Method for predicting concentration of dissolved gas in transformer oil and related device

Abstract

The invention discloses a method for predicting the concentration of dissolved gas in transformer oil and a related device, and belongs to the technical field of monitoring of gas in transformer oil, and the method comprises the following steps: carrying out modal decomposition on time sequence data of dissolved gas in oil before a to-be-predicted time point to obtain IMF components and trend components, and screening out effective IMF components; carrying out normalization processing on the effective IMF component and the trend component; for the normalized effective IMF components, extracting a sequence with an interception length of p in a sliding window mode as input, and training a CNN-LSTM model by taking a next moment value of the sequence as a label; training the trend component by adopting an ARIMA (Autoregressive Integrated Moving Average) model; intercepting a sub-sequence of which the tail end length is p of each effective IMF component, inputting the sub-sequence into the trained CNN-LSTM model, and inputting the normalized trend component into the trained ARIMA model to obtain a prediction result of each component; and carrying out reverse normalization on each component prediction result, and carrying out dynamic weighted fusion to generate a final prediction value, thereby improving the precision and stability of prediction of the concentration of the gas in the transformer oil.

Description

Technical Field

[0001] The present invention belongs to the technical field of gas monitoring in transformer oil, and in particular relates to a method for predicting the concentration of dissolved gas in transformer oil and a related device. Background Art

[0002] To proactively identify potential transformer failures and address them promptly, transformer fault prediction is necessary. Fault prediction involves predicting future system failures based on existing system status data, thereby understanding the development trend of the system's operating status in advance. Before a transformer failure occurs, the gas content in the transformer oil exhibits a certain regularity over time. Therefore, the type of failure can be predicted by predicting the content of fault-signaling gases dissolved in the transformer oil. This facilitates maintenance personnel in scheduling inspections and effectively avoids the occurrence or worsening of transformer failures. However, monitoring data for dissolved gases in transformer oil is often affected by noise, outliers, and nonlinearity, making it difficult for traditional statistical analysis methods to achieve ideal results. Summary of the Invention

[0003] The object of the present invention is to provide a method and related device for predicting the concentration of dissolved gas in transformer oil, so as to improve the accuracy and stability of the prediction of the gas concentration in transformer oil.

[0004] In order to achieve the above object, the present invention adopts the following technical solutions:

[0005] In a first aspect, the present invention provides a method for predicting the concentration of dissolved gas in transformer oil, comprising:

[0006] Perform modal decomposition on the time series data of dissolved gas in oil before the prediction time point to obtain n IMF components and 1 trend component, and select the effective IMF components based on energy proportion and spectrum stability;

[0007] Normalizing all the effective IMF components and trend components respectively;

[0008] For each normalized valid IMF component, a sequence of length p is extracted using a sliding window as the model input, and the next moment value of the sequence is used as the label to train the CNN-LSTM model; for the trend component, the ARIMA model is used for training;

[0009] Cut off the subsequence of length p at the end of each valid IMF component, input it into the trained CNN-LSTM model, input the normalized trend component into the trained ARIMA model, and obtain the prediction results of each component;

[0010] The prediction results of each component are denormalized and the final prediction value is generated through dynamic weighted fusion.

[0011] Furthermore, the screening of effective IMF components based on energy proportion and spectrum stability includes:

[0012] Calculate the energy proportion and spectrum stability of the IMF component;

[0013] The energy ratio calculation formula is:

[0014]

[0015] Among them, η i : energy proportion of the i-th component; c i (t): the amplitude of the i-th IMF component at time t; c i (t): the amplitude of the i-th IMF component at time t; i, j represent the component numbers;

[0016] The spectrum stability is calculated as follows:

[0017] σ i =std(|FFT(c i (t))|)

[0018] Among them, σ i : Spectral stability of the i-th IMF component; FFT(c i (t)) is the IMF component c i (t) Perform fast Fourier transform to obtain its frequency domain representation; |·| takes the amplitude (modulus) of the spectrum; std(·) is the standard deviation of the spectrum amplitude;

[0019] Set the energy threshold θ E and the spectrum stability threshold θ F , filter out i >θ E And σ i <θ F The IMF components of are taken as the effective IMF components.

[0020] Furthermore, the energy threshold is set to 3% of the total energy of the original signal; the spectrum stability threshold θ F Dynamically determined, set to the mean of the set of standard deviations of all IMF component spectra plus 1 times the standard deviation.

