SF6 humidity prediction method and device based on bidirectional long and short term memory neural network

Abstract

The invention discloses an SF6 humidity prediction method and device based on a bidirectional long and short term memory neural network, and the method comprises the steps: carrying out the Fourier transform and normalization processing of a light intensity signal and environment temperature data, and obtaining a normalized data set; performing data prediction on the normalized data set through a BiLSTM model to obtain an SF6 gas humidity value; wherein the BiLSTM model is obtained by performing back propagation training on the initial BiLSTM model according to the normalized data set. According to the SF6 humidity prediction method and device based on the bidirectional long and short term memory neural network, Fourier transform and normalization processing are carried out on light intensity signals and environment temperature data, key features can be extracted, dimensional differences can be eliminated, and periodic components can be identified; the initial BiLSTM model is trained by using the preprocessed data, prediction errors can be reduced, the SF6 gas humidity value can be accurately predicted according to the trained BiLSTM model, and the problem that it is difficult to accurately predict the SF6 gas humidity value according to limited samples can be solved.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to a method and device for predicting SF6 humidity based on a bidirectional long short-term memory neural network. Background Art

[0002] GIS equipment, or Gas Insulated Switchgear, is a high-voltage electrical device primarily used for power control and protection in power systems. If a GIS device experiences a leak, the humidity level of the SF6 gas will increase significantly. More seriously, SF6 gas decomposes under the action of an electric arc, and its decomposition products react with moisture in the air to produce highly corrosive acids, causing serious damage to the equipment. Therefore, accurately monitoring the humidity level of SF6 gas is crucial to ensuring the proper operation of GIS equipment. Humidity testing of SF6 gas in GIS equipment is typically conducted every three years. For equipment experiencing rapidly increasing humidity levels but not yet reaching the caution threshold, a simple and efficient linear prediction method is typically used to estimate the point at which the humidity level may exceed the caution threshold, allowing for proactive maintenance measures.

[0003] Although the linear prediction method is easy to use, its assumption of linear relationships faces uncertainties in practical applications. This is because changes in SF6 gas humidity are often affected by multiple nonlinear factors such as equipment aging, fluctuations in ambient temperature and humidity, and changes in leakage rate. These factors are difficult to fully and accurately capture and represent using linear models, resulting in the current linear prediction method's lack of accuracy and precision in predicting the growth of SF6 gas humidity in GIS equipment. In addition, the linear prediction method relies on the statistical laws of historical data to infer future trends. If the amount of data is insufficient, it may not be able to accurately capture the true relationship in the data, resulting in inaccurate prediction results. Summary of the Invention

[0004] The present invention provides an SF6 humidity prediction method and device based on a bidirectional long short-term memory neural network to solve the problem that it is difficult to accurately predict the humidity value of SF6 gas based on a limited number of samples.

[0005] To achieve the above objectives, the present application provides a SF6 humidity prediction method based on a bidirectional long short-term memory neural network, comprising:

[0006] Obtain light intensity signal and ambient temperature data;

[0007] Performing Fourier transform and normalization processing on the light intensity signal and ambient temperature data to obtain a normalized data set;

[0008] The normalized data set is predicted using a BiLSTM model to obtain a humidity value of the SF6 gas; wherein the BiLSTM model is obtained by backpropagation training of an initial BiLSTM model based on the normalized data set.

[0009] By performing Fourier transforms and normalization on the light intensity signal and ambient temperature data, the present invention effectively extracts key features from the data and eliminates dimensional differences. The Fourier transform helps identify periodic components in the data, while normalization ensures that the data is on a uniform scale, facilitating model processing. These preprocessing steps improve data quality, enabling better model learning and generalization even with limited sample sizes. Backpropagation training of the initial BiLSTM model using the normalized dataset continuously optimizes the model's parameters, ensuring a better fit to the data and reducing prediction error. Furthermore, the BiLSTM model is particularly well-suited for processing time series data because it can simultaneously consider both forward and reverse temporal dependencies. This capability gives BiLSTM a significant advantage in capturing long-term dependencies and complex patterns in the data. The trained BiLSTM model can predict the normalized dataset and output predicted SF6 gas humidity values. Because the model has learned the key features and patterns in the data, it can accurately predict humidity values ​​even with limited sample sizes.

[0010] Compared with the existing technology, the present invention can extract key features, eliminate dimensional differences, identify periodic components, and improve data quality by performing Fourier transform and normalization on the light intensity signal and ambient temperature data. Using this preprocessed data to train the initial BiLSTM model can reduce prediction errors, thereby accurately predicting the SF6 gas humidity value based on the trained BiLSTM model. This solves the problem of the difficulty in accurately predicting the SF6 gas humidity value based on a limited sample.

[0011] As a preferred solution, the light intensity signal and ambient temperature data are subjected to Fourier transform and normalization processing to obtain a normalized data set, specifically:

[0012] According to the light intensity signal, the gas concentration is obtained by Fourier transform, and the gas concentration is converted into relative humidity in combination with the ambient temperature data to obtain SF6 gas humidity time series data;

[0013] The SF6 gas humidity time series data is normalized to obtain the normalized data set.

[0014] This preferred solution can extract information related to gas concentration from the light intensity signal through Fourier transform, which is an effective signal processing method. By combining the ambient temperature data, the gas concentration can be converted into relative humidity, thereby obtaining time series data that more directly reflects the state of SF6 gas. This method helps to reduce noise and interference in the data and improve the quality of the data.

[0015] As a preferred solution, the SF6 gas humidity time series data is normalized to obtain the normalized data set, specifically:

[0016] replacing target abnormal values ​​in the SF6 gas humidity time series data according to a linear interpolation method, and filling missing values ​​in the SF6 gas humidity time series data, to obtain first data;

[0017] Normalizing the first data to obtain an initial normalized data set;

[0018] Inputting the initial normalized data set into a preset error correction model for prediction to obtain a test data set;

[0019] Correcting the test data set according to a preset model to obtain corrected humidity prediction data;

[0020] The corrected humidity prediction data is divided into the normalized data set including a training set, a validation set and a test set.

[0021] This preferred solution replaces outliers and fills missing values ​​using linear interpolation, ensuring data continuity and integrity. Furthermore, normalization scales the data to a consistent scale, eliminating dimensional differences between features. Predicting the normalized data using a pre-set error correction model and correcting the resulting test dataset further corrects for errors, improving data accuracy and reliability.

[0022] As a preferred solution, the test data set is corrected according to the preset model to obtain the corrected humidity prediction data, specifically:

[0023] Generating a multidimensional input sequence according to the error training set in the test data set and the initial normalized data set; wherein the error training set is obtained by performing data partitioning on the test data set;

[0024] Perform sequence prediction on the multidimensional input sequence according to a preset first BiLSTM model to obtain a prediction error value;

[0025] After denormalizing the prediction error value, the prediction error value is superimposed with the test data set to obtain the corrected humidity prediction data.

[0026] This preferred solution generates a multidimensional input sequence by introducing an error training set and an initial normalized dataset. Using the first BiLSTM model for sequence prediction, this approach captures the error patterns in the test dataset. Denormalizing the predicted error values ​​and then overlaying them with the test dataset corrects the errors in the original test data, thereby improving the accuracy of humidity prediction.

