SF6 liquefaction risk judgment method based on neural network

Abstract

The invention relates to an SF6 liquefaction risk judgment method based on a neural network. The SF6 liquefaction risk judgment method comprises the following steps: step 1, acquiring data of electrical equipment and a nearby environment; 2, constructing features based on the collected data; 3, constructing a neural network model, and carrying out training and parameter adjustment through features; 4, ensuring the generalization ability of the model by using a cross validation technology; 5, deploying the neural network model, and predicting the SF6 liquefaction risk; and 6, regularly collecting new data and retraining the model to adapt to environmental changes. According to the invention, through real-time monitoring, intelligent analysis and rapid response, a layer of additional safety guarantee is provided for operation and maintenance of electric power facilities, human errors and accidents are reduced, and stable operation of an electric power system and life safety of personnel are ensured.

Description

Technical Field

[0001] The invention belongs to the technical field of power supply equipment safety and discloses a SF6 liquefaction risk judgment method based on a neural network. Background Art

[0002] In the power industry, SF6 is a gas widely used in high-voltage electrical equipment as an insulating and arc-extinguishing medium. However, under conditions below the critical temperature, SF6 will liquefy as the pressure increases. This means that in high-voltage electrical equipment, especially in low-temperature environments (such as winter in high-latitude areas), SF6 gas may liquefy. The liquefaction of SF6 will cause its density and viscosity to increase, which will affect its performance as an insulating medium. The insulation strength of liquid SF6 is not as good as that of gaseous SF6. Therefore, if SF6 is liquefied in electrical equipment, it may not provide sufficient insulation effect, thereby increasing the risk of short circuit or failure of electrical equipment. After SF6 is liquefied, its arc-extinguishing ability will be greatly reduced, which may cause the arc to fail to be extinguished in time, causing fire or explosion inside the equipment. SF6 liquefaction will also cause additional stress on the mechanical structure of high-voltage electrical equipment. The volume of SF6 after liquefaction will decrease, but its weight will not change, which may cause additional pressure on the container wall and seals, which may cause damage or leakage of the equipment structure under long-term action. In addition, SF6 is a potent greenhouse gas with a greenhouse effect potential far higher than that of carbon dioxide. If SF6 leaks due to liquefaction or other reasons, it will have a negative impact on global warming. Therefore, preventing SF6 liquefaction is also for environmental protection considerations.

[0003] Therefore, there is an urgent need in this field to implement a more intelligent solution for SF6 liquefaction risk assessment. Summary of the invention

[0004] The present invention discloses a SF6 liquefaction risk judgment method based on a neural network, and the SF6 liquefaction risk judgment method based on a neural network comprises the following steps:

[0005] Step 1: Data collection of electrical equipment and surrounding environment;

[0006] Step 2: construct features based on the collected data;

[0007] Step 3: Build a neural network model and train and adjust parameters through features;

[0008] Step 5: deploy the neural network model and predict the SF6 liquefaction risk.

[0009] Furthermore, in step 1, the following steps are also included:

[0010] Step 11, directly collecting environmental parameters of ambient temperature, humidity and atmospheric pressure through sensors such as temperature and humidity sensors and barometers installed near the electrical equipment;

[0011] Step 12, through the historical liquefaction events, temperature and pressure installed on the SF6 circuit breaker or other electrical equipment, and directly collect the equipment parameters of internal temperature, pressure and current through the current sensor.

[0012] Furthermore, in step 1, the following steps are also included:

[0013] The environmental parameters and equipment parameters of historical cases are cleaned, and outliers and missing values ​​are removed to ensure data accuracy.

[0014] Furthermore, in step 2, the following steps are also included:

[0015] Step 21, extracting original features from the collected data, the original features including ambient temperature, ambient humidity, device internal temperature, device internal pressure and real-time current;

[0016] Step 22, performing statistical analysis on the original features to extract statistics reflecting data distribution characteristics, the statistical features including average current, current standard deviation, peak current, average ambient temperature, ambient temperature variance, and maximum internal pressure of the device;

[0017] Step 23, taking into account the time series of the data, extracting time series features that reflect the trend of data changes, the time series features include the first-order difference of the ambient temperature and the moving average of the internal pressure of the equipment;

[0018] Step 24, combining multiple original features to generate a combined feature;

[0019] Step 25, the original features, statistical features, time series features and combined features are sorted to generate a final feature vector.