[0021] Furthermore, for each normalized valid IMF component, a sequence of length p is extracted in a sliding window manner as a model input, and the next moment value of the sequence is used as a label. Training the CNN-LSTM model includes:

[0022] ① Assume that the standardized effective IMF component sequence is Construct samples by sliding window method, and the window length is set to p;

[0023] For each sample s∈[1,Np], construct:

[0024] Input sequence X s (feature):

[0025]

[0026] Prediction target y s (Label):

[0027]

[0028] ② Apply two layers of one-dimensional convolution operations to the input sequence Xs, and then perform a maximum pooling operation on the convolution output to obtain the sequence features;

[0029] ③ The sequence features obtained by CNN Input LSTM and take the hidden state h at the last moment U As LSTM output;

[0030] ④ Use the fully connected layer to represent the effective IMF component prediction value of the final output:

[0031]

[0032] in, is the model’s component prediction for the next time step, h U is the hidden state vector at the last time step U; W p and b p are weights and biases.

[0033] Furthermore, the ARIMA model expression is:

[0034] φ(B)(1-B) d y t =θ(B)ε t

[0035] Among them, y t is the residual component sequence value at time t; B is the backshift operator, satisfying By t =y t-1 ; d is the difference order, which is used to eliminate non-stationarity (such as trend); φ(B) is the autoregressive (AR) polynomial coefficient representation, which is used to model the linear relationship between the series and its lag terms; θ(B) is the moving average (MA) polynomial coefficient representation, which is used to model the dependence between the error terms of the series; ε t is a white noise disturbance term, which is assumed to be an independent and identically distributed random variable with mean 0 and variance constant.

[0036] Further, the time series data of the dissolved gas in the oil before the to-be-predicted time point is continuous sampling data of at least 100 sampling time points before the prediction point.

[0037] Further, the inverse normalization of the prediction results of each component and the generation of a final prediction value through dynamic weighting fusion include:

[0038] The normalized prediction value obtained for each valid IMF component i is inverse normalized:

[0039]

[0040] wherein, is the standardized value of the i-th component predicted by the model at time t; μ i , σ i : the mean and standard deviation of the i-th component calculated on the training set; The prediction value of the i-th component after inverse normalization is restored to the original dimension;

[0041] The inverse normalized prediction values of all valid IMF components are weighted and added according to the time point to obtain the final concentration prediction sequence:

[0042]

[0043] wherein, is the inverse normalized prediction value of the i-th IMF component or residual term, w i (t) is a dynamic weight; n is the total number of IMF components and trend components, the final concentration prediction value of the original sequence at time t;

[0044] The dynamic weight calculation formula is:

[0045]

[0046] wherein, MAE i (t) is the prediction error of the i-th sub-model.

[0047] In a second aspect, the present application provides a transformer oil dissolved gas concentration prediction system, comprising:

[0048] A modal decomposition module is configured to perform modal decomposition on the time series data of the dissolved gas in the oil before the to-be-predicted time point to obtain n IMF components and 1 trend component, and to screen valid IMF components based on energy proportion and spectral stability;

[0049] A normalization module is configured to perform normalization processing on all the valid IMF components and the trend component, respectively;

[0050] The training module is used to extract a sequence of length p from each normalized valid IMF component using a sliding window as the model input, and train the CNN-LSTM model using the next moment value of the sequence as the label. The trend component is trained using the ARIMA model.

[0051] The prediction module is used to intercept the subsequence of length p at the end of each valid IMF component, input it into the trained CNN-LSTM model, and input the normalized trend component into the trained ARIMA model to obtain the prediction results of each component;

[0052] The weighted fusion module is used to denormalize the prediction results of each component and generate the final prediction value through dynamic weighted fusion.

[0053] In a third aspect, the present invention provides an electronic device, comprising:

[0054] at least one processor; and,

[0055] a memory communicatively connected to the at least one processor; wherein,

[0056] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform a method for predicting dissolved gas concentration in transformer oil as described in any one of the first aspects of the present invention.

[0057] In a fourth aspect, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements a method for predicting dissolved gas concentration in transformer oil according to any one of the first aspects of the present invention.

