Early warning system and method for dissolved gas in transformer oil based on hox process

Abstract

The invention provides a system for early warning dissolved gas in transformer oil based on a Horkes process, and the system comprises a historical data fitting module which is used for dividing a historical concentration curve of specific dissolved gas according to a fluctuation period, selecting a rising curve from a valley value to a peak value in each period, carrying out the fitting of the rising curves of all periods, and carrying out the early warning of the dissolved gas in the transformer oil. A unified fitting function is obtained; the Horkes process modeling module is used for carrying out Horkes process modeling according to the fitting function of the rising curves of all the fluctuation periods to obtain a conditional intensity function for calculating the expected times of faults of the transformer corresponding to the dissolved gas; and the dissolved gas early warning module is used for calculating the expected times of faults, corresponding to the dissolved gas, of the transformer in the future time period according to the condition intensity function of the expected times, and giving out early warning in combination with an expected time threshold. According to the method, the time difference from fault occurrence to early warning is shortened, and the hysteresis problem caused by slow gas diffusion is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of online monitoring of power transformers in electrical engineering, and in particular to a Hawkes process-based early warning system and method for dissolved gas in transformer oil. Background Art

[0002] Online monitoring of dissolved gas in transformer oil is a key method for power transformer condition monitoring. Currently, the primary method used for fault warning is threshold comparison. However, this method has two significant limitations: First, the slow diffusion of gas in oil results in a significant lag in fault warnings; second, dissolved gas levels tend to decrease over time due to factors such as solubility equilibrium, affecting threshold determination.

[0003] Currently, most improved early warning methods use gas concentration prediction methods. Traditional trend fitting methods only extrapolate the current trend of change. The fault that causes the current trend change may have already occurred, and it cannot predict possible faults and probabilities in the future. In addition, the prediction method based on neural networks lacks interpretability, which is not conducive to fault diagnosis decision-making in actual engineering applications. Summary of the Invention

[0004] The purpose of the present invention is to address the defects of the prior art and provide a transformer oil dissolved gas early warning system based on the Hawkes process, comprising: a historical data fitting module, used to divide the historical fluctuation curve of the concentration of a certain dissolved gas in the transformer oil over time according to the fluctuation period, and in each divided fluctuation period, the rising curve of the fluctuation period is selected with the concentration valley of the dissolved gas as the starting point and the concentration peak of the dissolved gas as the end point, and the rising curves of all fluctuation periods are fitted to obtain the fitting function of the rising curves of all fluctuation periods; a Hawkes process modeling module, used to perform Hawkes process modeling based on the fitting function of the rising curves of all fluctuation periods to obtain a conditional strength function for calculating the expected number of transformer faults corresponding to the dissolved gas; a dissolved gas early warning module, used to calculate the expected number of transformer faults corresponding to the dissolved gas in a future time period based on the conditional strength function of the expected number, compare the expected number of transformer faults corresponding to the dissolved gas in the future time period with a set expected number threshold, and determine whether to issue an early warning based on the comparison result.

[0005] Furthermore, the historical data fitting module is also used to: Set the minimum peak-to-valley difference threshold, and record the dissolved gas concentration range in the rising curve of all fluctuation cycles with a peak-to-valley difference greater than the minimum peak-to-valley difference threshold as [V i ,P i ],V iis the trough value of the concentration of the dissolved gas of the rising curve of fluctuation cycle i i is the peak value of the concentration of the dissolved gas of the rising curve of fluctuation cycle i

