Transformer overheating protection method and system

Abstract

The invention provides a transformer overheating protection method and system. The method comprises the following steps: acquiring data acquired during operation of a transformer; performing time sequence data change rate analysis on the operation power of the cooling system to obtain a cooling activation degree index, and performing preprocessing to form a standard input tensor; calculating a unit heat rise efficiency index according to the standard input tensor, obtaining a dynamic residual error index based on the unit heat rise efficiency index and a deviation degree of a historical mean value of the unit heat rise efficiency index, and generating a cooling state label; inputting the standard input tensor and the cooling state label into a pre-constructed temperature prediction model, and generating a hot spot temperature prediction value of a future time period by integrating the cooling state label into a backbone network structure; and based on the hot spot temperature prediction value and the cooling state label of the future time period, matching with a preset thermal risk classification logic table, and feeding back an execution result to optimize a system closed loop.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of transformer overheat protection, and particularly relates to a transformer overheat protection method and system. BACKGROUND

[0002] With the continuous growth of power system load and the increase of instability of new energy access, the safe and stable operation of transformers as the core equipment of power transmission and distribution systems has been increasingly valued. In particular, a large amount of heat energy is generated during the operation of the transformer, and if it cannot be released in time and effectively, it will cause the temperature of the winding, iron core and insulation system to rise, eventually leading to insulation aging, thermal breakdown, and even serious power accidents. Therefore, as an important part of the safety of the power grid, the transformer overheat protection technology has become the focus of equipment manufacturers and operation units.

[0003] The existing transformer overheat protection methods can be roughly divided into two categories: one is a passive protection method with temperature threshold judgment as the core, which usually triggers an alarm or trip instruction when the measured temperature exceeds the set threshold through the temperature sensor installed near the oil tank or winding; the other is a prediction method based on thermal equivalent model for hotspot estimation, which is commonly used in IEC standard thermal modeling method. This method calculates the winding hotspot temperature through input variables such as load current and oil temperature, and then evaluates the thermal state of the transformer. However, these methods have obvious defects: first, the sensor distribution is limited and some hotspot areas cannot be directly measured, resulting in inaccurate estimation of hotspot temperature; second, the traditional thermal model relies on static parameters and empirical coefficients, making it difficult to adapt to the dynamic operating environment of the transformer, such as external climate change and cooling equipment performance degradation; third, these methods mostly use temperature exceeding as the basis for judgment, lacking the ability to distinguish the cause of temperature rise, and cannot distinguish whether the temperature rise is caused by load abnormality, cooling degradation or internal structure degradation, which is prone to misjudgment or omission. In addition, the current system usually lacks foresight, i.e. it cannot predict that the temperature will exceed the threshold, but only responds after the temperature has exceeded the threshold, thus losing the valuable prevention window.

[0004] More complex is that in actual operation, the temperature rise process of the transformer is often affected by non-electrical factors such as cooling system degradation, local load concentration, and oil blockage. These factors cause the device to still have hotspot aggregation or even thermal runaway even if the electrical load is within the normal range. However, the current overheat protection system cannot effectively identify and handle such implicit overheat risks, and often relies on manual experience for judgment, which is low in efficiency and poor in reliability. Therefore, there is an urgent need for a comprehensive system that can dynamically identify the state of the cooling system, accurately predict the winding hotspot temperature, analyze the cause of overheat and output effective protection strategies in time, in order to realize the leap from passive response to active identification + intelligent early warning. SUMMARY

[0005] The purpose of the present application is to propose a transformer overheat protection method and system, which effectively overcomes the core problems of passive monitoring, high false alarm rate, lack of interpretability and intervention ability in the prior art, and is suitable for modern smart grid scenarios with high requirements for transformer operating state.

[0006] In order to achieve the above purpose, in the first aspect of the present application, a transformer overheat protection method is provided, which comprises the following steps:

[0007] Collecting data collected during the operation of the transformer, including oil top layer temperature, ambient temperature, load power, cooling system operating power, and near three-minute temperature difference sequence; by analyzing the time series data change rate of the cooling system operating power, a cooling activation degree index is obtained, and the oil top layer temperature, ambient temperature, load power, cooling system operating power, near three-minute temperature difference sequence and cooling activation degree index are combined for pretreatment to form a standard input tensor; the cooling activation degree index is used to reflect the current state of the cooling system;

[0008] According to the standard input tensor, a unit heat rise efficiency index is calculated, a dynamic residual index is obtained based on the deviation degree of the unit heat rise efficiency index and its historical mean value, and a cooling state label is generated in combination with the cooling activation degree index, the cooling state label is used to represent the degradation degree of the cooling system;

