in that, Includes the following steps: The multi-dimensional operating status data of the transformer substation, including load current, ambient temperature and humidity, winding temperature, and oil tank temperature, are used to form a sample set to be queried. For each sample in the sample set to be queried, the feature sensitivity is set based on the gradient of the input of each dimension output by the benchmark temperature control prediction model. The weighted state distance between samples is calculated using the feature sensitivity to generate the local state concentration. The sample to be queried is input into a model group consisting of multiple temperature control prediction models with different structures. The gradient of the independent prediction results of each model with respect to each dimension of the input is extracted. When calculating the dispersion of the gradient of each model, the feature sensitivity is input for weighted summation. Combined with the rate of change of multidimensional operating status data, a composite uncertainty measure is generated. A final query score is generated by integrating a sparsity index that is inversely correlated with local state concentration and a composite uncertainty measure that is inversely weighted by local state concentration. The sample with the highest final score is selected as the target query sample. The current real-time operating state sample is obtained, and the temperature control bias is calculated based on the composite uncertainty measure and local state concentration of the sample. The temperature control bias is subtracted from the preset temperature target to obtain the real-time temperature control threshold. A temperature control command is triggered when the predicted temperature reaches the real-time temperature control threshold or the slope of future temperature change exceeds the preset benchmark.

2. The method according to claim 1, characterized in that, The acquired multi-dimensional operating status data, including the substation load current, ambient temperature and humidity, winding and oil tank temperatures, constitute the query sample set, including: The communication node module is synchronously woken up by a time trigger at a preset duration, the communication port data is read, the actual ampere values ​​of all transmission and distribution circuits of the transformer are extracted, and the single current characteristic value is obtained. The ambient dry temperature and water mist relative humidity coefficient were extracted as real-time parameters of the surrounding environment. The temperature of the heating coil and the oil tank, which are read in real time from the internal pre-embedded and surface contact detection elements, are used as the true indicator characteristics of the internal temperature detection. By aligning different types of detection results from the same period side-by-side, a full-feature sequence state group is formed and placed into the corresponding variable database to obtain the feature matrix of the parameter vector feature group to be combined, which is used to construct the data extraction library query parameters.

3. The method according to claim 1, characterized in that, For each sample in the set of samples to be queried, the feature sensitivity is set based on the gradient of the input dimensions output by the benchmark temperature control prediction model, including: Extract the input vector from the sample set to be queried, substitute it into the benchmark temperature control prediction model for forward calculation, and obtain the single temperature measurement result and the temperature prediction sequence within the future preset time period. Perform partial derivative calculations on the single-temperature measurement results to obtain the partial derivative values ​​corresponding to the direction of each input parameter as the gradient of each dimension input; The absolute values ​​of the partial derivatives of each gradient are calculated and converted into absolute gradient values, which are then summarized into a sample absolute gradient vector. The absolute gradient values ​​calculated from the collected samples are extracted, summed, and averaged to obtain the average absolute gradient of the class. The feature sensitivity is obtained by dividing each absolute gradient mean by the sum of all absolute gradient mean values.

4. The method according to claim 1, characterized in that, The method of calculating the weighted state distance between samples using feature sensitivity to generate local state concentration includes: For each sample to be compared, a single extractable feature parameter of the sample on each attribute data is obtained, and the feature parameter is dimensionless. Using a single extracted feature parameter after dimensionless processing, perform a difference operation between the two to obtain the absolute difference result; Perform a weighted product operation between the absolute difference result and the corresponding feature sensitivity; Sum all weighted product results to obtain the weighted state distance; The local state concentration degree, representing the proximity of samples, is calculated based on the reciprocal or negative exponential function of the weighted state distance.

5. The method according to claim 1, characterized in that, The process involves inputting the sample to be queried into a model group composed of multiple temperature control prediction models with different structures, extracting the gradient of each model's independent prediction result with respect to each dimension of the input, and weighting and summing the input feature sensitivity when calculating the dispersion of each model's gradient. A composite uncertainty metric is then generated by combining the rate of change of multidimensional operating state data, including: Extract the gradients of the independent prediction results of multiple structurally different temperature control prediction models with respect to each dimension of the input, and calculate the variance of the gradient of each model in each dimension as the basic dispersion. The basic dispersion of each dimension is multiplied by the corresponding feature sensitivity and then summed to obtain the prediction divergence metric on the amplified sensitive feature dimension. Extract the rate of change of multidimensional operating status data within a preset time window, and calculate the comprehensive scalar value of the rate of change of each dimension; The metric value and the comprehensive scalar value are normalized and weighted and merged to generate a composite uncertainty metric.