[0006] Further, the historical data fitting module is further configured to: fit the rising curve of all fluctuation cycles by using a Logistic function, and the formula is as follows: wherein, f i (t) is the fitting function of the rising curve of fluctuation cycle i, t is time, a i is the peak-trough difference of the concentration of the dissolved gas of the rising curve of fluctuation cycle i, b i is the growth rate of the concentration of the dissolved gas of the rising curve of fluctuation cycle i, c i is the time point at which the growth rate of the concentration of the dissolved gas of the rising curve of fluctuation cycle i is the fastest, d i is the initial dissolved concentration of the gas of the rising curve of fluctuation cycle i initialize the above parameters: a i =P i -V i ; b i =(P i -V i ) / t; c i is set as the midpoint of [V i , P i ]; d i =V i ; establish the boundary constraints of the above parameters: 0 < a i ≤ 2(P i -V i ); 0 < b i ≤ 1; 0.9×V i ≤ d i ≤ 1.1×P i ; solve the optimal values of a i , b i , c i , d i by using the Levenberg-Marquardt algorithm, and the following objective function is used in the solving process: wherein, k is the sampling point number of the rising curve of fluctuation cycle i, t k is the time corresponding to the sampling point k of the rising curve of fluctuation cycle i, x i,kis the dissolved gas concentration in the rising curve of fluctuation cycle i at time t k N is the total number of sampling points in the time series of the rising curve of fluctuation cycle i.

[0007] Further, the conditional intensity function in the Hox process modeling module is specifically as follows: wherein l(t) is the conditional intensity function for calculating the expected number of dissolved gas faults of the transformer, μ is the inherent gas production rate of the transformer without external fault stimulation, only by natural aging or regular operation, α is a fault self-excitation coefficient, the fault self-excitation coefficient α represents the instantaneous enhancement intensity of a single historical fault on the current fault occurrence rate, β is a fault decay coefficient, the fault decay coefficient β controls the decay speed of the influence of historical faults, w i is the weight of the rising curve of fluctuation cycle i; By maximizing the likelihood function L(μ, α, β), the optimal values of μ, α, and β are solved: .

[0008] Further, in the dissolved gas early warning module, the formula for the expected number of faults corresponding to the dissolved gas of the transformer in the future time period is as follows: wherein E[N(Δt)] is the expected number of faults in the future time period [T, T+Δt], T is a future time, and Δt is the time period length; When E[N(Δt)]<1.5, monitoring is maintained, and no early warning is issued; when 1.5≤E[N(Δt)]<2, the monitoring period is shortened, and no early warning is issued; and when E[N(Δt)]≥2, an early warning for starting offline detection and diagnosis is issued.

[0009] Further, the dissolved gas early warning module is also used to calculate the growth amount of the gas in the transformer oil in the future time period according to the expected number of faults corresponding to the dissolved gas of the transformer in the future time period and the maximum concentration difference of the dissolved gas in the rising curves of all fluctuation cycles, and the growth amount calculation formula is as follows: R=E[N(Δt)]×a max wherein R is the growth amount of the gas in the transformer oil in the future time period, a max is the maximum concentration difference of the dissolved gas in the rising curves of all fluctuation cycles.

[0010] The application discloses a transformer oil dissolved gas early warning method based on the Hock process, and comprises the following steps: dividing the concentration fluctuation curve of a certain dissolved gas in transformer oil along with time into fluctuation periods, taking the concentration valley value of the certain dissolved gas as the starting point and the concentration peak value of the certain dissolved gas as the ending point in each divided fluctuation period, selecting the rising curve of the fluctuation period, fitting the rising curves of all the fluctuation periods to obtain a fitting function of the rising curves of all the fluctuation periods, modeling the Hock process according to the fitting function of the rising curves of all the fluctuation periods to obtain a conditional intensity function for calculating the expected number of times of transformer failures corresponding to the certain dissolved gas, comparing the expected number of times of transformer failures corresponding to the certain dissolved gas in a future time period with a set expected number of times threshold, and judging whether to issue a warning according to the comparison result.

[0011] Further, the specific method of selecting the rising curve of the fluctuation period in each divided fluctuation period with the concentration valley value of the certain dissolved gas as the starting point and the concentration peak value of the certain dissolved gas as the ending point is as follows: A minimum peak-valley difference threshold is set, and the concentration range of the certain dissolved gas in the rising curve of all the fluctuation periods with the peak-valley difference greater than the minimum peak-valley difference threshold is recorded as [V i ,P i ], V i is the concentration valley value of the certain dissolved gas in the rising curve of the fluctuation period i, and P i is the concentration peak value of the certain dissolved gas in the rising curve of the fluctuation period i.

