A method for intelligent monitoring of substations

Abstract

This invention relates to the field of power system automation and intelligent monitoring technology, and discloses an intelligent monitoring method for substations. The method includes: acquiring historical transformer operating data, real-time meteorological data of the substation, and future task planning information; performing time alignment and missing value imputation to construct a multi-dimensional time-series input matrix; inputting this matrix into an insulation aging state coupling evaluation model based on a physical information neural network to generate an insulation aging state index; and triggering graded early warnings based on the comparison result of this index with a preset threshold. This model integrates the thermo-electrical-humidity multi-physics coupling equation and attention-enhanced LSTM to achieve high-precision, low-false-alarm online monitoring. This invention significantly reduces the false alarm rate and improves the safety and economy of substation operation and maintenance.

Description

Technical Field

[0001] This invention belongs to the field of power system automation and intelligent monitoring technology, and specifically relates to an intelligent monitoring method for substations. Background Technology

[0002] With the deepening of smart grid construction, substations, as key hubs in the power system, directly impact the safety and reliability of the power grid through real-time monitoring and fault early warning capabilities. Currently, substation monitoring systems generally employ state assessment models based on historical data and fixed thresholds to determine the health status of core equipment such as transformers and circuit breakers. While these methods are effective under steady-state conditions, their static modeling paradigm struggles to capture the dynamic coupling between equipment status and external disturbances in complex and ever-changing real-world operating environments. This is especially true under conditions of high load fluctuations or extreme weather, where prediction accuracy significantly decreases.

[0003] Online assessment of transformer insulation aging is one of the core tasks of intelligent monitoring in substations. Traditional methods mainly rely on load current, oil temperature, and partial discharge signals to construct a lifespan degradation model, based on the fundamental assumption that the equipment degradation process is driven solely by internal electrical stress. However, in actual operation, micro-meteorological factors such as ambient temperature, humidity, and wind speed significantly affect the transformer's heat dissipation efficiency and insulation material performance, thereby altering the aging rate. When the load undergoes a sudden change, if the model does not simultaneously incorporate real-time meteorological information, it will lead to misjudgments of the insulation condition. This manifests as underestimating the aging risk under high temperature and humidity conditions, or over-predicting under low temperature and dry conditions, resulting in a persistently high false alarm rate.

[0004] Although some systems attempt to incorporate meteorological data as an auxiliary input, most employ static weighting or offline calibration methods, lacking a millisecond-level response mechanism to both meteorological and load disturbances. Furthermore, fixed-threshold fault diagnosis strategies cannot be dynamically adjusted to adapt to sudden environmental changes, making it difficult to balance sensitivity and robustness. Summary of the Invention

[0005] This invention provides an intelligent monitoring method for substations, aiming to solve the technical problem of high false alarm rates for insulation aging caused by traditional state prediction models neglecting real-time meteorological data when transformer loads change abruptly. Existing systems lack the ability to couple and model the transformer's operating state with dynamic environmental factors, failing to accurately reflect the aging process of insulation materials under real-world conditions. This invention constructs a multi-source heterogeneous data fusion architecture, introduces a three-domain coupling mechanism of meteorology, electrical systems, and thermodynamics, and establishes an insulation aging state assessment model driven by both physical constraints and temporal characteristics, achieving high-precision, low-false-alarm-rate online monitoring of transformer insulation aging status.

[0006] This invention provides a method for intelligent monitoring of substations, comprising:

[0007] Obtain the historical operating data sequence of the transformer, which includes load current, winding temperature, top oil temperature, partial discharge quantity, and dielectric loss factor;

[0008] Acquire real-time meteorological data of the area where the substation is located, including ambient temperature, relative humidity, atmospheric pressure, solar radiation intensity, and wind speed;

[0009] Obtain current task planning information, which includes the planned load curve, maintenance schedule, and scheduling instructions for the next 24 hours;

[0010] The historical operational data sequence, real-time meteorological data, and current task planning information are timestamped and missing values ​​are imputed to form a unified multi-dimensional time series input matrix.

[0011] The multidimensional time-series input matrix is ​​input into the pre-trained insulation aging state coupled evaluation model to generate the insulation aging state index at the current moment.

