A power distribution transformer heavy overload risk identification method, system, device, medium and product

Abstract

The application discloses a power distribution transformer heavy overload risk identification method, system, device, medium and product, relates to the technical field of power system operation and maintenance, and comprises the following steps: acquiring multi-source heterogeneous data of a power distribution transformer; constructing a hotspot temperature prediction model based on physical information residual learning; predicting the multi-source heterogeneous data based on the hotspot temperature prediction model to obtain a hotspot temperature prediction value of the power distribution transformer; calculating a multi-dimensional heavy overload risk index according to the hotspot temperature prediction value and in combination with static account data and dynamic operation monitoring data in the multi-source heterogeneous data; and comparing the multi-dimensional heavy overload risk index with a preset threshold to obtain a risk state identification result of the heavy overload of the power distribution transformer. The application can improve the heavy overload risk identification accuracy, and greatly improve the perception accuracy and active operation and maintenance capability of the power distribution transformer under complex working conditions.

Description

Technical Field

[0001] This application relates to the field of power system operation and maintenance technology, and in particular to a method, system, equipment, medium and product for identifying heavy overload risks of distribution transformers. Background Technology

[0002] As a critical node connecting the power transmission network and end users, the operating status of distribution transformers directly determines the reliability and security of the power supply network. With the deepening of energy transition, the large-scale integration of distributed photovoltaic power, the penetration of electric vehicle charging loads, and the widespread use of electrification equipment, the load characteristics of distribution transformers have changed from traditional periodic fluctuations to a form with strong randomness and high volatility, leading to frequent heavy overload events. Heavy overload is essentially a continuous thermal stress process; prolonged overload will accelerate the aging of oil-paper insulation and may even cause serious safety accidents.

[0003] Current assessment indicators for heavy overload of distribution transformers are singular and rigid: existing operation and maintenance models generally rely on current thresholds (such as load rate > 1.2 times) for judgment, ignoring the transformer's huge thermal inertia and the influence of external conditions such as ambient temperature and solar radiation. In winter low temperatures, even if the equipment is overloaded, it may still be within the safe temperature range, and relying solely on current criteria will cause "false alarms"; conversely, in summer high temperatures or poor ventilation scenarios, even if the rated current is not reached, the internal hot spots of the equipment may overheat, and traditional methods are prone to "missed alarms". Summary of the Invention

[0004] The purpose of this application is to provide a method, system, equipment, medium, and product for identifying heavy overload risks in distribution transformers, which can avoid misjudgments and improve the accuracy of heavy overload risk identification.

[0005] To achieve the above objectives, this application provides the following solution: Firstly, this application provides a method for identifying the risk of heavy overload in a distribution transformer, including: Acquire multi-source heterogeneous data of the distribution transformer; the multi-source heterogeneous data includes static ledger data, dynamic operation monitoring data, and environmental data.

[0006] A hotspot temperature prediction model based on physical information residual learning is constructed. The hotspot temperature prediction model includes a white-box physical model and a black-box deep learning model. The white-box physical model is used to calculate the physical reference temperature based on the IEC standard thermal circuit model, and the black-box deep learning model is used to obtain the residual correction term of the physical reference temperature based on a deep neural network.

[0007] The hotspot temperature prediction value of the distribution transformer is obtained by using multi-source heterogeneous data based on the hotspot temperature prediction model.

[0008] Based on the predicted hotspot temperature and combined with static ledger data and dynamic operation monitoring data from multi-source heterogeneous data, a multi-dimensional heavy overload risk index is calculated. The multi-dimensional heavy overload risk index includes: thermal risk index, electrical stress index and equipment health index.

[0009] By comparing the multidimensional heavy overload risk index with the preset threshold, the risk status identification result of the heavy overload of the distribution transformer is obtained.

[0010] Secondly, this application provides a system for identifying the risk of heavy overload of a distribution transformer, including: a data acquisition unit, a model building unit, a prediction unit, an index calculation unit, and a risk status identification unit.

[0011] The data acquisition unit is used to acquire multi-source heterogeneous data of the distribution transformer; the multi-source heterogeneous data includes static ledger data, dynamic operation monitoring data, and environmental data.

[0012] The model building unit is used to build a hotspot temperature prediction model based on physical information residual learning. The hotspot temperature prediction model includes a white-box physical model and a black-box deep learning model. The white-box physical model is used to calculate the physical reference temperature based on the IEC standard thermal circuit model, and the black-box deep learning model is used to obtain the residual correction term of the physical reference temperature based on a deep neural network.

[0013] The prediction unit is used to predict the hot spot temperature of the distribution transformer based on the hot spot temperature prediction model using multi-source heterogeneous data.

[0014] The indicator calculation unit is used to calculate multidimensional heavy overload risk indicators based on the hot spot temperature prediction value and combined with static ledger data and dynamic operation monitoring data in multi-source heterogeneous data. The multidimensional heavy overload risk indicators include: thermal risk indicators, electrical stress indicators and equipment health indicators.

[0015] The risk status identification unit is used to compare the multi-dimensional heavy overload risk index with the preset threshold to obtain the risk status identification result of the heavy overload of the distribution transformer.

[0016] Thirdly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the distribution transformer overload risk identification method described in any one of the above.

[0017] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the distribution transformer overload risk identification method described in any one of the above descriptions.

[0018] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the distribution transformer overload risk identification method described above.

[0019] According to the specific embodiments provided in this application, the following technical effects are disclosed: This application provides a method, system, device, medium, and product for identifying heavy overload risks in distribution transformers. By acquiring multi-source heterogeneous data from distribution transformers, it overcomes the limitations of a single data source and constructs a hotspot temperature prediction model based on physical information residual learning. This model provides a physical reference temperature that conforms to the laws of thermodynamics and compensates for deviations in the white-box physical model through residual correction terms, achieving a balance between physical consistency and data-driven accuracy. The multi-source heterogeneous data is used to predict hotspot temperatures based on the hotspot temperature prediction model, instantly outputting the predicted hotspot temperature value at the current moment, which is beneficial for automated and accurate assessment. Multi-dimensional heavy overload risk indicators are calculated based on the predicted hotspot temperatures. These indicators are compared with preset thresholds to obtain the risk status identification result. This breaks away from the traditional single-dimensional judgment mode that relies solely on current thresholds, assessing heavy overload risks from multiple dimensions, avoiding misjudgments, improving the accuracy of heavy overload risk identification, and significantly enhancing the perception accuracy and proactive maintenance capabilities of distribution transformers under complex operating conditions. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating a method for identifying the risk of heavy overload of a distribution transformer, provided as an embodiment of this application.

[0022] Figure 2 This is a flowchart illustrating a method for identifying the risk of heavy overload in a distribution transformer, as provided in an embodiment of this application.

[0023] Figure 3 This is a flowchart of a multi-source heterogeneous data acquisition method for a distribution transformer, provided as an embodiment of this application.

[0024] Figure 4 This is a flowchart illustrating a method for constructing a comprehensive evaluation index system for state assessment, as provided in one embodiment of this application.

[0025] Figure 5 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0026] 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.

[0027] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0028] The current analysis of heavy overload conditions in distribution transformers has the following obvious shortcomings: Physical model parameters drift significantly: Traditional IEC standard thermal circuit models heavily rely on static nameplate parameters (such as rated losses and thermal time constants) at the time of equipment delivery, and cannot perceive the aging state of equipment after long-term operation (such as sludge deposition and reduced heat dissipation efficiency due to radiator dust accumulation). Static models based on "new equipment" parameters often underestimate the temperature rise risk of older equipment and are difficult to accurately describe the nonlinear characteristics of oil viscosity changes with temperature.