[0027] As a preferred solution, the calculation formula for the SF6 gas humidity time series data is specifically:

[0028]

[0029] Among them, f(m) is the preset parameter, α0 is the center position of the gas absorption peak, l is the absorption length, I 2f (ν0) and I f (ν0) are the second harmonic component and the fundamental component respectively, and ν0 is the frequency of the center position of the gas absorption peak.

[0030] As a preferred solution, the BiLSTM model is obtained by back-propagation training of the initial BiLSTM model based on the normalized data set, specifically:

[0031] Inputting the training set in the normalized data set into the initial BiLSTM model for bidirectional time series processing to obtain a model output value;

[0032] Establishing a loss function based on the model output value and a preset real measurement, and calculating the gradient of the loss function with respect to the model parameters through a back propagation algorithm to obtain a first gradient value;

[0033] The parameters of the initial BiLSTM model are updated according to the first gradient value, and the parameters of the initial BiLSTM model are adjusted by a cuckoo optimization algorithm to obtain the optimized BiLSTM model.

[0034] This preferred solution uses a backpropagation algorithm to calculate the gradient of the loss function with respect to the model parameters and updates the model parameters accordingly, enabling rapid convergence to an optimal solution. This allows the BiLSTM model to learn the complex features and patterns in the data. Furthermore, the Cuckoo Optimization algorithm, as a global optimization algorithm, searches for optimal parameter configurations within the solution space.

[0035] As a preferred solution, the normalized data set is predicted using a BiLSTM model to obtain the SF6 gas humidity value, specifically:

[0036] Inputting the normalized data set into the BiLSTM model; wherein the BiLSTM model includes a forward propagation layer and a backward propagation layer;

[0037] According to the forward propagation layer, from the starting point to the end point of the sequence in the normalized data set, the hidden layer state and output value at each time step are calculated step by step to obtain a first output value;

[0038] According to the back propagation layer, from the end point to the starting point of the sequence in the normalized data set, the hidden layer state and output value at each time step are calculated step by step to obtain a second output value;

[0039] The first output value and the second output value are superimposed to obtain the SF6 gas humidity value.

[0040] The BiLSTM model in this preferred solution combines forward and backward propagation layers, simultaneously utilizing both forward and backward information in time series data to generate output values. This bidirectional information fusion helps the model more comprehensively understand the dynamic characteristics and contextual relationships in the data, thereby improving the accuracy and reliability of predictions. By superimposing the output values ​​of the forward and backward propagation layers, a more accurate prediction of SF6 gas humidity values ​​can be obtained. This superposition method integrates the contributions of forward and backward information, helping to reduce prediction errors and improve prediction accuracy.

[0041] The present application also provides an SF6 humidity prediction device based on a bidirectional long short-term memory neural network, comprising a data module, a normalization module and a prediction module;

[0042] Wherein, the data module is used to obtain light intensity signal and ambient temperature data;

[0043] The normalization module is used to perform Fourier transform and normalization processing on the light intensity signal and the ambient temperature data to obtain a normalized data set;

[0044] The prediction module is used to perform data prediction on the normalized data set through a BiLSTM model to obtain the SF6 gas humidity value; wherein, the BiLSTM model is obtained by backpropagation training of an initial BiLSTM model based on the normalized data set.

[0045] As a preferred solution, the normalization module includes a sequence unit and a normalization unit;

[0046] The sequence unit is configured to obtain the gas concentration by Fourier transform according to the light intensity signal, and convert the gas concentration into relative humidity in combination with the ambient temperature data to obtain SF6 gas humidity time series data;

[0047] The normalization unit is used to perform normalization processing on the SF6 gas humidity time series data to obtain the normalized data set.

[0048] As a preferred solution, the normalization unit includes a processing subunit, a normalization subunit, a prediction subunit, a correction subunit and a division subunit;

[0049] The processing subunit is configured to replace target abnormal values ​​in the SF6 gas humidity time series data and fill missing values ​​in the SF6 gas humidity time series data according to a linear interpolation method to obtain first data;

[0050] The normalization subunit is used to perform normalization processing on the first data to obtain an initial normalized data set;

[0051] The prediction subunit is used to input the initial normalized data set into a preset error correction model for prediction to obtain a test data set;

[0052] The correction subunit is used to correct the test data set according to a preset model to obtain corrected humidity prediction data;

[0053] The division subunit is used to divide the corrected humidity prediction data into the normalized data set including a training set, a validation set and a test set. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 1 is a flow chart of a method for predicting SF6 humidity based on a bidirectional long short-term memory neural network according to an embodiment of the present application;

[0055] Figure 2 This is a schematic diagram of the BiLSTM model structure provided in the embodiment of the present application;

[0056] Figure 3 This is a flow chart of the SF6 humidity prediction model based on a bidirectional long short-term memory neural network provided in an embodiment of the present application;

[0057] Figure 4 It is a structural schematic diagram of the SF6 humidity prediction device based on the bidirectional long short-term memory neural network provided in an embodiment of the present application. DETAILED DESCRIPTION

[0058] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0059] In the description of this application, it should be understood that the terms "first," "second," "third," and "fourth" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of the technical features indicated. Thus, a feature defined as "first," "second," "third," and "fourth" may explicitly or implicitly include one or more of such features. In the description of this application, unless otherwise specified, "several" means two or more.

[0060] The core of the SF6 humidity prediction method based on a bidirectional long short-term memory neural network provided in the embodiment of the present application is to accurately predict the humidity value of SF6 gas using finite samples, thereby reducing the probability of gradient disappearance or gradient explosion during the training process.

[0061] Example 1:

[0062] See also Figure 1 The embodiment of the present application provides an SF6 humidity prediction method based on a bidirectional long short-term memory neural network, including S1 to S3, and the specific implementation steps are as follows:

[0063] S1. Obtain light intensity signal and ambient temperature data.

[0064] Step S1 of the embodiment of the present application is specifically as follows:

[0065] In a test environment or simulation environment related to GIS equipment, light intensity signals and ambient temperature data are obtained through intelligent instruments and monitoring systems combined with expression sets;

[0066] The expression set includes the first, second, and third equations. The first equation describes the light intensity attenuation caused by gas absorption, specifically describing the relationship between the incident and outgoing light intensities based on the Lambert-Beer law. The second equation is used to adjust the laser wavelength so that it periodically scans around the absorption peak center frequency v0. The third equation describes the modulated incident light intensity and includes a time-dependent modulation term.

[0067] The first formula is:

[0068] I(v)=I0(v)exp[-α(v)cl]

[0069] The second formula is:

[0070] v=v c +αsinωt

[0071] The third formula is:

[0072] I(t)=I0(l+ηsinωt)

[0073] Where I(v) is the outgoing light intensity, I0(v) is the incident light intensity, α(v) is the gas absorption coefficient, c is the gas concentration, and l is the absorption length;

[0074] ν represents the output wavelength of the laser, ν c is the center frequency of the laser; α is the frequency modulation amplitude; η is the light intensity modulation depth, ω is the frequency, and I(t) represents the output light intensity of the laser at a certain moment t.

