Intelligent monitoring and early warning method for power transmission and transformation equipment

Abstract

The invention relates to the field of early warning of power transmission and transformation equipment, in particular to an intelligent monitoring and early warning method for power transmission and transformation equipment, which comprises the following steps of: constructing a comprehensive fault vector by using acquired data, a discharge energy vector, a discharge temperature vector and an overheating risk coefficient; enabling the comprehensive fault vector to pass through a gas concentration prediction model based on an LSTNet architecture so as to generate a concentration prediction value of the fault gas molecules; and constructing a multi-task loss function based on a weighted loss item, a cross entropy loss item, a prediction uncertainty item and a Rogers ratio constraint item, and performing training optimization on the gas concentration prediction model through the multi-task loss function. According to the method, multi-scale prediction of short-term fluctuation characteristics and long-term trend characteristics of fault gas is taken into consideration, physical constraint of a model output result is realized by combining a multi-task loss function of a Rogers ratio constraint term, and early warning of an insulation fault of power transmission and transformation equipment in a germination stage is realized. And early and accurate monitoring and early warning of the power transmission and transformation equipment are realized.

Description

Technical Field

[0001] This invention relates to the field of early warning for power transmission and transformation equipment, and in particular to a method for intelligent monitoring and early warning of power transmission and transformation equipment. Background Technology

[0002] Power transmission and transformation equipment, especially oil-immersed power transformers and oil-immersed cables, are the core hubs of the power grid's power generation, transmission, transformation, and distribution chain. Their operating status directly determines the reliability and security of the power grid. According to State Grid's fault statistics, outages caused by faults such as insulation aging, partial discharge, and overheating account for over 60% of all outages of oil-immersed power transmission and transformation equipment. Once a fault occurs, it not only causes regional power outages but may also trigger a chain reaction of disasters such as equipment burnout and grid oscillations. Therefore, achieving early fault identification and accurate early warning for power transmission and transformation equipment is a core requirement for monitoring and early warning systems for such equipment.

[0003] Currently, dissolved gas analysis (DGA) is the most widely used core method in fault monitoring technology for oil-immersed power transmission and transformation equipment. Its principle is based on the fact that internal equipment faults cause the decomposition of insulating oil and solid insulating materials, producing characteristic fault gases such as C2H2, C2H6, CH4, C2H4, and H2. Furthermore, the gas composition is correlated with the fault type (discharge, overheating) and fault intensity (energy, temperature). However, the early warning and accuracy of existing DGA-based monitoring and early warning technologies need improvement, neglecting key requirements of actual operation and maintenance scenarios. For example, existing algorithm models do not incorporate classic fault diagnosis rules. For instance, the Rogers ratio is a classic rule for DGA fault diagnosis, effectively distinguishing the gas ratio ranges for different types of faults. The current models' lack of Rogers ratio inclusion may lead to predictions that contradict physical laws, such as predicting extremely high C2H2 concentrations without a discharge fault, resulting in poor generalization ability.

[0004] Therefore, how to address the shortcomings of DGA monitoring and early warning technology for power transmission and transformation equipment in terms of insufficient utilization of the correlation characteristics between fault gas components and insulation faults, as well as the significant limitations in multi-scale prediction and output result constraints, in order to meet the needs of power transmission and transformation equipment for early and accurate monitoring and early warning, is a technical problem that needs to be solved. Summary of the Invention

[0005] To this end, this invention provides an intelligent monitoring and early warning method for power transmission and transformation equipment. By constructing discharge energy vectors and discharge temperature vectors through the ratio relationships of multiple fault gas molecule data, abstract gas data is transformed into quantitative features characterizing the physical nature of the fault. This fully utilizes the correlation features between fault gas components and insulation faults. Furthermore, the LSTNet architecture model takes into account both short-term fluctuation characteristics and long-term trend characteristics of fault gases for multi-scale prediction. By combining Rogers ratio constraint terms with a multi-task loss function, physical constraints are imposed on the model output results. This enables early warning of insulation faults in power transmission and transformation equipment at the nascent stage, achieving early and accurate monitoring and early warning of power transmission and transformation equipment.

[0006] To achieve the above objectives, this invention proposes an intelligent monitoring and early warning method for power transmission and transformation equipment, comprising: The discharge energy vector of the insulation fault is constructed by summing multiple collected data of multiple fault gas molecules monitored by power transmission and transformation equipment. The discharge temperature vector is constructed based on the ratio of multiple collected data. The overheating risk coefficient is calculated based on the discharge energy vector and the discharge temperature vector. The collected data, the discharge energy vector, the discharge temperature vector and the overheating risk coefficient are combined to construct a comprehensive fault vector. The comprehensive fault vector is used to generate predicted concentration values ​​of fault gas molecules through a gas concentration prediction model based on the LSTNet architecture. A weighted loss term is constructed based on the temperature order of the faulty gas molecules. A cross-entropy loss term is constructed based on the predicted concentration value and the sample true concentration value of the faulty gas molecules. A prediction uncertainty term is constructed based on the statistical value of the sample true concentration value. A multi-task loss function is constructed based on the weighted loss term, the cross-entropy loss term, the prediction uncertainty term, and the Rogers ratio constraint term. The gas concentration prediction model is then trained and optimized using the multi-task loss function.