[0009] The standard input tensor and the cooling state label are input into a pre-constructed temperature prediction model, the cooling state label is integrated into the backbone network structure, the prediction path is dynamically adjusted, and the hotspot temperature prediction value of the future period is generated;

[0010] Based on the hotspot temperature prediction value of the future period and the cooling state label, by matching with a preset thermal risk classification logic table, a risk level is obtained and a protection strategy is generated, the protection strategy includes cooling instructions, load adjustment and alarm behavior;

[0011] The protection strategy is executed, control instructions are issued to the cooling system and load scheduling module, and the execution result is fed back to optimize the system closed loop.

[0012] Further, the pretreatment includes: time alignment and denoising processing of multi-frequency sampling data, and generating a short-term temperature difference sequence through a sliding window, the near three-minute temperature difference sequence is used to reflect the dynamic deviation trend of oil temperature and ambient temperature.

[0013] Further, the unit heat rise efficiency index calculated according to the standard input tensor specifically includes:

[0014] The oil top layer temperature, ambient temperature and load power are acquired, performance index analysis is performed, and an unit heat rise efficiency index is obtained, which is used for measuring the temperature rise degree presented by the system under unit load.

[0015] Further, the cooling state label includes normal cooling state, slight degradation and severe degradation.

[0016] Further, the structure of the temperature prediction model includes:

[0017] a backbone modeling layer, which processes the standard input tensor by a lightweight recurrent network, and learns potential nonlinear dynamic rules;

[0018] a cooling state adjustment layer, which integrates the cooling state label into the backbone network structure, dynamically adjusts the prediction path, and enables the model to adaptively select different internal prediction structures or adjust prediction weights under different cooling states;

[0019] The prediction function of the temperature prediction model is represented as:

[0020]

[0021] wherein, is the winding hotspot temperature after δ minutes in the future; f θ (X t ) is the baseline output of the temperature prediction model under the default cooling state, and the input is the standard input tensor; is a cooling state adjustment coefficient, which takes 0 when the cooling state label S c (t) = 0, takes 1 when S c (t) = 2, and represents the residual adjustment intensity; R(X t ) is a residual network, which is used for expressing the temperature rise offset modeling under cooling degradation.

[0022] Further, the loss function L of the temperature prediction model is represented as:

[0023]

[0024] wherein, T hotspot (t+δ) is the real future hotspot temperature, which is calculated by the field winding hotspot calculation system or the IEEE thermal model; is a cooling state response regularization term, which measures the sensitivity of the model prediction to the change of the cooling state label; is a main loss, which measures the prediction deviation by using mean square error; and λ is a regularization weight parameter.

[0025] Further, the hotspot temperature prediction value based on the future period and the cooling state label are matched with a preset thermal risk classification logic table to obtain a risk level and generate a protection strategy, specifically including:

[0026] The safety upper limit temperature of the current device model is obtained, and the over-limit ratio of the predicted temperature relative threshold value is calculated based on the hotspot temperature prediction value of the future period and the safety upper limit temperature;

[0027] According to the over-limit ratio of the predicted temperature relative threshold value and the cooling state label, a risk level is obtained by matching with a preset thermal risk classification logic table, including no risk, load impact risk, cooling degradation risk and coupling risk;

[0028] A protection strategy is generated according to the risk level.

[0029] Further, the protection strategy is executed, control instructions are issued to the cooling system and load scheduling module, and the execution result is fed back to optimize the system closed loop, specifically including:

[0030] The protection strategy is obtained, and the device control signal is mapped through a mapping function; if the response is successful, the device control signal is executed; if no confirmation feedback is received from the execution device within the set confirmation window after the device control signal is issued, the system enters a protection rollback process and triggers a default action.

[0031] In a second aspect of the present application, a transformer overheat protection system is provided, the system comprising:

[0032] A data acquisition module is used to acquire data collected during the operation of the transformer, including the oil top layer temperature, the ambient temperature, the load power, the cooling system operating power, and the near three-minute temperature difference sequence; the cooling activation index is obtained by analyzing the time series data change rate of the cooling system operating power, and the oil top layer temperature, the ambient temperature, the load power, the cooling system operating power, the near three-minute temperature difference sequence and the cooling activation index are preprocessed to form a standard input tensor; the cooling activation index is used to reflect the current state of the cooling system;

[0033] A cooling state recognition module is used to calculate the unit heat rise efficiency index according to the standard input tensor, obtain the dynamic residual index based on the deviation of the unit heat rise efficiency index and its historical mean value, and generate the cooling state label in combination with the cooling activation index, the cooling state label being used to represent the degradation degree of the cooling system;

[0034] A prediction module is used to input the standard input tensor and the cooling state label into a pre-constructed temperature prediction model, dynamically adjust the prediction path by integrating the cooling state label into the backbone network structure, and generate the hotspot temperature prediction value of the future period.