[0058] Compared with the prior art, the present invention has at least the following beneficial technical effects:

[0059] (1) Introducing the IMF component screening mechanism to improve the quality of predictive modeling

[0060] Compared with the existing method of indiscriminately modeling all CEEMDAN components, the present invention adaptively screens by combining energy proportion and spectral stability, retaining only the components that make key contributions to signal structure and predictability, thus avoiding the errors caused by redundant modeling.

[0061] (2) Adopt heterogeneous model matching strategy to achieve targeted modeling optimization

[0062] The present invention selects the most suitable modeling structure according to the component frequency characteristics, adopts the CNN-LSTM model for high-frequency components, and uses ARIMA for linear fitting of trend components, which solves the problem that the traditional unified model is difficult to simultaneously model the characteristics of the entire frequency band, and realizes multi-scale perceptual modeling of complex nonlinear sequences.

[0063] (3) Propose a residual-driven dynamic weighted fusion mechanism to enhance prediction robustness

[0064] Compared with traditional fixed weighted or average fusion methods, the present invention dynamically adjusts the output weights according to the historical prediction errors of each sub-model, so that the final prediction results have a natural inhibitory ability on low-quality model outputs, thereby improving the accuracy and stability of the overall prediction.

[0065] (4) Has broad adaptability and application prospects

[0066] The method of the present invention has broad engineering application value and promotion prospects in scenarios such as power transformer fault warning and status assessment. By implementing the method of the present invention, operation and maintenance personnel can predict the transformer status, reduce the impact of faults on the power system, ensure stable operation of the power grid, and improve operation and maintenance efficiency and equipment service life. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 A schematic flow chart of a method for predicting dissolved gas concentration in transformer oil provided by the present invention;

[0068] Figure 2 is the filtered IMF component; the horizontal axis is the time sequence number, and the vertical axis is the signal amplitude;

[0069] Figure 3 is the prediction result of the model of the present invention;

[0070] Figure 4 Predict results for the ANN model;

[0071] Figure 5 Predict results for the LSTM model;

[0072] Figure 6 A structural block diagram of a system for predicting dissolved gas concentration in transformer oil provided by an embodiment of the present invention;

[0073] Figure 7 A block diagram of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0074] In the following, the technical solutions in the embodiments of the present application will be described clearly and completely with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all the other embodiments obtained by a person of ordinary skill in the art without creative work should belong to the protection scope of the present application.

[0075] It should be noted that the terms "first", "second", and the like in the specification and claims of the present application and the above-described drawings are used to distinguish similar objects, and do not necessarily indicate a specific order or a chronological sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or device including a series of steps or units does not necessarily have to be limited to the clearly listed steps or units, but can include other steps or units not clearly listed or inherent to the process, method, product, or device.

[0076] The present application will be described in further detail below with reference to the drawings:

[0077] Embodiment 1

[0078] With reference to Figure 1 , the present embodiment provides a method for predicting the concentration of dissolved gas in transformer oil, comprising the following steps:

[0079] S1, using a complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) algorithm to decompose the time series of dissolved gas in oil before the to-be-predicted time point, to obtain n IMF components with different frequency bands and a trend component (residual), and introducing an IMF component screening mechanism, wherein the IMF components include medium-high frequency components and low frequency components.

[0080] S2, normalizing the trend component (residual) and the screened IMF components.

[0081] S3, modeling: the IMF components and the trend component are collectively referred to as components, and a sub-model is built for each component, the sub-model structures are the same, and the parameters are different.

[0082] For each modal component after normalization, heterogeneous prediction modeling is adopted. For the screened IMF components, 1D-CNN is used to fully mine the normalized data to obtain the deep features of the sequence and form feature vectors for input into the trained LSTM for prediction; ARIMA is used for fitting the trend component.

[0083] S4. After normalization, each modal component is trained separately for each sub-model.

[0084] S5. Take the time before the predicted time point as input, perform CEEMDAN modal decomposition, input it into the trained model, and obtain the prediction results of each component.

[0085] S6. Denormalize the prediction results of each component sequence and perform dynamic weighted fusion to obtain the final prediction result.

[0086] S1 includes the following steps:

[0087] Step 1.1, add noise: superimpose different Gaussian white noises on the time series x(t) of dissolved gas concentration in a certain oil, and use EMD decomposition to obtain the first layer of IMF, x(t) is recorded as the original signal;

[0088]

[0089] Among them, M is the number of samples in the Gaussian white noise set, p is the sample number, and w q (t) is the pth standard Gaussian white noise, and ε1 is the noise weight.