[0075] In this embodiment, S1 can extract information related to gas concentration from the light intensity signal through Fourier transform, which is an effective signal processing method. By combining the ambient temperature data, the gas concentration can be converted into relative humidity, thereby obtaining time series data that more directly reflects the state of SF6 gas. This method helps to reduce noise and interference in the data and improve the quality of the data.

[0076] S2. Perform Fourier transform and normalization on the light intensity signal and ambient temperature data to obtain a normalized data set.

[0077] Step S2 of the embodiment of the present application includes S2.1 to S2.3, wherein S2.1 is a process of obtaining SF6 gas humidity time series data using tunable semiconductor laser absorption spectroscopy technology, S2.2 is a process of obtaining an initial normalized data set by preprocessing the data, and S2.3 is a process of performing error correction on the data to obtain a normalized data set, specifically:

[0078] S2.1. Based on Tunable Diode Laser Absorption Spectroscopy (TDLAS), the gas concentration is obtained by Fourier transform according to the light intensity signal.

[0079] The gas concentration is converted into relative humidity by combining the atmospheric temperature data in the ambient temperature data, and dynamic pressure compensation is performed at the same time. Finally, a SF6 gas humidity parameter data set arranged in chronological order is generated, that is, SF6 gas humidity time series data containing multi-dimensional information of time-relative humidity-temperature-pressure is formed. Among them, SF6 (sulfur hexafluoride) is a colorless, odorless, non-toxic inert gas composed of sulfur and fluorine atoms, with a density about 5 times that of air.

[0080] Furthermore, the calculation process for obtaining the gas concentration can be expressed as follows:

[0081] ①Absorption spectrum analysis and harmonic component extraction:

[0082] According to the definition of absorption coefficient, under weak absorption conditions, Where α(v)cl represents the degree of laser absorption by the gas under the conditions of a given wavelength v, gas concentration c, and absorption length l. At the center of the absorption peak, v = v0, so the third equation can be written as the fourth equation:

[0083]

[0084] And, the formula for the absorption coefficient is:

[0085]

[0086] Among them, α0 is the center position of the gas absorption peak, γ L is the half-width at half-height of the Lorentz lineshape.

[0087] ②Gas concentration inversion:

[0088] Performing Fourier expansion on the fourth equation can yield the fundamental component and the second harmonic component. The ratio of the two can be expressed as a ratio formula:

[0089]

[0090] It can be seen that the ratio of the fundamental component to the second harmonic component is independent of the light intensity. In actual testing, this ratio can be used to determine the concentration of the gas to be measured. Therefore, by transforming the ratio formula, we can obtain the calculation formula for the concentration of the gas to be measured, that is, the calculation formula for the SF6 gas humidity time series data:

[0091]

[0092] Among them, f(m) is the preset parameter, α0 is the center position of the gas absorption peak, l is the absorption length, I 2f (ν0) and I f (ν0) are the second harmonic component and the fundamental component respectively, and ν0 is the frequency of the center position of the gas absorption peak.

[0093] In this embodiment S2.1, TDLAS technology is an important branch of laser spectroscopy absorption method. Its principle is to utilize the selective absorption characteristics of specific gas molecules to specific frequency lasers. By carefully adjusting the emission wavelength of the semiconductor laser, it is precisely aligned with the characteristic absorption wavelength of the target gas, thereby triggering the absorption effect of the gas molecules, causing the intensity of the incident laser to attenuate. The gas concentration and the attenuation degree of the laser are positively correlated, that is, the higher the gas concentration, the more significant the attenuation of the laser. Therefore, by accurately measuring the attenuation of the laser intensity, the gas concentration value can be reversely calculated;

[0094] Furthermore, when GIS equipment experiences an abnormal leak, SF6 gas concentration can rise dramatically. Using TDLAS technology, by precisely tuning the semiconductor laser wavelength to match the characteristic absorption wavelength of SF6, gas molecules are stimulated to absorb, weakening the incident laser intensity. By measuring the real-time decay of laser intensity, the gas concentration can be quickly inferred. Furthermore, based on the selective absorption of laser light by gas molecules, TDLAS technology accurately identifies and measures the target gas, effectively eliminating interference from other gases.

[0095] S2.2. Replace target outliers in the SF6 gas humidity time series data and fill missing values ​​in the SF6 gas humidity time series data using a linear interpolation method to obtain first data. The linear interpolation method is to establish a linear relationship between two valid data points before and after the missing value or outlier, and perform interpolation calculations using a first-order straight line equation at that location to estimate and replace the outlier or missing data.

[0096] The first data and related influencing factors are normalized and the data are normalized to [0, 1] to obtain an initial normalized data set.

[0097] The normalization formula is:

[0098]

[0099] Among them, x′ is the normalized value; x is the actual value; x max and x min are the maximum and minimum values ​​of the variable, respectively.

[0100] In this embodiment S2.2, outliers are replaced and missing values ​​are filled according to the linear interpolation method, thereby ensuring the continuity and integrity of the data. Furthermore, the normalization process scales the data to the same scale, eliminating the dimensional differences between different features, ensuring that different variables have the same impact on the humidity of SF6 gas, and ensuring the convergence speed and improving the model accuracy.

[0101] S2.3. Input the initial normalized data set into the preset error correction model for prediction to obtain a test data set.

[0102] The test dataset is divided into an error training set and an error test set in a ratio of 8:2. The error training set is compared with the initial normalized dataset to generate a multidimensional input sequence.

[0103] Based on the multidimensional input sequence, the future trend of SF6 gas humidity is predicted using the Bidirectional Long Short-Term Memory (BiLSTM) model to obtain the prediction error value.

[0104] After denormalizing the prediction error value, it is superimposed with the test data set to obtain the corrected humidity prediction data.

[0105] The corrected humidity prediction data is divided into a normalized dataset including a training set, a validation set, and a test set.

[0106] In this embodiment S2.3, by introducing an error training set and an initial normalized data set to generate a multidimensional input sequence, and using a bidirectional long short-term memory (BiLSTM) model for sequence prediction, it is possible to capture the error patterns in the test data set. Denormalizing the predicted error values ​​and superimposing them with the test data set can correct the errors in the original test data, thereby improving the prediction accuracy of the humidity value and better ensuring the safe operation of electrical equipment.

[0107] Furthermore, accurately measuring SF6 gas humidity is crucial, as it directly impacts the safety and reliability of the equipment. However, factors such as ambient temperature, gas impurity content, and measurement equipment quality can significantly impact measurement results. Error correction technology can effectively compensate for these factors, ensuring accurate measurement results.

[0108] S3. Use the BiLSTM model to predict the normalized data set and obtain the SF6 gas humidity value. The BiLSTM model is obtained by backpropagation training of the initial BiLSTM model based on the normalized data set.

[0109] Step S3 of the embodiment of the present application includes S3.1 to S3.2, wherein S3.1 is the process of training the initial BiLSTM model to obtain the BiLSTM model, and S3.2 is the process of predicting the SF6 gas humidity value based on the BiLSTM model, specifically:

[0110] S3.1. Input the training set from the normalized dataset into the initial BiLSTM model for bidirectional time series processing to obtain the model output value. The initial BiLSTM model, also known as the Bidirectional Long Short-Term Memory model, is an enhanced form of the Long Short-Term Memory (LSTM) network.