[0035] a strategy generation module configured to obtain a risk level and generate a protection strategy by matching with a preset thermal risk classification logic table based on the hotspot temperature prediction value and the cooling state label of the future period, the protection strategy including cooling instructions, load adjustment, and alarm behavior;

[0036] an execution module configured to execute the protection strategy, issue control instructions to the cooling system and load scheduling module, and feed back execution results to optimize system closed loop.

[0037] The present application has at least the following beneficial technical effects:

[0038] To solve the problems of low prediction accuracy of hotspot temperature, difficulty in identifying cooling system abnormalities, inability to explain the cause of overheating, and lack of proactive protection capability in existing overheating protection methods, the present application proposes a transformer overheating protection system based on multi-dimensional state modeling and thermal behavior fusion reasoning mechanism. The system aims to unify thermal risk identification and designs multiple modules with synergistic effects, which can dynamically model the changing trend of heat dissipation performance and accurately infer the temperature rise behavior of the winding hotspot in combination with the device operating parameters, thereby identifying potential thermal risks in advance. The present application establishes a key intermediate variable that can reflect the change of cooling capacity and integrates it into the hotspot temperature prediction mechanism, achieving classification identification of temperature rise causes and dynamic correction of hotspot prediction results. The system not only completes early warning before the temperature exceeds the limit, but also points out the main source of risk, whether it is cooling degradation or load fluctuation, thereby providing a scientific basis for subsequent protection actions such as cooling scheduling, load redistribution, or fault alarm. The design of the present application ensures a logical closed loop of the technical path, making prediction, judgment, and response mutually coupled, effectively overcoming the core problems of passive monitoring, high false alarm rate, lack of interpretability, and intervention capability in existing technologies, and is suitable for modern smart grid scenarios with high requirements for transformer operating state. BRIEF DESCRIPTION OF DRAWINGS

[0039] The present application is further illustrated by the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present application. For ordinary skilled persons in the art, other drawings can be obtained without creative labor based on the following drawings.

[0040] Figure 1 A flowchart of the transformer overheating protection method of the present application. DETAILED DESCRIPTION

[0041] Embodiments of the present application are described below in the detailed description and illustrated in the accompanying drawings by way of examples which, as will become apparent to those skilled in the art upon reading the present document, can be implemented in various ways. The embodiments described below are examples only, and are not to be construed as limiting the present application.

[0042] As shown in the drawings, Figure 1 The transformer overheat protection method provided by the embodiments of the present application comprises:

[0043] S1, collecting data collected during transformer operation, including oil top layer temperature, ambient temperature, load power, cooling system running power, and near three-minute temperature difference sequence; by analyzing the time series data change rate of the cooling system running power, a cooling activation index is obtained, and the oil top layer temperature, ambient temperature, load power, cooling system running power, near three-minute temperature difference sequence and cooling activation index are preprocessed to form a standard input tensor; the cooling activation index is used to reflect the current state of the cooling system.

[0044] Specifically, the purpose of this step is to build a unified and structured input tensor X t as the bottom input structure for modeling and decision-making of the entire overheat protection system in transformer operation monitoring.

[0045] Among them, the data collected during the operation of the transformer includes: oil top layer temperature T oil (t), ambient temperature T env (t), load power P(t), cooling system running power C(t), near three-minute temperature difference sequence ΔT t-k (k=1,2,3) and cooling response index Because the sampling frequencies of different sensors are different, for example, the oil temperature is sampled every 1 minute, and the ambient temperature is sampled every 5 minutes, linear interpolation and moving average are used for time alignment and denoising.

[0046] Among them, the short-term temperature difference characteristic sequence is calculated as follows to form the core input channel of thermal inertia trend:

[0047] ΔT t-k = T oil (t-k)-T env (t-k), k=1,2,3

[0048] This variable is different from the common first-order difference It more directly expresses the thermal deviation state of the transformer in the environmental background, and can avoid the model instability caused by the amplification of small temperature measurement errors. For example, during a certain operation process, the environmental temperature is 30℃, and the oil temperature rises from 54℃ to 58℃, at this time, the continuous increase of ΔT can express the system in the heat accumulation trend more than the temperature difference.