[0012] Based on the comparison result between the insulation aging state index and the preset threshold, a graded early warning signal is triggered and output to the monitoring terminal.

[0013] As one embodiment of the present invention, the acquisition of the historical operating data sequence of the transformer specifically includes:

[0014] The three-phase load current on the high-voltage and low-voltage sides of the transformer is collected by a current transformer at a sampling frequency of once per second.

[0015] The temperature of the winding hot spot is obtained by an embedded fiber optic temperature sensor, with a spatial resolution of one measuring point per meter and a temporal resolution of once every 10 seconds.

[0016] The top oil temperature is obtained through an oil temperature sensor, with a sampling cycle of once per minute;

[0017] The partial discharge pulse signal in the grounding wire of the bushing end screen is monitored by a high-frequency current sensor, with a bandwidth covering 300 kHz to 150 MHz.

[0018] The dielectric loss factor of the main insulation system is periodically measured using a dielectric loss tester, with a test cycle of once every 4 hours.

[0019] As one embodiment of the present invention, the acquisition of real-time meteorological data of the area where the substation is located specifically includes:

[0020] Ambient temperature is collected by a weather station deployed on top of the substation perimeter wall; the sensor accuracy is ±0.5 degrees Celsius.

[0021] Relative humidity is obtained using a capacitive humidity sensor, with a measurement range of 5% to 95% and an accuracy of ±2%.

[0022] Atmospheric pressure is obtained through a piezoresistive barometric pressure sensor with a range of 800 to 1100 hPa and a resolution better than 0.1 hPa.

[0023] The total solar radiation intensity on a horizontal surface is measured using a silicon photodiode array sensor with a spectral response range of 400 to 1100 nanometers.

[0024] The wind speed at a height of 10 meters above the ground is obtained using an ultrasonic anemometer, with a measurement range of 0 to 60 meters per second, and the starting wind speed is below 0.5 meters per second.

[0025] As one embodiment of the present invention, the timestamp alignment and missing value imputation of historical operational data sequences, real-time meteorological data, and current task planning information specifically includes:

[0026] Using Coordinated Universal Time (UTC) as the base time source, all data streams are synchronized at the nanosecond level.

[0027] The missing winding temperature and top oil temperature data were reconstructed using cubic spline interpolation, with the interpolation interval not exceeding 5 minutes.

[0028] For cases where partial discharge data is missing for more than 30 consecutive seconds, a Markov chain Monte Carlo imputation method based on the statistical distribution of adjacent time periods is adopted.

[0029] All data are resampled to a uniform time grid with a time step of 1 minute, forming a time-series input vector with a dimension of 64.

[0030] As one embodiment of the present invention, the insulation aging state coupled evaluation model is a hybrid architecture based on a physical information neural network. Its input layer receives the multi-dimensional time-series input matrix, the hidden layer includes a thermal-electrical-wet coupled differential equation constraint module and an attention-enhanced long short-term memory network module, and the output layer generates an insulation aging state index.

[0031] The thermo-electric-humidity coupled differential equation constraint module embeds the following physical constraints based on the transformer thermal circuit model, moisture diffusion equation, and electric field distribution theory:

[0032] The temperature rise of the winding is proportional to the square of the load current, and the proportionality coefficient is determined by the cooling method.

[0033] The moisture content in insulating paper dynamically changes with ambient humidity and oil temperature, and the migration rate follows the Arrhenius equation.

[0034] The partial discharge initiation voltage is modulated by the trace water content in the oil and the temperature, and the modulation function is in the form of exponential decay.

[0035] The attention-enhanced long short-term memory network module includes a dual-channel structure. The first channel processes electrical and thermal time-series features, and the second channel processes meteorological and task planning features. The two channels are dynamically weighted in the time step dimension through an additive attention mechanism.

[0036] The insulation aging state index is defined as a dimensionless quantity between 0 and 1. The closer the value is to 1, the more severe the insulation aging. This index is jointly optimized by the backpropagation algorithm and the model parameters and physical constraint weights.