[0111] Using the fitting coefficient R 2 , the loss function is calculated based on the model output value and the preset real measurement value;

[0112] Calculate the gradient of the loss function with respect to the model parameters through the back propagation algorithm to obtain the first gradient value;

[0113] Update the parameters of the initial BiLSTM model according to the first gradient value, and adjust the parameters of the initial BiLSTM model through the cuckoo optimization algorithm;

[0114] Repeat the above steps until the performance improvement of the model on the validation set in the normalized dataset stabilizes or reaches the preset iteration limit, thereby obtaining an optimized BiLSTM model.

[0115] The calculation formula of the fitting coefficient is:

[0116]

[0117] Among them, n is the amount of data, y t Indicates the true value of SF6 gas humidity at time t, represents the average value of the true value at time t, Indicates the corresponding predicted value.

[0118] In addition, the following describes the process of adjusting the initial BiLSTM model using the Cuckoo optimization algorithm:

[0119] During the parameter optimization process, the algorithm follows the following three core assumptions:

[0120] ① Single parameter optimization and random host selection: In each iteration, each cuckoo optimizes only a subset of parameters and places its "egg" (i.e., the optimized solution) in a randomly selected host nest, thereby simulating the fine-tuning of a single parameter.

[0121] ②Optimal retention strategy: During the search process, once a better nest location (i.e., a better parameter combination) is found, the algorithm will immediately retain the location for the next round of optimization to ensure the correctness and efficiency of the optimization direction.

[0122] ③ Number of host nests and discovery probability: The number of host nests is fixed, and the probability Pa of each nest being discovered by a cuckoo and used for incubation is between 0 and 1. This setting aims to balance the breadth and depth of the search.

[0123] Based on the above assumptions, the algorithm uses a global walk mechanism to continuously update the nest locations in the population, promoting iterative evolution and optimization of the population. The specific steps are as follows:

[0124] ① Initialization phase:

[0125] 1) Parameter encoding: First, the parameters of the initial Bi-LSTM model (such as weights and biases) are encoded as a vector of cuckoo positions; where each cuckoo position represents a specific set of parameter configurations.

[0126] 2) Population initialization: Randomly generate a set of cuckoo positions (i.e., parameter configurations) as the initial population; the population size can be determined based on the complexity of the problem and computing resources.

[0127] 3) Fitness function: Define a fitness function to evaluate the quality of each cuckoo position (parameter configuration); the fitness function can be the loss function value of the model on the validation set, such as mean square error, cross entropy, etc.

[0128] ②Iterative optimization stage:

[0129] 1) Fitness evaluation: For each cuckoo position in the current population, the initial Bi-LSTM model is used to make predictions and its fitness value is calculated according to the fitness function.

[0130] 2) Select the optimal nest: record the optimal cuckoo position (i.e., the optimal parameter configuration) and its fitness value in the current population.

[0131] 3) Global Walk (Lévy Flight): For each cuckoo position in the population, a new search path is generated based on the Lévy flight principle and its position is updated. The new position represents a new parameter configuration. The Lévy flight principle ensures a long-tail distribution of search paths, facilitating both large-scale jumps and small-scale, refined searches in the search space.

[0132] The formula for position update is:

[0133]

[0134] The search path for the bird's nest optimization must obey the Lévy flight principle:

[0135]

[0136] in, Indicates the position of the cuckoo after the update, represents the position of the cuckoo in the tth generation; α represents the step size factor; λ is the Lévy index; L(λ) is the step size of the random optimization search; μ and ν are both randomly generated values.

[0137] 4) Host nest update: A portion of host nests are randomly selected for update with a certain probability Pa. If the fitness value of the newly generated nest location (i.e., the new parameter configuration) is better than the current host nest location, it will be replaced.

[0138] 5) Iteration Termination Condition: Repeat the above steps until a comprehensive search of all potential nest locations is completed. The algorithm terminates when the maximum number of iterations is reached or the fitness value drops below a preset threshold, indicating that the global search is complete.

[0139] ③Output optimal parameter configuration:

[0140] When the iteration termination condition is met, the current optimal cuckoo position (i.e., the optimal parameter configuration) is output as the final parameter of the Bi-LSTM model.

[0141] In this embodiment, S3.1 uses a backpropagation algorithm to calculate the gradient of the loss function with respect to the model parameters and updates the model parameters accordingly, enabling rapid convergence to an optimal solution. This allows the BiLSTM model to learn the complex features and patterns in the data. Furthermore, the Cuckoo Optimization algorithm, as a global optimization algorithm, searches for optimal parameter configurations within the solution space. This algorithm not only ensures a comprehensive search process but also increases the likelihood of finding a globally optimal solution.

[0142] S3.2. For the BiLSTM model, determine the number of layers and the number of units (or neurons) in each layer, and clarify the specific structure of the input and output layers. Then, use a random initialization strategy to set the initial values ​​of the BiLSTM model weights. Furthermore, after the initialization, the BiLSTM model includes a forward propagation layer and a backward propagation layer.

[0143] Input the test set in the normalized dataset into the BiLSTM model; according to the forward propagation layer, from the starting point to the end point of the sequence in the normalized dataset (i.e., from time step 1 to t), gradually calculate the hidden layer state and output value at each time step to obtain the first output value;

[0144] The backpropagation layer calculates the hidden layer state and output value at each time step, from the end point to the starting point of the sequence in the normalized dataset (i.e., from time step t to 1), to obtain the second output value. Furthermore, at each time step, the forward and backward propagation layers independently manage and update their respective state information. Furthermore, the forward LSTM processes the sequence data in the current direction, while the backward LSTM processes the sequence data in the opposite direction. LSTM stands for Long Short-Term Memory.

[0145] The first and second output values ​​are superimposed to obtain the SF6 gas humidity value. Based on this SF6 gas humidity value, the GIS equipment operation and maintenance strategy is adjusted to prevent damage due to excessive humidity. The SF6 gas humidity value is an estimate of the humidity trend over a period of time or a predicted humidity value at a specific time in the future.

[0146] Using the root mean square error E RMSE To evaluate the accuracy of SF6 gas humidity value, if the root mean square error E RMSE If the value is greater than the preset threshold, the network parameters of the BiLSTM model are adjusted.

[0147] Among them, the mathematical expressions of the forward propagation layer and the back propagation layer are:

[0148] h t =f(W1·x t +W2·h t-1 );

[0149] o t =g(W4·x t +W6·h t ′)

[0150] The formula for calculating the root mean square error is:

[0151]

[0152] Among them, h t 、h t ′ are the forward and backward hidden layer states respectively; W1-W6 are shared weights, x t is the network input, o t is the network output, f and g are network functions;

[0153] y t Indicates the true value of SF6 gas humidity at time t, represents the corresponding predicted value; ρ represents the number of training samples.