[0020] Furthermore, in step 23, the first-order difference of the ambient temperature is:

[0021] ΔT env (t) = T env (t)-T env (t-1)

[0022] in:

[0023] ΔT env (t) is the first-order difference of the ambient temperature at time t, T env (t) is the ambient temperature at time t, T env (t-1) is the ambient temperature at time t-1;

[0024] The moving average pressure inside the equipment is:

[0025] MA(P inner ,window size )

[0026] =(P inner (t)+P inner (t-1)+...+P inner (t-window size

[0027] +1)) / window size

[0028] MA(P inner ,window size ) is P inner The moving average of the window size is window size , P inner (t) is the internal pressure of the equipment at time t, window size is the size of the moving window, that is, the number of continuous data points involved in the calculation.

[0029] Furthermore, in step 24, the combined feature includes the temperature difference obtained by subtracting the ambient temperature from the internal temperature of the device and then comparing the values:

[0030] (T inner -T env ) / T inner

[0031] T inner is the internal temperature of the device, T env is the ambient temperature.

[0032] Furthermore, in step 25, the final feature vector is:

[0033] [T env ,H env ,T inner ,P inner ,T envmean ,P innermax ,T envdiff ,(T inner -T env ) / T inner ,i avg ,i peak ,I std ,ΔT env (t),MA(P inner ,window size )].

[0034] Furthermore, in step 3, the following steps are also included:

[0035] Step 31, constructing a neural network model;

[0036] The final feature vectors and corresponding historical liquefaction events are used to generate training sets, test sets, and validation sets;

[0037] The neural network model is a long short-term memory network, including an input layer, an LSTM and an output layer;

[0038] The input layer is used to obtain the final feature vector [T env ,H env ,T inner ,P inner ,T envmean ,P innermax ,T envdiff ,(T inner -T env ) / T inner ,I avg ,I peak ,I std ,ΔT env (t),MA(P inner ,window size )];

[0039] The LSTM includes the first to fourth LSTM layers connected in sequence; the first layer LSTM has an input dimension of 13 and an output dimension of 128, the second to fourth LSTM layers have an input dimension of the output dimension of the previous layer, and the output dimension decreases layer by layer;

[0040] The input dimension of the fourth LSTM layer is the output dimension of the third LSTM layer, and the output dimension is 1. The sigmoid activation function is used to map the output to between 0 and 1.

[0041] The output layer includes a fully connected layer, which is connected to the fourth LSTM layer and uses a sigmoid activation function to output a probability value between 0 and 1, indicating the probability of SF6 liquefaction;

[0042] Step 32, the learning rate is selected as 0.01, and the Adam optimizer is used to train the neural network model with the cross entropy loss as the objective function;

[0043] Step 33, using the test set and the validation set to evaluate and verify the performance of the model;

[0044] Step 34, based on the evaluation and verification results, decide whether to readjust the learning rate, number of LSTM layers, etc. for retraining. When at least the loss functions on the training set and the verification set tend to be stable, the training is considered to have met the requirements.

[0045] Furthermore, the SF6 liquefaction risk judgment method based on neural network further includes the following steps between step 3 and step 5:

[0046] Step 4: Use cross-validation techniques to ensure the generalization ability of the model; evaluate model performance through indicators such as accuracy, recall, and F1 score.

[0047] Furthermore, the SF6 liquefaction risk judgment method based on neural network further includes the following steps:

[0048] Step 6: Regularly collect new data and retrain the model to adapt to environmental changes; continuously monitor model performance to ensure its accuracy and effectiveness.

[0049] The beneficial effects achieved by the present invention are:

[0050] The present invention provides an additional layer of safety for the operation and maintenance of power facilities through real-time monitoring, intelligent analysis and rapid response, which helps to reduce human errors and accidents and ensure the stable operation of the power system and the life safety of personnel.

