A method, medium and system for predicting the content range of dissolved gases in transformer oil

Abstract

The present invention discloses a method, medium and system for predicting the content interval of dissolved gases in transformer oil, including: collecting the content of dissolved gases in transformer oil at historical time to obtain a first sequence of the content of dissolved gases in transformer oil; decomposing the first sequence by using an adaptive noise complete ensemble empirical mode decomposition algorithm to obtain a plurality of second sequences; inputting each second sequence into a trained point prediction model corresponding to each second sequence, and outputting a third sequence corresponding to each second sequence; adding the elements with the same sorting in all the third sequences to obtain a point prediction result of the content of dissolved gases in transformer oil within a preset time after historical time; and using a nonparametric kernel density estimation algorithm for the point prediction result to calculate an interval prediction result of the content of dissolved gases in transformer oil within the preset time. The present invention introduces interval prediction, which can quantitatively analyze the changes of gases in oil caused by external uncertain factors, and helps to improve the accuracy of state estimation.

Description

Technical Field

[0001] The present invention relates to the technical field of predicting the content of dissolved gases in transformer oil, and particularly to a method, medium and system for predicting the content range of dissolved gases in transformer oil. Background Art

[0002] During the operation of a transformer, various gases are generated due to reasons such as electrical aging, thermal aging, and faults and dissolved in the transformer oil. The potential fault risks of the transformer can be detected by using the dissolved gas analysis technology in the oil. At present, many transformers have obtained historical data on the content of dissolved gases in the oil by installing on-line monitoring devices for oil chromatography. Therefore, by predicting the future change trend of the dissolved gases in the oil, it is helpful to predict the operating state of the transformer, detect potential faults, and reasonably arrange maintenance.

[0003] At present, the methods for predicting the content of dissolved gases in transformer oil are mainly divided into two categories: statistical methods and artificial intelligence prediction methods. Since the time series data of the dissolved gases in transformer oil is non-linear and non-stationary, the statistical prediction model has certain limitations in predicting a long gas content sequence. Compared with the statistical method, the artificial intelligence algorithm has certain advantages in time series data prediction. However, most of the existing research belongs to point prediction algorithms. Due to the fact that the content of dissolved gases in transformer oil may be affected by external random factors and limited by the error of oil chromatography monitoring, the actual monitoring data has certain uncertainty. Summary of the Invention

[0004] Embodiments of the present invention provide a method, medium and system for predicting the content range of dissolved gases in transformer oil to solve the problem that the point prediction algorithm in the prior art has uncertainty in predicting the content of dissolved gases in transformer oil.

[0005] In a first aspect, a method for predicting the content range of dissolved gases in transformer oil is provided, including:

[0006] Collect the content of dissolved gases in transformer oil at historical time to obtain a first sequence of the content of dissolved gases in transformer oil;

[0007] Decompose the first sequence by using the adaptive noise complete ensemble empirical mode decomposition algorithm to obtain a plurality of second sequences;

[0008] Input each second sequence into a trained point prediction model corresponding to each second sequence, and output a third sequence corresponding to each second sequence;

[0009] Sum up the elements with the same sorting in all the third sequences to obtain a point prediction result of the content of dissolved gases in transformer oil within a preset time after the historical time;

[0010] Using the non-parametric kernel density estimation algorithm for the point prediction results, an interval prediction result of the dissolved gas content in transformer oil within a preset time is calculated.

[0011] In a second aspect, a computer-readable storage medium is provided, on which computer program instructions are stored; when the computer program instructions are executed by a processor, the method for predicting the interval of the dissolved gas content in transformer oil as described in the embodiments of the first aspect above is implemented.

[0012] In a third aspect, a system for predicting the interval of the dissolved gas content in transformer oil is provided, including: the computer-readable storage medium as described in the embodiments of the second aspect above.