[0154] To apply this application example, please refer to Figure 2 , Figure 2 This is a schematic diagram of the BiLSTM model structure provided in the embodiment of this application, showing the structure of the BiLSTM model. Figure 2 As shown in the figure, the BiLSTM model can effectively overcome the limitations of traditional LSTM networks in processing insufficient information by constructing a bidirectional network structure containing forward and backward propagation paths, thereby improving data utilization efficiency and prediction performance.

[0155] In this embodiment S3.2, the BiLSTM model combines forward and backward propagation layers, and can simultaneously utilize the forward and backward information in the time series data to obtain output values. This bidirectional information fusion method helps the model more comprehensively understand the dynamic characteristics and contextual relationships in the data to generate outputs at each time point, thereby improving the accuracy and reliability of the prediction. By superimposing the output values ​​of the forward and backward propagation layers, a more accurate prediction result of the SF6 gas humidity value can be obtained. This superposition method combines the contributions of forward and backward information, helps to reduce prediction errors, effectively overcomes the limitations of traditional LSTM in processing missing information, and improves data utilization and prediction efficiency.

[0156] In addition, the long short-term memory network, as an upgraded version of the recurrent neural network, performs well in processing time series data, especially data containing long time intervals and delayed events, and effectively solves the problem of gradient vanishing or exploding in traditional RNNs. However, the standard LSTM only processes data in one direction during training, which is not efficient for data such as SF6 gas humidity in GIS equipment, which has a long collection cycle and limited data volume. For this reason, a bidirectional long short-term memory (BiLSTM) network was proposed. BiLSTM is composed of two LSTM networks, which can simultaneously utilize the past and future information of the time series to improve prediction accuracy. Therefore, in this embodiment, BiLSTM can combine current, historical and future SF6 gas humidity data to achieve more accurate prediction and correction of humidity.

[0157] To apply this application example, please refer to Figure 3 , Figure 3 This is a flow chart of the SF6 humidity prediction model based on the bidirectional long short-term memory neural network provided in an embodiment of the present application, which shows the process of predicting SF6 gas humidity based on the bidirectional long short-term memory neural network in this embodiment.

[0158] Overall, this application has the following beneficial effects:

[0159] This application applies Fourier transforms and normalization to the light intensity signal and ambient temperature data, effectively extracting key features and eliminating dimensional differences. The Fourier transform helps identify periodic components in the data, while normalization ensures the data is on a uniform scale, facilitating model processing. These preprocessing steps improve data quality, enabling better model learning and generalization even with limited sample sizes. Backpropagation training of the initial BiLSTM model using the normalized dataset continuously optimizes the model's parameters, ensuring a better fit to the data and reducing prediction error. Furthermore, the BiLSTM model is particularly well-suited for processing time series data because it can simultaneously consider both forward and backward temporal dependencies, significantly enhancing its ability to capture long-term dependencies and complex patterns in the data. The trained BiLSTM model can predict the normalized dataset and output predicted SF6 gas humidity values. Because the model has learned key features and patterns in the data, it can accurately predict humidity values ​​even with limited sample sizes.

[0160] Example 2:

[0161] See also Figure 4 , an embodiment of the present application provides an SF6 humidity prediction device based on a bidirectional long short-term memory neural network, comprising a data module 10, a normalization module 20 and a prediction module 30;

[0162] The data module 10 is used to obtain light intensity signals and ambient temperature data;

[0163] A normalization module 20 is used to perform Fourier transform and normalization processing on the light intensity signal and the ambient temperature data to obtain a normalized data set;

[0164] The prediction module 30 is used to perform data prediction on the normalized data set through the BiLSTM model to obtain the SF6 gas humidity value; wherein the BiLSTM model is obtained by back-propagation training of the initial BiLSTM model based on the normalized data set.

[0165] In one embodiment, the data module 10 is specifically:

[0166] In a test environment or simulation environment related to GIS equipment, light intensity signals and ambient temperature data are obtained through intelligent instruments and monitoring systems combined with expression sets;

[0167] The expression set includes the first, second, and third equations. The first equation describes the light intensity attenuation caused by gas absorption, specifically describing the relationship between the incident and outgoing light intensities based on the Lambert-Beer law. The second equation is used to adjust the laser wavelength so that it periodically scans around the absorption peak center frequency v0. The third equation describes the modulated incident light intensity and includes a time-dependent modulation term.

[0168] The first formula is:

[0169] I(v)=I0(v)exp[-α(v)cl]

[0170] The second formula is:

[0171] v=v c +αsinωt

[0172] The third formula is:

[0173] I(t)=I0(l+ηsinωt)

[0174] Where I(v) is the outgoing light intensity, I0(v) is the incident light intensity, α(v) is the gas absorption coefficient, c is the gas concentration, and l is the absorption length;

[0175] ν represents the output wavelength of the laser, ν c is the center frequency of the laser; α is the frequency modulation amplitude; η is the light intensity modulation depth, ω is the frequency, and I(t) represents the output light intensity of the laser at a certain moment t.

[0176] The data module 10 of this embodiment can extract information related to gas concentration from the light intensity signal through Fourier transform, which is an effective signal processing method. By combining the ambient temperature data, the gas concentration can be converted into relative humidity, thereby obtaining time series data that more directly reflects the state of SF6 gas. This method helps to reduce noise and interference in the data and improve the quality of the data.

[0177] In one embodiment, the normalization module 20 includes a sequence unit, a processing subunit, a normalization subunit, a prediction subunit, a correction subunit, and a partitioning subunit;

[0178] The sequence unit is a process for obtaining SF6 gas humidity time series data based on tunable semiconductor laser absorption spectroscopy technology. The processing subunit and normalization subunit are processes for obtaining an initial normalized data set through data preprocessing. The prediction subunit, correction subunit, and partitioning subunit are processes for performing error correction on the data to obtain a normalized data set. Specifically:

[0179] A sequence unit is used to obtain gas concentration through Fourier transform according to the light intensity signal based on Tunable Diode Laser Absorption Spectroscopy (TDLAS) technology;

[0180] The sequence unit is also used to convert gas concentration into relative humidity in combination with the atmospheric temperature data in the ambient temperature data, and perform dynamic pressure compensation at the same time, ultimately generating a SF6 gas humidity parameter data set arranged in chronological order, that is, forming SF6 gas humidity time series data containing multi-dimensional information of time-relative humidity-temperature-pressure; among them, SF6 (sulfur hexafluoride) is a colorless, odorless, non-toxic inert gas composed of sulfur and fluorine atoms, with a density about 5 times that of air.

[0181] Furthermore, the calculation process for obtaining the gas concentration can be expressed as follows:

[0182] ①Absorption spectrum analysis and harmonic component extraction:

[0183] According to the definition of absorption coefficient, under weak absorption conditions, Where α(v)cl represents the degree of laser absorption by the gas under the conditions of a given wavelength v, gas concentration c, and absorption length l. At the center of the absorption peak, v = v0, so the third equation can be written as the fourth equation:

[0184]

[0185] And, the formula for the absorption coefficient is:

[0186]

[0187] Among them, α0 is the center position of the gas absorption peak, γ L is the half-width at half-height of the Lorentz lineshape.