[0012] Further, the specific method of fitting the rising curves of all the fluctuation periods to obtain a fitting function of the rising curves of all the fluctuation periods is as follows: The rising curves of all the fluctuation periods are fitted by using a Logistic function, and the formula is as follows: Wherein, f i (t) is the fitting function of the rising curve of the fluctuation period i, t is time, a i is the peak-valley difference value of the concentration of the certain dissolved gas in the rising curve of the fluctuation period i, b i is the growth rate of the concentration of the certain dissolved gas in the rising curve of the fluctuation period i, c i is the time point at which the growth rate of the concentration of the certain dissolved gas in the rising curve of the fluctuation period i is the fastest, d i is the initial dissolved concentration of the certain gas in the rising curve of the fluctuation period i. The above parameters are initialized as follows: a i=P i -V i ; b i =(P i -V i ) / t;c i Set to [V i ,P i ] midpoint; d i =V i ; Establish bounds for the above parameters: 0 i ≤2(P i -V i );0 i ≤1; 0.9×V i ≤d i ≤1.1×P i ; Use the Levenberg-Marquardt algorithm to solve a i 、b i 、c i d i The optimal value of , the following objective function is used in the solution process: Among them, k is the sampling point number of the time series in the rising curve of fluctuation period i, t k is the time corresponding to sampling point k in the rising curve of fluctuation period i, x i,k is the rising curve of the fluctuation period i at time t k is the dissolved gas concentration, and N is the total number of sampling points in the time series of the rising curve of fluctuation period i.

[0013] Furthermore, the specific method of performing Hawkes process modeling based on the fitting function of the rising curve of all fluctuation cycles to obtain the conditional intensity function for calculating the expected number of transformer faults corresponding to the dissolved gas is: Where l(t) is the conditional intensity function used to calculate the expected number of transformer faults corresponding to the dissolved gas, μ is the inherent gas production rate of the transformer without external fault stimulation, only caused by natural aging or normal operation, α is the fault self-excitation coefficient, which represents the instantaneous enhancement intensity of a single historical fault on the current fault occurrence rate, β is the fault attenuation coefficient, which controls the attenuation rate of the impact of historical faults, and w i is the weight of the rising curve of fluctuation period i; By maximizing the likelihood function L(μ,α,β), we can find the optimal values ​​of μ, α, and β:​​ .

[0014] Furthermore, the method of calculating the expected number of transformer faults corresponding to the dissolved gas in a future time period based on the conditional intensity function of the expected number of times, comparing the expected number of transformer faults corresponding to the dissolved gas in the future time period with a set expected number threshold, and determining whether to issue an early warning based on the comparison result is as follows: Where E[N(Δt)] is the expected number of failures in the future period [T, T+Δt], T is a certain time in the future, and Δt is the length of the period; When E[N(Δt)]<1.5, monitoring is maintained and no warning is issued; when 1.5≤E[N(Δt)]<2, the monitoring cycle is shortened and no warning is issued; when E[N(Δt)]≥2, a warning is issued to start offline detection and diagnosis.

[0015] Furthermore, based on the expected number of transformer faults corresponding to the dissolved gas in the future time period and the maximum concentration difference of the dissolved gas in the rising curve of all fluctuation cycles, the growth amount of the gas in the transformer oil in the future time period is calculated. The specific method is: R=E[N(Δt)]×a max Where R is the growth amount of the gas in the transformer oil in the future time period, a max It is the maximum concentration difference of the dissolved gas in the rising curve of all fluctuation cycles.

[0016] A computer-readable medium having a computer program / instruction stored thereon, wherein the computer program / instruction executes the above-mentioned Hawkes process-based early warning method for dissolved gas in transformer oil when running.

[0017] An electronic device, characterized in that it includes: a memory and a processor, the memory and the processor are communicatively connected to each other, the memory stores a computer program / instructions, and the processor executes the above-mentioned Hawkes process-based dissolved gas early warning method in transformer oil by executing the computer program / instructions.