[0007] Furthermore, the gas concentration prediction model includes a multi-timescale parallel convolution unit, a gated recurrent unit, an attention unit, a regression unit, and a prediction output unit. The process of generating the predicted concentration values ​​and concentration prediction statistics of fault gas molecules through the gas concentration prediction model includes: The comprehensive fault vector is used to generate multi-timescale fused features by setting up multi-timescale parallel convolution units with dynamic convolution kernels. The multi-timescale fused features are passed through a gated loop unit to generate hidden states for multiple time steps; The hidden states of all time steps are amplified at critical fault moments through attention units to generate a temporally amplified fault vector; The time-series enhanced fault vector is regressed and mapped to all time steps by a regression unit to generate a linearly mapped fault vector. The linearly mapped fault vector and the time-enhanced fault vector are passed through the prediction output unit to generate a predicted concentration value of the fault gas molecules.

[0008] Furthermore, the multi-timescale parallel convolutional unit includes a global pooling layer, a kernel generator layer, and a convolution operation layer. The process of generating multi-timescale fused features through the multi-timescale parallel convolutional unit includes: The integrated fault vector is sequentially passed through a global pooling layer and a kernel generator layer to learn the time scale of discharge energy and discharge temperature changes in order to generate the dynamic convolution kernel. The comprehensive fault vector is subjected to dynamic time-scale convolution operations through multiple parallel branches of the convolution operation layer to generate feature vectors at multiple time scales. The feature vectors from all time scales are concatenated to generate the multi-time scale fused feature.

[0009] Furthermore, the attention unit includes an attention layer and a normalization layer, and the process of generating a temporal enhancement fault vector through the attention unit includes: The hidden states of the query time step and the hidden states of the final time step are fused and mapped through an attention layer to generate a time attention scalar for multiple time steps. The time attention scalars of all time steps are passed through a normalization layer to generate attention weights for multiple time steps; The attention weights and hidden states at all time steps are weighted and summed to generate the temporal augmentation fault vector.

[0010] Furthermore, the process of generating concentration prediction values ​​through the prediction output unit includes: The time-enhanced fault vector is passed through the fully connected layer of the prediction output unit to generate a nonlinear fault mapping vector; The linear fault mapping vector and the nonlinear fault mapping vector are superimposed to generate the concentration prediction value.

[0011] In particular, by using a global pooling layer and a kernel generator layer, the model automatically learns the time scale of changes in discharge energy and discharge temperature, and generates dynamic convolutional kernels that adapt to the current fault state. This avoids the suppression of features at non-adaptive scales by fixed convolutional kernels. Furthermore, by weighted summation of attention weights and hidden states, a time-enhanced fault vector is generated, which filters and amplifies features at critical moments and suppresses features at irrelevant stationary moments. This improves the model's sensitivity to fault inflection points. Finally, by superimposing linear and nonlinear mapping vectors, the model's prediction bias in abrupt fault scenarios is reduced.

[0012] Furthermore, the statistical values ​​include the mean, variance, and standard deviation. The process of constructing the prediction uncertainty term based on the statistical values ​​of the true concentration values ​​of the sample includes: The squared error of the predicted concentration value, the mean of the actual concentration value of the sample, and the variance of the actual concentration value of the sample is calculated to construct the mean prediction determination sub-item; The standard deviation of the true concentration values ​​of the sample is calculated through regularization to construct an uncertain constraint sub-term; The mean prediction determination sub-item and the uncertainty constraint sub-item are summed to construct the prediction uncertainty term.

[0013] Furthermore, the process of constructing a multi-task loss function based on the weighted loss term, cross-entropy loss term, prediction uncertainty term, and Rogers ratio constraint term includes: The Rogers ratio constraint term is constructed based on the ratio of the predicted concentrations of multiple faulty gas molecules not falling within the Rogers ratio range. The multi-task loss function is constructed by weighting and summing the weighted loss term, cross-entropy loss term, prediction uncertainty term, and Rogers ratio.

[0014] Furthermore, the fault gas molecules include ethane, ethylene, acetylene, and methane, and the discharge energy vector includes the total energy generated by the gas, the rate of energy generation by the gas, and the rate of acceleration of energy generation by the gas. The process of constructing the discharge energy vector of the insulation fault based on the sum of multiple collected data of multiple fault gas molecules includes: Based on the sum of the collected values ​​of ethane, ethylene, acetylene, and methane, the total energy generated by the generated gas is calculated. The first and second derivatives of the total energy generated by the gas in the time dimension are then calculated to determine the rate of energy generation and the acceleration rate of energy generation.