[0029] Lack of multi-dimensional comprehensive assessment capabilities: Existing monitoring methods are mostly limited to monitoring instantaneous electrical quantities, lacking dynamic quantitative analysis of transformer "thermal tolerance margin" and "insulation life loss". Maintenance personnel have difficulty knowing how long the equipment can operate safely under the current overload, and cannot fully explore the short-term overload potential of the transformer while ensuring safety. This leads to maintenance decisions that are either too conservative, resulting in wasted resources, or too aggressive, causing equipment burnout.

[0030] This application can achieve accurate soft measurement of hot spot temperature and multi-dimensional "thermal-electrical-life" status assessment by deeply fusing multi-source heterogeneous data, constructing a physical information residual learning model, and dynamically calculating insulation life loss. This solves the problems of parameter drift and single judgment index in traditional models, and improves the accuracy of overload warning for distribution transformers and the scientific nature of operation and maintenance decisions.

[0031] like Figure 1 As shown, a method for identifying the risk of heavy overload of a distribution transformer is provided, including the following steps 201 to 205.

[0032] Step 201: Obtain multi-source heterogeneous data of the distribution transformer; the multi-source heterogeneous data includes static ledger data, dynamic operation monitoring data, and environmental data.

[0033] Step 202: Construct a hotspot temperature prediction model based on physical information residual learning. The hotspot temperature prediction model includes a white-box physical model and a black-box deep learning model. The white-box physical model is used to calculate the physical reference temperature based on the IEC standard thermal circuit model, and the black-box deep learning model is used to obtain the residual correction term of the physical reference temperature based on a deep neural network.

[0034] In another exemplary embodiment of this application, the black-box deep learning model employs a long short-term memory network based on an attention mechanism, including an input layer, an LSTM temporal coding layer, an attention weight calculation layer, and a residual output layer.

[0035] The LSTM temporal coding layer encodes the input feature tensor in chronological order to obtain the hidden layer state at each historical time step.

[0036] The attention weight calculation layer calculates the attention score and attention weight based on the hidden layer state of each historical time step, and then performs weighted aggregation on the hidden layer state of the historical time step based on the attention weight to obtain the context vector.

[0037] The residual output layer concatenates the context vector with the single-step fused feature vector at the current time step to form the residual mapping input, and outputs the residual correction term through a fully connected mapping.

[0038] In another exemplary embodiment of this application, the loss function of the hotspot temperature prediction model is: .

[0039] in, For loss function, The number of training samples. For the first The input feature tensor of each training sample This refers to the residual correction term output by the black-box deep learning model. For the first The residual labels of each training sample, where λ is the regularization coefficient and Θ is the set of network parameters.

[0040] The residual label is determined by the difference between the winding hot spot temperature label at the corresponding target time and the physical reference temperature output by the IEC physical thermal circuit model, and is constructed based on the winding hot spot temperature label and the physical reference temperature. t Time-based residual labels : .

[0041] in, The winding hotspot temperature label is for the target moment. The physical reference temperature is the output of the IEC physical thermal circuit model at the target time.

[0042] Step 203: The multi-source heterogeneous data is used to predict the hot spot temperature based on the hot spot temperature prediction model to obtain the predicted value of the hot spot temperature of the distribution transformer.

[0043] Step 204: Based on the predicted hotspot temperature and combined with the static ledger data and dynamic operation monitoring data in the multi-source heterogeneous data, calculate the multi-dimensional heavy overload risk index. The multi-dimensional heavy overload risk index includes: thermal risk index, electrical stress index and equipment health index.

[0044] In another exemplary embodiment of this application, at least one of the following is used: hot spot temperature prediction, relative aging rate, remaining safe operating time, and bubble generation temperature safety margin.

[0045] The remaining safe operating time is determined by the time difference between the moment when the predicted winding hot spot temperature first reaches the preset safe temperature limit and the current moment.

[0046] The formula for calculating the relative aging rate is: .

[0047] in, For a moment The relative aging rate, These are insulation aging characteristic parameters. The reference hotspot temperature represents the relative aging rate. The corresponding hot spot temperature at that time This represents the temperature shift during the conversion from Celsius temperature to thermodynamic temperature. For a moment Hotspot temperature prediction values.

[0048] Under the condition that the current load and environmental conditions remain unchanged, or that the load and environmental data are known within the preset short-term prediction window, the hot spot temperature prediction model is run recursively according to the sampling period to obtain the winding hot spot temperature prediction sequence for subsequent times; the time difference between the moment when the winding hot spot temperature prediction sequence first reaches the preset hot spot temperature limit or the preset bubble generation temperature and the current moment is determined as the remaining safe operating time.

[0049] The formula for calculating the safety margin of the bubble generation temperature is as follows: .

[0050] in, For a moment The temperature safety margin for bubble generation The preset temperature for bubble generation.

[0051] In another exemplary embodiment of this application, the electrical stress parameters include at least one of the following: load factor, three-phase unbalance, zero-sequence current, and voltage harmonic distortion rate.

[0052] The formula for calculating the load rate is: .

[0053] .

[0054] .

[0055] in, For a moment load rate, For a moment The three-phase average current, Rated current, , , They are time points RMS value of three-phase current Rated capacity of distribution transformer This is the rated line voltage.

[0056] The formula for calculating the three-phase unbalance is: .

[0057] in, For a moment The three-phase imbalance.

[0058] In another exemplary embodiment of this application, the formula for calculating the device health index is: .

[0059] .

[0060] .

[0061] .

[0062] .

[0063] in, For equipment health indicators, , , , All are weighting coefficients. This is the normalized value of the equipment's service life. This is the normalized value of the number of historical overheating events within the statistical period. This is the normalized value of the number of historical defects within the statistical period. This is the normalized value of the number of historical severe defects within the statistical period. For the service life of the equipment, For design life, To count the number of historical overheating events within the statistical period, The preset normalized baseline value for the number of superheating cycles is used. For the number of historical defects, The default defect count is the normalized baseline value. This refers to the number of serious defects in history. This is the preset normalized baseline value for the number of severe defects.

[0064] Step 205: Compare the multidimensional heavy overload risk index with the preset threshold to obtain the risk status identification result of the heavy overload of the distribution transformer.

[0065] The risk status identification results include a risk warning level identifier (or status identifier), an over-limit indicator, and an over-limit amplitude. The risk warning level identifier includes Level I warning (emergency overload), Level II warning (severe overload), Level III warning (heavy load monitoring), and Level IV warning (mild attention).

[0066] By implementing steps 201 to 205 above, this application overcomes the limitations of a single data source by acquiring multi-source heterogeneous data of the distribution transformer, and constructs a hotspot temperature prediction model based on physical information residual learning. This model can provide a physical reference temperature that conforms to the laws of thermodynamics, and can also compensate for the deviation of the white-box physical model through residual correction terms, thus achieving a unity of physical consistency and data-driven accuracy. The multi-source heterogeneous data is used to predict the hotspot temperature based on the hotspot temperature prediction model, and the predicted hotspot temperature value at the current moment is output in real time, which is conducive to automated and accurate assessment. Based on the predicted hotspot temperature value, a multi-dimensional heavy overload risk index is calculated. The multi-dimensional heavy overload risk index is compared with a preset threshold to obtain the risk status identification result. This breaks the traditional single-dimensional judgment mode that only relies on the current threshold, and assesses the heavy overload risk from multiple dimensions, avoiding misjudgment, improving the accuracy of heavy overload risk identification, and significantly improving the perception accuracy and proactive operation and maintenance capability of the distribution transformer under complex operating conditions.

[0067] The specific content of the method for identifying the heavy overload risk of distribution transformers in this application is as follows, including the following steps.