[0090] Step 1.2, calculate the residual signal: remove the first layer of IMF from the original signal to obtain the new residual signal r1:

[0091] r1=x(t)-IMF1

[0092] Step 1.3, recursively calculate IMF: apply new noise to the residual signal r1, perform EMD decomposition again, and obtain the second layer of IMF:

[0093]

[0094] Step 1.4, repeat the above process until the remaining signal becomes a monotonic trend term (i.e., it cannot be decomposed any further). Ultimately, the original signal can be expressed as the sum of all IMF components and one trend component:

[0095]

[0096] Among them, IMFj is the jth IMF component obtained by decomposition, and rN is the residual signal left after the nth decomposition, that is, the trend component.

[0097] Step 1.5, in order to improve the efficiency of modeling and reduce the redundant modeling of invalid components, the application adds a joint component screening mechanism based on energy proportion and spectral stability after obtaining all the IMF components, and screens out effective IMF components:

[0098] The energy proportion is the proportion of the energy of the i-th IMF component in the total energy, which represents the proportion of the energy of the i-th component in the total energy of all IMF components, and is used to reflect the contribution of the component to the signal structure.

[0099]

[0100] Wherein, η i : the energy proportion of the i-th component; c i (t): the amplitude of the i-th IMF component at time t. c i (t): the amplitude of the i-th IMF component at time t. i and j represent the serial number of the component;

[0101] The spectral stability is measured by the standard deviation of the FFT transform spectrum of the component. The more intense the spectrum fluctuates, the more likely it is to be noise:

[0102] σ i =std(|FFT(c i (t))|)

[0103] Wherein, σ i : the spectral stability of the i-th IMF component; FFT(c i (t)) is the fast Fourier transform of the IMF component c i (t), which obtains its frequency domain representation; |·| takes the amplitude (modulus) of the spectrum; std(·) is the standard deviation of the spectrum amplitude;

[0104] Set threshold θ E , θ F , only keep the IMF components that satisfy η i > θ E and σ i < θ F , and the rest are rejected. This mechanism can effectively remove non-contributing components and noise strong components, and improve the stability of the model.

[0105] In the application, in order to ensure that the component screening process can both eliminate invalid or disturbance components and not damage effective feature information, the following double-threshold screening strategy is proposed:

[0106] Energy threshold θ E : set to 3% of the total energy of the original signal, that is, only keep the IMF components with energy proportion greater than 3%;

[0107] Spectral stability threshold θ F : Dynamically determined, set to the mean of the set of standard deviations of all IMF component spectra plus 1 standard deviation, that is:

[0108] θ F =mean(σ i )+1.0·std(σ i )

[0109] mean(·) is the mean;

[0110] Through the above method, redundant or pseudo-modal components can be adaptively identified and eliminated, thereby improving the modeling effect of the model on predictable components.

[0111] In step S2, the decomposed modal components are normalized using z-score normalization, also known as standard deviation normalization, which converts the data into a distribution with a mean of 0 and a standard deviation of 1. The specific expression is as follows:

[0112]

[0113] Where z is the normalized value of the subsequence component; x is the value of the subsequence component before normalization; μ and δ are the mean and standard deviation of the original data of the subsequence component, respectively. The subsequence component includes the trend component and the filtered IMF component.

[0114] In step S3

[0115] (1) For each valid IMF component sequence that is screened and retained, a convolutional neural network-long short-term memory network (CNN-LSTM) model is used.

[0116] The specific steps are as follows:

[0117] ① Assume that the time series of each effective IMF component after standardization is The samples are constructed by sliding window method, and the window length is set to p, which can be 7.

[0118] For each sample s∈[1,Np], construct:

[0119] Input sequence X s (feature):

[0120]

[0121] Prediction target y s (Label):

[0122]

[0123] ② Apply two layers of one-dimensional convolution operations to the input sequence Xs. Each layer includes convolution operation, ReLU activation and maximum pooling. Taking the lth layer as an example, the convolution operation is defined as follows:

[0124]

[0125] Where: τ is the time step index (position within the window); c is the cth channel number in the lth layer of one-dimensional convolution; K is the convolution kernel size; is the weight of the l-th layer channel c at the m-th position; is the bias term corresponding to the c channel, x τ+m-1 is the value of the input sequence Xs at the τ+m-1 position, is the value obtained after the convolution operation.