[0013] In this way, the embodiments of the present invention not only obtain the point prediction results of the gas content, but also introduce interval prediction, and can generate prediction intervals of the dissolved gas content in the oil under different confidence levels, thereby quantitatively analyzing the changes in the gas in the oil caused by external uncertain factors, which helps to improve the accuracy of state estimation. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments of the present invention. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0015] Figure 1 is a flowchart of the method for predicting the interval of the dissolved gas content in transformer oil according to the embodiments of the present invention;

[0016] Figure 2 is a flowchart of the adaptive noise complete ensemble empirical mode decomposition algorithm according to the embodiments of the present invention;

[0017] Figure 3 is a schematic diagram of the neuron structure of the long short-term memory network point prediction model according to the embodiments of the present invention;

[0018] Figure 4 is a schematic diagram of the gray wolf level and position update strategy of the gray wolf algorithm according to the embodiments of the invention;

[0019] Figure 5 is a flowchart of the gray wolf algorithm for optimizing the long short-term memory network point prediction model according to the embodiments of the present invention;

[0020] Figure 6 is a schematic diagram of multiple second sequences obtained by decomposing using the adaptive noise complete ensemble empirical mode decomposition algorithm according to a preferred embodiment of the present invention;

[0021] Figure 7It is a schematic diagram of the optimization process of the grey wolf algorithm in a preferred embodiment of the present invention;

[0022] Figure 8 It is a schematic diagram of the prediction result of the CO content point in a preferred embodiment of the present invention;

[0023] Figure 9 It is a schematic diagram for comparing probability density function curves in a preferred embodiment of the present invention. Among them, (a) is a schematic diagram for comparing probability density function curves of different kernel functions, and (b) is a schematic diagram for comparing probability density function curves of different window widths;

[0024] Figure 10 It is a schematic diagram of the prediction result of the CO content interval in a preferred embodiment of the present invention. Specific embodiments

[0025] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0026] The embodiments of the present invention disclose a method for predicting the content interval of dissolved gases in transformer oil. As Figure 1 shown, the method includes the following steps:

[0027] Step S1: Collect the content of dissolved gases in transformer oil at historical times to obtain a first sequence of the content of dissolved gases in transformer oil.

[0028] Specifically, it can be collected once every preset time. For example, it is collected once every four hours. The span of historical time can be determined according to the actual situation. For example, collect the CO content dissolved in transformer oil for 30 days.

[0029] Step S2: Use the complete ensemble empirical mode decomposition with adaptive noise algorithm to decompose the first sequence to obtain a plurality of second sequences.

[0030] The complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) algorithm is a time series signal processing method, which is developed from the empirical mode decomposition (EMD). EMD can adaptively decompose time series data into several independent intrinsic mode functions (IMFs) and a residual component, but it is prone to mode mixing phenomenon. The ensemble empirical mode decomposition (EEMD) effectively overcomes the mode mixing phenomenon by introducing uniformly distributed white noise multiple times during the decomposition process to mask the noise of the original signal itself; however, it will cause noise residue and increase the error of component reconstruction. As Figure 2As shown, CEEMDAN makes further improvements on the basis of EEMD. During the IMF decomposition process, white noise is adaptively added, and each IMF component is calculated through the final residue signal, which can almost eliminate the reconstruction error. Compared with EEMD, it has advantages such as completeness and rapidity.

[0031] This step specifically includes the following process:

[0032] (1) Add white noise to the first sequence K times to obtain K groups of noise sequences.

[0033] Specifically, add according to the following formula:

[0034] x i (t) = x(t) + ε 0 δ i (t).

[0035] Among them, x i (t) represents the noise sequence obtained after adding white noise for the i-th time, x(t) represents the first sequence, and δ i (t) represents the white noise added for the i-th time, and ε 0 represents the weight coefficient of the white noise added this time, which is the ratio of the standard deviation of the amplitude of the white noise to that of the original signal, and i = 1, 2,..., K. The mean value of the white noise added each time is 0.

[0036] (2) Use the CEEMDAN algorithm to decompose each group of noise sequences to obtain multiple IMF components of each group of noise sequences.

[0037] (3) Calculate the mean value of the first IMF component of all noise sequences.

[0038] The formula for the mean value of the first IMF component of all noise sequences is as follows:

[0039]

[0040] Among them, I 1 (t) represents the mean value of the first IMF component of all noise sequences, and I 1 i(t) represents the first IMF component of the noise sequence obtained after adding white noise for the i-th time.

[0041] (4) Subtract the mean value of the first IMF component of all noise sequences from the first sequence to obtain the first residue sequence.

[0042] The specific formula is as follows:

[0043] r 1 (t) = x(t) - I 1 (t).

[0044] Among them, r 1 (t) represents the first residue sequence.

[0045] (5) Repeat the above steps for the residue sequence to obtain other IMF components. The CEEMDAN decomposition ends until the extreme value of the residue sequence is less than or equal to 2.