[0068] Step 1: Obtain multi-source heterogeneous data of the distribution transformer. The multi-source heterogeneous data includes static ledger data from asset management data sources, dynamic operation monitoring data from operation monitoring data sources, and environmental data from meteorological and environmental data sources. The data is then preprocessed to obtain standardized feature vectors.

[0069] The process involves acquiring multi-source heterogeneous data from distribution transformers and preprocessing it to obtain standardized feature vectors. Specifically, this includes: acquiring static ledger data, including equipment model, core material, winding structure, insulation heat resistance level, transformer weight, total oil volume, and commissioning date; acquiring dynamic operation monitoring data, which comes from operation monitoring data sources, including at least one of metering automation systems, electricity consumption information collection systems, distribution automation systems, smart meters, distribution area integration terminals, or distribution transformer monitoring terminals. The dynamic operation monitoring data includes three-phase current, three-phase voltage, zero-sequence current, and total voltage harmonic distortion rate. The load factor and three-phase unbalance are calculated from the three-phase current, three-phase voltage, and rated parameters of the distribution transformer. When the operation monitoring data source directly provides the load factor or three-phase unbalance, it is used as verification data or backup input data. Meteorological environmental data is acquired from meteorological environmental data sources, including at least one of meteorological station interfaces, third-party meteorological platform interfaces, regional micro-meteorological stations, local temperature and humidity sensors, solar radiation sensors, or wind speed sensors. The meteorological environmental data includes at least one of ambient temperature, solar radiation intensity, wind speed, and humidity.

[0070] When preprocessing the above data, instead of simply splicing data from different sources, a unified time axis is first established. Specifically, the sampling time of the dynamic operation monitoring data output by the metering automation system is used as the reference time axis; for meteorological environmental data with different sampling periods, linear interpolation or proximity-based sampling is used to resample to the reference time axis; for static ledger data that does not change with short-term operating status, it is first categorized or numerically processed, and then copied at each sampling time to serve as equipment status characteristics for subsequent calculations.

[0071] At each sampling time Construct a single-step fusion feature vector: .

[0072] in, S This represents static ledger data. E s ( S This represents the static equipment characteristics obtained after encoding the equipment model, core material, winding structure, and insulation heat resistance level. Indicates the first The dynamic operating characteristics at each sampling time include the three-phase current, three-phase voltage, zero-sequence current and total voltage harmonic distortion rate collected, as well as the load rate and three-phase unbalance calculated based on the three-phase current, three-phase voltage and the rated parameters of the distribution transformer. Indicates the first Meteorological environmental characteristics at each sampling time, including ambient temperature, solar radiation intensity, and wind speed; This represents derived features calculated from the original data, including the square of the load rate, the load change rate, and the ambient temperature change rate; Norm(·) represents normalization.

[0073] The load factor and three-phase imbalance are calculated based on operational monitoring data during the multi-source heterogeneous data preprocessing stage and used as input features for the hotspot temperature prediction model. In the subsequent risk status identification stage, the load factor, three-phase imbalance, zero-sequence current, and total voltage harmonic distortion rate at the corresponding time points are reused to construct electrical stress indices. The predicted hotspot temperature is used to calculate thermal risk indices and is not used as input for calculating the load factor and three-phase imbalance.

[0074] To adapt to long short-term memory networks that incorporate attention mechanisms, the current moment is... and its predecessor The single-step fusion feature vectors from each historical sampling moment are stacked in chronological order to form the input feature tensor.

[0075] Using the sampling time of dynamic operation monitoring data as the reference time axis, time alignment, outlier cleaning, missing value completion, feature encoding, and normalization are performed on static ledger data, dynamic operation monitoring data, and meteorological environment data. Specifically, static ledger data is encoded as equipment status features and copied in the time dimension, meteorological environment data is resampled to the reference time axis, and load rate, three-phase imbalance, zero-sequence current, and harmonic distortion characteristics are calculated from dynamic operation monitoring data.

[0076] At each sampling time Constructing a single-step fusion feature vector ,in, This is the sampling time index. For the target prediction time... The single-step fusion feature vector at the current moment is denoted as ,Right now for The single-step fusion feature vector corresponding to the time step. The target prediction time step. and its predecessor The single-step fused feature vectors from each historical sampling moment are stacked in chronological order to obtain the input feature tensor: .

[0077] in, The length of the historical time window. m The dimension of the feature vector fused in a single step. The input feature tensor is used as input to the long short-term memory network for the fusion attention mechanism. Thus, It also includes information on inherent equipment attributes, operating electrical quantities, meteorological boundary conditions, and historical load changes, used to characterize the time lag characteristics of hot spot temperatures in distribution transformers.

[0078] Step 2: Construct a hotspot temperature prediction model based on physical information residual learning. The model includes a white-box physical model and a black-box deep learning model. Input the standardized feature vector into the hotspot temperature prediction model, calculate the physical reference temperature and nonlinear residual correction term in parallel, and obtain the real-time hotspot temperature prediction value of the distribution transformer winding by superposition.

[0079] This includes constructing a hotspot temperature prediction model based on physical information residual learning, and inputting feature vectors into the model for calculation, specifically including: The input feature tensor Input the residual correction network and calculate the predicted winding hot spot temperature using the following formula: .

[0080] .

[0081] in, For a moment Predicted values ​​for winding hot spot temperature; The physical reference temperature is calculated based on the IEC thermal circuit model; optional, Based on the IEC standard thermal circuit model, using the load current I and ambient temperature Calculated; This is the temperature correction term output by the Attention-LSTM residual correction network; The input feature tensor is obtained by time alignment, feature encoding, normalization, and sliding time window stacking of multi-source heterogeneous data; This is the output of the black-box deep learning part, representing the processing of input features using a deep neural network. The residual correction term obtained by nonlinear mapping is used to compensate for errors caused by parameter drift in the physical model and environmental simplification.

[0082] Black-box deep learning Employing an attention-based long short-term memory network structure, the network does not directly output the winding hotspot temperature, but instead outputs a residual correction term relative to a physical reference temperature. .

[0083] The black-box deep learning part It includes an input layer, an LSTM temporal coding layer, an attention weight calculation layer, and a residual output layer.

[0084] The LSTM temporal coding layer encodes the input feature tensor in chronological order to obtain the hidden layer state at each historical time step: .

[0085] in, For the first The hidden layer state vector corresponding to each sampling time step. LSTM(·) is the temporal coding function of the Long Short-Term Memory network, used to fuse the feature vector at the current sampling time step and the hidden layer state at the previous sampling time step, and output the hidden layer state at the current sampling time step. For the first The single-step fusion feature vector at each sampling time point For the first The LSTM hidden layer state vector corresponding to each sampling time step This serves as an index for sampling moments within a historical time window. Predict the time for the current target. This represents the length of the historical time window.

[0086] The attention weight calculation layer calculates the attention score and attention weight based on the state of each hidden layer: .

[0087] .

[0088] in, This serves as the summation index during the Softmax normalization process; For the first time window in history The unnormalized attention score corresponding to the sampling time step is used to characterize the first sampling time step. Hidden layer state at each sampling time Relative to the target prediction time The degree of relevance; The attention score projection vector; This is the hidden layer state weight matrix, used to weight the hidden layer states. Mapped to the attention computation space; This is the single-step fusion feature weight matrix at the current time, used to fuse the single-step feature vector at the current time. Mapped to the attention computation space; This is the attention bias vector; , , and All of these are network parameters obtained through training; For the first The LSTM hidden layer state vector corresponding to each sampling time; Predicting the target time The corresponding current single-step fusion feature vector; For the first Normalized attention weights corresponding to each sampling time; Index of sampling times within the historical time window; This serves as the summation index during the attention weight normalization process; For the first The unnormalized attention score corresponding to each sampling time; The length of the historical time window; It is the hyperbolic tangent activation function; It is an exponential function.