[0126] Then perform a maximum pooling operation on the convolution output, with a pooling window size of S:

[0127]

[0128] in, The output of the max pooling operation, the last one Denoted as feature map P;

[0129] The feature map P obtained after two layers of convolution and pooling is flattened into a one-dimensional feature vector z:

[0130] z=Flatten(P)

[0131] ③LSTM time series modeling. The sequence features obtained by CNN Enter LSTM, which consists of four main parts: input gate, forget gate, output gate and cell state. The number of hidden units is set to H, and the input at time step u is z u , the hidden state at the previous moment is h u-1 , the cell state is C u-1 The forget gate determines which information in the cell state should be retained and which should be discarded. It receives the current input and the hidden state of the previous moment as input, and maps it to a value between 0 and 1 through a Sigmoid activation function. Among them, a value close to 0 indicates that the corresponding cell state information will be forgotten, and a value close to 1 indicates that the information will be retained. The calculation formula of the forget gate is as follows:

[0132] f u =σ(W f ·[h u-1 ,z u ]+b f )

[0133] Where σ(·) is the activation function, W fand b f is the weight matrix and bias term of the forget gate, f u is the output of the forget gate.

[0134] The input gate is responsible for controlling how much information in the current input will be updated to the cell state. It also receives and as input, calculates an update ratio through the Sigmoid function, and transforms the current input through a Tanh activation function, and then multiplies the two to obtain the information that needs to be updated to the cell state. The calculation formula is as follows:

[0135] i u =σ(W i ·[h u-1 ,z u ]+b i )

[0136]

[0137] Among them, i u is the output of the input gate, is the candidate memory content, W i With W C are the weight matrices of the input gate and candidate state, b i with b C is the corresponding bias term, tanh(·) is used to generate candidate states, and its value range is (-1,1).

[0138] Combining the forget gate and the input gate, the update formula of the unit state is:

[0139]

[0140] Among them, C u is the current memory state, and ⊙ represents element-wise product, i.e., element-wise multiplication.

[0141] The output gate determines which information in the cell state will be output as the hidden state at the current moment. It receives and as input, calculates an output ratio through the Sigmoid function, and then multiplies it with the cell state processed by the Tanh activation function to obtain the hidden state at the current moment. The calculation formula of the output gate is as follows:

[0142] o u =σ(W o ·[h u-1 ,z u ]+b o )

[0143] h u =o u ⊙tanh(C u )

[0144] Among them, u is the output gate result, W o with b o are the weight matrix and bias vector of the output gate respectively.

[0145] Take the hidden state h at the last moment U As LSTM output.

[0146] ④Finally, a fully connected layer is used to represent the effective IMF component prediction value of the final output:

[0147]

[0148] in, is the model’s component prediction for the next time step, h U is the hidden state vector at the last time step U. p and b p are weights and biases.

[0149] (2) For the residual component, the ARIMA model is used for fitting:

[0150] φ(B)(1-B) d y t =θ(B)ε t

[0151] Among them, y t is the residual component sequence value at time t; B is the backshift operator, satisfying By t =y t-1 ; d is the difference order, which is used to eliminate non-stationarity (such as trend); φ(B) is the autoregressive (AR) polynomial coefficient representation, which is used to model the linear relationship between the series and its lag terms; θ(B) is the moving average (MA) polynomial coefficient representation, which is used to model the dependence between the error terms of the series; ε t is a white noise disturbance term, which is assumed to be an independent and identically distributed machine variable with a mean of 0 and a constant variance.

[0152] In step S4, the models (CNN-LSTM model and ARIMA model) constructed from the above subsequences are trained. The subsequences include effective IMF components and trend components.

[0153] The filtered effective components are input into the CNN-LSTM network; the trend components are input into the ARIMA network. The output of each model is the predicted value of the component at the next time point. The loss function used is mean squared error (MSE):

[0154]

[0155] Where M is the number of samples, is the true value of the kth sample, is the predicted value of the kth sample.

[0156] Optimization algorithm: For neural network models, the Adam optimizer is used to update the parameters, and for ARIMA models, the model parameters are estimated through maximum likelihood estimation.

[0157] In step S5, the period before the time point to be predicted is used as input, and CEEMDAN modal decomposition and normalization are performed according to the method of steps 1-2 to obtain normalized subsequence components. The normalized subsequence components are input into the trained model to obtain the predicted values ​​of each component.