[0188] ②Gas concentration inversion:

[0189] Performing Fourier expansion on the fourth equation can yield the fundamental component and the second harmonic component. The ratio of the two can be expressed as a ratio formula:

[0190]

[0191] It can be seen that the ratio of the fundamental component to the second harmonic component is independent of the light intensity. In actual testing, this ratio can be used to determine the concentration of the gas to be measured. Therefore, by transforming the ratio formula, we can obtain the calculation formula for the concentration of the gas to be measured, that is, the calculation formula for the SF6 gas humidity time series data:

[0192]

[0193] Among them, f(m) is the preset parameter, α0 is the center position of the gas absorption peak, l is the absorption length, I 2f (ν0) and I f (ν0) are the second harmonic component and the fundamental component respectively, and ν0 is the frequency of the center position of the gas absorption peak.

[0194] In the sequence unit of this embodiment, TDLAS technology is an important branch of laser spectroscopy absorption method. Its principle is to utilize the selective absorption characteristics of specific gas molecules to specific frequency lasers. By carefully adjusting the emission wavelength of the semiconductor laser, it is precisely aligned with the characteristic absorption wavelength of the target gas, thereby triggering the absorption effect of the gas molecules, causing the intensity of the incident laser to attenuate. The gas concentration is positively correlated with the attenuation degree of the laser, that is, the higher the gas concentration, the more significant the attenuation of the laser. Therefore, by accurately measuring the attenuation of the laser intensity, the gas concentration value can be reversely calculated;

[0195] Furthermore, when GIS equipment experiences an abnormal leak, SF6 gas concentration can rise dramatically. Using TDLAS technology, by precisely tuning the semiconductor laser wavelength to match the characteristic absorption wavelength of SF6, gas molecules are stimulated to absorb, weakening the incident laser intensity. By measuring the real-time decay of laser intensity, the gas concentration can be quickly inferred. Furthermore, based on the selective absorption of laser light by gas molecules, TDLAS technology accurately identifies and measures the target gas, effectively eliminating interference from other gases.

[0196] a processing subunit, configured to replace target outliers in the SF6 gas humidity time series data and fill missing values ​​in the SF6 gas humidity time series data according to a linear interpolation method, to obtain first data; wherein the linear interpolation method is to establish a linear relationship between two valid data points before and after the missing value or outlier, and perform interpolation calculation using a first-order straight line equation at the position to estimate and replace the outlier or missing data;

[0197] The normalization subunit is used to normalize the first data and related influencing factors, and normalize the data to [0, 1] to obtain an initial normalized data set.

[0198] The normalization formula is:

[0199]

[0200] Among them, x′ is the normalized value; x is the actual value; x max and x min are the maximum and minimum values ​​of the variable, respectively.

[0201] In this embodiment, the processing subunit and the normalization subunit replace outliers and fill missing values ​​according to the linear interpolation method, thereby ensuring the continuity and integrity of the data. Furthermore, the normalization process scales the data to the same scale, eliminating the dimensional differences between different features, ensuring that different variables have the same impact on the humidity of SF6 gas, and ensuring the convergence speed and improving the model accuracy.

[0202] The prediction subunit is used to input the initial normalized data set into the preset error correction model for prediction to obtain a test data set.

[0203] The correction subunit is used to divide the test data set into an error training set and an error test set in a ratio of 8:2, and generate a multi-dimensional input sequence by comparing the error training set with the initial normalized data set;

[0204] The correction subunit is also used to predict the future change trend of SF6 gas humidity through a bidirectional long short-term memory (BiLSTM) model based on the multidimensional input sequence to obtain a prediction error value;

[0205] The correction subunit is also used to denormalize the prediction error value and superimpose it with the test data set to obtain the corrected humidity prediction data.

[0206] The partitioning subunit is used to divide the corrected humidity prediction data into a normalized data set including a training set, a validation set and a test set.

[0207] The prediction subunit, correction subunit, and partitioning subunit of this embodiment generate a multidimensional input sequence by introducing an error training set and an initial normalized data set, and use a bidirectional long short-term memory (BiLSTM) model for sequence prediction, which can capture the error pattern in the test data set. Denormalizing the predicted error value and superimposing it with the test data set can correct the errors in the original test data, thereby improving the prediction accuracy of the humidity value and better ensuring the safe operation of electrical equipment.

[0208] Furthermore, accurately measuring SF6 gas humidity is crucial, as it directly impacts the safety and reliability of the equipment. However, factors such as ambient temperature, gas impurity content, and measurement equipment quality can significantly impact measurement results. Error correction technology can effectively compensate for these factors, ensuring accurate measurement results.

[0209] In one embodiment, the prediction module 30 includes a model unit and a prediction unit; wherein the model unit is a process of training the initial BiLSTM model to obtain the BiLSTM model, and the prediction unit is a process of predicting the SF6 gas humidity value based on the BiLSTM model;

[0210] The model unit is used to input the training set in the normalized dataset into the initial BiLSTM model for bidirectional time series processing to obtain the model output value. The initial BiLSTM model, also known as the Bidirectional Long Short-Term Memory model, is an enhanced form of the Long Short-Term Memory (LSTM) network.

[0211] Model unit, also used to use the fitting coefficient R 2 , the loss function is calculated based on the model output value and the preset real measurement value;

[0212] The model unit is further configured to calculate the gradient of the loss function with respect to the model parameters through a back-propagation algorithm to obtain a first gradient value;

[0213] The model unit is further configured to update parameters of the initial BiLSTM model according to the first gradient value and adjust the parameters of the initial BiLSTM model using a cuckoo optimization algorithm;

[0214] The model unit is also used to repeatedly execute the above steps until the performance improvement of the model on the validation set in the normalized dataset becomes stable or reaches a preset iteration limit, thereby obtaining an optimized BiLSTM model.

[0215] The calculation formula of the fitting coefficient is:

[0216]

[0217] Among them, n is the amount of data, y t Indicates the true value of SF6 gas humidity at time t, represents the average value of the true value at time t, Indicates the corresponding predicted value.

[0218] In addition, the following describes the process of adjusting the initial BiLSTM model using the Cuckoo optimization algorithm:

[0219] During the parameter optimization process, the algorithm follows the following three core assumptions:

[0220] ① Single parameter optimization and random host selection: In each iteration, each cuckoo optimizes only a subset of parameters and places its "egg" (i.e., the optimized solution) in a randomly selected host nest, thereby simulating the fine-tuning of a single parameter.

[0221] ②Optimal retention strategy: During the search process, once a better nest location (i.e., a better parameter combination) is found, the algorithm will immediately retain the location for the next round of optimization to ensure the correctness and efficiency of the optimization direction.

[0222] ③ Number of host nests and discovery probability: The number of host nests is fixed, and the probability Pa of each nest being discovered by a cuckoo and used for incubation is between 0 and 1. This setting aims to balance the breadth and depth of the search.

[0223] Based on the above assumptions, the algorithm uses a global walk mechanism to continuously update the nest locations in the population, promoting iterative evolution and optimization of the population. The specific steps are as follows:

[0224] ① Initialization phase:

[0225] 1) Parameter encoding: First, the parameters of the initial Bi-LSTM model (such as weights and biases) are encoded as a vector of cuckoo positions; where each cuckoo position represents a specific set of parameter configurations.