[0051] The present invention is based on the collection of data such as ambient temperature, humidity and atmospheric pressure, as well as historical liquefaction events, temperature, pressure and current on SF6 circuit breakers or other electrical equipment, and fully considers different types of data directly and indirectly related to SF6 liquefaction, and incorporates all data related to SF6 liquefaction into the neural network to the greatest extent, so that the risk of SF6 liquefaction can be judged more comprehensively and accurately. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is a flow chart of a neural network-based SF6 liquefaction risk judgment method proposed by the present invention. DETAILED DESCRIPTION

[0053] The present invention will be further described below in conjunction with specific embodiments, and the advantages and features of the present invention will become clearer as the description proceeds. However, these embodiments are exemplary only and do not constitute any limitation to the scope of the present invention. It should be understood by those skilled in the art that the details and forms of the technical solution of the present invention may be modified or replaced without departing from the spirit and scope of the present invention, but these modifications and replacements all fall within the scope of protection of the present invention.

[0054] like Figure 1 As shown, a method for judging the risk of SF6 liquefaction based on a neural network comprises the following steps:

[0055] Step 1, data collection;

[0056] In step 1, the following steps are specifically included:

[0057] Step 11, environmental parameter collection:

[0058] Direct collection: Data such as ambient temperature, humidity, and atmospheric pressure are directly collected through sensors such as temperature and humidity sensors and barometers installed near electrical equipment.

[0059] Step 12, equipment parameter collection:

[0060] Direct collection: directly collect internal temperature, pressure, current and other data through historical liquefaction events, temperature, pressure and current sensors installed on SF6 circuit breakers or other electrical equipment.

[0061] Indirect collection: Historical operation data can be obtained from the equipment's SCADA system or other similar systems and analyzed.

[0062] In step 1, the following may also be included:

[0063] Step 13: Collect environmental parameters and equipment parameters from historical cases:

[0064] The environmental parameters and equipment parameters of historical cases are cleaned to remove outliers and missing values ​​to ensure data accuracy.

[0065] Step 2, feature construction;

[0066] In step 2, the following steps are specifically included:

[0067] Step 21, extracting original features: features extracted from the collected data, such as ambient temperature, humidity, air pressure, internal temperature, pressure, current, etc.

[0068] Step 22, generate statistical features: perform statistical analysis on the above features and extract statistics that reflect data distribution characteristics, such as mean, variance, standard deviation, maximum value, minimum value, etc.

[0069] Step 23, generate time series features: Consider the time series of the data and extract features that reflect the trend of data changes, such as first-order difference, moving average, seasonal index, etc.

[0070] Step 24, combining features: combining multiple original features to generate new features. For example, subtracting the ambient temperature from the internal temperature of the device and then comparing them to obtain a feature that reflects the temperature difference.

[0071] Specifically, a specific example of building a feature set:

[0072] The following data were collected:

[0073] Ambient temperature (T env )

[0074] Ambient humidity (Henv )

[0075] Internal temperature of the device (T inner )

[0076] Equipment internal pressure (P inner )

[0077] Historical liquefaction events (whether liquefaction occurred)

[0078] The following current-related data were also collected:

[0079] Average current (I avg )

[0080] Peak current (I peak )

[0081] Current standard deviation (I std )

[0082] Among them, average current: calculates the average current over a period of time, reflecting the average load level of the equipment.

[0083] Peak current: records the maximum current value within a period of time, reflecting the instantaneous load condition of the equipment.

[0084] Current standard deviation: reflects the degree of dispersion of a set of current data relative to the average value, that is, the severity of current fluctuation. The larger the standard deviation, the greater the current fluctuation, and vice versa.