[0015] Furthermore, the discharge temperature vector includes an oil crack vector, a medium-temperature superheat vector, and a high-temperature vector. The process of constructing the discharge temperature vector based on the ratio of multiple collected data includes: An oil cracking vector is constructed based on the ratio of ethane to methane, a medium-temperature superheating vector is constructed based on the ratio of ethylene to ethane, and a high-temperature vector is constructed based on the ratio of acetylene to ethylene.

[0016] Furthermore, the process of calculating the overheating risk coefficient based on the discharge energy vector and discharge temperature vector includes: The difference between the intermediate-temperature overheating vector and the intermediate-temperature overheating threshold is used to construct a temperature risk term through a sigmoid function; The discharge energy vector is passed through a first ReLU function to construct a trend risk term; The gas-generated energy acceleration rate is passed through a second ReLU function to construct an acceleration risk term; The overheating risk coefficient is constructed by multiplying the temperature risk term, the trend risk term, and the acceleration risk term.

[0017] In particular, the total energy generated by the gas is constructed based on the sum of the collected values ​​of ethane, ethylene, acetylene and methane to reflect the total energy released by the fault and the rate and acceleration of its change. Based on the gas ratio mapping of the temperature range of insulating oil decomposition, the energy transformation and temperature characteristics of insulation faults in power transmission and transformation equipment are accurately quantified, and the overheating risk coefficient enables the quantitative calculation of insulation fault risk.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: the present invention constructs discharge energy vector and discharge temperature vector by means of the ratio relationship of multiple fault gas molecule collection data, transforms abstract gas data into quantitative features that characterize the physical nature of the fault, realizes full utilization of the correlation features between fault gas components and insulation faults, and uses the LSTNet architecture model to take into account the multi-scale prediction of fault gas short-term fluctuation characteristics and long-term trend characteristics. By combining Rogers ratio constraint term multi-task loss function to realize physical constraints on model output results, so as to realize early warning of insulation faults in power transmission and transformation equipment at the bud stage, and realize early and accurate monitoring and early warning of power transmission and transformation equipment.

[0019] In particular, this invention automatically learns the time scale of discharge energy and discharge temperature changes through a global pooling layer and a kernel generator layer, generating a dynamic convolution kernel that adapts to the current fault state. This avoids the suppression of features at non-adaptive scales by fixed convolution kernels. Furthermore, by weighted summation of attention weights and hidden states, a time-enhanced fault vector is generated, which filters and amplifies features at critical moments and suppresses features at irrelevant stationary moments. This improves the model's sensitivity to fault inflection points. Finally, by superimposing linear and nonlinear mapping vectors, the prediction bias of the model in abrupt fault scenarios is reduced.

[0020] In particular, this invention constructs the total energy of gas generation based on the sum of the collected values ​​of ethane, ethylene, acetylene and methane to reflect the total energy scale of fault release and the rate and acceleration of its change. Based on the gas ratio mapping of the temperature range of insulating oil decomposition, it realizes the accurate quantification of energy transformation and temperature characteristics of insulation faults in power transmission and transformation equipment, and the overheating risk coefficient enables the quantitative calculation of insulation fault risk. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the intelligent monitoring and early warning method for power transmission and transformation equipment according to an embodiment of the present invention. Figure 2 This is a flowchart illustrating the gas concentration prediction model of the intelligent monitoring and early warning method for power transmission and transformation equipment according to an embodiment of the present invention. Figure 3This is a schematic diagram illustrating the construction process of the multi-task loss function in the intelligent monitoring and early warning method for power transmission and transformation equipment according to an embodiment of the present invention; Figure 4 This is a schematic diagram illustrating the process of constructing a comprehensive fault vector in the intelligent monitoring and early warning method for power transmission and transformation equipment according to an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0023] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0024] It should be noted that in the description of this invention, the terms "upper", "lower", "left", "right", "inner", "outer", etc., which indicate directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.

[0025] Furthermore, it should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0026] like Figures 1 to 4 As shown, this invention provides an intelligent monitoring and early warning method for power transmission and transformation equipment. It constructs discharge energy and discharge temperature vectors by using the ratio relationships of multiple fault gas molecule data, transforming abstract gas data into quantitative features characterizing the physical nature of the fault. This fully utilizes the correlation features between fault gas components and insulation faults. Furthermore, it employs an LSTNet architecture model to perform multi-scale predictions that consider both short-term fluctuations and long-term trends of the fault gas. By combining a Rogers ratio constraint term with a multi-task loss function, it achieves physical constraints on the model output, enabling early warning of insulation faults in power transmission and transformation equipment at their nascent stage, thus realizing early and accurate monitoring and early warning of power transmission and transformation equipment.