[0089] For the same target prediction time ,satisfy ,and .

[0090] The hidden layer states corresponding to each sampling time within the historical time window are weighted and aggregated according to the attention weights to obtain the target prediction time. Corresponding context vector: .

[0091] in, This is the context vector.

[0092] The residual output layer will be composed of the input feature tensor. The obtained context vector , and the target prediction time The corresponding current single-step fusion feature vector The data is then concatenated to form the residual mapping input.

[0093] .

[0094] The residual correction term is output through a fully connected mapping. : .

[0095] in, , , , For the network parameters obtained during training, The residual correction term is a nonlinear activation function used to compensate for temperature calculation errors in the IEC physical thermal circuit model caused by parameter drift, meteorological boundary simplification, and equipment aging.

[0096] The final predicted winding hot spot temperature is: .

[0097] in, The reference hot spot temperature output by the IEC physical thermal circuit model. This is the residual correction term output by the black-box deep learning part. This structure enables the LSTM network to learn nonlinear temperature rise biases not covered by the physical model, including changes in meteorological boundaries, degradation of heat dissipation conditions, equipment aging, and historical load hysteresis effects.

[0098] The black-box deep learning part of this application A long short-term memory (LSTM) network structure based on an attention mechanism is employed. This network is built upon existing LSTM timing units and attention weight calculation units, but instead of directly predicting winding hotspot temperatures, it is embedded within a physical information residual learning framework to predict temperature errors in the IEC physical thermal circuit model. Its output is a residual correction term. The final hot spot temperature is obtained by superimposing the physical reference temperature and the residual correction term.

[0099] Black-box deep learning The training samples consist of input feature tensors and hotspot temperature labels. For any training time... Input features , tag as .in, This includes static ledger characteristics, dynamic operational characteristics, meteorological environment characteristics, and derived time series characteristics; The winding hot spot temperature at the corresponding moment can be obtained by fiber optic temperature measurement device, temperature rise test or multi-physics field coupling simulation model.

[0100] The training process of the hotspot temperature prediction model in this application includes: Establish a training sample set, wherein a single sample in the training sample set is represented as ,in, The input feature tensor is obtained by time alignment, feature encoding, normalization, and sliding time window stacking of static ledger data, dynamic operation monitoring data, and meteorological environment data. For time The corresponding winding hot spot temperature label.

[0101] The winding hot spot temperature label The data are derived from measured data of distribution transformers with fiber optic temperature measurement, temperature rise test data, or multi-physics field coupling simulation data of electro-magnetic-thermal-fluid fields.

[0102] Input the load current and ambient temperature from the sample into the IEC physical thermal circuit model to calculate the physical reference temperature: .

[0103] in, For a moment The physical reference temperature of the winding hot spot is calculated based on the IEC standard thermal circuit model. This is a physical temperature calculation function built based on the IEC standard thermal circuit model. It is used to calculate the physical reference temperature of the winding hot spot based on the equivalent load current, ambient temperature, and pre-configured thermal characteristic parameters of the distribution transformer. For a moment Input the equivalent load current of the physical temperature calculation function, wherein the equivalent load current is determined based on the effective value of the three-phase current or the load factor; For a moment The ambient temperature; the thermal characteristic parameters include at least one of rated current, rated loss, temperature rise parameter and thermal time constant.

[0104] When using the recursive IEC thermal circuit calculation, the physical temperature calculation function also updates the physical reference temperature of the winding hot spot at the current moment based on the oil temperature state and winding thermal state at the previous sampling time.

[0105] Residual labels are constructed based on winding hot spot temperature labels and physical reference temperatures: .

[0106] With the input feature tensor As input to the model, with the residual labels To supervise the labeling, train the black-box deep learning part. Make its output satisfy: .

[0107] in, For a moment The hot spot temperature residual correction term output by the black-box deep learning model is used to compensate for the prediction deviation between the physical reference temperature calculated by the IEC physical thermal circuit model and the winding hot spot temperature label. For black-box deep learning models, it is a non-linear mapping function; For time The input feature tensor constructed for the target prediction time; For a moment The corresponding residual label is determined by the difference between the winding hot spot temperature label and the physical reference temperature output by the IEC physical thermal circuit model.

[0108] The training loss function is: .

[0109] in, N The number of training samples. For network parameter set, This is the regularization coefficient.

[0110] During training, multi-physics field coupled simulation data is used for pre-training, and then measured data or temperature rise test data of pilot distribution transformers with fiber optic temperature measurement are used for parameter fine-tuning to adapt the model parameters to actual operating noise and individual equipment differences.

[0111] This application does not directly use during training. Instead of using the black-box network as the output target, the physical reference temperature at the same moment is first calculated using the IEC physical thermal circuit model, and then residual labels are constructed; For input, To supervise the training of the black-box deep learning part with labeled input, the residual correction term of the network output must satisfy... The corresponding training loss function is: Through this training method, the IEC model provides a temperature benchmark, and the deep learning network only learns the deviation between the physical model and the real hot spot temperature, reducing the training sample size requirement and improving the physical consistency of the model output.

[0112] Step 3: Based on the real-time hot spot temperature prediction value, the insulation parameters in the static ledger data, and the preset temperature limit, calculate the thermal risk assessment parameters of the distribution transformer. The thermal risk assessment parameters include the relative aging rate, the remaining safe operating time, and the safety margin of the bubble generation temperature.

[0113] Based on the real-time hotspot temperature prediction and the insulation parameters in the static ledger data, the insulation life loss is calculated. Specifically, the insulation aging reaction rate model (i.e., the modified Arrhenius reaction rate equation) is used to calculate the relative aging rate, as shown in the following formula: .

[0114] in, For a moment The relative aging rate represents the multiple of the insulation aging rate at the current hot spot temperature relative to the reference hot spot temperature; For a moment Real-time hotspot temperature prediction values, in °C; These are insulation aging characteristic parameters, the values ​​of which are determined by the type of insulation material, heat resistance class, or test calibration results; The reference hotspot temperature represents the relative aging rate. The corresponding hotspot temperature at that time; This represents the temperature shift from Celsius temperature to thermodynamic temperature.

[0115] Step four: Based on the real-time hotspot temperature prediction values ​​obtained in step two and the thermal risk assessment parameters obtained in step three, and combined with dynamic operation monitoring data and static ledger data, construct thermal risk indicators, electrical stress indicators, and equipment health indicators to form a distribution transformer state feature vector; compare the state feature vector with a preset threshold set to obtain the risk state identification result of the distribution transformer.

[0116] Based on the real-time hotspot temperature prediction, load operation data, and equipment ledger data, thermal risk indicators, electrical stress indicators, and equipment health indicators are calculated to form a distribution transformer state feature vector, which is then compared with a preset threshold set, specifically including: Based on the load operation data, meteorological environment data, and static ledger data within the current time and subsequent prediction window, the hot spot temperature prediction model constructed in step two is used to recursively obtain the winding hot spot temperature prediction sequence for subsequent time periods; the time difference between the moment when the winding hot spot temperature prediction sequence first reaches the preset safe temperature limit and the current time is determined as the remaining safe operating time.

[0117] Define preset safe temperature limits: .

[0118] Then, the remaining safe operating time is determined based on the future hotspot temperature prediction sequence: .

[0119] In engineering implementation, it can be calculated according to discrete sampling times as follows: .

[0120] Among them, if in the prediction window If no temperature exceeds the limit within the prediction window, output "Temperature limit not reached within the prediction window", or mark it as: .