[0018] The beneficial effects of the present invention are: 1. This invention uses the self-excitation characteristics of historical fault events to model the Hawkes process and utilizes the conditional intensity function λ(t) to dynamically calculate the expected number of faults E[N(Δt)] in the future time period. Compared with the traditional threshold comparison method that relies on passive early warning of current concentration exceeding the standard, this method can predict the probability of future faults in advance, significantly shortening the time difference from fault occurrence to early warning, and overcoming the hysteresis problem caused by slow gas diffusion.

[0019] 2. By segmenting the rising curve in historical data, starting at the valley concentration and ending at the peak concentration, and accurately fitting it using a logistic function, we focus only on the concentration rise phase caused by the fault. This method effectively eliminates the natural downward trend of gas due to solubility equilibrium and eliminates the risk of misjudgment caused by concentration fluctuations in traditional threshold methods.

[0020] 3. Use the Hawkes process model to replace the neural network black box model, solve the parameters by maximizing the likelihood function, and make the prediction process transparent and traceable. At the same time, combine the expected number of failures E[N(Δt)] and the maximum concentration difference a max , calculate the future gas growth R, provide dual quantitative basis for "whether to issue an early warning" and "change the monitoring cycle", and support accurate operation and maintenance decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 This is a system block diagram of the present invention.

[0022] Figure 2 Flow chart of the method of the present invention.

[0023] Figure 3 This is the historical fluctuation curve of the C2H2 historical concentration of a transformer over time in the past 50 days.

[0024] Figure 4 This is a schematic diagram of the segmentation points of the historical fluctuation curve of C2H2 historical concentration over time after data smoothing and data segmentation.

[0025] Figure 5 This is a schematic diagram after the rising curve is fitted. DETAILED DESCRIPTION

[0026] In order to make the technical problems, technical solutions and beneficial effects to be solved by this application more clearly understood, this application is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0027] The application is based on Hawkes Process theory, realizes the prediction and early warning of future gas production failure by modeling the time sequence correlation of historical gas production events. Hawkes Process is a point process model with self-excitation characteristics, which can capture the triggering effect and aggregation phenomenon between events. Its core feature is that historical events can enhance the intensity of future events, and this influence decays exponentially with time and superimposes. Related transformer failure cases show that from occasional gas production events to high frequency and high concentration gas production, and even to serious failure, the triggering effect and aggregation phenomenon are also presented.

[0028] Embodiment 1 As Figure 1 described, a transformer oil dissolved gas early warning system based on Hawkes Process includes: A historical data fitting module is configured to divide the historical fluctuation curve of the concentration of a certain dissolved gas in transformer oil changing with time into fluctuation periods, select the rising curve of each fluctuation period with the concentration valley of the certain dissolved gas as the starting point and the concentration peak of the certain dissolved gas as the end point, and fit the rising curves of all fluctuation periods to obtain a fitting function of the rising curves of all fluctuation periods. A Hawkes Process modeling module is configured to model Hawkes Process according to the fitting function of the rising curves of all fluctuation periods to obtain a conditional intensity function for calculating the expected number of transformer failures corresponding to the certain dissolved gas. A dissolved gas early warning module is configured to calculate the expected number of transformer failures corresponding to the certain dissolved gas in a future time period according to the conditional intensity function of the expected number, compare the expected number of transformer failures corresponding to the certain dissolved gas in the future time period with a set expected number threshold, and determine whether to issue a warning according to the comparison result.

[0029] (1) The historical data fitting module is further configured to: Collect the historical fluctuation curve of the concentration of H2, CH4, C2H2 and other dissolved gases in transformer oil changing with time (at least 1 month of historical data, as shown in Figure 3 ), and pre-process the data, such as removing values deviating more than 3 times the variance within a certain time range and filling in missing values by difference. The data of each component can be processed separately, such as C2H2 mainly focusing on discharge failure, CH4 mainly focusing on overheat failure, and H2 assisting in judging the correlation with other gas production.