[0046] The formula for the j-th IMF component is as follows:

[0047]

[0048] Among them, I j (t) represents the j-th IMF component, and E j (·) represents the j-th IMF component obtained by EMD decomposition. ε j-1 represents the weight coefficient of the white noise added in the (j - 1)-th time, and r j-1 (t) represents the (j - 1)-th residue sequence.

[0049] Through the above process, when the decomposition ends, the first sequence is decomposed into m IMF components and 1 residue, specifically as follows:

[0050]

[0051] The m IMF components and 1 residue obtained by decomposition are multiple second sequences.

[0052] Step S3: Input each second sequence into the trained point prediction model corresponding to each second sequence, and output the third sequence corresponding to each second sequence.

[0053] The point prediction model adopted in the embodiment of the present invention is a long short-term memory network (LSTM) point prediction model. The single neuron structure of the long short-term memory network point prediction model is as Figure 3 shown. On the basis of the recurrent neural network (RNN), it replaces the hidden layer with a gated memory unit to control the influence degree of the current moment information on the previous information. Figure 3 In it, x t represents the input at time t, c t-1 and c t represent the memory unit states at times t - 1 and t respectively, h t-1 and h t represent the neuron outputs at times t - 1 and t respectively. The gated memory unit includes three gated structures: a forget gate, an input gate, and an output gate. The specific functions are as follows:

[0054] (1) Forget gate: The forget gate reads h t-1 and x t , and outputs f t (f tBetween 0 and 1, where 0 represents "completely forgotten" and 1 represents "completely retained") as c t-1 's weight, controlling c t-1 's degree of forgetting, thus realizing c t-1 influencing c t , and the specific formula is as follows:

[0055] f t =σ(W f ·[h t-1 ,x t +b f ).

[0056] Among them, σ(·) represents the sigmoid activation function, W f and b f respectively represent the weight matrix and bias term of the forget gate.

[0057] (2) Input gate: The input gate reads h t-1 and x t , and respectively obtains the weights i t and g t through the sigmoid layer and the tanh layer, realizing h t-1 and x t influencing c t , and the specific formula is as follows:

[0058] i t =σ(W i ·[h t-1 ,x t +b i ).

[0059] g t =σ(W g ·[h t-1 ,x t +b g ).

[0060] W i and b i respectively represent the weight matrix and bias term of the sigmoid layer, and W g and b g respectively represent the weight matrix and bias term of the tanh layer.

[0061] From the above three formulas, it can be seen that the state c t of the memory unit is jointly determined by c t-1 , h t-1 and x t as follows:

[0062] c t =f t ·c t-1+i t ·g t 。

[0063] (3) Output gate: The output gate obtains h through the sigmoid layer t-1 and x t weights o t , and obtains the information of c through the tanh layer t , thus comprehensively combining h t-1 , x t and c t to obtain the neuron output by integrating the information of the three parameters. The specific formula is as follows:

[0064] o t =σ(W o ·[h t-1 , x t +b o ).

[0065] h t =o t ·tanh(c t ).

[0066] Among them, W o and b o respectively represent the weight matrix and bias term of the sigmoid layer.

[0067] The embodiment of the present invention constructs an LSTM model based on the Adam solver. Its parameters mainly include two categories: structural parameters and hyperparameters, as shown in Table 1 specifically.

[0068] Table 1 LSTM parameters

[0069] Parameter Value Input Feature Dimension 1 Number of Hidden Layer Neurons n Stacking Layers 1 Number of Fully Connected Layers 1 Number of Iterations l Gradient Threshold 1 Initial Learning Rate α Learning Rate Update Period l / 2 Learning Rate Update Factor 0.2 Solver Adam

[0070] In addition, the embodiment of the present invention pre-trains a long short-term memory network point prediction model. Each time the long short-term memory network point prediction model is trained, the Grey Wolf Optimization (GWO) algorithm is used to optimize the number of hidden layer neurons n and the initial learning rate α of the long short-term memory network point prediction model. That is, the GWO algorithm first selects a pair of values of the number of hidden layer neurons n and the initial learning rate α, and then inputs them into the LSTM model for training to obtain a prediction result. If the prediction result does not meet the requirements, the GWO algorithm is returned to select another pair of values of the number of hidden layer neurons n and the initial learning rate α, and input them into the LSTM model for training to obtain a prediction result. Repeat this process. After multiple cycles, when the output prediction result meets the requirements, a corresponding pair of optimal values of the number of hidden layer neurons n and the initial learning rate α can be obtained as the parameters of the trained LSTM model.