[0037] As one embodiment of the present invention, the training process of the insulation aging state coupling evaluation model includes:

[0038] Collect full lifecycle operation data of no less than 500 transformers of the same type under different geographical and climatic zones and different load modes, and construct a training dataset;

[0039] The measured value of the degree of polymerization of the insulating paper was marked for each transformer, serving as the target label for supervised learning;

[0040] The degree of polymerization is mapped to the insulation aging state index, and the mapping relationship is an exponential decay function. When the degree of polymerization is less than 200, the index value is greater than 0.8.

[0041] A physical residual term is introduced into the loss function. This term is the sum of squared residuals of the thermal-electric-humidity coupling differential equation at the discrete time step, and its weight is adaptively adjusted by Lagrange multipliers.

[0042] A phased training strategy is adopted. In the first phase, only the neural network parameters are trained, and in the second phase, the neural network parameters and the unknown coefficients in the physical equations are jointly optimized.

[0043] As one embodiment of the present invention, triggering a graded early warning signal based on the comparison result of the insulation aging state index and a preset threshold specifically includes:

[0044] The first-level warning threshold is set at 0.6. When the insulation aging status index exceeds this value for the first time, a yellow warning signal is generated, prompting an increase in the frequency of inspections.

[0045] The threshold for a Level 2 warning is set at 0.75. If the index exceeds this value for more than two hours, an orange warning signal will be generated, and it is recommended to adjust the load distribution.

[0046] The threshold for Level 3 early warning is set at 0.85. When the instantaneous value of the index exceeds this value or is higher than 0.8 for 4 consecutive hours, a red early warning signal is generated, triggering the emergency shutdown plan.

[0047] All warning signals are accompanied by a confidence score, which is calculated by the uncertainty estimation module output by the model. The uncertainty estimation is based on the standard deviation of the results of multiple forward propagations during the inference phase using the Monte Carlo dropout method.

[0048] As one embodiment of the present invention, the method further includes an online model update mechanism:

[0049] Once new laboratory test results for the degree of polymerization of insulating paper are received, the sample will be added to the incremental training set.

[0050] The model is fine-tuned using an elastic weight consolidation algorithm to prevent catastrophic forgetting;

[0051] The updated model version needs to undergo cross-validation to ensure that the prediction error on historical data does not exceed 110% of the original model;

[0052] The model update process is completed locally on the edge computing node, and after the update is completed, it is synchronized to the substation monitoring master station.

[0053] As one embodiment of the present invention, the construction of the multidimensional time-series input matrix further includes an abnormal data filtering step:

[0054] For load current data, transient spikes with amplitudes exceeding 150% of the rated current and durations of less than 10 milliseconds are removed.

[0055] For partial discharge data, interference signals with spectra concentrated near the power frequency and its odd harmonics are filtered out;

[0056] For meteorological data, exclude abnormal readings where wind speed changes exceed 10 meters per second and there is no corresponding weather system to support them;

[0057] Anomaly determination is based on a pre-set physical rationality rule base, which includes transformer electromagnetic transient characteristics, meteorological evolution laws, and equipment response delay characteristics.

[0058] As one embodiment of the present invention, the insulation aging state coupled evaluation model is deployed in the substation edge computing unit. This unit uses a domestically produced processor chip, which has a floating-point operation capability of no less than 1 trillion times per second, a memory capacity of no less than 16 gigabytes, and supports model inference latency of less than 500 milliseconds.

[0059] The edge computing unit is connected to the station monitoring system via gigabit Ethernet and uses a power-specific security protocol for encrypted data transmission.

[0060] The model inference results are encapsulated in a structured data format, including timestamps, device identifiers, aging indexes, confidence levels, and ranking of the contributions of major influencing factors.

[0061] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0062] 1. This invention constructs a coupled evaluation model of insulation aging state with physical mechanism constraints by deeply integrating transformer electrical operation data, thermal state parameters and real-time meteorological environment information, which solves the problem of false alarm of insulation aging caused by neglecting the coupling effect of dynamic environment in traditional methods.

