Switch cabinet condensation prediction method fusing thermodynamic characteristics and time series deep learning

Abstract

The application provides a switch cabinet condensation prediction method combining thermodynamic characteristics and time series deep learning, relates to the technical field of switch cabinet condensation prediction based on deep learning, and aims to optimize the switch cabinet condensation prediction method by combining key thermodynamic characteristics, and comprises the following steps: continuously collecting original data and performing data preprocessing; obtaining various thermodynamic characteristic data by thermodynamic characteristic extraction, and then extracting various dynamic interaction features between the characteristic data; obtaining a characteristic vector matrix by characteristic coding and splicing; based on the characteristic vector matrix, a sliding time window method is used to construct the coding splicing vector into a reconstructed coding splicing vector; a time series deep model is constructed, the time series deep model is used for realizing deep fusion of thermodynamic characteristics and time series information, and a prediction result is generated according to the input reconstructed coding splicing vector. The application has the advantages of improving the prediction accuracy, generalization ability and adaptability.

Description

Technical Field

[0001] This invention relates to the field of deep learning-based condensation prediction technology for switchgear, and more specifically, to a method for predicting condensation in switchgear that integrates thermodynamic characteristics and time-series deep learning. Background Technology

[0002] Switchgear, as a crucial piece of equipment in power systems, undertakes vital functions such as high-voltage circuit safety isolation, power distribution, and overload and short-circuit protection. The safety of its operating environment directly impacts the stability and reliability of the power system. However, condensation inside switchgear is a significant cause of decreased insulation performance, accelerated corrosion, and malfunctions. Accurately predicting condensation in switchgear helps in taking preventative protective measures, ensuring safe equipment operation, and reducing power outage risks and maintenance costs.

[0003] Currently, research on condensation prediction in switchgear is still in its early stages, and existing technical solutions have significant limitations: First, most methods rely on empirical rules or simple statistical models, making it difficult to deeply characterize the complex mechanisms and dynamic evolution of multi-physics coupling involved in condensation formation; second, some studies only model based on single physical quantities such as ambient temperature and humidity, ignoring key thermodynamic characteristics (such as dew point temperature and thermal conductivity) and their temporal evolution patterns in the condensation process, resulting in low prediction accuracy and poor stability; third, traditional methods lack the ability to deeply mine time series data, failing to effectively capture the temporal dependence and potential nonlinear relationships in the condensation formation process, greatly limiting the model's generalization ability and practical engineering application effects.

[0004] Therefore, it is necessary to optimize the condensation prediction method for switchgear by combining key thermodynamic characteristics in order to improve the accuracy, generalization ability and adaptability of the prediction. Summary of the Invention

[0005] The purpose of this invention is to provide a method for predicting condensation in switchgear that integrates thermodynamic features and temporal deep learning, which improves the accuracy, generalization ability and adaptability of the prediction.

[0006] This invention is achieved through the following technical solution: A method for predicting condensation in switchgear by integrating thermodynamic features and temporal deep learning includes the following steps: Raw data is continuously collected and preprocessed. The raw data includes ambient temperature, relative humidity, atmospheric pressure, switch cabinet surface temperature, switch cabinet internal temperature, historical load, and historical condensation status. Multiple thermodynamic feature data are obtained through thermodynamic feature extraction, and then multiple dynamic interaction features between feature data are extracted. The feature vector matrix is ​​obtained through feature encoding and concatenation. The feature vector matrix includes encoded and concatenated vectors of multiple raw data acquisition times within a time series length. The encoded and concatenated vectors include multiple mapping features obtained by mapping multiple concatenated features based on the corresponding times. The multiple concatenated features include all raw data, thermodynamic feature data, and dynamic interaction features. Based on the feature vector matrix, the sliding time window method is used to construct the reconstructed encoded concatenation vector from the encoded concatenation vector; A temporal deep model is constructed to achieve deep fusion of thermodynamic features and temporal information, and to generate prediction results based on the input reconstructed encoding concatenation vector.