[0071] Grey Wolf Optimizer (GWO) is a swarm intelligence optimization algorithm that simulates the hunting behavior of grey wolf packs. It has the characteristics of simple structure, fast convergence speed, and high search efficiency. GWO divides the grey wolf population into four levels: α, β, δ, and ω, following the Figure 4 population hierarchy shown, corresponding to the four levels of the optimal solution, the sub-optimal solution, the third-best solution, and other solutions. Among them, the three grey wolves α, β, and δ are responsible for guiding the ω grey wolves towards the position of the prey (global optimal solution). The optimization process is divided into the following two steps:

[0072] (1) Prey recognition (searching for the optimal solution) step:

[0073] The specific algorithm for this step is as follows:

[0074] D = |C? X p (t) X(t)|

[0075] X(t + 1) = X p (t) - A·D

[0076] A = 2ar 1 - a

[0077] C = 2r 2

[0078] where D represents the distance between the prey and the grey wolf, X represents the position of the grey wolf, X p represents the position of the prey, t represents the number of iterations, A and C both represent coefficient vectors, a represents the convergence factor, a = 2(1 - m / M), m represents the current iteration number of GWO, and M represents the upper limit of the iteration number. As the number of iterations increases from 2 to 0, r 1 and r 2 are random vectors between 0 and 1.

[0079] (2) Prey capture (determining the optimal solution) step:

[0080] After the grey wolves recognize the prey, the α wolf will guide the entire wolf pack to surround the prey, and the positions of the α, β, and δ wolves gradually approach the position of the prey, thereby determining the optimal solution. The position update strategy is as Figure 4 shown, and the mathematical model is as follows:

[0081]

[0082]

[0083]

[0084] In the formula, D α , D β and D δ represent the distances between the individual X and α, β, and δ, C1 , C 2 , C 3 and A 1 , A 2 , A 3 have the same meanings as the aforementioned C and A, X α , X β , X δ respectively represent the positions of α, β, and δ wolves, X(t + 1) represents the position of individual X at the next moment, and X 1 , X 2 , X 3 are intermediate variables in the calculation process.

[0085] Since LSTM realizes the minimization of the error function by continuously updating network parameters during training, the setting of its initial parameters has a great impact on the training results. Usually, the empirical method is relied on to set and adjust each parameter, which has strong uncertainty and it is difficult to determine the optimal parameters of the model. The number of hidden layer neurons n and the initial learning rate α of LSTM have a significant impact on the prediction results of time series data. Therefore, this paper uses GWO to optimize the two key parameters n and α of LSTM, and the overall algorithm flow chart is as Figure 5 shown. During training, the training samples are divided into a training set and a test set. The ratio of the samples in the training set and the test set can be 4:1.

[0086] Step S4: Sum the elements with the same sorting in all the third sequences to obtain the point prediction result of the content of dissolved gases in transformer oil within the preset time after the historical time.

[0087] For example, if a total of two third sequences are generated, the elements in one third sequence are a1, a2, a3,..., aN in sequence, and the elements in the other third sequence are b1, b2, b3,..., bN in sequence, then the superimposed point prediction result is: a1 + b1, a2 + b2, a3 + b3,..., aN + bN.

[0088] Generally, the span of the historical time has a corresponding relationship with the predictable preset time. For example, in the embodiment of the present invention, the proportional relationship between the span of the historical time and the preset time is 5:1, that is, by using the samples of 30 days of historical time, the results within the next 6 days can be predicted.

[0089] Step S5: Use the non - parametric kernel density estimation algorithm for the point prediction result to calculate the interval prediction result of the content of dissolved gases in transformer oil within the preset time.

[0090] Specifically, the formula for the probability density of the non - parametric kernel density estimation algorithm is as follows:

[0091]

[0092] Among them, represents the probability density, p = {p 1 , p 2 , …, p N}, represents the error value of the point prediction result relative to the true result, p i represents the i-th point prediction error sample point, h represents the window width (h > 0), N represents the total number of point prediction result sample points, k(p, h) represents the kernel function, and both the error value p of the point prediction result relative to the true result and the window width h are independent variables of the kernel function.