[0158] In step S6

[0159] Perform denormalization on the normalized forecast values ​​obtained for each effective IMF component i:

[0160]

[0161] in, is the normalized value of the i-th component predicted by the model at time t. i , σ i : The mean and standard deviation of the i-th component calculated on the training set. The predicted value of the i-th component after denormalization is restored to its original dimension.

[0162] The denormalized prediction values ​​of all n IMF components are weighted and added together at each time point to obtain the final concentration prediction sequence:

[0163]

[0164] in, is the denormalized forecast value of the ith IMF component or residual term, w i (t) is the dynamic weight. n is the total number of IMF components (including residual terms), The final concentration prediction value of the original sequence at time t.

[0165] Dynamic weight definition:

[0166]

[0167] Among them, MAE i (t) is the prediction error of the i-th sub-model.

[0168] Simulation experiment

[0169] This paper uses online oil chromatography data from a 500kV transformer from 2016 to 2018 as the research object for simulation experiments and performance evaluation. This data set comprises 1,350 data sets, with a 12-hour detection cycle. The experiment used data on hydrogen (H2) concentration in dissolved oil. The first 1,094 data sets served as training samples, and the last 256 data sets served as test samples. In other words, one data set includes both the detection time and the corresponding hydrogen concentration.

[0170] In order to verify the effect of the prediction model of the present invention, the data after data cleaning is used.

[0171] 1) Modal decomposition of gas sequences

[0172] First, CEMMDAN is used to decompose the time series consisting of the first 1094 data sets. In this process, a white noise H with an amplitude of 0.005 is introduced into the original time series, and 100 integrated trials are used for decomposition to reduce mode aliasing and improve the stability of the decomposition. Finally, the H2 series is decomposed into a set of intrinsic mode functions (IMFs) and a trend component (residual signal). The component screening mechanism is used to obtain the filtered effective IMF components, such as Figure 2 The seven groups of effective IMF components in the figure have different frequencies, reflecting the fluctuation characteristics of different time scales.

[0173] The effective IMF component subsequences obtained after modal decomposition, screening, and normalization are input into the CNN-LSTM model for training and prediction to obtain the prediction model of each subsequence. The trend component is fitted with the ARIMA model. Finally, the prediction results of each effective IMF component and trend component sequence are denormalized and dynamically weighted fused to obtain the final prediction result. The prediction results of the model of the present invention on the test set are as follows: Figure 3 shown.

[0174] In order to further verify the accuracy of the model of the present invention, based on the same sample, the traditional LSTM and ANN models are used for prediction, and the results are as follows: Figure 4 、 5 The comparison results are shown in Table 1. Three commonly used evaluation indicators are introduced: root mean square error (RMSE), mean absolute error (MAE) and mean absolute percentage error (MAPE).

[0175] The calculation formulas are as follows:

[0176]

[0177] Where: y i represents the dissolved gas concentration measured at the i-th time point, is the model prediction value at the corresponding moment; n represents the number of test samples.

[0178] Table 1 Comparison of model prediction performance

[0179] Prediction model RMSE MAE MAPE Invention model 0.5648 0.3566 1.84％ LSTM 0.8885 0.61883 2.91％ ANN 0.9173 0.6530 3.07％

[0180] It can be seen from Table 1 that the root mean square error of the prediction model constructed and trained by the present invention is reduced by 36.4% compared with LSTM and by 38.4% compared with ANN; the mean absolute error is reduced by 42.4% compared with LSTM and by 45.4% compared with ANN; the mean absolute percentage error is reduced by 36.8% compared with LSTM and by 40.1% compared with ANN.

[0181] Example 2

[0182] See also Figure 6 In this embodiment, a system for predicting dissolved gas concentration in transformer oil is provided, comprising:

[0183] The modal decomposition module is used to perform modal decomposition on the time series data of dissolved gas in oil before the prediction time point, obtain n IMF components and 1 trend component, and screen out the effective IMF components based on energy proportion and spectrum stability;

[0184] A normalization module, configured to perform normalization processing on all the effective IMF components and trend components respectively;

[0185] The training module is used to extract a sequence of length p from each normalized valid IMF component using a sliding window as the model input, and train the CNN-LSTM model using the next moment value of the sequence as the label. The trend component is trained using the ARIMA model.