[0063] 2. This invention introduces a thermal-electrical-humidity multi-physics coupled differential equation as a hard constraint for the neural network to ensure that the model output conforms to the basic physical laws of transformer insulation aging; at the same time, it adopts an attention mechanism to dynamically quantify the influence weight of meteorological factors on the aging process, so that the model can adapt to the operating environment of different climatic regions.

[0064] 3. The edge deployment architecture and online update mechanism of this invention ensure the accuracy and adaptability of the model during long-term operation, providing reliable technical support for condition-based maintenance of power equipment. Attached Figure Description

[0065] Figure 1 This is a schematic diagram of the overall technical solution architecture of the intelligent monitoring method for substations proposed in this invention;

[0066] Figure 2 This is a schematic diagram of the core principle framework of the insulation aging state coupling evaluation model in this invention;

[0067] Figure 3 This is a flowchart illustrating the logical process framework for multi-source heterogeneous data fusion and time alignment processing in this invention.

[0068] Figure 4 This is a schematic diagram illustrating the interaction principle between the thermo-electric-humidity multiphysics coupled differential equation constraint module and the attention-enhanced LSTM dual-channel structure in this invention.

[0069] Figure 5 This is a logical flowchart of the graded early warning triggering mechanism and confidence assessment in this invention;

[0070] Figure 6 This is a schematic diagram of the multi-level interaction relationship and data flow of the edge computing unit deployment architecture and model online update mechanism in this invention. Detailed Implementation

[0071] Please refer to Figures 1 to 6This invention provides an intelligent monitoring method for substations, aiming to solve the technical problem of high false alarm rates for insulation aging caused by traditional state prediction models neglecting real-time meteorological data when transformer loads change abruptly. Existing systems lack the ability to couple and model the transformer's operating state with dynamic environmental factors, failing to accurately reflect the aging process of insulation materials under real-world conditions. This invention constructs a multi-source heterogeneous data fusion architecture, introduces a meteorological-electrical-thermal three-domain coupling mechanism, and establishes an insulation aging state assessment model driven by both physical constraints and temporal characteristics, achieving high-precision, low-false-alarm-rate online monitoring of transformer insulation aging status.

[0072] The intelligent monitoring method for substations includes the following steps: acquiring historical operating data sequences of transformers; acquiring real-time meteorological data of the area where the substation is located; acquiring current task planning information; performing timestamp alignment and missing value interpolation on the above three types of data to form a unified multi-dimensional time-series input matrix; inputting the matrix into a pre-trained insulation aging state coupled evaluation model to generate the insulation aging state index at the current moment; and triggering a graded early warning signal and outputting it to the monitoring terminal based on the comparison result of the index and a preset threshold.

[0073] In the aforementioned steps, acquiring the historical operating data sequence of the transformer specifically includes: collecting the three-phase load current on the high-voltage and low-voltage sides of the transformer through a current transformer, with a sampling frequency of once per second; acquiring the winding hot spot temperature through an embedded fiber optic temperature sensor, with a spatial resolution of one measuring point per meter and a temporal resolution of once every 10 seconds; acquiring the top oil temperature through an oil temperature sensor, with a sampling period of once per minute; monitoring the partial discharge pulse signal in the bushing end screen grounding wire through a high-frequency current sensor, with a bandwidth covering 300 kHz to 150 MHz; and periodically measuring the dielectric loss factor of the main insulation system through a dielectric loss tester, with a test period of once every 4 hours. All sensors have undergone national metrological certification, and their output signals undergo preliminary filtering and analog-to-digital conversion through the station's synchronous sampling unit. The converted digital signal is time-stamped using the IEEE 1588 precision time protocol and transmitted to the data buffer of the edge computing unit through the station control layer network.

[0074] The acquisition of real-time meteorological data for the substation area specifically includes: collecting ambient temperature data via a meteorological station deployed atop the substation perimeter wall, with a sensor accuracy of ±0.5 degrees Celsius; acquiring relative humidity data via a capacitive humidity sensor, with a measurement range of 5% to 95% and an accuracy of ±2%; acquiring atmospheric pressure data via a piezoresistive barometric pressure sensor, with a range of 800 to 1100 hPa and a resolution better than 0.1 hPa; measuring the total solar radiation intensity on the horizontal surface via a silicon photodiode array sensor, with a spectral response range of 400 to 1100 nanometers; and acquiring wind speed data at a height of 10 meters above the ground via an ultrasonic anemometer, with a measurement range of 0 to 60 meters per second and an initial wind speed below 0.5 meters per second. The meteorological station has a built-in self-calibration module that performs a zero-point drift correction once daily at 2:00 AM and writes the correction coefficients into the metadata header. Meteorological data is packaged and uploaded every minute, with each data packet containing a timestamp, original value, correction value, and sensor health status code.