[0085] Construct the following features:

[0086] Original feature: T env ,H env ,T inner ,P inner

[0087] Statistical characteristics: T env The mean value T envmean , Variance T envdiff , P inner The maximum value P innermax , I std That is, the current standard deviation

[0088] Timing characteristics: T env The first-order difference ΔT env (t), P inner Moving average MA(P inner ,window size )

[0089] ΔT env (t) = T env (t)-T env (t-1)

[0090] in:

[0091] ΔT env (t) is the first-order difference of the ambient temperature at time t, T env (t) is the ambient temperature at time t, T env (t-1) is the ambient temperature at time t-1;

[0092] T env The first-order difference of is used to highlight the trend of ambient temperature;

[0093] MA(P inner ,window size )

[0094] =(P inner (t)+P inner (t-1)+...+P inner (t-window size

[0095] +1)) / window size

[0096] MA(P inner ,window size ) is P inner The moving average of the window size is window size , P inner (t) is the internal pressure of the equipment at time t, window size is the size of the moving window, that is, the number of continuous data points involved in the calculation, MA(P inner ,window size ) is used to highlight the long-term trend of pressure inside the equipment;

[0097] Combined features: (T inner -T env ) / T inner

[0098] The final feature vector can be expressed as:

[0099] [T env ,H env ,T inner ,P inner ,T envmean ,P innermax ,T envdiff ,(T inner

[0100] -T env )

[0101] / T inner ,Iavg ,I peak ,I std ,ΔT env (t),MA(P inner ,window size )]

[0102] Step 3: Build a neural network model

[0103] Step 31, model selection: Select a suitable neural network architecture, such as BP neural network, convolutional neural network (CNN) or long short-term memory network (LSTM), depending on the type and complexity of the input data.

[0104] Data partitioning: Divide the data into training set, validation set and test set.

[0105] Model training: Use the training set data to train the neural network model.

[0106] 1. It is necessary to clarify which neural network to use, or how to select a neural network based on the conditions;

[0107] Preferably, a long short-term memory network (LSTM) is selected to better capture the temporal dynamic changes in the SF6 liquefaction process.

[0108] Example model structure:

[0109] Input layer: receives the 13-dimensional feature vector [T env ,H env ,T inner ,P inner ,T envmean ,P innermax ,T envdiff ,(T inner -T env ) / T inner ,i avg ,i peak ,I std ,ΔT env (t),MA(P inner ,window size )]

[0110] LSTM: Multi-layer LSTM layers are used to extract time series features. Each LSTM layer contains multiple LSTM units, each of which is responsible for processing data for one time step.

[0111] Here is a specific example of a multi-layer LSTM layer, for example, a total of 4 LSTM layers:

[0112] The above input layer to the first layer LSTM

[0113] Input dimension: 13

[0114] Output dimension: Depending on the data complexity and computing resources, the output dimension is expanded compared to the input dimension, for example, it can be set to 128.

[0115] Subsequent LSTM layers, including the second, third, and fourth LSTM layers

[0116] Input dimension: the output dimension of the previous LSTM layer.

[0117] Output dimension: Compared with the first LSTM layer, the output dimension of subsequent LSTM layers gradually decreases. For example, the output dimension of the second LSTM layer is 64, the output dimension of the third LSTM layer is 32, and the output dimension of the fourth LSTM layer is 16.

[0118] Output Layer

[0119] Input dimension: The output dimension of the fourth layer LSTM.

[0120] Output dimension: 1, indicating the probability of SF6 liquefaction.

[0121] The activation function is: sigmoid, which maps the output to between 0 and 1.

[0122] The output layer includes a fully connected layer: the output of the fourth LSTM layer in the multi-layer LSTM layer is mapped to a fully connected layer for classification.

[0123] Output layer: By using the sigmoid activation function, it outputs a probability value between 0 and 1, indicating the probability of SF6 liquefaction.

[0124] Regarding the training of the above LSTM model, for the above 13-dimensional feature vector [T env ,H env ,T inner ,P inner ,T envmean ,P innermax ,T envdiff ,(T inner -T env ) / T inner ,I avg ,I peak ,I std ,ΔT env (t),MA(P inner ,window size )], normalize it, divide it into training set, validation set and test set according to the ratio of 70%:20%:10% or 70%, 15%, 15%, and perform training according to the following steps:

[0125] Step 32, training: select a learning rate of 0.01, use the Adam optimizer, and use the cross entropy loss as the objective function to train the model.

