A power distribution network transformer oil temperature protection threshold self-learning method and system

Abstract

The present application belongs to the technical field of power equipment monitoring and protection, and discloses a power distribution network transformer oil temperature protection threshold self-learning method and system. The method comprises: acquiring multi-dimensional state data of the power distribution network transformer in real time; inputting the data into a reinforcement learning model, and obtaining an oil temperature protection threshold adjustment amount through decision-making based on a state space, an action space and a reward function; wherein the state space is constructed based on the multi-dimensional state data, the action space corresponds to the oil temperature protection threshold adjustment amount, and the reward function is a weighted sum of safety indexes, economy indexes and stability indexes; obtaining the oil temperature protection threshold adjustment amount output by the model at the current time to update the oil temperature protection threshold; and when the real-time acquired oil temperature of the power distribution network transformer exceeds the updated oil temperature protection threshold, performing a protection action. The method provided by the present application can achieve adaptive optimization of the oil temperature protection threshold while taking into account equipment safety and operation economy, and effectively improve the accuracy and adaptability of equipment protection.

Description

Technical Field

[0001] This invention belongs to the field of power equipment monitoring and protection technology, specifically relating to a self-learning method and system for the protection threshold of transformer oil temperature in a power distribution network. Background Technology

[0002] Currently, distribution transformers, as core equipment in the power system, directly affect the safety and stability of the power grid due to their operational reliability. During operation, losses generated in the windings and core of oil-immersed distribution transformers are converted into heat, causing the transformer oil temperature to rise. Excessively high oil temperatures accelerate the aging of insulation materials, shorten the transformer's service life, and in severe cases, even lead to burn-out accidents. Therefore, accurately setting the oil temperature protection threshold and taking timely protective measures is of great significance.

[0003] However, existing methods for setting transformer oil temperature protection thresholds typically rely on manual experience or fixed rules, lacking self-learning capabilities and the ability to continuously optimize and adjust strategies based on historical operating data, making them ill-suited to complex and ever-changing operating conditions. Furthermore, existing methods fail to fully utilize multi-dimensional state information for comprehensive decision-making, and there is a lack of coordinated linkage between protection thresholds and load regulation, making it difficult to ensure both equipment safety and power supply reliability. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a self-learning method and system for oil temperature protection threshold of distribution network transformers. By constructing a state space containing multi-dimensional operating states and designing a reward function that integrates safety, economy, and stability, a reinforcement learning model is used to achieve autonomous optimization and dynamic adjustment of the oil temperature protection threshold, effectively improving the accuracy and adaptability of equipment protection.

[0005] This invention provides the following technical solution: The primary objective of this invention is to provide a self-learning method for the oil temperature protection threshold of a distribution network transformer, comprising: Real-time acquisition of multi-dimensional status data of transformers in the distribution network; The multidimensional state data is input into a reinforcement learning model, which makes decisions based on a state space, an action space, and a reward function to obtain the oil temperature protection threshold adjustment amount; wherein, the state space is constructed based on the multidimensional state data, the action space corresponds to the oil temperature protection threshold adjustment amount, and the reward function is a weighted sum of safety indicators, economic indicators, and stability indicators; Obtain the current oil temperature protection threshold adjustment amount output by the reinforcement learning model, and update the oil temperature protection threshold. When the real-time measured oil temperature of the distribution network transformer exceeds the updated oil temperature protection threshold, a protection action is executed.

[0006] As a further improvement of the present invention, the multi-dimensional state data includes top oil temperature, load current, ambient temperature, ambient humidity, oil level, partial discharge, oil temperature change rate, and load current change rate. Multi-dimensional state information can more comprehensively reflect the operating conditions of the transformer, thereby improving the accuracy of model decision-making.

[0007] As a further improvement of the present invention, the calculation formula of the reward function is as follows:

[0008]

[0009]

[0010]

[0011] in, R For the reward function, , , These are safety indicators, economic indicators, and stability indicators. , , All are weighting coefficients. The top oil temperature, This is the oil temperature protection threshold. The absolute safe upper limit of temperature, , All are penalty coefficients. For load current, Rated current, This is the adjustment amount for the oil temperature protection threshold. This is the load carrying capacity bonus coefficient. This is the threshold fluctuation penalty coefficient. This is the threshold stationarity penalty coefficient.

[0012] The designed reward function supports load demand as much as possible while ensuring equipment safety and suppressing frequent fluctuations in the threshold, thus achieving a balance between safety and economy.