[0030] The Savitzky-Golay filter (SG filter) is used to eliminate high-frequency noise in the data. It is a smoothing filter method based on local polynomial least squares fitting. Its core idea is to perform polynomial fitting on the data within a sliding window and replace the original data with the fitted value, thereby smoothing the noise while retaining the high-order features of the signal (such as peaks, inflection points, etc.) as much as possible. For a discrete time series x={x1,x2,…,x N}, its expression is: Among them, y t is the filtered gas concentration data, t is time, K is the half window width (total window width W=2K+1), c k The polynomial coefficients determined for the least-squares fit. Recommended parameters: 11-point window width, 3rd-order polynomial. Because the SG filter cannot be calculated at the data boundaries (the first and last K points), use a smaller window size (e.g., linear extrapolation) at the boundary points.

[0031] For the filtered dissolved gas concentration fluctuation curve over time, a minimum peak-to-valley difference threshold is set, and the dissolved gas concentration range in the rising curve of all fluctuation cycles with a peak-to-valley difference greater than the minimum peak-to-valley difference threshold is recorded as [V i ,P i ],V i is the valley value of the dissolved gas concentration in the rising curve of fluctuation period i, P i is the peak value of the dissolved gas concentration in the rising curve of fluctuation period i. A fluctuation period refers to a segment of the historical fluctuation curve that first rises and then falls, or first rises and then flattens. The historical fluctuation curve includes multiple fluctuation periods, each of which corresponds to a transformer fault.

[0032] Collect the historical time series data of dissolved gas concentration in oil of a transformer in the last 50 days. Taking C2H2 as an example, after removing obvious outliers and completing the data, the data trend is as follows: Figure 3 As shown, the horizontal axis unit is h, and the vertical axis unit is ppm. It can be seen that the data has an increasing trend and stabilizes at around 1.0ppm. If the conventional trend fitting prediction method is used, it may be considered that the transformer fault development tends to be stable. The local peaks and valleys of the filtered data are detected, and the data is divided according to the peaks and valleys. The minimum peak-valley difference threshold (such as 5% concentration baseline value) is set, and the minimum distance between adjacent peaks and valleys is set to 3 to avoid noise interference. Figure 4 As shown, the peak point is marked with "×" and the valley point is marked with "●". After segmentation, the data is divided into rising and falling segments.

[0033] Traditional methods are sensitive to small fluctuations in gas concentration and can easily lead to misjudgment due to noise or natural fluctuations ("the downward trend caused by dissolved gas balance affects threshold judgment" in the background technology). The present invention eliminates small fluctuations (such as <5% of the baseline value) through the threshold, retaining only the significant concentration increase segments caused by real faults, thereby reducing the false alarm rate.

[0034] (2) The historical data fitting module is also used to: The Logistic function is used to fit the rising curves of all fluctuation cycles. The Logistic function (also known as the Sigmoid function) is a classic S-shaped growth curve. The concentration change of dissolved gas in oil usually shows a trend of "slow growth-rapid rise-saturation stability", which is highly consistent with the S-shaped growth characteristics of the Logistic function. The formula is as follows: Among them, f i (t) is the fitting function of the rising curve of the fluctuation period i, t is time, a i is the peak-to-valley difference of the dissolved gas concentration in the rising curve of fluctuation period i, b i is the growth rate of the dissolved gas concentration in the rising curve of the fluctuation period i, c i is the time point when the concentration of the dissolved gas increases fastest in the rising curve of the fluctuation period i, d i is the initial dissolved concentration of the gas in the rising curve of fluctuation period i; Initialize the above parameters: a i =P i -V i ; b i =(P i -V i ) / t;c i Set to [V i ,P i ] midpoint; d i =V i ; Establish bounds for the above parameters: 0 i ≤2(P i -V i );0 i ≤1; 0.9×V i ≤d i ≤1.1×P i ; Use the Levenberg-Marquardt algorithm to solve a i 、b i 、c i d​​i The optimal value of , the following objective function is used in the solution process: Among them, k is the sampling point number of the time series in the rising curve of fluctuation period i, t k is the time corresponding to sampling point k in the rising curve of fluctuation period i, x i,k is the rising curve of the fluctuation period i at time t k is the dissolved gas concentration, and N is the total number of sampling points in the time series of the rising curve of fluctuation period i.

[0035] After fitting the rising curve, Figure 5 shown.