[0121] in, For a moment The remaining safe operating time, representing the time from time [time]. The time interval between the start and the first time the predicted winding hot spot temperature reaches the preset safe temperature limit; For a moment Predicted values ​​for winding hot spot temperature; For self-time The predicted duration of the start; Preset safe temperature limits; This is the upper limit of the winding hot spot temperature determined according to the insulation heat resistance class, operating procedures, or test calibration results; θ b The preset bubble generation temperature; The sampling period; k For predicting the step size index; K This represents the maximum number of prediction steps within the prediction window.

[0122] For predicting hotspot temperatures in the future, the hotspot temperature prediction model constructed in step two is used to recursively obtain the prediction input feature tensor formed by subsequent load operation data and meteorological environment data. When subsequent load operation data and meteorological environment data are unavailable, the prediction input feature tensor can be generated based on the assumption that the current load operation conditions and current meteorological environment conditions continue.

[0123] Based on the real-time hotspot temperature prediction value Relative aging rate Remaining safe operating time and the safety margin of temperature for bubble generation Construct a thermal risk index vector: .

[0124] The hotspot temperature safety margin is as follows: .

[0125] In the formula, The preset bubble generation temperature is determined based on the type of insulating oil, water content, insulation class, or test calibration results.

[0126] The load factor is calculated based on the three-phase current, three-phase voltage, and rated parameters of the distribution transformer from the dynamic operation monitoring data. and three-phase imbalance ; Obtain or calculate zero-sequence current and voltage harmonic total distortion Construct an electrical stress index vector: .

[0127] Among them, load rate Calculate using the following formula: .

[0128] .

[0129] .

[0130] Three-phase imbalance Formula calculation: .

[0131] in , , These are the effective values ​​of the three-phase currents. The rated capacity of the distribution transformer. Rated line voltage, This is the rated current.

[0132] When the monitoring terminal directly outputs zero-sequence current For the operation monitoring terminal at time The effective value of the output zero-sequence current; when the operation monitoring terminal provides three-phase current phasors, the zero-sequence current is calculated according to the following formula: .

[0133] .

[0134] in, For a moment The zero-sequence current phasor; For a moment The effective value of the zero-sequence current; , and They are time points The A-phase, B-phase, and C-phase current phasors; |·| represents the magnitude of the phasor.

[0135] When the monitoring terminal collects the neutral line current At that time, the zero-sequence current can be determined according to... Sure.

[0136] When the monitoring terminal directly outputs the total harmonic distortion rate of each phase voltage, the maximum value among the three-phase voltage harmonic distortion rates is used as... When the monitoring terminal provides voltage harmonic components of each order, the total harmonic distortion rate of each phase voltage is calculated according to the following formula: .

[0137] .

[0138] in, For a moment The p Total harmonic distortion of phase voltage; p For the phase index, the value is... a , b or c ; For a moment No. p The fundamental effective value of the phase voltage; For a moment No. p Xiangdi q Effective value of subharmonic voltage;q For harmonic order; Q The highest harmonic order used in the calculation of total harmonic distortion; For a moment The maximum value of the total harmonic distortion rate of the three-phase voltage is used to characterize the voltage harmonic stress of the most severe phase under the current operating condition.

[0139] The risk status identification results include status identifier, over-limit indicator name, indicator value, corresponding threshold, over-limit amplitude, remaining safe operating time, and output timestamp.

[0140] Optionally, the preset threshold set can be configured based on the equipment insulation level, operating procedures, enterprise operation and maintenance requirements, or test calibration results. The preset threshold set is used to limit the allowable range of each indicator in the state feature vector and is not used as a standalone manual judgment rule.

[0141] In one implementation, the state feature vector is: .

[0142] in, This is a real-time hotspot temperature prediction value. This refers to the relative aging rate. For the remaining safe operating time, To ensure a safety margin for hotspot temperatures, For load rate, For three-phase imbalance, It is the zero-sequence current. The total harmonic distortion of voltage. For equipment health indicators.

[0143] The state feature vector With preset threshold set Comparison is performed. When any indicator in the state feature vector reaches the corresponding threshold vector... When the limit condition is met, the system outputs the corresponding status flag. It also outputs the name of the over-limit indicator, the over-limit value, and the remaining safe operating time.

[0144] In one application example, the status identifier can be configured to display four status categories: light concern, heavy load monitoring, severe overload, and emergency overload. This display status is used for the operation and maintenance interface and does not limit the source, number, or handling method of the thresholds in the status identification method of this application.

[0145] This application presents a method for identifying heavy overload risks in distribution transformers based on physical information deep learning. This method integrates heterogeneous data from multiple sources, including equipment ledgers, real-time operation data, and meteorological data, to construct a gray-box prediction model based on "physical baseline + residual correction." It utilizes Attention-LSTM to capture the effects of thermal inertia and environmental nonlinearity, achieving precise soft measurement of winding hotspot temperatures. Furthermore, it combines insulation aging mechanisms to calculate dynamic thermal setpoints and lifespan losses, establishing a four-level comprehensive assessment system encompassing thermal risk, electrical stress, and equipment health. Through physical information residual learning, Arrhenius lifespan calculation, and multi-dimensional index comparison, it achieves a shift from single current monitoring to a comprehensive "thermal-electrical-lifespan" assessment, improving the distribution transformer's heavy overload perception capability and proactive maintenance level under complex operating conditions.

[0146] Please refer to Figure 2 , Figure 2 This is a flowchart illustrating the identification of heavy overload risk in distribution transformers based on physical information deep learning in some embodiments of this application. This method is used in power monitoring terminals or cloud servers. The comprehensive judgment method for heavy overload of distribution transformers based on physical information deep learning includes the following steps: S11. Collect static data from the PMS ledger of the distribution transformer, dynamic data from metering automation, and meteorological environmental data.

[0147] S12. Perform capacity verification and cleaning, missing value imputation and time granularity alignment preprocessing on multi-source heterogeneous data.

[0148] S13. Input the feature vector into the physical information residual learning model and calculate the IEC physical reference temperature and deep learning residual correction term in parallel.

[0149] S14, superimposed with physical reference and residual correction terms, outputs accurate soft measurement values ​​of hot spot temperature of distribution transformer windings.

[0150] S15. Calculate the dynamic thermal setpoint, remaining safe operating time, and insulation life loss by combining insulation type parameters.

[0151] S16. Construct a comprehensive assessment index system that includes thermal risk, electrical stress, and equipment health for condition assessment.

[0152] S17. Based on the assessment results, trigger the four-level early warning mechanism (Level I, II, III, and IV) and output differentiated operation and maintenance decision suggestions.

[0153] It should be noted that the core of this method lies in breaking through the limitations of traditional overload judgments that rely solely on current thresholds. Steps S13 to S14 utilize a "physical information residual learning" strategy, employing the IEC physical model to provide a baseline temperature, ensuring physical consistency in data-sparse regions. Simultaneously, a deep learning model captures nonlinear temperature deviations (residual terms) caused by environmental changes (such as sunlight and wind speed) and equipment aging (such as sludge deposition). This "gray box" modeling approach effectively solves the problem of prediction inaccuracies caused by parameter drift in traditional physical models. Step S15 further extends beyond simple temperature monitoring to the assessment of "lifespan" and "capacity," calculating how long the transformer can safely operate under current conditions (remaining safe operating time), thus achieving quantitative control over overload risks.

[0154] Please refer to Figure 3 , Figure 3 This is a flowchart illustrating multi-source heterogeneous data acquisition for distribution transformers, provided in an embodiment of this application. According to this embodiment, the acquisition of static PMS ledger data, dynamic metering automation data, and meteorological environmental data from the distribution transformer specifically includes: S21. Connect to the PMS asset management system to obtain static ledger data of distribution transformers, including equipment model, core material, winding structure, insulation heat resistance level, transformer weight, total oil volume and commissioning date.