[0013] As a further improvement of the present invention, the step of updating the oil temperature protection threshold based on the threshold adjustment amount includes: Obtain the basic oil temperature threshold, which is set according to the transformer's rated parameters or industry standards; The sum of the adjustments to the base oil temperature threshold and the oil temperature protection threshold is used as the updated oil temperature protection threshold.

[0014] As a further improvement of the present invention, the training of the reinforcement learning model includes: Construct a deep Q-network that includes an evaluation network and a target network; Acquire historical multidimensional state data and store it as training samples in the experience pool; Samples are sampled from the experience pool, the current Q-value of the current state action pair is calculated using the evaluation network, the target Q-value of the next state action pair is calculated using the target network, and the network parameters are updated based on the difference between the target Q-value and the current Q-value.

[0015] As a further improvement of the present invention, the method further includes: Load characteristics are extracted based on high-frequency sampling data, and load types are identified based on these characteristics; Based on the multidimensional state data, an oil temperature prediction result is obtained, which is used to characterize the oil temperature change trend within a preset time period in the future. Based on the oil temperature prediction results and the rate of temperature rise, a tiered adjustment strategy corresponding to the load type is matched and executed.

[0016] By identifying, matching, and executing differentiated adjustment strategies based on load type, the linkage between protection thresholds and load control is achieved, upgrading from passive protection to active regulation.

[0017] As a further improvement of the present invention, the method further includes: Acquire historical multidimensional status data and evaluate the transformer's health index based on the historical multidimensional status data; The basic oil temperature threshold is dynamically adjusted based on the aforementioned health index.

[0018] By introducing a health index throughout the entire equipment lifecycle and dynamically adjusting the protection benchmark based on the health index, the risks associated with using a unified benchmark for old and new equipment are resolved, thus achieving personalized protection.

[0019] The second objective of this invention is to provide a self-learning system for the oil temperature protection threshold of a distribution network transformer, used to implement the above method, comprising: The data acquisition unit is configured to acquire multi-dimensional status data of transformers in the distribution network in real time. An edge computing unit, connected to the data acquisition unit, is configured to input the multidimensional state data into a reinforcement learning model for inference to output an oil temperature protection threshold adjustment amount, and update the oil temperature protection threshold based on the oil temperature protection threshold adjustment amount. The execution unit, connected to the edge computing unit, is configured to perform a protection action when the real-time acquired oil temperature of the distribution network transformer exceeds the updated oil temperature protection threshold.

[0020] A third objective of this invention is to provide a computer device comprising at least one processing unit and at least one storage unit, wherein the storage unit stores a computer program, and when the program is executed by the processing unit, the processing unit performs the aforementioned method.

[0021] A fourth objective of this invention is to provide a computer-readable storage medium storing a computer program executable by an electronic device, which, when run on the electronic device, causes the electronic device to perform the above-described method.

[0022] Compared with the prior art, the beneficial effects of the present invention are as follows: By constructing a state space containing multi-dimensional operating states and designing a reward function that integrates safety, economy, and stability, a reinforcement learning model is used to achieve autonomous optimization and dynamic adjustment of the oil temperature protection threshold. This eliminates the need for manual setting of adjustment rules and overcomes the limitations of traditional methods that rely on expert experience. Furthermore, by comprehensively considering multi-dimensional state information in threshold decision-making, it can more accurately identify equipment operating risks and avoid erroneous or non-operational actions caused by improper rule settings.

[0023] The designed reward function can support the load demand as much as possible while ensuring equipment safety and suppressing frequent fluctuations in the threshold, thus achieving a balance between safety and economy.

[0024] By using load type identification and tiered adjustment mechanisms, the oil temperature protection threshold is linked with load control, upgrading from passive protection to active regulation. Attached Figure Description

[0025] Figure 1 A flowchart of a self-learning method for oil temperature protection threshold of distribution network transformers; Figure 2 A training framework diagram for reinforcement learning models. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] The present invention will now be described in further detail with reference to the accompanying drawings: Example 1: like Figure 1As shown, this embodiment provides a self-learning method for the oil temperature protection threshold of a distribution network transformer. By introducing a reinforcement learning mechanism, it utilizes multi-dimensional state data to achieve adaptive optimization of the oil temperature protection threshold, effectively solving the technical problems of threshold adjustment relying on human experience and lacking adaptive capability in the prior art.

[0028] S1, obtain historical multidimensional status data of distribution network transformers.