[0036] Traditional linear / polynomial fitting cannot capture the "S-shaped growth" characteristics of fault gas release (slow start → accelerated rise → saturation and stability). The present invention uses the Logistic function to naturally match the nonlinear growth process of fault gas production, and the physical meaning of the parameters is clear (such as b i The Levenberg-Marquardt algorithm is highly resistant to noise, ensuring reliable fitting results.

[0037] (3) In the Hawkes process modeling module, the specific formula of the conditional intensity function is as follows: Where l(t) is the conditional intensity function used to calculate the expected number of transformer faults corresponding to the dissolved gas, μ is the inherent gas production rate of the transformer without external fault stimulation, only caused by natural aging or normal operation, α is the fault self-excitation coefficient, which represents the instantaneous enhancement intensity of a single historical fault on the current fault occurrence rate, β is the fault attenuation coefficient, which controls the attenuation rate of the impact of historical faults, and w i is the weight of the rising curve of fluctuation period i; By maximizing the likelihood function L(μ,α,β), we can find the optimal values ​​of μ, α, and β: .

[0038] The starting point of each ascending segment t Vi Marked as a potential gas production failure event. Figure 4 As can be seen from the above, there are 9 rising segments in total, and the marked gas production event time points are {0, 156, 316, 436, 600, 744, 880, 968, 1016}, in hours.

[0039] The fitted parameters and calculated weights are shown in Table 1.

[0040] Table 1 By maximizing the likelihood function L(μ, α, β), we find the optimal values ​​of μ, α, and β using the EM algorithm or gradient descent. The total duration of T is 1196 hours. Based on the data in the table above, the fitting results are μ = 0.01, α = 0.50, and β = 0.01.

[0041] The existing neural network method lacks interpretability and cannot quantify the triggering effect of historical faults on future faults. i Fusion concentration change range a i and speed b i , fully characterizes the severity of historical faults. The exponential decay term e -β(t-tk) Accurately model the timeliness of fault impacts. μ (intrinsic gas generation rate), α (fault self-excitation intensity), and β (impact decay rate) have clear physical meanings, supporting root cause analysis. Maximizing the likelihood function ensures statistical optimality of parameter estimation.

[0042] (4) In the dissolved gas warning module, the formula for the expected number of transformer faults corresponding to the dissolved gas in the future time period is as follows: where E[N(Δt)] is the expected number of failures in the future time period [T, T+Δt], T is a future moment, and Δt is the length of the time period.

[0043] When E[N(Δt)] < 1.5, monitoring is maintained without issuing an early warning. When 1.5 ≤ E[N(Δt)] < 2, the monitoring cycle is shortened without issuing an early warning. When E[N(Δt)] ≥ 2, an early warning is issued to initiate offline testing and diagnosis. Further judgment is made based on whether R exceeds the threshold specified in relevant standards.

[0044] Based on the expected number of transformer faults corresponding to the dissolved gas in the future time period and the maximum concentration difference of the dissolved gas in the rising curve of all fluctuation cycles, the growth amount of the gas in the transformer oil in the future time period is calculated. The growth amount calculation formula is as follows: R=E[N(Δt)]×a max Where R is the growth amount of the gas in the transformer oil in the future time period, a max It is the maximum concentration difference of the dissolved gas in the rising curve of all fluctuation cycles.

[0045] By integrating the above data, we can predict the number of events in the next 168 hours to be E[N(Δt)]=2.27 times. When E[N(Δt)]≥2, an early warning should be issued to start offline detection and diagnosis. The maximum value of the fault amplitude a in the past 50 days is a max =0.402, R=2.27×0.402=0.91, it is predicted that the acetylene content will increase by about 0.91ppm in the next 168 hours.

[0046] Example 2 like Figure 2 As shown, a method for early warning of dissolved gas in transformer oil based on Hawkes process includes: The historical fluctuation curve of the concentration of a certain dissolved gas in transformer oil over time is divided according to the fluctuation period. In each divided fluctuation period, the rising curve of the fluctuation period is selected with the concentration valley of the dissolved gas as the starting point and the concentration peak of the dissolved gas as the end point. The rising curves of all fluctuation periods are fitted to obtain the fitting function of the rising curves of all fluctuation periods. Hawkes process modeling is performed based on the fitting function of the rising curve of all fluctuation cycles to obtain a conditional intensity function for calculating the expected number of transformer faults corresponding to the dissolved gas; Based on the conditional intensity function of the expected number of times, the expected number of transformer faults corresponding to the dissolved gas in the future time period is calculated. The expected number of transformer faults corresponding to the dissolved gas in the future time period is compared with the set expected number threshold, and a judgment is made based on the comparison result whether to issue an early warning.