[0155] S22. Connect to the operation monitoring data source and collect dynamic operation monitoring data of the distribution transformer. The operation monitoring data source includes at least one of the following: metering automation system, electricity consumption information collection system, distribution automation system, smart meter, distribution area integration terminal, or distribution transformer monitoring terminal. The dynamic operation monitoring data includes three-phase current, three-phase voltage, zero-sequence current, and total voltage harmonic distortion rate. Load rate and three-phase imbalance are not required to be directly collected data items, but are calculated from three-phase current, three-phase voltage, and rated parameters of the distribution transformer. When the operation monitoring data source has already output the load rate or three-phase imbalance, it is used as verification data or backup input data.

[0156] S23. Access meteorological environmental data sources to obtain environmental data of the area where the distribution transformer is located. The meteorological environmental data sources include meteorological station interfaces, third-party meteorological platform interfaces, regional micro-meteorological stations, local temperature and humidity sensors, solar radiation sensors, or wind speed sensors. The environmental data includes ambient temperature, solar radiation intensity, wind speed, and humidity.

[0157] S24. The static ledger data, dynamic operation monitoring data, and meteorological environmental data mentioned above are aggregated to construct a multi-source heterogeneous raw dataset for subsequent processing. Both the operation monitoring data source and the meteorological environmental data source can be implemented using existing data systems or existing acquisition terminals. The operation monitoring data source provides electrical quantity data during the operation of the distribution transformer. Its specific form is not limited to a single metering automation system; it can also be an electricity consumption information acquisition system, a distribution automation system, a smart meter, a distribution area fusion terminal, or a distribution transformer monitoring terminal. The meteorological environmental data source provides the external heat dissipation boundary conditions of the distribution transformer. Its specific form is not limited to a single meteorological data service interface; it can also be a meteorological station interface, a third-party meteorological platform interface, a regional micro-meteorological station, a local environmental sensor, a solar radiation sensor, or a wind speed sensor. The above data sources only serve as data acquisition channels; the core processing steps involve time alignment, feature encoding, normalization, and time-series tensor construction of the multi-source heterogeneous data.

[0158] It should be noted that the comprehensiveness of the data source is fundamental to ensuring the accuracy of the analysis. Static ledger data (S21) determines the basic thermal characteristics of the transformer (such as heat capacity and rated temperature rise); dynamic operating data (S22) includes the load factor square (…). The direct correlation between copper loss and heat generation is a core characteristic of heat source input; meteorological environmental data (S23) constitutes the boundary conditions for heat dissipation. In particular, the introduction of "solar radiation intensity" and "wind speed" can effectively correct the overheating risk assessment of outdoor box-type transformers under high temperature and high radiation in summer.

[0159] Please refer to Figure 4 , Figure 4 This is a flowchart illustrating a comprehensive assessment index system for condition evaluation provided in this application embodiment. The construction of the comprehensive assessment index system, which includes thermal risk, electrical stress, and equipment health, specifically includes: S31. Calculate the thermal risk index vector, which includes the predicted value of the winding hot spot temperature, the relative aging rate, the remaining safe operating time, and the safety margin of the bubble generation temperature; wherein, the safety margin of the bubble generation temperature is determined based on the difference between the preset bubble generation temperature and the predicted value of the winding hot spot temperature.

[0160] Specifically, the safety margin for bubble generation temperature is calculated using the following formula: .

[0161] in, For a moment The safety margin of the temperature for bubble generation; The preset bubble generation temperature can be determined based on the type of insulating oil, water content, insulation class, or test calibration results. For a moment Predicted values ​​for winding hot spot temperature.

[0162] when When >0, it indicates that the winding hot spot temperature is lower than the bubble formation temperature; when When the value is ≤0, it indicates that the winding hot spot temperature has reached or exceeded the bubble formation temperature. This yields the thermal risk index vector: .

[0163] in, For thermal risk index vectors, This refers to the relative aging rate. This represents the remaining safe operating time.

[0164] S32. Quantify electrical stress indicators. Based on the three-phase current, three-phase voltage and rated parameters of the distribution transformer in the dynamic operation monitoring data, calculate the current load rate and three-phase unbalance, and construct an electrical stress indicator vector by combining the zero-sequence current and voltage harmonic total distortion rate.

[0165] Specifically, set time The effective values ​​of the three-phase currents are respectively , , The rated capacity of the distribution transformer is The rated line voltage is Then the rated current is: .

[0166] The average three-phase current is: .

[0167] The load factor is calculated using the following formula: .

[0168] When the three-phase apparent power can be obtained , , At the same time, the load factor can also be calculated using the following formula: .

[0169] The three-phase current unbalance is calculated using the following formula: .

[0170] in, For a moment load rate, For a moment The three-phase current imbalance. Based on this, an electrical stress index vector is constructed: .

[0171] in, It is the zero-sequence current. This represents the total harmonic distortion rate of the voltage.

[0172] S33. Determine the equipment health indicators, calculate the equipment health correction factor based on the equipment's service life, historical overheating frequency, and historical defect frequency, and use the equipment health correction factor as the equipment health component in the state feature vector.

[0173] Specifically, let the service life of the equipment be... Design life is The number of historical overheating events within the statistical period is The number of historical defects is The number of serious historical defects is First, normalize each indicator: .

[0174] .

[0175] .

[0176] .

[0177] in, , , These are the preset normalized baseline values ​​for overheating cycles, defect cycles, and severe defect cycles, respectively.

[0178] Equipment health indicators are calculated using the following formula: .

[0179] The equipment health correction factor is obtained by the following formula: .

[0180] in, For equipment health correction factors, This is a correction factor. The larger the value, the higher the degree of equipment aging or defect accumulation, and the greater the equipment health correction factor. The equipment health correction factor is used to characterize the impact of equipment health status on the risk of heavy overload.

[0181] S34, combined with thermal risk indicators, electrical stress indicators and equipment health indicators, forms the state characteristic vector of the distribution transformer.

[0182] Specifically, the state feature vector is constructed using the following formula: .

[0183] in, For a moment The state feature vector; This is a vector of thermal risk indicators; This is the electrical stress index vector; For equipment health indicators.

[0184] In one implementation: .

[0185] .

[0186] Therefore, the state feature vector can be represented as: .

[0187] The above processing involves combining indicators without weighted summation. Each indicator retains its original physical meaning and is compared with its corresponding threshold.

[0188] S35. Compare the state feature vector with the preset threshold set item by item to determine the index that meets the limit condition.

[0189] The preset threshold set is represented as follows: .

[0190] in, This is the threshold vector corresponding to the j-th state identifier. The state feature vector... Each indicator and its corresponding threshold vector Compare item by item.

[0191] For hot spot temperature, relative aging rate, load rate, three-phase imbalance, zero-sequence current, total voltage harmonic distortion, and equipment health indicators, the upper limit comparison method is used: .

[0192] For the remaining safe operating time and the safety margin of bubble generation temperature, a lower limit comparison method is used: .

[0193] in, The first eigenvector in the state feature vector One indicator, Threshold vector The threshold corresponding to the indicator is set. When at least one indicator reaches the corresponding threshold condition, the corresponding status identifier is output, and the name of the indicator that exceeded the limit and the magnitude of the exceeded limit are recorded.

[0194] S36. The status recognition module outputs the status recognition result, which includes the status identifier, the name of the over-limit indicator, the indicator value, the corresponding threshold, the over-limit amplitude, the remaining safe running time, and the timestamp.