[0027] like Figure 1As shown in the figure, this embodiment proposes an intelligent monitoring and early warning method for power transmission and transformation equipment, including: The discharge energy vector of the insulation fault is constructed by summing multiple collected data of multiple fault gas molecules monitored by power transmission and transformation equipment. The discharge temperature vector is constructed based on the ratio of multiple collected data. The overheating risk coefficient is calculated based on the discharge energy vector and the discharge temperature vector. The collected data, the discharge energy vector, the discharge temperature vector and the overheating risk coefficient are combined to construct a comprehensive fault vector. The comprehensive fault vector is used to generate predicted concentration values ​​of fault gas molecules through a gas concentration prediction model based on the LSTNet architecture. A weighted loss term is constructed based on the temperature order of the faulty gas molecules. A cross-entropy loss term is constructed based on the predicted concentration value and the sample true concentration value of the faulty gas molecules. A prediction uncertainty term is constructed based on the statistical value of the sample true concentration value. A multi-task loss function is constructed based on the weighted loss term, the cross-entropy loss term, the prediction uncertainty term, and the Rogers ratio constraint term. The gas concentration prediction model is then trained and optimized using the multi-task loss function.

[0028] In particular, through the multi-task loss function and the gas concentration prediction model based on the LSTNet architecture, accurate prediction of multi-scale features can be achieved. During implementation, an early warning of excessive gas concentration in faulty systems can be generated 3 days in advance. Maintenance personnel can formulate shutdown and maintenance plans in advance, avoiding emergency repairs caused by traditional early warnings that exceed the standard.

[0029] like Figure 2 As shown, the gas concentration prediction model further includes a multi-timescale parallel convolution unit, a gated recurrent unit, an attention unit, a regression unit, and a prediction output unit. The process of generating the predicted concentration values ​​and concentration prediction statistics of fault gas molecules through the gas concentration prediction model includes: The comprehensive fault vector is used to generate multi-timescale fused features by setting up multi-timescale parallel convolution units with dynamic convolution kernels. The multi-timescale fused features are passed through a gated loop unit to generate hidden states for multiple time steps; The hidden states of all time steps are amplified at critical fault moments through attention units to generate a temporally amplified fault vector; The time-series enhanced fault vector is regressed and mapped to all time steps by a regression unit to generate a linearly mapped fault vector. The linearly mapped fault vector and the time-enhanced fault vector are passed through the prediction output unit to generate a predicted concentration value of the fault gas molecules.

[0030] In particular, by processing the comprehensive fault vector through dynamic convolution kernels, the feature scales of both short-term sudden faults and long-term cumulative faults can be covered simultaneously. Short-term sudden faults include gas concentration fluctuations during partial discharge, while long-term cumulative faults include gradual changes in gas concentration under overheating faults.

[0031] like Figure 2 As shown, the multi-timescale parallel convolutional unit further includes a global pooling layer, a kernel generator layer, and a convolution operation layer. The process of generating multi-timescale fused features through the multi-timescale parallel convolutional unit includes: The integrated fault vector is sequentially passed through a global pooling layer and a kernel generator layer to learn the time scale of discharge energy and discharge temperature changes in order to generate the dynamic convolution kernel. The comprehensive fault vector is subjected to dynamic time-scale convolution operations through multiple parallel branches of the convolution operation layer to generate feature vectors at multiple time scales. The feature vectors from all time scales are concatenated to generate the multi-time scale fused feature.

[0032] In particular, traditional multi-scale convolution relies on manually preset fixed scales. In this embodiment, global pooling is used to perceive the overall energy change rate and temperature rise and fall trend of the fault. The kernel generator dynamically outputs convolution kernels with matching scales, so that the convolution kernels can accurately adapt to the time characteristics of different faults. For example, the energy change of discharge faults is fast and the time scale is short, avoiding the mismatch between fixed kernels and fault characteristics.

[0033] Specifically, the process of generating multi-timescale fused features can be represented as:

[0034] In the formula, Let T represent the global average pooling feature vector, and let T represent all time steps of the comprehensive fault vector. This represents the i-th element of the composite fault vector. This represents the kernel weights of the dynamic convolution kernel in the k-th parallel branch. This indicates a tensor reshaping operation. This represents the kernel generator for the k-th parallel branch, with a fully connected network serving as the kernel generator. This represents the eigenvector of the k-th parallel branch, where k = 1, 2, 3. The first to third branches represent the feature vectors at the time scales of short-term mutations, medium-term trends, and long-term trends, respectively. express Activation function Represents the combined fault vector. This represents the bias term of the dynamic convolution kernel in the k-th parallel branch. This represents the eigenvector of the fourth parallel branch. AvgPooling represents the background features used to provide the overall fault vector. This represents a vector concatenation operation. This represents the fusion features across multiple time scales.