[0226] 2) Population initialization: Randomly generate a set of cuckoo positions (i.e., parameter configurations) as the initial population; the population size can be determined based on the complexity of the problem and computing resources.

[0227] 3) Fitness function: Define a fitness function to evaluate the quality of each cuckoo position (parameter configuration); the fitness function can be the loss function value of the model on the validation set, such as mean square error, cross entropy, etc.

[0228] ②Iterative optimization stage:

[0229] 1) Fitness evaluation: For each cuckoo position in the current population, the initial Bi-LSTM model is used to make predictions and its fitness value is calculated according to the fitness function.

[0230] 2) Select the optimal nest: record the optimal cuckoo position (i.e., the optimal parameter configuration) and its fitness value in the current population.

[0231] 3) Global Walk (Lévy Flight): For each cuckoo position in the population, a new search path is generated based on the Lévy flight principle and its position is updated. The new position represents a new parameter configuration. The Lévy flight principle ensures a long-tail distribution of search paths, facilitating both large-scale jumps and small-scale, refined searches in the search space.

[0232] The formula for position update is:

[0233]

[0234] The search path for the bird's nest optimization must obey the Lévy flight principle:

[0235]

[0236] in, Indicates the position of the cuckoo after the update, represents the position of the cuckoo in the tth generation; α represents the step size factor; λ is the Lévy index; L(λ) is the step size of the random optimization search; μ and ν are both randomly generated values.

[0237] 4) Host nest update: A portion of host nests are randomly selected for update with a certain probability Pa. If the fitness value of the newly generated nest location (i.e., the new parameter configuration) is better than the current host nest location, it will be replaced.

[0238] 5) Iteration Termination Condition: Repeat the above steps until a comprehensive search of all potential nest locations is completed. The algorithm terminates when the maximum number of iterations is reached or the fitness value drops below a preset threshold, indicating that the global search is complete.

[0239] ③Output optimal parameter configuration:

[0240] When the iteration termination condition is met, the current optimal cuckoo position (i.e., the optimal parameter configuration) is output as the final parameter of the Bi-LSTM model.

[0241] The model unit in this embodiment uses a backpropagation algorithm to calculate the gradient of the loss function with respect to the model parameters and updates the model parameters accordingly, enabling rapid convergence to an optimal solution. This allows the BiLSTM model to learn the complex features and patterns in the data. Furthermore, the Cuckoo Optimization algorithm, as a global optimization algorithm, searches for optimal parameter configurations within the solution space; this algorithm not only ensures a comprehensive search process but also increases the likelihood of finding the global optimal solution.

[0242] The prediction unit is used to determine the number of layers and the number of units (or neurons) in each layer of the BiLSTM model, and to clarify the specific structure of the input and output layers. Subsequently, a random initialization strategy is used to set the initial values ​​of the weights in the BiLSTM model. Moreover, after the initial setting, the BiLSTM model includes a forward propagation layer and a backward propagation layer.

[0243] The prediction unit is also used to input the test set in the normalized dataset into the BiLSTM model; according to the forward propagation layer, from the starting point to the end point of the sequence in the normalized dataset (i.e., from time step 1 to t), the hidden layer state and output value at each time step are calculated step by step to obtain the first output value;

[0244] The prediction unit is also used to calculate the hidden layer state and output value at each time step from the end point to the starting point of the sequence in the normalized dataset (i.e., from time step t to 1) based on the backpropagation layer to obtain the second output value. Furthermore, at each time step, the forward and backpropagation layers independently manage and update their respective state information. Furthermore, the forward LSTM processes the sequence data in the current direction, while the backward LSTM processes the sequence data in the opposite direction. LSTM stands for Long Short-Term Memory.

[0245] The prediction unit is further configured to superimpose the first and second output values ​​to obtain an SF6 gas humidity value. The SF6 gas humidity value is used to adjust the GIS equipment's operation and maintenance strategy based on the SF6 gas humidity value to prevent damage to the equipment due to excessive humidity. The SF6 gas humidity value is an estimate of the humidity trend over a period of time, or a predicted humidity value at a specific point in the future.

[0246] The prediction unit is also used to use the root mean square error E RMSE To evaluate the accuracy of SF6 gas humidity value, if the root mean square error E RMSE If the value is greater than the preset threshold, the network parameters of the BiLSTM model are adjusted.

[0247] Among them, the mathematical expressions of the forward propagation layer and the back propagation layer are:

[0248] h t =f(W1·xt +W2·h t-1 );

[0249] o t =g(W4·x t +W6·h t ′)

[0250] The formula for calculating the root mean square error is:

[0251]

[0252] Among them, h t 、h t ′ are the forward and backward hidden layer states respectively; W1-W6 are shared weights, x t is the network input, o t is the network output, f and g are network functions;

[0253] y t Indicates the true value of SF6 gas humidity at time t, represents the corresponding predicted value; ρ represents the number of training samples.

[0254] To apply this application example, please refer to Figure 2 , Figure 2 This is a schematic diagram of the BiLSTM model structure provided in the embodiment of this application, showing the structure of the BiLSTM model. Figure 2 As shown in the figure, the BiLSTM model can effectively overcome the limitations of traditional LSTM networks in processing insufficient information by constructing a bidirectional network structure containing forward and backward propagation paths, thereby improving data utilization efficiency and prediction performance.

[0255] In the prediction unit of this embodiment, the BiLSTM model combines the forward propagation layer and the backward propagation layer, and can simultaneously use the forward and backward information in the time series data to obtain the output value. This two-way information fusion method helps the model more comprehensively understand the dynamic characteristics and contextual relationships in the data to generate the output at each time point, thereby improving the accuracy and reliability of the prediction. By superimposing the output values ​​of the forward propagation layer and the backward propagation layer, a more accurate prediction result of the SF6 gas humidity value can be obtained. This superposition method combines the contributions of forward and backward information, helps to reduce prediction errors, effectively overcomes the limitations of traditional LSTM in processing missing information, and improves data utilization and prediction efficiency.

[0256] In addition, the long short-term memory network, as an upgraded version of the recurrent neural network, performs well in processing time series data, especially data containing long time intervals and delayed events, and effectively solves the problem of gradient vanishing or exploding in traditional RNNs. However, the standard LSTM only processes data in one direction during training, which is not efficient for data such as SF6 gas humidity in GIS equipment, which has a long collection cycle and limited data volume. For this reason, a bidirectional long short-term memory (BiLSTM) network was proposed. BiLSTM is composed of two LSTM networks, which can simultaneously utilize the past and future information of the time series to improve prediction accuracy. Therefore, in this embodiment, BiLSTM can combine current, historical and future SF6 gas humidity data to achieve more accurate prediction and correction of humidity.

[0257] To apply this application example, please refer to Figure 3 , Figure 3 This is a flow chart of the SF6 humidity prediction model based on the bidirectional long short-term memory neural network provided in an embodiment of the present application, which shows the process of predicting SF6 gas humidity based on the bidirectional long short-term memory neural network in this embodiment.