[0007] Preferably, the method for obtaining multiple thermodynamic feature data through thermodynamic feature extraction is as follows: Extraction dew point temperature : ; Where T is the ambient temperature and RH is the relative humidity. and It is a constant, with values ​​of 17.625 and 243.04 respectively; Extracting thermal conductivity : ; in, The temperature coefficient of thermal conductivity The standard thermal conductivity coefficient of the cabinet material at 25℃; Extraction wet bulb temperature : ; Extracting temperature and humidity features: Calculating the change in ambient temperature between two adjacent raw data acquisition times. Relative humidity change And the difference between the surface temperature of the switch cabinet and the ambient temperature. ; Extracting thermal performance indicators : ; Where d is the thickness of the cabinet material. The thermal conductivity of the cabinet material. This refers to the specific heat capacity of the cabinet material.

[0008] Preferably, the method for extracting multiple dynamic interaction features between feature data is as follows: Extracting temperature-dew point interaction features: ; Extracting humidity-dew point interaction features: ; Extracting heat conduction-gradient interaction features: ; Extracting the humidity-heat conduction interaction characteristics: ; Extracting temperature-humidity coupling features: ; Extracting the dew point-temperature gradient interaction features: .

[0009] Preferably, the method for obtaining the feature vector matrix through feature encoding and concatenation is as follows: The continuous data in the initial data are normalized using the Min-Max normalization method; The raw data, thermodynamic feature data, and dynamic interaction features at each time point are concatenated to obtain the concatenated vector matrix for that time point. The concatenated vector matrix is ​​then concatenated in chronological order to obtain the multidimensional time series feature matrix. : ; ; in, The length of the time series. To find the transpose of a matrix, D is the feature dimension. Let be the concatenated vector matrix at time t. arrive These are the M types of splicing features used for splicing at time t; Features for splicing A mapping method is used to uniformly transform the feature space to the same dimension, m=1,2,…,M:

[0010] in, The mapping feature is the m-th type of splicing feature. This is the weight matrix. It is the bias vector, and ReLU is the activation function; Perform coded feature concatenation representation: ; in, This is the encoded concatenated vector obtained by concatenating the encoded features at time t. This is a join function.

[0011] Preferably, the method for constructing time series samples using the sliding time window method is as follows: Set the time window length to L, and concatenate the encoded vectors at time t in the feature vector matrix. Construct as a reconstructed encoding concatenation vector : .

[0012] Preferably, the method for constructing a temporal deep model includes: The reconstructed encoding concatenation vector is input into the feature fusion layer. The importance weight of each concatenation feature is calculated by the feature weighting module based on the attention mechanism. The reconstructed encoding concatenation vector is then weighted and fused to obtain the temporal window feature. Based on the temporal window feature, the ability to capture long temporal dependencies is enhanced in the Transformer encoder; Global average pooling is performed on the output of the Transformer encoder in the fully connected layer to obtain the global feature vector; In the output layer, the global feature vector is mapped to the condensation probability.

[0013] Preferably, the method for constructing the feature fusion layer is as follows: Obtain the initial attention score for the m-th splicing feature at time t. : ; in, It is a row vector, and tanh is a non-linear activation function. and Let b be the linear transformation matrix, and b be the bias term. Initial attention score for all spliced ​​features at time t Normalization is performed to obtain the importance weights of the m-th splicing feature at time t. : ; We perform a weighted summation to obtain the time-series window features at time t. : .

[0014] Preferably, the Transformer encoder includes an encoder input layer and an encoder structure, and the construction method includes: The method for constructing the encoder input layer is as follows: The fused feature matrix is ​​input into the encoder input layer, and temporal information is injected through positional encoding. The encoder structure comprises three encoder blocks, each of which includes a multi-head self-attention mechanism, a feedforward neural network, and a time-adaptive gated residual connection. The construction method is as follows: Eight attention heads are set at the multi-head self-attention mechanism to capture temporal dependencies at different time scales and output attention feature maps through parallel computation. The feedforward neural network consists of two fully connected layers with ReLU activation function and one normalization layer, which realizes non-linear transformation of the attention feature map; In the time-adaptive gated residual connection, a time-adaptive gating unit is introduced to dynamically adjust the residual connection weights through the gating network: ; in, The output of the Transformer encoder, This is the output of the feedforward neural network. For the input of the encoder block, For gating networks.

[0015] Preferably, the method for mapping the global feature vector to condensation probability in the output layer is as follows: ; in, This represents the probability of condensation. For the global feature vector, and These are the output layer weights and output layer biases, respectively.