[0035] Furthermore, the attention unit includes an attention layer and a normalization layer, and the process of generating a temporal enhancement fault vector through the attention unit includes: The hidden states of the query time step and the hidden states of the final time step are fused and mapped through an attention layer to generate a time attention scalar for multiple time steps. The time attention scalars of all time steps are passed through a normalization layer to generate attention weights for multiple time steps; The attention weights and hidden states at all time steps are weighted and summed to generate the temporal augmentation fault vector.

[0036] In particular, the hidden state of the final time step integrates the full temporal information of the fault from its initial stage to the current stage, serving as a global summary of the fault process, while the hidden state of the query time step only reflects the local features of a single moment. When the two are fused and mapped, the characteristics of key time steps such as early warning moments and sudden deterioration moments of the fault can be accurately identified.

[0037] Specifically, the process of generating time-enhanced fault vectors can be represented as:

[0038] In the formula, This represents the hidden state at multiple time steps generated by the gated recurrent unit (GRU). Representing the first to the last hidden state Elements at each time step This represents the time attention scalar at the i-th time step. This represents the transpose of the attention weight vector. , The linear transformation weights of the historical hidden states and the convolution linear transformation weights of the final hidden states are respectively... express Activation function This represents the hidden state at the i-th time step. This represents the hidden state (Query) at the final time step. This represents a learnable bias term. express The activation function is used for the temporal attention scalar from the first time step to the final time step. Normalize, This represents a timing-enhanced fault vector.

[0039] Furthermore, the process of generating concentration prediction values ​​through the prediction output unit includes: The time-enhanced fault vector is passed through the fully connected layer of the prediction output unit to generate a nonlinear fault mapping vector; The linear fault mapping vector and the nonlinear fault mapping vector are superimposed to generate the concentration prediction value.

[0040] In particular, by using a global pooling layer and a kernel generator layer, the model automatically learns the time scale of changes in discharge energy and discharge temperature, and generates dynamic convolutional kernels that adapt to the current fault state. This avoids the suppression of features at non-adaptive scales by fixed convolutional kernels. Furthermore, by weighted summation of attention weights and hidden states, a time-enhanced fault vector is generated, which filters and amplifies features at critical moments and suppresses features at irrelevant stationary moments. This improves the model's sensitivity to fault inflection points. Finally, by superimposing linear and nonlinear mapping vectors, the model's prediction bias in abrupt fault scenarios is reduced.

[0041] Specifically, the process of generating concentration prediction values ​​and concentration prediction statistics can be represented as follows:

[0042] In the formula, Represents a linearly mapped fault vector. Represents a nonlinear fault mapping vector. This indicates the number of time steps removed from the final time step. This represents the timing-enhanced fault vector at time step tk. This represents the lag autoregressive weight matrix at the k-th time step. This represents the bias term of the autoregressive component. This represents the sequence of predicted concentration values ​​at time t. This indicates that for the future time step t+1 to the t-th... The concentration prediction value sequence of time steps includes concentration prediction value elements, each of which includes multiple predicted concentration values ​​of fault gas molecules C2H2, C2H6, CH4, and C2H4.

[0043] Specifically, the hyperparameters of the gas concentration prediction model based on the LSTNet architecture include: a time window length of 30 days for input data, a prediction step size of 7 days for output data, multi-scale convolutional kernel sizes of 2, 4, 8, and 16, a GRU hidden layer dimension of 64, an attention mechanism dimension of 32, a training batch size of 32, an initial learning rate of 1e-3, and the Adam optimizer is used for training optimization.

[0044] like Figure 3As shown, further, the statistical values ​​include the mean and standard deviation. The process of constructing the prediction uncertainty term based on the statistical values ​​of the true concentration values ​​of the sample includes: The squared error of the predicted concentration value, the mean of the actual concentration value of the sample, and the standard deviation of the actual concentration value of the sample is calculated to construct the mean prediction determination sub-item; The standard deviation of the true concentration values ​​of the sample is calculated through regularization to construct an uncertain constraint sub-term; The mean prediction determination sub-item and the uncertainty constraint sub-item are summed to construct the prediction uncertainty term.

[0045] In particular, by calculating the concentration prediction value, the mean of the actual concentration of the sample, and the variance of the actual concentration of the sample, the prediction bias is bound to the overall distribution characteristics of the sample, so as to avoid the limitation of single-point bias.

[0046] Specifically, the prediction uncertainty can be expressed as:

[0047] In the formula, This represents the uncertainty in the prediction, and N represents the sample size. This represents the predicted concentration value. This represents the mean of the true concentration values ​​of the sample. The standard deviation of the true concentration value of the sample. This represents the calculation of the squared Euclidean distance. This indicates that the mean forecast determines the sub-item, which is used to calculate the squared error. This represents an uncertain constraint term used to regularize the standard deviation.