[0093] Through the above steps, the method for predicting the interval of dissolved gas content in transformer oil proposed in the embodiments of the present invention can be implemented, generating prediction intervals under different confidence levels, thereby quantitatively analyzing the changes in the gases in transformer oil due to external uncertain factors, which helps to improve the accuracy of state estimation.

[0094] The embodiments of the present invention also disclose a computer-readable storage medium, on which computer program instructions are stored; when the computer program instructions are executed by a processor, the method for predicting the interval of dissolved gas content in transformer oil as described in the above embodiments is implemented.

[0095] The embodiments of the present invention also disclose a system for predicting the interval of dissolved gas content in transformer oil, including: the computer-readable storage medium as described in the above embodiments.

[0096] Next, a specific application example is used to further illustrate the technical solution of the embodiments of the present invention.

[0097] Application Example

[0098] First, use CEEMDAN to decompose the first sequence of dissolved gases in transformer oil to obtain the second sequences of each first sequence. Taking the CO gas as an example, the second sequences generated by CEEMDAN decomposition are as Figure 6 shown.

[0099] Secondly, construct an LSTM prediction model for each second sequence, and use GWO to optimize the number of hidden layer neurons n and the initial learning rate α of LSTM. The parameters of GWO are shown in Table 2.

[0100] Table 2 GWO Parameter Settings

[0101]

[0102]

[0103] Taking the optimization of the parameters of an LSTM prediction model for a second sequence as an example, the GWO optimization process is as Figure 7As shown in the figure, the abscissa in the figure is the optimization range of the number of hidden layer neurons, and the ordinate is the optimization range of the initial learning rate. The root mean square error of each iteration result is used as the fitness. Figure 7 Among them, the scatter points represent the distribution of 20 ω gray wolves during 20 iterations, and the three trajectories represent the hunting trajectories of the α, β, and δ wolves. It can be seen from the figure that during the iteration process, the ω gray wolves will first cover a large range of the value space, and then gradually gather at the optimal solution coordinates. As the positions of the gray wolves are gradually updated during the iteration process, the α, β, and δ wolves finally gather at (114.428, 0.038), indicating that the fitness is the highest at this coordinate. Therefore, the number of hidden layer neurons of this LSTM point prediction model is taken as 114, and the initial learning rate is taken as 0.038. The above optimization is performed on the LSTM prediction models of all the second sequences of gases to obtain the point prediction models of each second sequence.

[0104] The third sequence is predicted by using each second sequence through the point prediction model, and the third sequences are re-superimposed to form the gas sequence prediction result. Taking the CO gas as an example, the point prediction result is as Figure 8 shown.

[0105] The root mean square error y RMSE and the prediction accuracy y RA are selected as two indicators to evaluate the point prediction results. The calculation formulas are as follows:

[0106]

[0107]

[0108] Among them, x real represents the actual monitoring value sequence of the gas content, x real (i) represents the i-th sample point, x pred represents the point prediction result sequence of the gas content, x pred (i) represents the i-th sample point, and n represents the sample length of the gas content sequence.

[0109] The smaller the root mean square error y RMSE and the closer the prediction accuracy y RA is to 1, the better the model prediction effect. It can be calculated that the root mean square error of the point prediction result of the method proposed in the embodiment of the present invention is 4.275, and the prediction accuracy is 0.990. From the perspective of the root mean square error, the prediction result of the method proposed in the embodiment of the present invention has a very small deviation from the actual value, with an average deviation of only about 1.3%, and the prediction accuracy reaches 99.0%.

[0110] When conducting transformer condition assessment based on dissolved gases in oil, if only point prediction results are used, the uncertainty error of point prediction will be fully transmitted to the condition assessment results, affecting the accuracy of condition assessment. To characterize the uncertainty of oil chromatogram gas content prediction, based on the point prediction results, this invention uses non-parametric kernel density estimation method to generate gas content prediction intervals at different confidence levels. During the kernel density estimation process, the selection of window width h and kernel function k has a direct impact on the estimation results of probability density. Regarding the window width h, this paper uses the MISE criterion to select the optimal value of window width h, and selects the kernel function k shown in Table 3 for comparison, where I(·) is an indicator function. When |p| ≤ 1, I(|p| ≤ 1) = 1, otherwise I(|p| ≤ 1) = 0.