[0016] Preferably, when training the time-series deep model, the dataset is divided into a training set, a validation set, and a test set in a 7:2:1 ratio. The training set is used for updating model parameters, the validation set is used for hyperparameter tuning, and the test set is used for model performance evaluation. The binary cross-entropy loss function is used to measure the difference between the predicted values ​​and the true labels during model performance evaluation. :

[0017] in, For real labels, The predicted probability output by the time-series deep model. This represents the total number of samples in the test set.

[0018] The technical solution of the present invention has at least the following advantages and beneficial effects: This invention achieves high accuracy in predicting condensation in switchgear and adaptability to complex operating conditions by constructing a multi-dimensional time-series feature representation that integrates environmental perception data and thermodynamic features, combined with deep learning. Compared with traditional methods that rely on empirical rules, single physical quantities, or simple statistical models, this invention effectively solves the problems of insufficient characterization of condensation formation mechanism, limited capture of time-series dependence, and poor scene adaptability. It can stably adapt to diverse operating environments such as high temperature and high humidity, and drastic day-night temperature differences. This invention unifies and fuses raw data, thermodynamic feature data, and dynamic interaction features, enabling time-series deep models to learn under multi-source and multi-scale feature conditions, effectively improving their generalization ability under different load levels and complex environments. This invention automatically completes feature fusion and predictive reasoning through a time-series deep model, reducing reliance on manual experience rules and fixed thresholds. This helps guide the early implementation of anti-condensation measures such as dehumidification and heating, thereby improving the safety and reliability of switchgear operation. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the switchgear condensation prediction method that integrates thermodynamic features and temporal deep learning, as provided in Embodiment 1 of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0021] Example This embodiment provides a method for predicting condensation in switchgear by integrating thermodynamic features and temporal deep learning. (See reference...) Figure 1 This includes the following steps: Step S1: Continuously collect raw data and perform data preprocessing. The raw data includes ambient temperature, relative humidity, atmospheric pressure, switch cabinet surface temperature, switch cabinet internal temperature, historical load, and historical condensation status. In the preprocessing, outliers can be removed using the three-standard-deviation method, and missing data can be filled using the linear interpolation method. This facilitates the application of the preprocessed raw data to subsequent steps to obtain more accurate prediction results.

[0022] Step S2: Obtain various thermodynamic feature data through thermodynamic feature extraction, and then extract various dynamic interaction features between the feature data.

[0023] In this embodiment, the method for obtaining various thermodynamic feature data through thermodynamic feature extraction is as follows: (2.1) Extraction of dew point temperature This characteristic directly reflects the critical temperature condition for condensation to occur: ; Where T is the ambient temperature and RH is the relative humidity. and It is a constant, with values ​​of 17.625 and 243.04 respectively.

[0024] (2.2) Estimate the thermal conductivity coefficient by combining the cabinet material (such as cold-rolled steel plate, stainless steel), cabinet thickness and ambient temperature. This feature characterizes the effect of heat transfer within the cabinet on internal humidity condensation: ; in, The temperature coefficient of thermal conductivity The standard thermal conductivity coefficient of the cabinet material at 25℃.

[0025] (2.3) Wet bulb extraction temperature This parameter primarily assesses the air's humidity saturation and evaporative cooling potential. .

[0026] (2.4) Extracting temperature and humidity features: Calculate the change in ambient temperature between two adjacent raw data acquisition times. Relative humidity change And the difference between the surface temperature of the switch cabinet and the ambient temperature. These parameters can reflect the triggering effect of dynamic changes in temperature and humidity on condensation formation.

[0027] (2.5) Extract thermal performance indicators based on the specific heat capacity and thermal conductivity of the switchgear material. : ; Where d is the thickness of the cabinet material. The thermal conductivity of the cabinet material. This refers to the specific heat capacity of the cabinet material.

[0028] The purpose of this implementation in extracting multiple dynamic interaction features between feature data is to fuse the correlation information of different features through element-wise multiplication. Based on the above parameters, the specific calculation method is as follows: Extracting temperature-dew point interaction features: It is used to reflect the coupling relationship between temperature change and critical condensation temperature.

[0029] Extracting humidity-dew point interaction features: It is used to reflect the coupling relationship between humidity changes and critical condensation temperature.