[0029] In this embodiment, the multidimensional status data includes top oil temperature, load current, ambient temperature, ambient humidity, oil level, partial discharge quantity, oil temperature change rate, and load current change rate. Top oil temperature data is collected in real time by an oil temperature sensor installed on the top of the transformer; load current data is collected by a current transformer; oil level data is collected by an oil level sensor; partial discharge quantity is collected by a partial discharge sensor; and ambient temperature and humidity are measured by temperature and humidity sensors.

[0030] The introduction of multi-dimensional state data can more comprehensively depict the operating conditions of the transformer, providing rich information support for subsequent model decisions. For example, the introduction of ambient temperature can help the model identify seasonal differences and avoid malfunctions caused by high summer temperatures; the introduction of oil temperature change rate helps identify transient processes such as load changes. The acquisition frequency of multi-dimensional state data can be set according to actual protection needs. In this embodiment, it is set to be collected once per minute, continuously collecting historical operating data for 6 months.

[0031] The collected data were preprocessed as follows: outliers were removed using the 3σ principle; missing values ​​were filled using linear interpolation; each feature was scaled to the [0,1] interval using the Min-Max normalization method; and the time series data were time-aligned to ensure that the data from each sensor at the same time corresponded.

[0032] S2, Reinforcement Learning State Space Construction.

[0033] The preprocessed data is mapped to a state space S. In this embodiment, the state space contains nine dimensions: top oil temperature. Load current ambient temperature Degree, ambient humidity Oil level Partial discharge quantity Oil temperature change rate Load current change rate Time feature coding t (Seasons: 1-4 represent spring, summer, autumn and winter; Time period: 0-23 represent hours).

[0034] S3, Reinforcement Learning Action Space Definition.

[0035] The action space A is defined as the oil temperature protection threshold adjustment amount. In this embodiment, the oil temperature base threshold... Set to 85℃ (according to GB / T 1094.2 standard, the top oil temperature limit for oil-immersed transformers is 95℃, and 85℃ is taken as the starting point for the warning threshold). Maximum allowable adjustment range. The threshold is set to 5℃, meaning it can be dynamically adjusted within the range of 80℃ to 90℃. The action space is discretized into 11 actions, represented as {-5℃, -4℃, ..., 0℃, ..., +4℃, +5℃}.

[0036] S4, Reward Function Design.

[0037] reward function R For safety indicators Economic indicators and stability indicators The weighted sum is expressed as:

[0038] The weighting coefficients are set as follows: , , .

[0039] Safety indicators Represented as:

[0040] in, Top oil temperature; This is the oil temperature protection threshold. The absolute safe upper limit is set at 95℃; , All are penalty coefficients, and , When the oil temperature exceeds the current threshold but does not exceed the absolute safe temperature limit, a minor penalty is imposed; when the oil temperature exceeds the absolute safe temperature limit, a major penalty is imposed.

[0041] Economic indicators Represented as:

[0042] in, For load current, This refers to the rated current (transformer rated value). This is the adjustment amount for the oil temperature protection threshold; This is a load-bearing incentive coefficient, designed to encourage load-bearing under safe conditions. In this embodiment... ; This is the threshold fluctuation penalty coefficient, used to penalize large fluctuations in the threshold. In this embodiment... .

[0043] Stability Indicators Represented as:

[0044] in, , These represent the oil temperature protection thresholds at time t and time t-1, respectively. This is a threshold stability penalty coefficient used to suppress frequent threshold jumps. In this embodiment... =0.1.

[0045] S5 uses historical data to train a reinforcement learning model.

[0046] In this embodiment, the reinforcement learning model is a deep Q-network (DQN), and the specific architecture is as follows: Input layer: 9 neurons, corresponding to the dimension of the state space; Hidden layer 1: 128 neurons, ReLU activation function; Hidden layer 2: 64 neurons, ReLU activation function; Output layer: 11 neurons, corresponding to the Q-values ​​of 11 actions in the action space.

[0047] Training parameters are set as follows: experience pool capacity is 10,000 records; batch size is 32; discount factor. The learning rate is 0.001; the target network is updated every 100 steps; the exploration rate of the greedy strategy decreases linearly from 1.0 to 0.1.

[0048] Training samples were constructed using six months of historical data collected by S1. Each sample included the current state s, the action a, the reward r, and the next state s′. A total of 5000 episodes were trained, each containing 1440 time steps (corresponding to 24 hours, 1 step per minute). The training process is as follows: Figure 2 As shown.