[0111] Table 3 Expressions of Different Kernel Functions

[0112]

[0113] Figure 9 Shows the probability density functions (PDFs) under different kernel functions and different window widths. Comparing with the frequency histogram, it can be seen that when the Gaussian kernel function is used as the kernel function and the window width is 0.7533, the adaptability is the best. Therefore, the window width is taken as 0.7533 and the Gaussian kernel function is selected as the kernel function.

[0114] After selecting the optimal window width and kernel function, use KDE to generate the prediction interval of gas content. Taking CO as an example, its prediction interval is as Figure 10 shown.

[0115] To evaluate the interval prediction results, select the interval coverage rate P ICP and the interval average width P INAW two indicators to evaluate the interval prediction results, and the calculation methods are shown as follows:

[0116]

[0117]

[0118] Among them, n represents the data length, ε i represents a Boolean variable. When the actual value is within the prediction interval, ε i is 1, otherwise it is 0; R = x real-max -x real-min represents the difference between the maximum and minimum values of the true value, and ΔP represents the width of the prediction interval.

[0119] The larger the interval coverage rate P ICP , the better the coverage effect of the interval on the actual value. The smaller the interval average width P INAW , the narrower the interval. The evaluation index results of the prediction intervals at different confidence levels are shown in Table 4.

[0120] Table 4 Evaluation of Different Confidence Intervals

[0121] Confidence Interval <![CDATA[P ICP > <![CDATA[P INAW > 98.000 0.958 0.046 90.000 0.917 0.028 80.000 0.792 0.021 70.000 0.667 0.017 60.000 0.583 0.013

[0122] From Figure 10 and Table 4, it can be seen that the method proposed in the present invention can reliably depict the change trend and confidence interval of the gas content in transformer oil. Generally speaking, if you want to increase the coverage rate of the interval, it will inevitably lead to an increase in the interval width. Therefore, when applying the interval prediction results of dissolved gases in oil for condition assessment, the two indicators can be comprehensively considered, and the interval prediction results at a certain confidence level can be selected according to the actual situation, striving to achieve a higher coverage rate while the interval is relatively narrow. As can be seen from Table 4, the coverage rate of the prediction interval at 90% confidence level exceeds 90%, and at the same time the interval width is less than 3%, which can better describe the uncertainty of point prediction, thus providing a more complete and accurate data basis for subsequent transformer condition assessment.

[0123] In summary, the embodiment of the present invention not only obtains the point prediction results of gas content, but also introduces interval prediction, and can generate prediction intervals of dissolved gas content in oil at different confidence levels, so as to quantitatively analyze the changes of gases in oil caused by external uncertain factors, which helps to improve the accuracy of state estimation.

[0124] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or substitutions, which should all be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims

1. A method for predicting the interval of dissolved gas content in transformer oil, characterized in that, it includes: Collect the dissolved gas content in transformer oil at historical time to obtain the first sequence of the dissolved gas content in transformer oil; Use the adaptive noise complete ensemble empirical mode decomposition algorithm to decompose the first sequence to obtain multiple second sequences; Input each of the second sequences into the trained point prediction model corresponding to each of the second sequences, and output the third sequence corresponding to each of the second sequences; Sum the elements with the same sorting in all the third sequences to obtain the point prediction result of the dissolved gas content in transformer oil within the preset time after the historical time; Use the non-parametric kernel density estimation algorithm for the point prediction result to calculate the interval prediction result of the dissolved gas content in transformer oil within the preset time; The point prediction model is a long short-term memory network point prediction model; Each time the long short-term memory network point prediction model is trained, use the grey wolf algorithm to optimize the number of hidden layer neurons and the initial learning rate of the long short-term memory network point prediction model; The formula for the probability density of the non-parametric kernel density estimation algorithm is as follows: ； Among them, represents the probability density, p represents the error value of the point prediction result relative to the true result, represents the i th point prediction result, h represents the window width, , N represents the total number of point prediction results, represents the kernel function; The kernel function of the non-parametric kernel density estimation algorithm is a Gaussian kernel function.

2. A computer-readable storage medium, characterized in that: Computer program instructions are stored on the computer-readable storage medium; when the computer program instructions are executed by a processor, the method for predicting the interval of dissolved gas content in transformer oil as described in claim 1 is implemented.

3. A system for predicting the interval of dissolved gas content in transformer oil, characterized in that, it includes: The computer-readable storage medium as described in claim 2.

Images