[0030] Extracting heat conduction-gradient interaction features: It is used to reflect the synergistic effect of heat transfer and temperature gradient.

[0031] Extracting the humidity-heat conduction interaction characteristics: This is used to reflect the mechanism by which heat transfer exacerbates condensation under high humidity conditions.

[0032] Extracting temperature-humidity coupling features: It is used to reflect the dynamic characteristics of synchronous fluctuations in temperature and humidity.

[0033] Extracting the dew point-temperature gradient interaction features: It is used to reflect the degree of matching between the critical condensation temperature and the actual temperature difference conditions.

[0034] Step S3: Obtain the feature vector matrix through feature encoding and concatenation. The feature vector matrix includes the encoded and concatenated vectors of multiple original data acquisition times within a time series length. The encoded and concatenated vectors include multiple mapped features obtained by mapping multiple concatenated features based on the corresponding times. The multiple concatenated features include all original data, thermodynamic feature data, and dynamic interaction features.

[0035] As a preferred embodiment, the method for obtaining the feature vector matrix through feature encoding and concatenation is as follows: Step S301: Normalize the continuous data in the initial dataset using the Min-Max normalization method to avoid model training bias caused by differences in feature dimensions. Continuous data refers to specific numerical values ​​that vary over an interval, such as parameters like temperature and humidity.

[0036] The Min-Max normalization method is as follows: ; in, and These are the values ​​before and after normalization, respectively. and They are respectively The minimum and maximum values ​​of the corresponding parameters.

[0037] Step S302: Concatenate the original data, thermodynamic feature data, and dynamic interaction features at each time point to obtain the concatenated vector matrix at the corresponding time point. Concatenate the concatenated vector matrix in time series order to obtain the multidimensional time series feature matrix. : ; ; in, The length of the time series. To find the transpose of a matrix, D is the feature dimension. Let be the concatenated vector matrix at time t. arrive These represent the M splicing features used for splicing at time t.

[0038] Step S303: Apply splicing features A mapping method is used to uniformly transform the feature space to the same dimension, m=1,2,…,M:

[0039] in, The mapping feature is the m-th type of splicing feature. This is the weight matrix. It is the bias vector, and ReLU is the activation function. Step S304: Perform coded feature concatenation representation: ; in, This is the encoded concatenated vector obtained by concatenating the encoded features at time t. This is a join function.

[0040] Step S4: Based on the feature vector matrix, the sliding time window method is used to construct the reconstructed encoded concatenated vector from the encoded concatenated vector. The method is as follows: Set the time window length to L, and concatenate the encoded vectors at time t in the feature vector matrix. Construct as a reconstructed encoding concatenation vector : .

[0041] Step S5: Construct a time-series deep model. The time-series deep model is used to achieve deep fusion of thermodynamic features and time-series information, and generate prediction results based on the input reconstructed encoding concatenation vector.

[0042] As a preferred embodiment, the temporal deep model includes a feature fusion layer, a Transformer encoder, a fully connected layer, and an output layer connected in sequence, and its construction method includes: The reconstructed encoding concatenation vector is input into the feature fusion layer. The importance weight of each concatenation feature is calculated by the feature weighting module based on the attention mechanism. The reconstructed encoding concatenation vector is then weighted and fused to obtain the temporal window feature. Based on the temporal window feature, the ability to capture long temporal dependencies is enhanced in the Transformer encoder; Global average pooling is performed on the output of the Transformer encoder in the fully connected layer to obtain the global feature vector; In the output layer, the global feature vector is mapped to the condensation probability.

[0043] Based on the above structure, the specific construction methods for each layer are as follows: First, the method for constructing the feature fusion layer is as follows: Obtain the initial attention score for the m-th splicing feature at time t. : ; in, It is a row vector, and tanh is a non-linear activation function. and Let b be the linear transformation matrix, and b be the bias term. Initial attention score for all spliced ​​features at time t Normalization is performed to obtain the importance weights of the m-th splicing feature at time t. : ; We perform a weighted summation to obtain the time-series window features at time t. : .