[0075] Obtaining current task planning information specifically includes: receiving the planned load curve for the next 24 hours through the scheduling automation system interface, which defines the expected active and reactive power for each time period in 15-minute granularity; obtaining maintenance arrangements for the next 72 hours through the production management system interface, including equipment outage plans, test projects, and personnel configuration; and receiving real-time scheduling instructions through the energy management system, including load adjustment commands, voltage control targets, and emergency tripping plans. All task planning information is stored in a structured format, including a unique task identifier, effective time window, operation type, and priority level. Upon receiving new task planning information, the edge computing unit immediately parses its content and updates its local task queue, while simultaneously triggering the rescheduling mechanism of the data fusion module.

[0076] The specific process of aligning historical operational data sequences, real-time meteorological data, and current task planning information with timestamps and imputing missing values ​​is as follows: First, using Coordinated Universal Time (UTC) as the reference time source, all data streams are synchronized at the nanosecond level. The synchronization process is initiated by the station's master clock server. After receiving the synchronization pulse, each data source device resets its internal counter to zero and starts timing, ensuring that the timestamps of all data have a unified starting point.

[0077] Secondly, cubic spline interpolation is used to reconstruct the missing winding temperature and top oil temperature data, with an interpolation interval not exceeding 5 minutes. If more than 5 consecutive time points are missing, the data for that time period is deemed unreliable and marked as invalid, and is not included in subsequent modeling. Thirdly, for cases where partial discharge data is missing for more than 30 consecutive seconds, a Markov chain Monte Carlo imputation method based on the statistical distribution of adjacent time periods is adopted. This method first constructs the conditional probability distribution of partial discharge under normal operating conditions, and then uses the observations before and after the missing data as boundary conditions to generate an imputation sequence that conforms to this distribution. Fourthly, all data is resampled to a unified time grid with a time step of 1 minute, forming a time-series input vector with a dimension of 64. During the resampling process, for high-frequency signals such as load current, a moving average method is used to reduce the frequency; for low-frequency signals such as dielectric loss factor, the last effective value is retained to fill the intermediate gaps.

[0078] After data preprocessing, the multidimensional time-series input matrix is ​​input into a pre-trained insulation aging state coupled evaluation model. This model is a hybrid architecture built on a physical information neural network. Its input layer receives the multidimensional time-series input matrix, its hidden layer includes a thermo-electrical-humidity coupled differential equation constraint module and an attention-enhanced long short-term memory network module, and its output layer generates an insulation aging state index. The thermo-electrical-humidity coupled differential equation constraint module embeds the following physical constraints based on the transformer thermal circuit model, moisture diffusion equation, and electric field distribution theory:

[0079] First, the winding temperature rise is directly proportional to the square of the load current, and the proportionality coefficient is determined by the cooling method. This relationship can be expressed as:

[0080]

[0081] in, For winding temperature, This is the effective value of the load current. For AC resistance, For the winding heat capacity, The top oil temperature, The thermal time constant from the winding to the oil. Indicates time.

[0082] Second, the moisture content in the insulating paper dynamically migrates with varying ambient humidity and oil temperature, and the migration rate follows the Arrhenius equation. Moisture migration flux. Represented as:

[0083]

[0084] in, Pre-exponential factor, For activation energy, The gas constant is Oil temperature This represents the partial pressure of water vapor corresponding to the ambient relative humidity. This represents the equilibrium partial pressure corresponding to dissolved water in the oil.

[0085] Third, the partial discharge initiation voltage is modulated by both the trace water content in the oil and the temperature, and the modulation function is in an exponential decay form:

[0086]

[0087] in, The starting voltage in the dry clean oil, and This is an empirical coefficient. This refers to the trace water content in the oil.