[0186] The prediction module is used to intercept the subsequence of length p at the end of each valid IMF component, input it into the trained CNN-LSTM model, and input the normalized trend component into the trained ARIMA model to obtain the prediction results of each component;

[0187] The weighted fusion module is used to denormalize the prediction results of each component and generate the final prediction value through dynamic weighted fusion.

[0188] All relevant contents of each step involved in the embodiment of the aforementioned method for predicting dissolved gas concentration in transformer oil can be referred to the functional description of the functional modules corresponding to the system for predicting dissolved gas concentration in transformer oil in the embodiment of the present invention, and will not be repeated here.

[0189] Example 3

[0190] ReferenceFigure 7 The embodiment provides an electronic device, which comprises a processor and a memory connected through a bus; the memory is used for storing a computer program, the computer program comprises program instructions, and the processor is used for executing the program instructions stored in the computer storage medium. The processor can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc., which are the computing core and control core of the terminal, and are suitable for implementing one or more instructions, and are particularly suitable for loading and executing one or more instructions in the computer storage medium to implement a corresponding method flow or a corresponding function; the processor in the embodiment of the present application can be used for the operation of the transformer oil dissolved gas concentration prediction method. The bus can be a peripheral component interconnect (PCI) bus or an extended industry standard architecture (EISA) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For the convenience of representation, Figure 7 In the figure, only one line is used to represent the bus, but it does not mean that there is only one bus or only one type of bus.

[0191] Embodiment 4

[0192] This embodiment provides a storage medium, specifically a computer-readable storage medium (Memory). The computer-readable storage medium is a memory device in an electronic device for storing programs and data. It is understood that the computer-readable storage medium herein may include both built-in storage media in the electronic device and, of course, extended storage media supported by the electronic device. The computer-readable storage medium provides storage space, which stores the terminal's operating system. Furthermore, the storage space also stores one or more instructions suitable for being loaded and executed by a processor. These instructions may be one or more computer programs (including program code). It should be noted that the computer-readable storage medium herein may be a high-speed RAM memory or a non-volatile memory, such as at least one disk storage device. The processor may load and execute one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the method for predicting dissolved gas concentration in transformer oil in the above-mentioned embodiment.

[0193] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0194] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the steps in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0195] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1The function specified in one or more boxes.

[0196] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0197] Example 5

[0198] This embodiment provides a computer program product, including a non-volatile computer-readable storage medium, wherein the non-volatile computer-readable storage medium stores a computer program product, and when the computer program is executed by a processor, the steps of the method described in each embodiment of the present application are implemented.

[0199] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with the relevant laws, regulations and standards of relevant countries and regions.

[0200] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification and features of different embodiments or examples without contradiction.

[0201] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered by the scope of protection of the claims of the present invention.

Claims

1. A method for predicting dissolved gas concentration in transformer oil, characterized in that: include: Perform modal decomposition on the time series data of dissolved gas in oil before the prediction time point to obtain n IMF components and 1 trend component, and select the effective IMF components based on energy proportion and spectrum stability; Normalizing all the effective IMF components and trend components respectively; For each normalized valid IMF component, a sequence of length p is extracted using a sliding window as the model input, and the next moment value of the sequence is used as the label to train the CNN-LSTM model; For the trend component, the ARIMA model is used for training; Cut off the subsequence of length p at the end of each valid IMF component, input it into the trained CNN-LSTM model, input the normalized trend component into the trained ARIMA model, and obtain the prediction results of each component; The prediction results of each component are denormalized and the final prediction value is generated through dynamic weighted fusion.

2. The method for predicting dissolved gas concentration in transformer oil according to claim 1, wherein: The screening of effective IMF components based on energy proportion and spectrum stability includes: Calculate the energy proportion and spectrum stability of the IMF component; The energy ratio calculation formula is: Among them, η i : energy proportion of the i-th component; c i (t): the amplitude of the i-th IMF component at time t; c i (t): the amplitude of the i-th IMF component at time t; i, j represent the component numbers; The spectrum stability is calculated as follows: σ i =std(|FFT(c i (t))|) Among them, σ i : Spectral stability of the i-th IMF component; FFT(c i (t)) is the IMF component c i (t) Perform fast Fourier transform to obtain its frequency domain representation; |·| takes the amplitude (modulus) of the spectrum; std(·) is the standard deviation of the spectrum amplitude; Set the energy threshold θ E and the spectrum stability threshold θ F , filter out i >θ E And σ i <θ F The IMF components of are taken as the effective IMF components.