[0049] At the same time, this step introduces an innovative input variable for cooling dynamic behavior modeling-cooling activation index defined as:

[0050]

[0051] This variable quantifies the relative intervention amplitude of the cooling system at the current time, that is, whether the cooling equipment is just activated, increased power or in stable operation at a certain time point. For example, when the transformer load suddenly rises, the oil temperature rises, and the system control activates the standby fan, C(t) will appear a step rise, at this time is obviously greater than 0, indicating that the cooling intervention action has occurred, and this feature has strong explanatory power for judging whether the cooling responds in time.

[0052] Finally, the standard input tensor is:

[0053]

[0054] This tensor is not only a simple splicing of variables, but also a feature combination optimized by thermal physical structure. All variables have uniform units (temperature in ℃, power in kW), consistent sampling time steps, and compatibility with different transformer models. Through this structure input method, the present application not only solves the heterogeneous problem of data sources, but also models the thermal capacity effect and cooling responsiveness in the transformer operation as input features, which is ignored in most existing AI models. It provides consistent, physically interpretable and engineering deployment value data basis for all subsequent modules of the patent.

[0055] S2, calculate the unit heat rise efficiency index based on the standard input tensor, obtain the dynamic residual index based on the deviation degree of the unit heat rise efficiency index and its historical mean value, and generate a cooling state label in combination with the cooling activation index, the cooling state label is used to represent the degradation degree of the cooling system.

[0056] Specifically, this step is based on the standard input tensor X t, the purpose is to extract and quantify whether the current cooling capacity of the transformer is effective, so as to judge the potential non-load driven overheating risk. The traditional protection system usually only pays attention to whether the temperature exceeds the threshold, but ignores the key fact that under the same load, the cooling performance decline may cause faster temperature rise or even thermal runaway, and such phenomenon is quite hidden in the oil temperature data. Therefore, the goal of this step is not only to identify the temperature rise state, but also to build a cooling state label S c (t) that can be propagated, explained and adjusted to serve subsequent prediction and response decisions.

[0057] From the X t of step one, the invention selects three key variables as the core input of this step: oil temperature T oil (t), ambient temperature T env (t) and load power P(t), builds a unit heat rise efficiency index η t , which is used to measure the temperature rise degree presented by the system under unit load. Its basic definition is:

[0058]

[0059] Where T oil (t) is the top layer temperature of the oil tank, which is collected by the temperature sensor in real time; T env (t) is the ambient temperature, which is measured by the outdoor monitoring point; P(t) is the current transformer load power, unit is kW; ∈ is a small constant, usually 0.1, used to prevent the denominator from being zero when the load is low.

[0060] This index reflects the temperature rise effect of unit load. If η t continuously rises under the condition that the load does not change or changes slowly, it means that the cooling system may have the trend of efficiency decline or response delay.

[0061] But since η t is a transient index, it is greatly affected by external interference (such as sudden change of ambient temperature, load peak, etc.), and it is difficult to be directly used for system judgment. Therefore, the invention introduces a trend residual regulation mechanism to build a dynamic residual index δ t :

[0062]

[0063] Where μ t = mean(η t-w:t-1 ) represents the sliding mean value of the past w time points (such as 15 minutes); is the time variation rate of heat rise efficiency (η t - η t-1The trend is reflected by the growth trend of the average temperature rise, and λ is the trend adjustment coefficient, which controls the sensitivity of the system to the accelerated temperature rise, and is usually taken as 0.5-1.0.

[0064] By comparing the current efficiency with the deviation of its historical average and rising trend, δ t The cooling performance deterioration can be more stable and forward-looking.

[0065] In order to enhance the physical interpretation ability, the cooling activation degree index constructed in step one is introduced Reflect whether the cooling system is started or enhanced at present. When is very low and δ t is significantly positive, it indicates that the cooling system does not respond; when is very high and δ t is still positive, it indicates that the cooling system responds but does not work. Both of the two cases essentially indicate cooling degradation, but the causes are different, and this identification ability will provide clear guidance for the downstream model f θ .

[0066] Finally, the cooling state label S c (t) is generated in this step, which takes the following values:

[0067] S c (t) = 0: the cooling state is normal, δ t is close to zero, stable;

[0068] S c (t) = 1: slight degradation, δ t has a deviation, has a response but is insufficient;

[0069] S c (t) = 2: severe degradation, δ t continues to have a large positive deviation, ineffective or non-responsive.