[0048] Furthermore, the process of constructing a multi-task loss function based on the weighted loss term, cross-entropy loss term, prediction uncertainty term, and Rogers ratio constraint term includes: The Rogers ratio constraint term is constructed based on the ratio of the predicted concentrations of multiple faulty gas molecules not falling within the Rogers ratio range. The multi-task loss function is constructed by weighting and summing the weighted loss term, cross-entropy loss term, prediction uncertainty term, and Rogers ratio.

[0049] In particular, by using the Rogers ratio constraint term, the objective law of gas molecule ratio under insulation fault is forced to be learned by the model, thus solving the defect of traditional loss functions that emphasize error and neglect physical logic.

[0050] Specifically, the faulty gas molecules in the plurality of molecules include C2H2, C2H6, CH4, and C2H4, wherein C2H2 and C2H4 are high-temperature faulty gases, and CH4 is a low-temperature faulty gas. The temperature order of the faulty gas molecules is as follows: The weighted loss term is:

[0051] In the formula, temperature This represents the temperature weight of the gas in the i-th temperature order. , Let i and j represent the order of the faulty gas molecule, respectively. This represents the loss term, where N represents the sample size. These represent the predicted concentration and the actual concentration of the sample, respectively. This represents the Frobenius norm.

[0052] Specifically, the cross-entropy loss term can be expressed as:

[0053] In the formula, This represents the cross-entropy loss term, where C represents the total number of elements in the concentration prediction sequence. This represents the i-th concentration prediction value in the concentration prediction value sequence. This represents the i-th true concentration value in the sequence of true concentration values ​​of the sample.

[0054] Specifically, the Rogers ratio constraint term can be expressed as:

[0055] In the formula, These represent multiple ratios of the predicted concentrations of various faulty gas molecules. This represents a rule constraint on the range of fault gas molecule concentration ratios in the Rogers ratio method, namely, a constraint ratio r greater than L and less than U. For Rogers ratio regulations that only specify the maximum or minimum value, only this constraint needs to be set. or , This represents the Rogers ratio constraint term, which is the summation of multiple rule constraints to perform an AND operation on the rule constraints. The Rogers ratio method rules include partial discharge rules, which are as follows: .

[0056] Specifically, the multi-task loss function can be expressed as: In the formula, These represent the weighted loss term, cross-entropy loss term, prediction uncertainty term, and Rogers ratio, respectively. , This indicates the weighting, which is preferably 1.0, 0.2, or 0.1.

[0057] like Figure 4As shown, further, the fault gas molecules include ethane (C2H6), ethylene (C2H4), acetylene (C2H2), and methane (CH4), and the discharge energy vector includes the total energy generated by the gas, the rate of energy generated by the gas, and the rate of acceleration of energy generated by the gas. The process of constructing the discharge energy vector of the insulation fault based on the sum of multiple collected data of multiple fault gas molecules includes: Based on the sum of the collected values ​​of ethane, ethylene, acetylene, and methane, the total energy generated by the generated gas is calculated. The first and second derivatives of the total energy generated by the gas in the time dimension are then calculated to determine the rate of energy generation and the acceleration rate of energy generation.

[0058] In particular, by summing the collected values ​​of ethane, ethylene, acetylene, and methane to generate the total energy of the generated gases, the overall intensity of oil cracking caused by the fault can be directly reflected, breaking through the limitations of traditional methods that rely on the concentration of a single gas to judge the scale of the fault.

[0059] Specifically, the process of generating total energy from a gas can be represented as:

[0060] In the formula, This represents the total energy produced by the gas. These represent the collected data for methane, ethylene, acetylene, and ethane, respectively.

[0061] like Figure 4 As shown, further, the discharge temperature vector includes an oil crack vector, a medium-temperature superheat vector, and a high-temperature vector. The process of constructing the discharge temperature vector based on the ratio of multiple collected data includes: An oil cracking vector is constructed based on the ratio of ethane to methane, a medium-temperature superheating vector is constructed based on the ratio of ethylene to ethane, and a high-temperature vector is constructed based on the ratio of acetylene to ethylene.

[0062] In particular, by using the low-temperature vector of the oil crack vector and the medium-temperature and high-temperature vector of the medium-temperature overheat vector, a collaborative diagnosis of insulation faults in oil-immersed power transmission and transformation equipment with full temperature coverage was achieved.

[0063] Specifically, the process of constructing the oil crack vector, the intermediate-temperature superheat vector, and the high-temperature vector can be expressed as:

[0064] In the formula, These represent the oil crack vector, the intermediate temperature superheat vector, and the high temperature vector, respectively. These represent the collected data for methane, ethylene, acetylene, and ethane, respectively. This represents the minimum value to prevent the denominator from being zero.