[0258] Overall, this application has the following beneficial effects:

[0259] This application applies Fourier transforms and normalization to the light intensity signal and ambient temperature data, effectively extracting key features and eliminating dimensional differences. The Fourier transform helps identify periodic components in the data, while normalization ensures the data is on a uniform scale, facilitating model processing. These preprocessing steps improve data quality, enabling better model learning and generalization even with limited sample sizes. Backpropagation training of the initial BiLSTM model using the normalized dataset continuously optimizes the model's parameters, ensuring a better fit to the data and reducing prediction error. Furthermore, the BiLSTM model is particularly well-suited for processing time series data because it can simultaneously consider both forward and backward temporal dependencies, significantly enhancing its ability to capture long-term dependencies and complex patterns in the data. The trained BiLSTM model can predict the normalized dataset and output predicted SF6 gas humidity values. Because the model has learned key features and patterns in the data, it can accurately predict humidity values ​​even with limited sample sizes.

[0260] Example 3:

[0261] An embodiment of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium includes a stored computer program, wherein when the computer program is executed, the device where the computer-readable storage medium is located is controlled to execute the SF6 humidity prediction method based on the bidirectional long short-term memory neural network;

[0262] Wherein, the SF6 humidity prediction method based on the bidirectional long short-term memory neural network, if implemented in the form of a software functional unit and used as an independent product, can be stored in a computer-readable storage medium. Based on this understanding, the present invention implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, it can implement the steps of the above-mentioned various method embodiments. Wherein, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium may include: any entity or device that can carry the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal and software distribution medium, etc.

[0263] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. The SF6 humidity prediction method based on bidirectional long short-term memory neural network is characterized by: include: Obtain light intensity signal and ambient temperature data; Performing Fourier transform and normalization processing on the light intensity signal and ambient temperature data to obtain a normalized data set; The normalized data set is predicted using a BiLSTM model to obtain a humidity value of the SF6 gas; wherein the BiLSTM model is obtained by backpropagation training of an initial BiLSTM model based on the normalized data set.

2. The SF6 humidity prediction method based on a bidirectional long short-term memory neural network according to claim 1, characterized in that: The light intensity signal and ambient temperature data are subjected to Fourier transformation and normalization to obtain a normalized data set, specifically: According to the light intensity signal, the gas concentration is obtained by Fourier transform, and the gas concentration is converted into relative humidity in combination with the ambient temperature data to obtain SF6 gas humidity time series data; The SF6 gas humidity time series data is normalized to obtain the normalized data set.

3. The SF6 humidity prediction method based on a bidirectional long short-term memory neural network according to claim 2, characterized in that: The SF6 gas humidity time series data is normalized to obtain the normalized data set, specifically: replacing target abnormal values ​​in the SF6 gas humidity time series data according to a linear interpolation method, and filling missing values ​​in the SF6 gas humidity time series data, to obtain first data; Normalizing the first data to obtain an initial normalized data set; Inputting the initial normalized data set into a preset error correction model for prediction to obtain a test data set; Correcting the test data set according to a preset model to obtain corrected humidity prediction data; The corrected humidity prediction data is divided into the normalized data set including a training set, a validation set and a test set.

4. The SF6 humidity prediction method based on a bidirectional long short-term memory neural network according to claim 3, characterized in that: The test data set is corrected according to the preset model to obtain the corrected humidity prediction data, specifically: Generating a multidimensional input sequence according to the error training set in the test data set and the initial normalized data set; wherein the error training set is obtained by performing data partitioning on the test data set; Perform sequence prediction on the multidimensional input sequence according to a preset first BiLSTM model to obtain a prediction error value; After denormalizing the prediction error value, the prediction error value is superimposed with the test data set to obtain the corrected humidity prediction data.

5. The SF6 humidity prediction method based on a bidirectional long short-term memory neural network according to any one of claims 2 to 4, characterized in that: The calculation formula of the SF6 gas humidity time series data is specifically: Among them, f(m) is the preset parameter, α0 is the center position of the gas absorption peak, l is the absorption length, I 2f (v0) and I f (v0) are the second harmonic component and the fundamental component respectively, and v0 is the frequency of the center position of the gas absorption peak.

6. The SF6 humidity prediction method based on a bidirectional long short-term memory neural network according to claim 1, characterized in that: The BiLSTM model is obtained by back-propagation training of the initial BiLSTM model based on the normalized data set, specifically: Inputting the training set in the normalized data set into the initial BiLSTM model for bidirectional time series processing to obtain a model output value; Establishing a loss function based on the model output value and a preset real measurement, and calculating the gradient of the loss function with respect to the model parameters through a back propagation algorithm to obtain a first gradient value; The parameters of the initial BiLSTM model are updated according to the first gradient value, and the parameters of the initial BiLSTM model are adjusted by a cuckoo optimization algorithm to obtain the optimized BiLSTM model.

7. The SF6 humidity prediction method based on a bidirectional long short-term memory neural network according to claim 1, characterized in that: The normalized data set is predicted using the BiLSTM model to obtain the SF6 gas humidity value, specifically: Inputting the normalized data set into the BiLSTM model; wherein the BiLSTM model includes a forward propagation layer and a backward propagation layer; According to the forward propagation layer, from the starting point to the end point of the sequence in the normalized data set, the hidden layer state and output value at each time step are calculated step by step to obtain a first output value; According to the back propagation layer, from the end point to the starting point of the sequence in the normalized data set, the hidden layer state and output value at each time step are calculated step by step to obtain a second output value; The first output value and the second output value are superimposed to obtain the SF6 gas humidity value.

8. SF6 humidity prediction device based on bidirectional long short-term memory neural network, characterized in that: Includes data module, normalization module and prediction module; Wherein, the data module is used to obtain light intensity signal and ambient temperature data; The normalization module is used to perform Fourier transform and normalization processing on the light intensity signal and the ambient temperature data to obtain a normalized data set; The prediction module is used to perform data prediction on the normalized data set through a BiLSTM model to obtain the SF6 gas humidity value; wherein, the BiLSTM model is obtained by backpropagation training of an initial BiLSTM model based on the normalized data set.

9. The SF6 humidity prediction device based on a bidirectional long short-term memory neural network according to claim 8, characterized in that: The normalization module includes a sequence unit and a normalization unit; The sequence unit is configured to obtain the gas concentration by Fourier transform according to the light intensity signal, and convert the gas concentration into relative humidity in combination with the ambient temperature data to obtain SF6 gas humidity time series data; The normalization unit is used to perform normalization processing on the SF6 gas humidity time series data to obtain the normalized data set.

10. The SF6 humidity prediction device based on a bidirectional long short-term memory neural network according to claim 9, characterized in that: The normalization unit includes a processing subunit, a normalization subunit, a prediction subunit, a correction subunit and a division subunit; The processing subunit is configured to replace target abnormal values ​​in the SF6 gas humidity time series data and fill missing values ​​in the SF6 gas humidity time series data according to a linear interpolation method to obtain first data; The normalization subunit is used to perform normalization processing on the first data to obtain an initial normalized data set; The prediction subunit is used to input the initial normalized data set into a preset error correction model for prediction to obtain a test data set; The correction subunit is used to correct the test data set according to a preset model to obtain corrected humidity prediction data; The division subunit is used to divide the corrected humidity prediction data into the normalized data set including a training set, a validation set and a test set.

Images