[0070] The label will be directly transmitted to the hotspot prediction model f θ in the next step as part of the model input to participate in conditional modeling. Compared with the traditional method, this step does not rely on the artificially set temperature rise threshold, but models the causal relationship between the structure variable η t , the trend residual δ t and the response index , truly realizing the organic integration of dynamic identification, classified explanation and data-driven control of the cooling state. This mechanism also enables the system to have adaptive explanation ability, which can explicitly model the changes in cooling mechanism in thermal behavior, supporting the closed-loop control of subsequent prediction logic and protection response strategy.

[0071] S3, input the standard input tensor and cooling state label into a pre-constructed temperature prediction model, dynamically adjust the prediction path by integrating the cooling state label into the backbone network structure, and generate the hotspot temperature prediction value of the future period.

[0072] Specifically, the present step takes X t and S c (t) as input to construct a cooling state guided temperature prediction model f θ , and the output target is , i.e., the prediction of the winding hotspot temperature after δ minutes. The core difficulty of this prediction task is that temperature change is not only affected by load but also by cooling performance (usually non-directly measurable); if the model cannot perceive the difference in cooling state, it may mistakenly believe that all temperature rises are caused by load, leading to misjudgment or even omission. However, the present application has explicitly labeled the cooling state label S c (t) in step two, which provides a key fulcrum for constructing a state-aware prediction model.

[0073] To realize the state-aware prediction mechanism, the present application proposes a conditional adjustment type prediction network structure f θ , which consists of two parts: (a) the backbone modeling layer, which processes the thermal behavior sequence X t with a lightweight recurrent network (such as GRU or time convolution module) to learn the underlying nonlinear dynamic law; (b) the cooling state adjustment layer, which integrates S c (t) into the backbone network structure to dynamically adjust the prediction path, so that the model can adaptively select different internal prediction structures or adjust the prediction weights under different cooling states. For example, when S c (t) = 0, the model should trust that the cooling system is working effectively; when S c (t) = 2, the model should enhance the attention to temperature rise trend and be biased towards early warning.

[0074] To specifically model this structure adjustment effect, the present application introduces a state-guided gating structure, so that the prediction function f θ is mathematically expressed as follows:

[0075]

[0076] where, is the prediction of the winding hotspot temperature after δ minutes; f θ (X t ) is the baseline output of the model under the default cooling state, with X t as input, containing T oil (t), T env (t), P(t), C(t), ΔT t-k, etc. is the cooling state adjustment coefficient, when S c (t) = 0, S c (t) = 2, indicates the residual adjustment strength; R(X t ) is a residual network used to express the temperature rise offset modeling in cooling degradation.

[0077] In order to improve the stability of the predicted output and prevent the model from blindly relying on historical temperature trends and ignoring cooling signals, the present application introduces an innovative loss regularization term, called state variation consistency constraint, to encourage the model to be more sensitive to temperature prediction when the cooling state changes sharply, and more conservative when the state is stable. Its loss function form is as follows:

[0078]

[0079] Where T hotspot (t+δ) is the real future hot spot temperature, which is calculated by the on-site winding hot spot calculation system or the IEEE thermal model; the first term is the main loss, which measures the prediction error using mean square error; the second term is the cooling state response regularization term, which measures the sensitivity of the model prediction to the change of S c (t); λ is the regularization weight parameter, usually set to 0.05-0.2.

[0080] This structure enables the model to have a thermophysical adjustment channel, thereby realizing the sensitivity and interpretability of the temperature prediction behavior to the cooling state. As an example in a real engineering: in some grid monitoring data, the present application found that in two oil temperature rising events, the load change amplitude was similar, but once was due to the cooling system not restarting in time after a short power failure, and the other was due to the external environment temperature rising in normal operation. The traditional black box model cannot distinguish between the two scenarios, but the present patent method can correctly identify the first one as a cooling degradation type heat rise by the label change of S c (t), so as to make a hot spot temperature rise judgment earlier and trigger the warning mechanism in advance.

[0081] S4, based on the hot spot temperature prediction value and the cooling state label of the future period, matching with a preset thermal risk classification logic table, obtaining a risk level and generating a protection strategy, the protection strategy including cooling instructions, load adjustment and alarm behavior.

[0082] Specifically, the goal of this step is to classify and judge the system thermal risk and generate a response strategy based on the hot spot temperature prediction result and the cooling state label. This step will and S c (t) two structural variables are input into the rule judgment system to form an interpretable risk level output R(t) and an actual executable protection strategy A(t).