[0195] Specifically, the state recognition module is a program module deployed in a power monitoring terminal, edge computing device, or cloud server, and is executed by a processor. The state recognition module receives state feature vectors. and preset threshold set and output the state recognition result. : .

[0196] in, To match the preset threshold vector Corresponding status indicator; The set of names of indicators that exceed the limit conditions; This refers to the current or predicted value of the indicator corresponding to the out-of-limit indicator. For the corresponding threshold; This exceeds the limit. Remaining safe operating time; This is the output time.

[0197] For indicators with upper limit constraints, the over-limit value is calculated using the following formula: .

[0198] For indicators that employ lower limit constraints, the exceedance value is calculated using the following formula: .

[0199] in, State feature vector The Middle One indicator, Threshold vector The threshold corresponding to this indicator. When When ≥0, it means that the indicator has reached the corresponding limit condition.

[0200] When multiple indicators reach the limit conditions at the same time, the status recognition module outputs the names of all over-limit indicators and their over-limit values; when multiple status indicators meet the conditions at the same time, it outputs the status indicator with the highest preset priority, while retaining the information of other over-limit indicators that meet the conditions.

[0201] It should be noted that the comprehensive assessment system embodies a multi-dimensional evaluation concept integrating "thermal, electrical, and lifespan." In S31, the thermal risk index is not a single temperature value, but rather a combination of the predicted winding hot spot temperature, relative aging rate, remaining safe operating time, and the safety margin for bubble generation temperature. The safety margin for bubble generation temperature is the preset bubble generation temperature. Predicted values ​​of winding hot spot temperature The difference. This difference characterizes the remaining temperature margin between the current hotspot temperature and the bubble generation boundary. Preset bubble generation temperature. The threshold can be determined based on the type of insulating oil, water content, insulation class, or test calibration results. S33 introduces an equipment health correction factor, which automatically lowers the warning threshold for older equipment with a long service life or a history of overheating, reflecting the principle of differentiated management. S35's four-level warning system (I, II, III, IV) transforms rigid numerical indicators into visualized risk levels. For example: Level IV warning (mild overload): only normal inspection is required; Level III warning (heavy load monitoring): business expansion applications need to be suspended; Level II warning (severe overload): load transfer needs to be prepared; Level I warning (emergency overload): load disconnection or shutdown is required immediately. This tiered mechanism guides maintenance personnel to take precise intervention measures, avoiding overly conservative maintenance or risky operation.

[0202] The technical solutions in this application will be clearly and completely described below with reference to a specific embodiment.

[0203] One oil-immersed distribution transformer within the jurisdiction of a power supply bureau was selected as the subject of the study.

[0204] Device object: Model S13-M-400 / 10, rated capacity 400kVA, rated current The insulation system is Class A (heat resistant to 105℃), and the reference hot spot temperature is 110℃.

[0205] Equipment status: Commissioned in 2010, with a service life of 16 years. Inspection records show moderate dust accumulation on the radiator surface.

[0206] Environmental conditions: Time set to 14:00 on August 15th (typical high-temperature period in summer), ambient temperature... Solar radiation intensity (Strong sunlight, directly hitting the fuel tank), wind speed (Still wind).

[0207] Electrical operation: Three-phase current monitoring value , , Calculate the average load current. .

[0208] Data acquisition and preprocessing: The system receives the aforementioned multi-source heterogeneous data and first performs alignment and cleaning. The current load rate is then calculated. K : .

[0209] Construct the current moment Standardized feature vectors (After normalization): .

[0210] Hotspot temperature soft measurement based on physical information residual learning calls the physical information residual learning model and executes the following parallel computation and superposition logic: .

[0211] Physical baseline calculation (white box) ): Using the IEC 60076-7 standard thermal circuit model, input K=1.2 and Because the IEC model only considers standard heat dissipation conditions and ignores solar radiation ( Due to the influence of dust accumulation and aging (Age), the calculated baseline hotspot temperature is relatively low. .

[0212] Residual correction calculation (black box) ): eigenvectors Input a pre-trained Attention-LSTM network. The model identifies and captures the high volatility of historical loads through the attention mechanism. The nonlinear characteristics of =900 (fuel tank temperature rise due to strong sunlight) and Age=16 (increased thermal resistance) are used to output a positive temperature correction value: .

[0213] The aforementioned residual correction term is the nonlinear mapping output of the deep learning model based on multidimensional input features. From a physical interpretability perspective, this value mainly consists of three parts: firstly, strong solar radiation ( The additional temperature rise is approximately 8.4℃ due to three factors: first, the increased thermal resistance caused by the equipment's service life (16 years) and radiator dust accumulation, resulting in a temperature rise of approximately 4℃; and second, the computational residual caused by the nonlinear characteristics of oil flow under heavy overload, resulting in a temperature rise of approximately 1.4℃. This "physical attribution" verifies that the corrective role of the deep learning model in the "gray box" architecture is consistent with actual physical laws.

[0214] Superimposed output: .

[0215] Note: This result indicates that although the physical model considers the temperature safe ( However, the actual temperature has exceeded the standard due to environmental and aging factors.

[0216] Insulation life loss and overload capacity assessment: Based on predicted hotspot temperatures The relative aging rate is calculated using the insulation aging reaction rate model: .

[0217] In this embodiment, the transformer adopts an oil-paper insulation structure, taking... , , .when hour: .

[0218] This result indicates that 1 hour of operation under the current conditions corresponds to approximately 1.88 hours of insulation aging loss at the baseline hot spot temperature.

[0219] A comprehensive assessment and tiered early warning system is constructed, and rules are matched accordingly. Electrical stress: K =1.25(1.1< K <1.3 interval).

[0220] Heat risk: (belong (Interval).

[0221] Analysis logic: If we only look at the load rate K This should trigger a Level III warning (overload monitoring).

[0222] But look at the hotspot temperature It has reached the threshold of Level II warning (severe overload, long-term emergency response limit).

[0223] Based on the principle of "maximizing risk", the system ultimately determined it to be a Level II warning.

[0224] Decision output: Warning level: Level II.

[0225] Maintenance recommendation: "Severe thermal overload detected, insulation aging rate 1.88 times. It is recommended to transfer 20% of the load via the tie switch within 2 hours, or turn on the forced air cooling device."

[0226] Comparative analysis with traditional methods: To verify the advancement of the method in this application, its performance was compared with that of the existing "traditional current threshold method" and "standard IEC model method" in the same scenario. The results are shown in Table 1.

[0227] Table 1. Comparison between this application and traditional methods

[0228] This application first constructs a multi-source heterogeneous standardized feature vector by integrating static data from equipment ledgers, dynamic data from operation monitoring, and meteorological environmental data of the distribution transformer. Second, it employs a physical information residual learning strategy, using the IEC standard thermal circuit model to calculate the physical reference temperature, and predicts nonlinear residual correction terms through a long short-term memory network with an attention mechanism. Accurate soft measurement of winding hotspot temperatures is achieved by superimposing the reference temperature and residual. Next, based on the predicted hotspot temperature and insulation type parameters, the real-time overload capacity, remaining safe operating time, and insulation life loss of the transformer are dynamically calculated. Finally, thermal risk indicators, electrical stress indicators, and equipment health indicators are calculated to form a distribution transformer state feature vector. This state feature vector is then compared with a preset threshold set to output state identifiers, over-limit indicators, and over-limit amplitudes. This application effectively solves the prediction bias problem caused by parameter drift and environmental simplification in traditional models, realizing a shift from single current monitoring to a comprehensive "thermal-electrical-life" assessment, significantly improving the perception accuracy and proactive maintenance capabilities of distribution transformers under complex operating conditions for heavy overload risks.