[0065] Furthermore, the process of calculating the overheating risk coefficient based on the discharge energy vector and discharge temperature vector includes: The difference between the intermediate-temperature overheating vector and the intermediate-temperature overheating threshold is used to construct a temperature risk term through a sigmoid function; The discharge energy vector is passed through a first ReLU function to construct a trend risk term; The gas-generated energy acceleration rate is passed through a second ReLU function to construct an acceleration risk term; The overheating risk coefficient is constructed by multiplying the temperature risk term, the trend risk term, and the acceleration risk term.

[0066] Specifically, the process of constructing the overheating risk coefficient can be expressed as:

[0067] In the formula, Indicates the overheating risk factor. Represents the intermediate temperature superheat vector. This indicates the intermediate temperature overheating threshold, preferably 150℃. , These represent the discharge energy vector and the gas generation energy acceleration rate, respectively.

[0068] In particular, the total energy generated by the gas is constructed based on the sum of the collected values ​​of ethane, ethylene, acetylene and methane to reflect the total energy released by the fault and the rate and acceleration of its change. Based on the gas ratio mapping of the temperature range of insulating oil decomposition, the energy transformation and temperature characteristics of insulation faults in power transmission and transformation equipment are accurately quantified, and the overheating risk coefficient enables the quantitative calculation of insulation fault risk.

[0069] In this embodiment, discharge energy and discharge temperature vectors are constructed by the ratio relationships of multiple fault gas molecule data. This transforms abstract gas data into quantitative features characterizing the physical nature of the fault, fully utilizing the correlation features between fault gas components and insulation faults. The LSTNet architecture model considers both short-term fluctuations and long-term trends of the fault gas in multi-scale prediction. A multi-task loss function incorporating Rogers ratio constraints physically constrains the model output, enabling early warning of insulation faults in power transmission and transformation equipment at their nascent stage, achieving early and accurate monitoring and early warning. Through global pooling layers and kernel generator layers, the model automatically learns the time scale of discharge energy and temperature changes, generating dynamic convolutional kernels adapted to the current fault state. This avoids the suppression of features at non-adaptive scales by fixed convolutional kernels. A time-enhanced fault vector is generated through the weighted summation of attention weights and hidden states, filtering and amplifying key moment features while suppressing irrelevant stationary moment features. This improves the model's sensitivity to fault inflection points. Furthermore, the superposition of linear and nonlinear mapping vectors reduces prediction bias in abrupt fault scenarios. By constructing the total energy generated by gases based on the sum of the collected values ​​of ethane, ethylene, acetylene, and methane, the total energy released by the fault and the rate and acceleration of its change are reflected. Based on the gas ratio mapping of the temperature range of insulating oil decomposition, the energy transformation and temperature characteristics of insulation faults in power transmission and transformation equipment are accurately quantified, and the overheating risk coefficient enables the quantitative calculation of insulation fault risk.

[0070] Those skilled in the art will recognize that the modules and algorithm steps of the examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0071] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0072] The above description is merely a preferred embodiment of the present invention and is not intended to limit the 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 method for intelligent monitoring and early warning of power transmission and transformation equipment, characterized in that, include: The discharge energy vector of the insulation fault is constructed by summing multiple collected data of multiple fault gas molecules monitored by power transmission and transformation equipment. The discharge temperature vector is constructed based on the ratio of multiple collected data. The overheating risk coefficient is calculated based on the discharge energy vector and the discharge temperature vector. The collected data, the discharge energy vector, the discharge temperature vector and the overheating risk coefficient are combined to construct a comprehensive fault vector. The comprehensive fault vector is used to generate predicted concentration values ​​of fault gas molecules through a gas concentration prediction model based on the LSTNet architecture. A weighted loss term is constructed based on the temperature order of the faulty gas molecules. A cross-entropy loss term is constructed based on the predicted concentration value and the sample true concentration value of the faulty gas molecules. A prediction uncertainty term is constructed based on the statistical value of the sample true concentration value. A multi-task loss function is constructed based on the weighted loss term, the cross-entropy loss term, the prediction uncertainty term, and the Rogers ratio constraint term. The gas concentration prediction model is then trained and optimized using the multi-task loss function.

2. The intelligent monitoring and early warning method for power transmission and transformation equipment according to claim 1, characterized in that, The gas concentration prediction model includes a multi-timescale parallel convolutional unit, a gated recurrent unit, an attention unit, a regression unit, and a prediction output unit. The process of generating the predicted concentration values ​​and statistical values ​​of fault gas molecules through the gas concentration prediction model includes: The comprehensive fault vector is used to generate multi-timescale fused features by setting up multi-timescale parallel convolution units with dynamic convolution kernels. The multi-timescale fused features are passed through a gated loop unit to generate hidden states for multiple time steps; The hidden states of all time steps are amplified at critical fault moments through attention units to generate temporally amplified fault vectors; The time-series enhanced fault vector is regressed and mapped to all time steps by a regression unit to generate a linearly mapped fault vector. The linearly mapped fault vector and the time-enhanced fault vector are passed through the prediction output unit to generate a predicted concentration value of the fault gas molecules.