[0126] Step 33, Evaluation and Validation: Use the test set and validation set to evaluate and validate the performance of the model.

[0127] Step 34, parameter adjustment: Based on the evaluation and verification results, decide whether to readjust the learning rate, number of LSTM layers, etc. for retraining. When the loss functions on at least the training set and the verification set are stable, the training is considered to meet the requirements. A grid search strategy is used to find the optimal number of hidden layers, number of nodes (i.e., number of LSTM layers), and learning rate.

[0128] Step 4: Model Validation

[0129] Cross-validation: Use cross-validation techniques to ensure the generalization ability of the model.

[0130] Performance evaluation: In addition to the above loss functions, model performance can be further evaluated through indicators such as accuracy, recall, and F1 score.

[0131] Step 5: Implementation and deployment

[0132] Real-time monitoring: Integrate the model into the monitoring system to continuously collect data and make real-time predictions.

[0133] Alarm system: Once the model predicts an increased risk of liquefaction, an alarm is immediately triggered to notify maintenance personnel.

[0134] Decision Support: Provide decision support to operators to help them take appropriate preventive measures.

[0135] Step 6: Continue to optimize

[0136] Model updating: Regularly collect new data and retrain the model to adapt to changes in the environment.

[0137] Performance monitoring: Continuously monitor model performance to ensure its accuracy and effectiveness.

[0138] Since the present invention fully incorporates the data of ambient temperature, humidity and atmospheric pressure directly collected by sensors such as temperature and humidity sensors and barometers near the electrical equipment, and also incorporates the data of historical liquefaction events, temperature, pressure and current on the SF6 circuit breaker or other electrical equipment, as well as historical operation data, the SF6 liquefaction risk judgment method based on neural network disclosed in the present invention fully considers different types of data directly and indirectly related to SF6 liquefaction, and incorporates all data related to SF6 liquefaction into the neural network to the greatest extent, so that the SF6 liquefaction risk can be judged more comprehensively and accurately.

[0139] The present invention is not limited to the above-mentioned specific implementation modes. A person skilled in the art may implement the present invention in various other specific implementation modes according to the embodiments and the disclosure of the drawings. Therefore, any design that adopts the design structure and concept of the present invention and makes some simple transformations or changes shall fall within the scope of protection of the present invention.

Claims

1. A method for judging SF6 liquefaction risk based on neural network, characterized in that: The SF6 liquefaction risk judgment method based on neural network comprises the following steps: Step 1: Data collection of electrical equipment and surrounding environment; Step 2: construct features based on the collected data; Step 3: Build a neural network model and train and adjust parameters through features; Step 5: deploy the neural network model and predict the SF6 liquefaction risk.

2. The SF6 liquefaction risk judgment method based on neural network according to claim 1 is characterized in that: In step 1, the following steps are also included: Step 11, directly collecting environmental parameters of ambient temperature, humidity and atmospheric pressure through sensors such as temperature and humidity sensors and barometers installed near the electrical equipment; Step 12, through the historical liquefaction events, temperature and pressure installed on the SF6 circuit breaker or other electrical equipment, and directly collect the equipment parameters of internal temperature, pressure and current through the current sensor.

3. The SF6 liquefaction risk judgment method based on neural network according to claim 2 is characterized in that: In step 1, the following steps are also included: The environmental parameters and equipment parameters of historical cases are cleaned, and outliers and missing values ​​are removed to ensure data accuracy.

4. The SF6 liquefaction risk judgment method based on neural network according to claim 2 is characterized in that: In step 2, the following steps are also included: Step 21, extracting original features from the collected data, the original features including ambient temperature, ambient humidity, device internal temperature, device internal pressure and real-time current; Step 22, performing statistical analysis on the original features to extract statistics reflecting data distribution characteristics, the statistical features including average current, current standard deviation, peak current, average ambient temperature, ambient temperature variance, and maximum internal pressure of the device; Step 23, taking into account the time series of the data, extracting time series features that reflect the trend of data changes, the time series features include the first-order difference of the ambient temperature and the moving average of the internal pressure of the equipment; Step 24, combining multiple original features to generate a combined feature; Step 25, the original features, statistical features, time series features and combined features are sorted to generate a final feature vector.