3. The method for predicting dissolved gas concentration in transformer oil according to claim 2, wherein: The energy threshold is set to 3% of the total energy of the original signal; the spectrum stability threshold θ F Dynamically determined, set to the mean of the set of standard deviations of all IMF component spectra plus 1 times the standard deviation.

4. The method for predicting dissolved gas concentration in transformer oil according to claim 1, wherein: For each normalized valid IMF component, a sequence of length p is extracted in a sliding window manner as a model input, and the next moment value of the sequence is used as a label. Training the CNN-LSTM model includes: ① Assume that the standardized effective IMF component sequence is Construct samples by sliding window method, and the window length is set to p; For each sample s∈[1,Np], construct: Input sequence X s (feature): Prediction target y s (Label): ② Apply two layers of one-dimensional convolution operations to the input sequence Xs, and then perform a maximum pooling operation on the convolution output to obtain the sequence features; ③ The sequence features obtained by CNN Input LSTM and take the hidden state h at the last moment U As LSTM output; ④ Use the fully connected layer to represent the effective IMF component prediction value of the final output: in, is the model’s component prediction for the next time step, h U is the hidden state vector at the last time step U; W p and b p are weights and biases.

5. The method for predicting dissolved gas concentration in transformer oil according to claim 1, wherein: The ARIMA model expression is: φ(B)(1-B) d y t =θ(B)ε t Among them, y t is the residual component sequence value at time t; B is the backshift operator, satisfying By t =y t-1 ; d is the difference order, which is used to eliminate non-stationarity (such as trend); φ(B) is the autoregressive (AR) polynomial coefficient representation, which is used to model the linear relationship between the series and its lag terms; θ(B) is the moving average (MA) polynomial coefficient representation, which is used to model the dependence between the error terms of the series; ε t is a white noise disturbance term, which is assumed to be an independent and identically distributed random variable with mean 0 and variance constant.

6. The method for predicting dissolved gas concentration in transformer oil according to claim 1, wherein: The time series data of dissolved gas in oil before the time point to be predicted is continuous sampling data of at least 100 sampling moments before the prediction point.

7. The method for predicting dissolved gas concentration in transformer oil according to claim 1, wherein: The denormalization of the prediction results of each component and the generation of the final prediction value through dynamic weighted fusion include: Perform denormalization on the normalized forecast values ​​obtained for each effective IMF component i: in, is the normalized value of the i-th component predicted by the model at time t; μ i , σ i : The mean and standard deviation of the i-th component calculated on the training set; The predicted value of the i-th component after denormalization is restored to the original dimension; The denormalized prediction values ​​of all effective IMF components are weighted and added together according to the time point to obtain the final concentration prediction sequence: in, is the denormalized forecast value of the ith IMF component or residual term, w i (t) is the dynamic weight; n is the total number of IMF components and trend components, The final concentration prediction value of the original sequence at time t; The dynamic weight calculation formula is: Among them, MAE i (t) is the prediction error of the i-th sub-model.

8. A system for predicting dissolved gas concentration in transformer oil, characterized in that: include: The modal decomposition module is used to perform modal decomposition on the time series data of dissolved gas in oil before the prediction time point, obtain n IMF components and 1 trend component, and screen out the effective IMF components based on energy proportion and spectrum stability; A normalization module, configured to perform normalization processing on all the effective IMF components and trend components respectively; The training module is used to extract a sequence of length p from each normalized valid IMF component using a sliding window as the model input, and train the CNN-LSTM model using the next moment value of the sequence as the label. The trend component is trained using the ARIMA model. The prediction module is used to intercept the subsequence of length p at the end of each valid IMF component, input it into the trained CNN-LSTM model, and input the normalized trend component into the trained ARIMA model to obtain the prediction results of each component; The weighted fusion module is used to denormalize the prediction results of each component and generate the final prediction value through dynamic weighted fusion.

9. An electronic device, characterized in that: include: at least one processor; as well as, a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor. The instructions are executed by the at least one processor to enable the at least one processor to perform the method for predicting dissolved gas concentration in transformer oil according to any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method for predicting the concentration of dissolved gas in transformer oil according to any one of claims 1 to 7 is implemented.

Images