6. The method according to claim 1, characterized in that, The final query score is generated by combining a sparsity index that is inversely correlated with the local state concentration and a composite uncertainty measure that is inversely weighted by the local state concentration, including: The sparsity index is obtained by dividing the numerical constant 1 as the numerator and the local state concentration as the denominator. Using the reciprocal of the local state concentration as the adjustment coefficient, multiplying it with the composite uncertainty measure, we obtain the composite uncertainty measure with inverse weighted adjustment; The final query score is generated by adding the sparsity index to the composite uncertainty measure with inverse weighting adjustment.

7. The method according to claim 1, characterized in that, The process of acquiring the current real-time operating status sample, calculating the temperature control bias based on the composite uncertainty metric and local state concentration of the sample, and subtracting the temperature control bias from the preset temperature target to obtain the real-time temperature control threshold includes: Obtain the composite uncertainty measure and local state concentration of the current real-time running state sample; Multiply the composite uncertainty measure by a preset positive scaling factor to obtain the uncertainty bias component; Divide the numerical constant 1 by the local state concentration to obtain the reciprocal of the concentration, and multiply the reciprocal of the concentration by the preset inverse scaling factor to obtain the concentration bias component. The temperature control bias is obtained by adding the uncertainty bias component and the concentration bias component.

8. The method according to claim 1, characterized in that, The method of triggering a temperature control command when the predicted temperature reaches the real-time temperature control threshold or the slope of future temperature change exceeds a preset benchmark includes: Obtain the current real-time operating status sample, and obtain the predicted temperature of the current real-time operating status sample and the temperature prediction sequence within a preset time period from the benchmark temperature control prediction model. Determine whether the predicted temperature is greater than or equal to the real-time temperature control threshold; Calculate the slope of temperature change within a preset time period based on temperature prediction sequences; Determine if the slope of the temperature change is greater than a preset benchmark. If the predicted temperature is greater than or equal to the real-time temperature control threshold, or the temperature change slope is greater than the preset benchmark, then a temperature control command is generated and output to the temperature control actuator.

9. A new energy transformer substation automatic temperature control system based on deep learning, characterized in that, Includes the following modules: The generation module is used to acquire multi-dimensional operating status data of the transformer load current, ambient temperature and humidity, winding and oil tank temperature to form a sample set to be queried; for each sample in the sample set to be queried, the feature sensitivity is set based on the gradient of each dimension input output by the benchmark temperature control prediction model, and the weighted state distance between samples is calculated using the feature sensitivity to generate the local state concentration. The calculation module is used to input the sample to be queried into a model group consisting of multiple temperature control prediction models with different structures, extract the gradient of the independent prediction results of each model with respect to each dimension of input, input the feature sensitivity for weighted summation when calculating the dispersion of the gradient of each model, and generate a composite uncertainty measure by combining the rate of change of multidimensional operating status data. The calculation module is used to fuse a sparsity index that is inversely correlated with the local state concentration and a composite uncertainty measure adjusted by inverse weighting of the local state concentration to generate a final query score, and select the sample with the highest final score as the target query sample; obtain the current real-time operating state sample, calculate the temperature control bias based on the composite uncertainty measure and local state concentration of the sample, subtract the temperature control bias from the preset temperature target to obtain the real-time temperature control threshold; and trigger a temperature control command when the predicted temperature reaches the real-time temperature control threshold or the slope of future temperature change exceeds a preset benchmark.

10. The system according to claim 9, characterized in that, The acquired multi-dimensional operating status data, including the substation load current, ambient temperature and humidity, winding and oil tank temperatures, constitute the query sample set, including: The communication node module is synchronously woken up by a time trigger at a preset duration, the communication port data is read, the actual ampere values ​​of all transmission and distribution circuits of the transformer are extracted, and the single current characteristic value is obtained. The ambient dry temperature and water mist relative humidity coefficient were extracted as real-time parameters of the surrounding environment. The temperature of the heating coil and the oil tank, which are read in real time from the internal pre-embedded and surface contact detection elements, are used as the true indicator characteristics of the internal temperature detection. By aligning different types of detection results from the same period side-by-side, a full-feature sequence state group is formed and placed into the corresponding variable database to obtain the feature matrix of the parameter vector feature group to be combined, which is used to construct the data extraction library query parameters.

Images