[0044] Secondly, the Transformer encoder introduces a time-adaptive gated residual connection to enhance the ability to capture long-term time-dependent dependencies. Structurally, it includes an encoder input layer and an encoder structure, and the construction method includes: The method for constructing the encoder input layer is as follows: The fused feature matrix is ​​input into the encoder input layer, and temporal information is injected through positional encoding. The encoder structure comprises three encoder blocks, each of which includes a multi-head self-attention mechanism, a feedforward neural network, and a time-adaptive gated residual connection. The construction method is as follows: Eight attention heads are set at the multi-head self-attention mechanism to capture temporal dependencies at different time scales and output attention feature maps through parallel computation. The feedforward neural network consists of two fully connected layers with ReLU activation function and one normalization layer, which realizes non-linear transformation of the attention feature map; A temporally adaptive gating unit is introduced into the temporally adaptive gated residual connection. The residual connection weights are dynamically adjusted through the gating network to enhance the model's adaptability to scenarios with sudden temporal changes. ; in, The output of the Transformer encoder, This is the output of the feedforward neural network. For the input of the encoder block, For gating networks.

[0045] Finally, the method for mapping the global feature vector to the condensation probability in the output layer is as follows: ; in, This represents the probability of condensation. For the global feature vector, and These are the output layer weights and the output layer bias, respectively. Let P be the predetermined step size achieved at time t. Based on the output results, a probability threshold θ can be set. When p ≥ θ, it is determined that "condensation will occur at time t+P in the future", and when p < θ, it is determined that "condensation has not occurred".

[0046] In addition, when training the time-series deep model, the dataset is divided into training, validation, and test sets in a 7:2:1 ratio. The training set is used for updating model parameters, the validation set is used for hyperparameter tuning, and the test set is used for model performance evaluation. The binary cross-entropy loss function is used to measure the difference between the predicted values ​​and the true labels when evaluating model performance. :

[0047] in, For real labels, The predicted probability output by the time-series deep model. This represents the total number of samples in the test set. An early stopping strategy is employed during training (training stops if the validation set loss does not decrease for 10 consecutive rounds) to prevent overfitting of the model.

[0048] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting condensation in switchgear by integrating thermodynamic features and temporal deep learning, characterized in that, Includes the following steps: Raw data is continuously collected and preprocessed. The raw data includes ambient temperature, relative humidity, atmospheric pressure, switch cabinet surface temperature, switch cabinet internal temperature, historical load, and historical condensation status. Multiple thermodynamic feature data are obtained through thermodynamic feature extraction, and then multiple dynamic interaction features between feature data are extracted. The feature vector matrix is ​​obtained through feature encoding and concatenation. The feature vector matrix includes encoded and concatenated vectors of multiple raw data acquisition times within a time series length. The encoded and concatenated vectors include multiple mapping features obtained by mapping multiple concatenated features based on the corresponding times. The multiple concatenated features include all raw data, thermodynamic feature data, and dynamic interaction features. based on The feature vector matrix is ​​used to construct the reconstructed encoded concatenated vector using the sliding time window method; A temporal deep model is constructed to achieve deep fusion of thermodynamic features and temporal information, and to generate prediction results based on the input reconstructed encoding concatenation vector; The method for obtaining multiple thermodynamic feature data through thermodynamic feature extraction is as follows: Extraction dew point temperature : ； Where T is the ambient temperature and RH is the relative humidity. and It is a constant, with values ​​of 17.625 and 243.04 respectively; Extracting thermal conductivity : ； in, The temperature coefficient of thermal conductivity The standard thermal conductivity coefficient of the cabinet material at 25℃; Extraction wet bulb temperature : ； Extracting temperature and humidity features: Calculating the change in ambient temperature between two adjacent raw data acquisition times. Relative humidity change And the difference between the surface temperature of the switch cabinet and the ambient temperature. ; Extracting thermal performance indicators : ； Where d is the thickness of the cabinet material. The thermal conductivity of the cabinet material. The specific heat capacity of the cabinet material; The method for extracting multiple dynamic interaction features between feature data is as follows: Extracting temperature-dew point interaction features: ; Extracting humidity-dew point interaction features: ; Extracting heat conduction-gradient interaction features: ; Extracting the humidity-heat conduction interaction characteristics: ; Extracting temperature-humidity coupling features: ; Extracting the dew point-temperature gradient interaction features: .