5. The SF6 liquefaction risk judgment method based on neural network according to claim 4 is characterized in that: In step 23, the first-order difference of the ambient temperature is: ΔT env (t)=T env (t)-T env (t-1) in: ΔT env (t) is the first-order difference of the ambient temperature at time t, T env (t) is the ambient temperature at time t, T env (t-1) is the ambient temperature at time t-1; The moving average pressure inside the equipment is: MA(P inner , window size ) =(P inner (t)+P inner (t-1)+...+P inner (t-window size +1)) / window siee MA(P inner ,window size ) is P inner The moving average of the window size is window size , P inner (t) is the internal pressure of the equipment at time t, window size is the size of the moving window, that is, the number of continuous data points involved in the calculation.

6. The SF6 liquefaction risk judgment method based on neural network according to claim 4 is characterized in that: In step 24, the combined feature includes the temperature difference obtained by subtracting the ambient temperature from the internal temperature of the device and then comparing the values: (T inner -T env ) / T inner T inner is the internal temperature of the device, T env is the ambient temperature.

7. The SF6 liquefaction risk judgment method based on neural network according to claim 4 is characterized in that: In step 25, the final feature vector is: [T env ,H env ,T inner ,P inner ,T envmean ,P innermax ,T envdiff ,(T inner -T env ) / T inner ,I avg ,I peak ,I std ,ΔT env (t),MA(P inner ,window size )].

8. The SF6 liquefaction risk judgment method based on neural network according to claim 1 is characterized in that: In step 3, the following steps are also included: Step 31, constructing a neural network model; The final feature vectors and corresponding historical liquefaction events are used to generate training sets, test sets, and validation sets; The neural network model is a long short-term memory network, including an input layer, an LSTM and an output layer; The input layer is used to obtain the final feature vector [T env ,H env ,T inner ,P inner ,T envmean ,P innermax ,T envdiff ,(T inner -T env ) / T inner ,I avg ,I peak ,I std ,ΔT env (t),MA(P inner ,window size )]; The LSTM includes the first to fourth LSTM layers connected in sequence; the first layer LSTM has an input dimension of 13 and an output dimension of 128, the second to fourth LSTM layers have an input dimension of the output dimension of the previous layer, and the output dimension decreases layer by layer; The input dimension of the fourth LSTM layer is the output dimension of the third LSTM layer, and the output dimension is 1. The sigmoid activation function is used to map the output to between 0 and 1. The output layer includes a fully connected layer, which is connected to the fourth LSTM layer and uses a sigmoid activation function to output a probability value between 0 and 1, indicating the probability of SF6 liquefaction; Step 32, the learning rate is selected as 0.01, and the Adam optimizer is used to train the neural network model with the cross entropy loss as the objective function; Step 33, using the test set and the validation set to evaluate and verify the performance of the model; Step 34, based on the evaluation and verification results, decide whether to readjust the learning rate, number of LSTM layers, etc. for retraining. When at least the loss functions on the training set and the verification set tend to be stable, the training is considered to have met the requirements.

9. The SF6 liquefaction risk judgment method based on neural network according to claim 1 is characterized in that: The SF6 liquefaction risk judgment method based on neural network further includes the following steps between step 3 and step 5: Step 4: Use cross-validation techniques to ensure the generalization ability of the model; evaluate model performance through indicators such as accuracy, recall, and F1 score.

10. The SF6 liquefaction risk judgment method based on neural network according to claim 1 is characterized in that: The SF6 liquefaction risk judgment method based on neural network also includes the following steps: Step 6: Regularly collect new data and retrain the model to adapt to environmental changes; continuously monitor model performance to ensure its accuracy and effectiveness.

Images