[0049] After receiving multi-dimensional state data at the current moment, the trained reinforcement learning model, through internal inference logic, outputs a specific oil temperature protection threshold adjustment amount. This adjustment amount is a relative value, representing the degree of deviation from the base oil temperature threshold.

[0050] S6, online optimization and execution of oil temperature protection threshold.

[0051] The trained DQN model is deployed on an edge computing gateway (such as an ARM-based embedded device) to collect transformer data in real time. The current state is calculated every minute, and the model is input to obtain the optimal action (oil temperature protection threshold adjustment amount), updating the current oil temperature protection threshold. When the real-time oil temperature exceeds the updated oil temperature protection threshold, an early warning signal is triggered and sent to maintenance personnel via the communication module.

[0052] In this embodiment, the oil temperature protection threshold is dynamically updated, thus enabling more accurate identification of abnormal operating conditions and avoiding protection failure due to excessively high threshold settings or false protection activation due to excessively low threshold settings.

[0053] Example 2: Based on Example 1, this embodiment adds load type identification and hierarchical linkage functions to further optimize the coordination of protection and control.

[0054] Based on high-frequency sampling data (sampling frequency increased to 1 second / time), load characteristics are extracted, including RMS current, harmonic content, power factor, fluctuation rate, and start-stop characteristics.

[0055] The loads were divided into the following four categories using the K-means clustering algorithm: Type A, electric vehicle charging load, characterized by a steady increase over several hours followed by a steady decrease, with a power factor close to 1; Type B, air conditioning load, characterized by periodic fluctuations and a strong correlation with temperature; Type C, industrial impact load, characterized by: short-term high-current impact and high harmonic content; Type D, Residential Basic Load, characterized by morning and evening peak hours and fluctuating patterns.

[0056] Based on the oil temperature prediction results and the rate of temperature rise, the adjustment strategy is dynamically selected: When the oil temperature rises rapidly ( When the load type is identified as Type C, the first level of adjustment is initiated: the charging power is reduced from 7kW to 3.5kW through the power adjustment interface of the electric vehicle charging pile; When the oil temperature continues to rise but at a slow rate ( When the load type is identified as Type B, the second-level adjustment is initiated: the temperature setpoint is increased from 26℃ to 28℃ through the air conditioning group control system.

[0057] When the oil temperature approaches its absolute upper limit ( When this occurs, the third layer of regulation is activated: through the demand response system, a load reduction invitation is sent to the relevant industrial and commercial users.

[0058] Example 3: Based on Example 1, this embodiment adds a function that links health status assessment with remaining lifespan, enabling personalized protection throughout the entire lifecycle of the equipment.

[0059] Based on the thermal aging model of IEC 60076-7 standard, the health index of transformers is evaluated online.

[0060] Aging Accelerator The calculation formula is:

[0061] in, The hot spot temperature is estimated based on the top oil temperature, load current, and ambient temperature.

[0062] The formula for calculating the Health Index (HI) is as follows:

[0063] in, The initial health index (1.0 for new devices) is used. k The aging factor is... L To accumulate lifespan loss, .

[0064] The base oil temperature threshold is dynamically adjusted based on the Health Index (HI). :

[0065] in, For correction factors, This is a reference value for the health index.

[0066] Specifically: When HI > 0.8, it indicates that the transformer is in good health, and the correction factor is [value missing]. At this point, the basic oil temperature threshold can be appropriately increased to 87℃ to support a higher load; When 0.5 ≤ HI ≤ 0.8, it indicates that the transformer is of moderate aging, and the correction factor is [missing value]. At this point, the basic oil temperature threshold remains at 85℃; When HI < 0.5, it indicates that the transformer is severely aged, and the correction factor is [not specified]. At this point, the basic oil temperature threshold can be reduced to 77°C to enhance protection.

[0067] Example 4: This embodiment provides a self-learning system for the oil temperature protection threshold of a distribution network transformer, used to implement the methods provided in Embodiments 1 to 3, including: The data acquisition unit is configured to acquire multi-dimensional status data of the distribution network transformer in real time, including intelligent distribution transformer oil temperature sensor, current transformer, environmental sensor, oil level sensor and partial discharge sensor. An edge computing unit, connected to the data acquisition unit, is configured to input the multidimensional state data into a reinforcement learning model for inference to output an oil temperature protection threshold adjustment amount, and update the oil temperature protection threshold based on the oil temperature protection threshold adjustment amount. An execution unit, connected to the edge computing unit, is configured to perform a protection action when the real-time acquired oil temperature of the distribution network transformer exceeds the updated oil temperature protection threshold. The cloud platform, which communicates with the edge computing unit, is used to train and update the reinforcement learning model offline based on historical multidimensional state data, and to send the updated model parameters to the edge computing unit. Example 5: This embodiment provides a computer device, including at least one processing unit and at least one storage unit, wherein the storage unit stores a computer program, and when the program is executed by the processing unit, the processing unit performs the method provided in Embodiments 1 to 3.