3. The intelligent monitoring and early warning method for power transmission and transformation equipment according to claim 2, characterized in that, The multi-timescale parallel convolutional unit includes a global pooling layer, a kernel generator layer, and a convolution operation layer. The process of generating multi-timescale fused features through the multi-timescale parallel convolutional unit includes: The integrated fault vector is sequentially passed through a global pooling layer and a kernel generator layer to learn the time scale of discharge energy and discharge temperature changes in order to generate the dynamic convolution kernel. The comprehensive fault vector is subjected to dynamic time-scale convolution operations through multiple parallel branches of the convolution operation layer to generate feature vectors at multiple time scales. The feature vectors from all time scales are concatenated to generate the multi-time scale fused feature.

4. The intelligent monitoring and early warning method for power transmission and transformation equipment according to claim 2, characterized in that, The attention unit includes an attention layer and a normalization layer. The process of generating a temporal augmentation fault vector through the attention unit includes: The hidden states of the query time step and the hidden states of the final time step are fused and mapped through an attention layer to generate a time attention scalar for multiple time steps. The time attention scalars of all time steps are passed through a normalization layer to generate attention weights for multiple time steps; The attention weights and hidden states at all time steps are weighted and summed to generate the temporal augmentation fault vector.

5. The intelligent monitoring and early warning method for power transmission and transformation equipment according to claim 2, characterized in that, The process of generating concentration prediction values ​​through the prediction output unit includes: The time-enhanced fault vector is passed through the fully connected layer of the prediction output unit to generate a nonlinear fault mapping vector; The linear fault mapping vector and the nonlinear fault mapping vector are superimposed to generate the concentration prediction value.

6. The intelligent monitoring and early warning method for power transmission and transformation equipment according to claim 1, characterized in that, The statistical values ​​include the mean, variance, and standard deviation. The process of constructing the prediction uncertainty term based on the statistical values ​​of the true concentration values ​​of the sample includes: The squared error of the predicted concentration value, the mean of the actual concentration value of the sample, and the variance of the actual concentration value of the sample is calculated to construct the mean prediction determination sub-item; The standard deviation of the true concentration values ​​of the sample is calculated through regularization to construct an uncertain constraint sub-term; The mean prediction determination sub-item and the uncertainty constraint sub-item are summed to construct the prediction uncertainty term.

7. The intelligent monitoring and early warning method for power transmission and transformation equipment according to claim 1, characterized in that, The process of constructing a multi-task loss function based on the weighted loss term, cross-entropy loss term, prediction uncertainty term, and Rogers ratio constraint term includes: The Rogers ratio constraint term is constructed based on the ratio of the predicted concentrations of multiple faulty gas molecules not falling within the Rogers ratio range. The multi-task loss function is constructed by weighting and summing the weighted loss term, cross-entropy loss term, prediction uncertainty term, and Rogers ratio.

8. The intelligent monitoring and early warning method for power transmission and transformation equipment according to any one of claims 1 to 7, characterized in that, The fault gas molecules include ethane, ethylene, acetylene, and methane. The discharge energy vector includes the total energy generated by the gas, the rate of energy generation by the gas, and the rate of acceleration of energy generation by the gas. The process of constructing the discharge energy vector of the insulation fault based on the sum of multiple collected data of multiple fault gas molecules includes: Based on the sum of the collected values ​​of ethane, ethylene, acetylene, and methane, the total energy generated by the generated gas is calculated. The first and second derivatives of the total energy generated by the gas in the time dimension are then calculated to determine the rate of energy generation and the acceleration rate of energy generation.

9. The intelligent monitoring and early warning method for power transmission and transformation equipment according to claim 8, characterized in that, The discharge temperature vector includes an oil crack vector, a medium-temperature superheat vector, and a high-temperature vector. The process of constructing the discharge temperature vector based on the ratio of multiple collected data includes: An oil cracking vector is constructed based on the ratio of ethane to methane, a medium-temperature superheating vector is constructed based on the ratio of ethylene to ethane, and a high-temperature vector is constructed based on the ratio of acetylene to ethylene.

10. The intelligent monitoring and early warning method for power transmission and transformation equipment according to claim 9, characterized in that, The process of calculating the overheating risk coefficient based on the discharge energy vector and discharge temperature vector includes: The difference between the intermediate-temperature overheating vector and the intermediate-temperature overheating threshold is used to construct a temperature risk term through a sigmoid function; The discharge energy vector is passed through a first ReLU function to construct a trend risk term; The gas-generated energy acceleration rate is passed through a second ReLU function to construct an acceleration risk term; The overheating risk coefficient is constructed by multiplying the temperature risk term, the trend risk term, and the acceleration risk term.

Images