[0088] The attention-enhanced Long Short-Term Memory (LSTM) network module comprises a dual-channel structure. The first channel processes electrical and thermal timing characteristics, including load current, winding temperature, top oil temperature, partial discharge, and dielectric loss factor. The second channel processes meteorological and mission planning characteristics, including ambient temperature, relative humidity, atmospheric pressure, solar radiation intensity, wind speed, planned load curve, and scheduling instructions. The two channels are dynamically weighted along the time step dimension using an additive attention mechanism. Attention Weights The calculation formula is:

[0089]

[0090] in, and The two channels are at time steps respectively. The hidden state, For learnable parameters, and The two channels are at time steps respectively. The hidden state, For time step The total number.

[0091] The insulation aging state index is defined as a dimensionless quantity between 0 and 1, with values ​​closer to 1 indicating more severe insulation aging. This index is used to jointly optimize model parameters and physical constraint weights through a backpropagation algorithm, ensuring that the model output conforms to both data-driven principles and physical interpretability.

[0092] The training process of the insulation aging state coupled assessment model includes: collecting full life cycle operation data of no less than 500 transformers of the same type under different geographical climate zones and different load modes to construct a training dataset; labeling the measured value of insulation paper polymerization degree for each transformer as the target label for supervised learning; mapping the polymerization degree value to the insulation aging state index, with the mapping relationship being an exponential decay function, where the exponential value is greater than 0.8 when the polymerization degree is less than 200; introducing a physical residual term into the loss function, which is the sum of squared residuals of the thermo-electric-humidity coupled differential equation at discrete time steps, with its weight adaptively adjusted by Lagrange multipliers; and adopting a phased training strategy, where the first phase only trains the neural network parameters, and the second phase jointly optimizes the neural network parameters and the unknown coefficients in the physical equation.

[0093] The tiered early warning signals triggered based on the comparison between the insulation aging state index and preset thresholds include: a Level 1 warning threshold of 0.6, where a yellow warning signal is generated when the insulation aging state index exceeds this value for the first time, prompting an increase in inspection frequency; a Level 2 warning threshold of 0.75, where an orange warning signal is generated when the index exceeds this value for more than two hours, suggesting adjustments to load distribution; and a Level 3 warning threshold of 0.85, where a red warning signal is generated when the instantaneous value of the index exceeds this value or remains above 0.8 for four consecutive hours, triggering an emergency shutdown plan. All warning signals are accompanied by a confidence score, which is calculated by the uncertainty estimation module output by the model. The uncertainty estimation is based on the standard deviation of multiple forward propagation results during the inference phase using the Monte Carlo dropout method.

[0094] The method also includes an online model update mechanism: when a new laboratory test result of the degree of polymerization of insulating paper is received, the sample is added to the incremental training set; the model is fine-tuned using an elastic weight consolidation algorithm to prevent catastrophic amnesia; the updated model version needs to be cross-validated to ensure that the prediction error on historical data does not exceed 110% of the original model; the model update process is completed locally on the edge computing node and synchronized to the substation monitoring master station after the update is completed.

[0095] The construction of the multidimensional time-series input matrix also includes an abnormal data filtering step: for load current data, transient spikes with amplitudes exceeding 150% of the rated current and durations of less than 10 milliseconds are removed; for partial discharge data, interference signals with spectra concentrated near the power frequency and its odd harmonics are filtered out; for meteorological data, abnormal readings with wind speed changes exceeding 10 meters per second and no corresponding weather system support are excluded; the anomaly determination is based on a preset physical rationality rule base, which includes the electromagnetic transient characteristics of transformers, meteorological evolution laws, and equipment response delay characteristics.

[0096] The insulation aging state coupled assessment model is deployed in the substation edge computing unit. This unit uses a domestically produced processor chip with a floating-point operation capability of no less than 1 trillion operations per second, a memory capacity of no less than 16 gigabytes, and supports model inference latency of less than 500 milliseconds. The edge computing unit is connected to the substation monitoring system via gigabit Ethernet and uses a power-specific security protocol for encrypted data transmission. The model inference results are encapsulated in a structured data format, including timestamps, equipment identifiers, aging indexes, confidence levels, and ranking of the contributions of major influencing factors.