[0229] Based on the same inventive concept, this application also provides a distribution transformer overload risk identification system for implementing the aforementioned distribution transformer overload risk identification method. The solution provided by this system is similar to the implementation scheme described in the above method. Therefore, the specific limitations of one or more distribution transformer overload risk identification system embodiments provided below can be found in the limitations of the distribution transformer overload risk identification method described above, and will not be repeated here.

[0230] In one exemplary embodiment, a system for identifying the risk of heavy overload of a distribution transformer is provided, including: a data acquisition unit, a model building unit, a prediction unit, an index calculation unit, and a risk status identification unit.

[0231] The data acquisition unit is used to acquire multi-source heterogeneous data of the distribution transformer; the multi-source heterogeneous data includes static ledger data, dynamic operation monitoring data, and environmental data.

[0232] The model building unit is used to build a hotspot temperature prediction model based on physical information residual learning. The hotspot temperature prediction model includes a white-box physical model and a black-box deep learning model. The white-box physical model is used to calculate the physical reference temperature based on the IEC standard thermal circuit model, and the black-box deep learning model is used to obtain the residual correction term of the physical reference temperature based on a deep neural network.

[0233] The prediction unit is used to predict the hot spot temperature of the distribution transformer based on the hot spot temperature prediction model using multi-source heterogeneous data.

[0234] The indicator calculation unit is used to calculate multidimensional heavy overload risk indicators based on the hot spot temperature prediction value and combined with static ledger data and dynamic operation monitoring data in multi-source heterogeneous data. The multidimensional heavy overload risk indicators include: thermal risk indicators, electrical stress indicators and equipment health indicators.

[0235] The risk status identification unit is used to compare the multi-dimensional heavy overload risk index with the preset threshold to obtain the risk status identification result of the heavy overload of the distribution transformer.

[0236] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 5 As shown, the computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores data on the risk identification of overloaded distribution transformers. The I / O interfaces are used for information exchange between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a method for identifying the risk of overloaded distribution transformers.

[0237] Those skilled in the art will understand that Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0238] In one exemplary embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0239] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0240] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0241] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0242] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0243] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0244] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0245] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for identifying the risk of heavy overload in a distribution transformer, characterized in that, include: Acquire multi-source heterogeneous data of the distribution transformer; the multi-source heterogeneous data includes static ledger data, dynamic operation monitoring data, and environmental data; A hotspot temperature prediction model based on physical information residual learning is constructed, which includes a white-box physical model and a black-box deep learning model. The white-box physical model is used to calculate the physical reference temperature based on the IEC standard thermal circuit model, and the black-box deep learning model is used to obtain the residual correction term of the physical reference temperature based on the deep neural network. The hotspot temperature prediction value of the distribution transformer is obtained by using multi-source heterogeneous data based on the hotspot temperature prediction model. Based on the hotspot temperature prediction and combined with static ledger data and dynamic operation monitoring data from multi-source heterogeneous data, a multi-dimensional heavy overload risk index is calculated. The multi-dimensional heavy overload risk index includes: thermal risk index, electrical stress index and equipment health index. By comparing the multidimensional heavy overload risk index with the preset threshold, the risk status identification result of the heavy overload of the distribution transformer is obtained.

2. The method for identifying the risk of heavy overload in a distribution transformer according to claim 1, characterized in that, The black-box deep learning model employs a long short-term memory network based on an attention mechanism, comprising an input layer, an LSTM temporal coding layer, an attention weight calculation layer, and a residual output layer. The LSTM temporal coding layer encodes the input feature tensor in chronological order to obtain the hidden layer state at each historical time step; The attention weight calculation layer calculates the attention score and attention weight based on the hidden layer state of each historical time step, and performs weighted aggregation of the hidden layer state of the historical time step based on the attention weight to obtain the context vector; The residual output layer concatenates the context vector with the single-step fused feature vector at the current time step to form the residual mapping input, and outputs the residual correction term through a fully connected mapping.

3. The method for identifying the risk of heavy overload of a distribution transformer according to claim 1, characterized in that, The thermal risk indicators include at least one of the following: hot spot temperature prediction, relative aging rate, remaining safe operating time, and safety margin for bubble generation temperature. The formula for calculating the relative aging rate is: ； in, For a moment The relative aging rate, These are insulation aging characteristic parameters. The reference hotspot temperature represents the relative aging rate. The corresponding hot spot temperature at that time This represents the temperature shift during the conversion from Celsius temperature to thermodynamic temperature. For a moment Hotspot temperature prediction values; The formula for calculating the safety margin of the bubble generation temperature is as follows: ； in, For a moment The temperature safety margin for bubble generation The preset temperature for bubble generation.

4. The method for identifying the risk of heavy overload of a distribution transformer according to claim 1, characterized in that, The electrical stress indicators include at least one of the following: load factor, three-phase unbalance, zero-sequence current, and voltage harmonic distortion rate; The formula for calculating the load rate is: ； ； ； in, For a moment load rate, For a moment The three-phase average current, Rated current, , , They are time points The effective value of the three-phase current, The rated capacity of the distribution transformer This is the rated line voltage; The formula for calculating the three-phase unbalance is: ； in, For a moment The three-phase imbalance.

5. The method for identifying the risk of heavy overload of a distribution transformer according to claim 1, characterized in that, The formula for calculating the equipment health index is as follows: ； ； ； ； ； in, For equipment health indicators, , , , All are weighting coefficients. This is the normalized value of the equipment's service life. This is the normalized value of the number of historical overheating events within the statistical period. This is the normalized value of the number of historical defects within the statistical period. This is the normalized value of the number of historical severe defects within the statistical period. For the service life of the equipment, For design life, To count the number of historical overheating events within the statistical period, The preset normalized baseline value for the number of superheating cycles is used. For the number of historical defects, The default defect count is the normalized baseline value. This refers to the number of serious defects in history. This is the preset normalized baseline value for the number of severe defects.

6. The method for identifying the risk of heavy overload of a distribution transformer according to claim 1, characterized in that, The loss function of the hotspot temperature prediction model is: ； in, For loss function, The number of training samples. For the first The input feature tensor of each training sample This refers to the residual correction term output by the black-box deep learning model. For the first The residual labels of each training sample. λ Θ is the regularization coefficient; Θ is the set of network parameters.

7. A system for identifying the risk of heavy overload in a distribution transformer, characterized in that, include: The data acquisition unit is used to acquire multi-source heterogeneous data of the distribution transformer; the multi-source heterogeneous data includes static ledger data, dynamic operation monitoring data, and environmental data. The model building unit is used to build a hotspot temperature prediction model based on physical information residual learning. The hotspot temperature prediction model includes a white-box physical model and a black-box deep learning model. The white-box physical model is used to calculate the physical reference temperature based on the IEC standard thermal circuit model, and the black-box deep learning model is used to obtain the residual correction term of the physical reference temperature based on the deep neural network. The prediction unit is used to predict the hot spot temperature of the distribution transformer based on the hot spot temperature prediction model from multi-source heterogeneous data. The indicator calculation unit is used to calculate multi-dimensional heavy overload risk indicators based on the hot spot temperature prediction value and combined with static ledger data and dynamic operation monitoring data in multi-source heterogeneous data. The multi-dimensional heavy overload risk indicators include: thermal risk indicators, electrical stress indicators and equipment health indicators. The risk status identification unit is used to compare the multi-dimensional heavy overload risk indicators with preset thresholds to obtain the risk status identification result of the heavy overload of the distribution transformer.

8. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the method for identifying the heavy overload risk of a distribution transformer according to any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the method for identifying the heavy overload risk of a distribution transformer as described in any one of claims 1-6.

10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the method for identifying the heavy overload risk of a distribution transformer as described in any one of claims 1-6.

Images