[0083] Inputs include: Model output from Step 3, representing the system's predicted future hotspot temperature; S c (t): Cooling status label from Step 2, representing the current cooling capability condition, taking values 0 (normal), 1 (mild degradation), 2 (severe degradation); T safe : Safety upper limit temperature determined by device model, for example 110℃ for oil-immersed transformer.

[0084] To determine the risk level, first calculate the over-limit ratio of the predicted temperature relative to the threshold ρ(t):

[0085]

[0086] Where: ρ(t) represents the proportion of temperature prediction exceeding the safety upper limit; if ρ(t) ≤ 0, it indicates that there is no over-temperature risk at present; if ρ(t) > 0, it indicates that there is a temperature prediction over-limit risk, which needs to be further judged. According to the predicted temperature rise intensity (i.e. the numerical value of ρ(t)) and the cooling state S c (t), the present invention designs a rule-driven thermal risk classification logic table, outputting the risk level R(t), whose value meanings are as follows:

[0087] R(t) = 0: No risk, ρ(t) ≤ 0;

[0088] R(t) = 1: Load impact type risk, ρ(t) > 0 and S c (t) = 0;

[0089] R(t) = 2: Cooling degradation type risk, ρ(t) > 0 and S c (t) = 2;

[0090] R(t) = 3: Coupling type risk, ρ(t) > 0 and S c (t) = 1.

[0091] The advantage of this structure is that the judgment process is reproducible, the logic is clear, and it is easy to adjust or manually revise, especially suitable for power system engineering deployment needs.

[0092] On this basis, the system outputs a response strategy A(t) that the system can recognize based on R(t), whose structure is a three-tuple:

[0093] A(t) = [cooling instruction, load strategy, alarm behavior]

[0094] For example:

[0095] If R(t) = 0, A(t) = [cooling maintenance, normal load, no alarm];

[0096] If R(t) = 1, A(t) = [Auxiliary cooling enabled, Load down, Alert for operation monitoring];

[0097] If R(t) = 2, A(t) = [Forced fan start, Load maintained, Alarm + manual maintenance];

[0098] If R(t) = 3, A(t) = [Cooling augmentation, Load down, Record and alarm].

[0099] These actions are not suggestive text descriptions, but can be configured as actual control system output instructions (such as through remote terminals, PLC controllers, SCADA systems), with physical executability. Finally, R(t) and A(t) as the output of this step will be passed into Step Five to perform actual control response, complete the key jump from intelligent prediction to engineering execution, and ensure that the entire patent system forms a prediction-judgment-response closed loop.

[0100] S5, execute the protection strategy, and issue control instructions to the cooling system and load scheduling module, and feedback the execution results to optimize the system closed loop.

[0101] Specifically, the goal of this step is to implement the response strategy A(t) output by the previous step into actual control actions U(t), forming an executable and feedbackable protection response mechanism, thereby realizing the system closed loop from prediction and judgment to control execution.

[0102] The input is A(t), a ternary structure control strategy, from Step Four, defined as: A(t) = [cooling control instruction, load adjustment strategy, alarm behavior];

[0103] Each element is a structured action selected from a finite action set, for example:

[0104] Cooling instruction: ∈ {maintain cooling, fan augmentation, oil pump start, standby cooling switch};

[0105] Load strategy: ∈ {normal operation, short-term load, fixed 10% load reduction, step-down load};

[0106] Alarm behavior: ∈ {no alarm, log record, manual prompt, forced alarm, emergency shutdown}.

[0107] This step constructs a protection execution function to map the strategy structure A(t) to device control signals U(t):

[0108] U(t) = Ψ(A(t), Ω dev )

[0109] Wherein, U(t) is a control layer execution instruction (such as a start fan command, a load limiting request, and a SCADA alarm trigger); A(t) is a structure response strategy input; Ω dev represents a device layer execution capability and constraint parameter (such as a certain type of fan cannot be repeatedly started and stopped, and a maximum load reduction limit is 20%); Ψ(·) is a mapping function, which is based on a rule table or control logic, and converts a strategy action into a hardware instruction to ensure that each type of action can be executed.

[0110] To enhance the robustness and practicality of control, a response confirmation mechanism is introduced in this step. If no confirmation feedback is received from the execution device within a set confirmation window τ ack , the system enters a protection fallback process and triggers a default action U fail (t):

[0111]

[0112] Wherein, τ ack is a response confirmation time limit, generally 10-30 seconds; U fail (t) can be a forced alarm + cooling standby channel activation + load limiting 20%, to ensure system stability.