[0097] The edge computing unit internally includes a data access module, a preprocessing engine, a model inference core, an early warning decision-maker, and a communication interface. The data access module subscribes to the required data streams from the station control layer network and performs preliminary verification and caching. The preprocessing engine performs time alignment, missing value imputation, and anomaly filtering. The model inference core loads the insulation aging state coupled assessment model and performs forward computation. The early warning decision-maker generates early warning commands based on the comparison results between the model output and the threshold. The communication interface encapsulates the early warning signal and auxiliary information and sends it to the monitoring terminal, and receives model update commands from the main station.

[0098] In actual operation, the edge computing unit executes a complete inference process once per minute. Before each inference, the system checks for new task planning information; if so, it immediately updates the relevant fields in the input matrix. The aging index output by the model is recorded in a local database and used for subsequent trend analysis. When the confidence score falls below a preset lower limit (e.g., 0.3), the system automatically marks the inference result as suspicious and triggers a manual review process.

[0099] The method described in this embodiment deeply integrates transformer electrical operation data, thermal state parameters, and real-time meteorological information to construct a coupled assessment model for insulation aging status with physical mechanism constraints. This solves the problem of false alarms due to insulation aging caused by neglecting the coupling effect of dynamic environment in traditional methods. Experiments show that under load change conditions, this invention reduces the false alarm rate of insulation aging from 32% in traditional models to below 7%, and extends the average early warning time by 48 hours, significantly improving the safety and economy of substation operation and maintenance. In addition, the edge deployment architecture and online update mechanism of this invention ensure the accuracy and adaptability of the model in long-term operation, providing reliable technical support for condition-based maintenance of power equipment.

Claims

1. A method for intelligent monitoring of substations, characterized in that, include: Obtain the historical operating data sequence of the transformer, which includes load current, winding temperature, top oil temperature, partial discharge quantity, and dielectric loss factor; Acquire real-time meteorological data of the area where the substation is located, including ambient temperature, relative humidity, atmospheric pressure, solar radiation intensity, and wind speed; Obtain current task planning information, which includes the planned load curve, maintenance schedule, and scheduling instructions for the next 24 hours; The historical operational data sequence, real-time meteorological data, and current task planning information are timestamped and missing values ​​are imputed to form a unified multi-dimensional time series input matrix. The multidimensional time-series input matrix is ​​input into the pre-trained insulation aging state coupled evaluation model to generate the insulation aging state index at the current moment. Based on the comparison result between the insulation aging state index and the preset threshold, a graded early warning signal is triggered and output to the monitoring terminal; The insulation aging state coupled evaluation model is a hybrid architecture based on a physical information neural network. The input layer receives the multi-dimensional time-series input matrix, the hidden layer includes a thermal-electrical-humidity coupled differential equation constraint module and an attention-enhanced long short-term memory network module, and the output layer generates an insulation aging state index. The thermo-electric-humidity coupled differential equation constraint module embeds the following physical constraints based on the transformer thermal circuit model, moisture diffusion equation, and electric field distribution theory: The temperature rise of the winding is proportional to the square of the load current, and the proportionality coefficient is determined by the cooling method. The moisture content in insulating paper dynamically changes with ambient humidity and oil temperature, and the migration rate follows the Arrhenius equation. The partial discharge initiation voltage is modulated by the trace water content in the oil and the temperature, and the modulation function is in the form of exponential decay. The attention-enhanced long short-term memory network module includes a dual-channel structure. The first channel processes electrical and thermal time-series features, and the second channel processes meteorological and task planning features. The two channels are dynamically weighted in the time step dimension through an additive attention mechanism. The insulation aging state index is defined as a dimensionless quantity between 0 and 1, with the value being closer to 1 indicating a more severe degree of insulation aging. In the thermo-electric-humidity coupled differential equation constraint module, the relationship between the winding temperature rise and the square of the load current is expressed as: ； For winding temperature, This is the effective value of the load current. For AC resistance, For the winding heat capacity, The top oil temperature, The thermal time constant from the winding to the oil. Indicates time; In the thermo-electric-humidity coupled differential equation constraint module, the moisture migration flux in the insulating paper Represented as: ； in, Pre-exponential factor, For activation energy, The gas constant is... Oil temperature This represents the partial pressure of water vapor corresponding to the ambient relative humidity. This represents the equilibrium partial pressure corresponding to dissolved water in the oil. In the thermo-electric-humidity coupled differential equation constraint module, the partial discharge initiation voltage Represented as: ； in, The starting voltage in the dry clean oil. and This is an empirical coefficient. This refers to the trace water content in the oil; In the attention-enhanced long short-term memory network module, attention weights The calculation formula is: ； and The two channels are at time steps respectively. The hidden state, These are learnable parameters; and The two channels are at time steps respectively. The hidden state, For time steps The total number.