[0068] Example 6: This embodiment provides a computer-readable storage medium storing a computer program executable by an electronic device, which, when run on the electronic device, causes the electronic device to perform the methods provided in Embodiments 1 to 3.

[0069] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A self-learning method for the oil temperature protection threshold of a distribution network transformer, characterized in that, include: Real-time acquisition of multi-dimensional status data of transformers in the distribution network; The multidimensional state data is input into a reinforcement learning model, which makes decisions based on a state space, an action space, and a reward function to obtain the oil temperature protection threshold adjustment amount; wherein, the state space is constructed based on the multidimensional state data, the action space corresponds to the oil temperature protection threshold adjustment amount, and the reward function is a weighted sum of safety indicators, economic indicators, and stability indicators; Obtain the current oil temperature protection threshold adjustment amount output by the reinforcement learning model, and update the oil temperature protection threshold. When the real-time measured oil temperature of the distribution network transformer exceeds the updated oil temperature protection threshold, a protection action is executed.

2. The method according to claim 1, characterized in that, The multidimensional state data includes top oil temperature, load current, ambient temperature, ambient humidity, oil level, partial discharge, oil temperature change rate, and load current change rate.

3. The method according to claim 1, characterized in that, The formula for calculating the reward function is as follows: in, R For the reward function, , , These are safety indicators, economic indicators, and stability indicators. , , All are weighting coefficients. The top oil temperature, This is the oil temperature protection threshold. The absolute safe temperature limit, , All are penalty coefficients. For load current, Rated current, This is the adjustment amount for the oil temperature protection threshold. This is the load carrying capacity bonus coefficient. This is the threshold fluctuation penalty coefficient. This is the threshold stationarity penalty coefficient.

4. The method according to claim 1, characterized in that, The step of updating the oil temperature protection threshold based on the threshold adjustment amount includes: Obtain the basic oil temperature threshold, which is set according to the transformer's rated parameters or industry standards; The sum of the adjustments to the base oil temperature threshold and the oil temperature protection threshold is used as the updated oil temperature protection threshold.

5. The method according to claim 1, characterized in that, The training of the reinforcement learning model includes: Construct a deep Q-network that includes an evaluation network and a target network; Acquire historical multidimensional state data and store it as training samples in the experience pool; Samples are sampled from the experience pool, the current Q-value of the current state action pair is calculated using the evaluation network, the target Q-value of the next state action pair is calculated using the target network, and the network parameters are updated based on the difference between the target Q-value and the current Q-value.

6. The method according to claim 1, characterized in that, The method further includes: Load characteristics are extracted based on high-frequency sampling data, and load types are identified based on these characteristics; Based on the multidimensional state data, an oil temperature prediction result is obtained, which is used to characterize the oil temperature change trend within a preset time period in the future. Based on the oil temperature prediction results and the rate of temperature rise, a tiered adjustment strategy corresponding to the load type is matched and executed.

7. The method according to claim 1, characterized in that, The method further includes: Acquire historical multidimensional status data and evaluate the transformer's health index based on the historical multidimensional status data; The basic oil temperature threshold is dynamically adjusted based on the health index.

8. A self-learning system for oil temperature protection threshold of a distribution network transformer, used to implement the method described in any one of claims 1 to 7, characterized in that, include: The data acquisition unit is configured to acquire multi-dimensional status data of transformers in the distribution network in real time. An edge computing unit, connected to the data acquisition unit, is configured to input the multidimensional state data into a reinforcement learning model for inference to output an oil temperature protection threshold adjustment amount, and update the oil temperature protection threshold based on the oil temperature protection threshold adjustment amount. The execution unit, connected to the edge computing unit, is configured to perform a protection action when the real-time acquired oil temperature of the distribution network transformer exceeds the updated oil temperature protection threshold.

9. A computer device, characterized in that, It includes at least one processing unit and at least one storage unit, wherein the storage unit stores a computer program that, when executed by the processing unit, causes the processing unit to perform the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores a computer program executable by an electronic device, which, when run on the electronic device, causes the electronic device to perform the method as described in any one of claims 1 to 7.

Images