[0113] In actual application, the system can distribute actions through the following interfaces:

[0114] The control layer output U(t) can be connected to a DCS control module, a SCADA system, an intelligent PLC, or an IEC61850 protocol interface;

[0115] The system timestamps and records each execution action and feedback as a data basis for subsequent operation and maintenance auditing and fault analysis.

[0116] Finally, the control execution result U(t) is written into a protection system log synchronously, completing the closed-loop response logic of the application, and realizing a truly intelligent and executable overheat protection system.

[0117] The embodiment of the application also provides a transformer overheat protection system, which comprises:

[0118] A data acquisition module is configured to acquire data acquired during transformer operation, including an oil top layer temperature, an ambient temperature, a load power, a cooling system operation power, and a near three-minute temperature difference sequence; a cooling activation degree index is obtained by performing time series data change rate analysis on the cooling system operation power; the oil top layer temperature, the ambient temperature, the load power, the cooling system operation power, the near three-minute temperature difference sequence, and the cooling activation degree index are preprocessed to form a standard input tensor; and the cooling activation degree index is used to reflect the current state of the cooling system.

[0119] The cooling state recognition module is configured to calculate a unit heat rise efficiency index according to the standard input tensor, obtain a dynamic residual index based on a deviation degree of the unit heat rise efficiency index and a historical mean value of the unit heat rise efficiency index, and generate a cooling state label in combination with the cooling activation index, the cooling state label being used to represent a degradation degree of the cooling system;

[0120] The prediction module is configured to input the standard input tensor and the cooling state label into a pre-constructed temperature prediction model, dynamically adjust a prediction path by integrating the cooling state label into a backbone network structure, and generate a hotspot temperature prediction value of a future period;

[0121] The strategy generation module is configured to obtain a risk level and generate a protection strategy by matching the hotspot temperature prediction value of the future period and the cooling state label with a preset thermal risk classification logic table, the protection strategy including a cooling instruction, a load adjustment, and an alarm behavior.

[0122] The execution module is configured to execute the protection strategy, issue a control instruction to a cooling system and a load scheduling module, and feed back an execution result to optimize a system closed loop.

[0123] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the system, device and unit described above can refer to the corresponding process in the foregoing method embodiments, which will not be repeated here.

[0124] In several embodiments provided in the present application, it should be understood that the disclosed system, device and method can be implemented in other ways. For example, the device embodiments described above are merely schematic, for example, the division of the units is only a logical function division, and actual implementation can have another division manner, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the units shown or discussed can be indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.

[0125] If the functions are implemented in the form of software function units and sold or used as independent products, they can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application essentially or the parts that contribute to the prior art or parts of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0126] Although the embodiments of the present application have been shown and described, those skilled in the art can understand that various changes, modifications, replacements and deformations can be made to the embodiments without departing from the principles and purposes of the present application, and the scope of the present application is defined by the claims and their equivalents.

Claims

1. A method for overheat protection of a transformer, characterized in that, The method includes: Data collected during transformer operation includes top oil temperature, ambient temperature, load power, cooling system operating power, and a three-minute temperature difference sequence. A cooling activation index is obtained by performing time-series data change rate analysis on the cooling system operating power. This data, combined with the top oil temperature, ambient temperature, load power, cooling system operating power, three-minute temperature difference sequence, and cooling activation index, is preprocessed to form a standard input tensor. The cooling activation index reflects the current state of the cooling system. The unit heat rise efficiency index is calculated based on the standard input tensor. A dynamic residual index is obtained based on the deviation of the unit heat rise efficiency index and its historical mean. Combined with the cooling activation index, a cooling status label is generated. The cooling status label is used to characterize the degree of degradation of the cooling system. The standard input tensor and cooling state label are input into the pre-built temperature prediction model. By integrating the cooling state label into the backbone network structure, the prediction path is dynamically adjusted to generate hotspot temperature prediction values ​​for future periods. Based on the predicted hotspot temperature and cooling status labels for the future period, the risk level is obtained and a protection strategy is generated by matching it with a preset thermal risk classification logic table. The protection strategy includes cooling commands, load adjustment and alarm behavior. The protection strategy is executed by sending control commands to the cooling system and load scheduling module, and the execution results are fed back to optimize the system closed loop.

2. The transformer overheat protection method according to claim 1, characterized in that, The preprocessing includes: performing time alignment and noise reduction on the multi-frequency sampled data, and generating a short-term temperature difference sequence through a sliding window. The temperature difference sequence of the past three minutes is used to reflect the dynamic offset trend between oil temperature and ambient temperature.