2. The intelligent monitoring method for substations according to claim 1, characterized in that, Obtain the historical operating data sequence of the transformer, including: The three-phase load current on the high-voltage and low-voltage sides of the transformer is collected by a current transformer at a sampling frequency of once per second. The temperature of the winding hot spot is obtained by an embedded fiber optic temperature sensor, with a spatial resolution of one measuring point per meter and a temporal resolution of once every 10 seconds. The top oil temperature is obtained through an oil temperature sensor, with a sampling cycle of once per minute; The partial discharge pulse signal in the grounding wire of the bushing end screen is monitored by a high-frequency current sensor, with a bandwidth covering 300 kHz to 150 MHz. The dielectric loss factor of the main insulation system is periodically measured using a dielectric loss tester, with a test cycle of once every 4 hours.

3. The intelligent monitoring method for substations according to claim 2, characterized in that, Obtain real-time meteorological data for the area where the substation is located, including: Ambient temperature is collected by a weather station deployed on top of the substation perimeter wall; the sensor accuracy is ±0.5 degrees Celsius. Relative humidity is obtained using a capacitive humidity sensor, with a measurement range of 5% to 95% and an accuracy of ±2%. Atmospheric pressure is obtained through a piezoresistive barometric pressure sensor with a range of 800 to 1100 hPa and a resolution better than 0.1 hPa. The total solar radiation intensity on a horizontal surface is measured using a silicon photodiode array sensor with a spectral response range of 400 to 1100 nanometers. The wind speed at a height of 10 meters above the ground is obtained using an ultrasonic anemometer, with a measurement range of 0 to 60 meters per second, and the starting wind speed is below 0.5 meters per second.

4. The intelligent monitoring method for substations according to claim 3, characterized in that, The historical operational data sequence, real-time meteorological data, and current task planning information are timestamped and missing values ​​are imputed, including: Using Coordinated Universal Time (UTC) as the base time source, all data streams are synchronized at the nanosecond level. The missing winding temperature and top oil temperature data were reconstructed using cubic spline interpolation, with the interpolation interval not exceeding 5 minutes. For cases where partial discharge data is missing for more than 30 consecutive seconds, a Markov chain Monte Carlo imputation method based on the statistical distribution of adjacent time periods is adopted. All data are resampled to a uniform time grid with a time step of 1 minute, forming a time-series input vector with a dimension of 64.

5. The intelligent monitoring method for substations according to claim 4, characterized in that, A graded early warning signal is triggered based on the comparison result between the insulation aging state index and a preset threshold, including: The first-level warning threshold is set to 0.

6. When the insulation aging state index exceeds the first-level warning threshold for the first time, a yellow warning signal is generated. The threshold for a Level 2 warning is set at 0.

75. When the index exceeds the Level 2 warning threshold for more than two hours, an orange warning signal is generated. The threshold for Level 3 warning is set at 0.

85. A red warning signal is generated when the instantaneous value of the index exceeds the Level 3 warning threshold or is higher than 0.8 for 4 consecutive hours. All warning signals are accompanied by a confidence score, which is calculated by the uncertainty estimation module output by the model. The uncertainty estimation is based on the standard deviation of the results of multiple forward propagations during the inference phase using the Monte Carlo dropout method.

Images