2. The method for predicting condensation in switchgear by integrating thermodynamic features and temporal deep learning according to claim 1, characterized in that, The method for obtaining the feature vector matrix through feature encoding and concatenation is as follows: The continuous data in the initial data are normalized using the Min-Max normalization method; The raw data, thermodynamic feature data, and dynamic interaction features at each time point are concatenated to obtain the concatenated vector matrix for that time point. The concatenated vector matrix is ​​then concatenated in chronological order to obtain the multidimensional time series feature matrix. : ； ； in, The length of the time series. To find the transpose of a matrix, D is the feature dimension. Let be the concatenated vector matrix at time t. arrive These are the M types of splicing features used for splicing at time t; Features for splicing A mapping method is used to uniformly transform the feature space to the same dimension, m=1,2,…,M: in, The mapping feature is the m-th type of splicing feature. This is the weight matrix. It is the bias vector, and ReLU is the activation function; Perform coded feature concatenation representation: ； in, This is the encoded concatenated vector obtained by concatenating the encoded features at time t. This is a join function.

3. The method for predicting condensation in switchgear by integrating thermodynamic features and temporal deep learning according to claim 2, characterized in that, The method for constructing time series samples using the sliding time window method is as follows: Set the time window length to L, and concatenate the encoded vectors at time t in the feature vector matrix. Construct as a reconstructed encoding concatenation vector : 。 4. The method for predicting condensation in switchgear by integrating thermodynamic features and temporal deep learning according to claim 3, characterized in that, The method for constructing a temporal deep model includes: The reconstructed encoding concatenation vector is input into the feature fusion layer. The importance weight of each concatenation feature is calculated by the feature weighting module based on the attention mechanism. The reconstructed encoding concatenation vector is then weighted and fused to obtain the temporal window feature. Based on the temporal window feature, the ability to capture long temporal dependencies is enhanced in the Transformer encoder; Global average pooling is performed on the output of the Transformer encoder in the fully connected layer to obtain the global feature vector; In the output layer, the global feature vector is mapped to the condensation probability.

5. The method for predicting condensation in switchgear by integrating thermodynamic features and temporal deep learning according to claim 4, characterized in that, The method for constructing the feature fusion layer is as follows: Obtain the initial attention score for the m-th splicing feature at time t. : ； in, It is a row vector, and tanh is a non-linear activation function. and Let b be the linear transformation matrix, and b be the bias term. Initial attention score for all spliced ​​features at time t Normalization is performed to obtain the importance weights of the m-th splicing feature at time t. : ； We perform a weighted summation to obtain the time-series window features at time t. : 。 6. The method for predicting condensation in switchgear by integrating thermodynamic features and temporal deep learning according to claim 5, characterized in that, The Transformer encoder includes an encoder input layer and an encoder structure, and its construction method includes: The method for constructing the encoder input layer is as follows: The fused feature matrix is ​​input into the encoder input layer, and temporal information is injected through positional encoding. The encoder structure comprises three encoder blocks, each of which includes a multi-head self-attention mechanism, a feedforward neural network, and a time-adaptive gated residual connection. The construction method is as follows: Eight attention heads are set at the multi-head self-attention mechanism to capture temporal dependencies at different time scales and output attention feature maps through parallel computation. The feedforward neural network consists of two fully connected layers with ReLU activation function and one normalization layer, which realizes non-linear transformation of the attention feature map; In the time-adaptive gated residual connection, a time-adaptive gating unit is introduced to dynamically adjust the residual connection weights through the gating network: ； in, The output of the Transformer encoder, This is the output of the feedforward neural network. For the input of the encoder block, For gating networks.

7. The method for predicting condensation in switchgear by integrating thermodynamic features and temporal deep learning according to claim 6, characterized in that, The method for mapping the global feature vector to condensation probability in the output layer is as follows: ； in, This represents the probability of condensation. For the global feature vector, and These are the output layer weights and output layer biases, respectively.

8. The method for predicting condensation in switchgear by integrating thermodynamic features and temporal deep learning according to claim 1, characterized in that, When training the time-series deep model, the dataset is divided into training, validation, and test sets in a 7:2:1 ratio. The training set is used for updating model parameters, the validation set is used for hyperparameter tuning, and the test set is used for model performance evaluation. The binary cross-entropy loss function is used to measure the difference between the predicted values ​​and the true labels during model performance evaluation. : in, For real labels, The predicted probability output by the time-series deep model. This represents the total number of samples in the test set.

Images