3. The transformer overheat protection method according to claim 1, characterized in that, The calculation of the unit heat rise efficiency index based on the standard input tensor specifically includes: The oil top layer temperature, ambient temperature, and load power are obtained, and performance index analysis is performed to obtain the unit heat rise efficiency index, which is used to measure the degree of temperature rise of the system under unit load.

4. The transformer overheat protection method according to claim 1, characterized in that, The cooling status labels include normal cooling status, slight degradation, and severe degradation.

5. A transformer overheat protection method according to claim 1, characterized in that, The structure of the temperature prediction model includes: The backbone modeling layer processes the standard input tensor with a lightweight recursive network to learn the underlying nonlinear dynamic laws; The cooling state adjustment layer integrates the cooling state label into the backbone network structure and dynamically adjusts the prediction path, enabling the model to adaptively select different internal prediction structures or adjust prediction weights under different cooling states. The prediction function of the temperature prediction model is expressed as: in, It predicts the winding hot spot temperature δ minutes from now; f θ (X t This is the baseline output of the temperature prediction model under default cooling conditions, with the input being the standard input tensor; It is the cooling state adjustment coefficient, when the cooling state label S c When (t) = 0, take the value 0, S c When (t) = 2, it takes the value 1, indicating the residual adjustment strength; R(X) t ) is a residual network used to represent the temperature rise offset modeling during cooling degradation.

6. A transformer overheat protection method according to claim 5, characterized in that, The loss function L of the temperature prediction model is expressed as: Among them, T hotspot (t+δ) represents the actual future hot spot temperature, which is derived from the field winding hot spot estimation system or the IEEE thermal model. The cooling state response regularization term measures the model's prediction sensitivity to changes in the cooling state label. The main loss is calculated using mean squared error to measure prediction bias; λ is the regularization weighting parameter.

7. A transformer overheat protection method according to claim 1, characterized in that, The predicted hotspot temperature and cooling status labels based on the future time period are matched with a preset thermal risk classification logic table to obtain the risk level and generate a protection strategy, specifically including: Obtain the safe upper limit temperature for the current equipment model, and calculate the over-limit ratio of the predicted temperature relative to the threshold based on the predicted hotspot temperature for future periods and the safe upper limit temperature; Based on the over-limit ratio of the predicted temperature relative threshold and the cooling status label, the risk level is obtained by matching it with the preset thermal risk classification logic table, including no risk, load impact risk, cooling degradation risk and coupling risk. A protection strategy is generated based on the risk level.

8. A transformer overheat protection method according to claim 1, characterized in that, The execution of the protection strategy involves sending control commands to the cooling system and load scheduling module, and feeding back the execution results to optimize the system closed loop. Specifically, this includes: The protection strategy is obtained and mapped to a device control signal through a mapping function; if the response is successful, the device control signal is executed; if no confirmation feedback is received from the executing device within the set confirmation window after the device control signal is sent, the system enters the protection rollback process and triggers the default action.

9. A transformer overheat protection system, characterized in that, The system includes: The data acquisition module is used to collect data during transformer operation, including top oil temperature, ambient temperature, load power, cooling system operating power, and the temperature difference sequence over the past three minutes. By performing time-series data change rate analysis on the cooling system operating power, a cooling activation index is obtained. This index, combined with the top oil temperature, ambient temperature, load power, cooling system operating power, the temperature difference sequence over the past three minutes, and the cooling activation index, undergoes preprocessing to form a standard input tensor. The cooling activation index reflects the current state of the cooling system. The cooling status identification module is used to calculate the unit heat rise efficiency index based on the standard input tensor, obtain the dynamic residual index based on the deviation of the unit heat rise efficiency index and its historical mean, and generate a cooling status label by combining the cooling activation index. The cooling status label is used to characterize the degree of degradation of the cooling system. The prediction module is used to input the standard input tensor and cooling state label into a pre-built temperature prediction model. By integrating the cooling state label into the backbone network structure, the prediction path is dynamically adjusted to generate hotspot temperature prediction values ​​for future periods. The strategy generation module is used to obtain the risk level and generate a protection strategy based on the predicted hot spot temperature value and cooling status label of the future time period by matching it with a preset thermal risk classification logic table. The protection strategy includes cooling instructions, load adjustment and alarm behavior. The execution module is used to execute the protection strategy, send control commands to the cooling system and load scheduling module, and provide feedback on the execution results to optimize the system closed loop.

Images