[0047] (1) The specific method for selecting the rising curve of each fluctuation period by taking the valley value of the concentration of the dissolved gas as the starting point and the peak value of the concentration of the dissolved gas as the end point is as follows: Collect the historical fluctuation curve of the concentration of dissolved gases such as H2, CH4, C2H2 in transformer oil over time (historical data within at least 1 month, such as Figure 3 ), and preprocess the data, such as removing values ​​with deviations exceeding three times the variance within a certain time range and filling missing values ​​using interpolation. Data for each component can be processed separately, for example, focusing on discharge failures for C2H2, overheating failures for CH4, and using H2 to assist in determining correlations with other gas production.

[0048] The Savitzky-Golay filter (SG filter) is used to eliminate high-frequency noise in the data. It is a smoothing filter method based on local polynomial least squares fitting. Its core idea is to perform polynomial fitting on the data within a sliding window and replace the original data with the fitted value, thereby smoothing the noise while retaining the high-order features of the signal (such as peaks, inflection points, etc.) as much as possible. For a discrete time series x={x1,x2,…,x N}, its expression is: Among them, y t is the filtered gas concentration data, t is time, K is the half window width (total window width W=2K+1), c k The polynomial coefficients determined for the least-squares fit. Recommended parameters: 11-point window width, 3rd-order polynomial. Because the SG filter cannot be calculated at the data boundaries (the first and last K points), use a smaller window size (e.g., linear extrapolation) at the boundary points.

[0049] For the filtered dissolved gas concentration fluctuation curve over time, a minimum peak-to-valley difference threshold is set, and the dissolved gas concentration range in the rising curve of all fluctuation cycles with a peak-to-valley difference greater than the minimum peak-to-valley difference threshold is recorded as [V i ,P i ],V i is the valley value of the dissolved gas concentration in the rising curve of fluctuation period i, P i It is the peak value of the dissolved gas concentration in the rising curve of the fluctuation period i.

[0050] Collect the historical time series data of dissolved gas concentration in oil of a transformer in the last 50 days. Taking C2H2 as an example, after removing obvious outliers and completing the data, the data trend is as follows: Figure 3 As shown, the horizontal axis unit is h, and the vertical axis unit is ppm. It can be seen that the data has an increasing trend and stabilizes at around 1.0ppm. If the conventional trend fitting prediction method is used, it may be considered that the transformer fault development tends to be stable. The local peaks and valleys of the filtered data are detected, and the data is divided according to the peaks and valleys. The minimum peak-valley difference threshold (such as 5% concentration baseline value) is set, and the minimum distance between adjacent peaks and valleys is set to 3 to avoid noise interference. Figure 4 As shown, the peak point is marked with "×" and the valley point is marked with "●". After segmentation, the data is divided into rising and falling segments.

[0051] (2) The specific method of fitting the rising curves of all fluctuation periods to obtain the fitting function of the rising curves of all fluctuation periods is: The rising curve of each fluctuation cycle is fitted by using a Logistic function, which is a classical S-shaped growth curve. The concentration of dissolved gas in oil usually presents a trend of "slow growth-rapid rise-saturation stability", which is highly consistent with the S-shaped growth characteristics of the Logistic function. The formula is as follows: wherein, f i (t) is the fitting function of the rising curve of fluctuation cycle i, t is time, a i is the peak-to-valley difference of the concentration of the dissolved gas in the rising curve of fluctuation cycle i, b i is the growth rate of the concentration of the dissolved gas in the rising curve of fluctuation cycle i, c i is the time point at which the growth rate of the concentration of the dissolved gas in the rising curve of fluctuation cycle i is the fastest, d i is the initial dissolved concentration of the gas in the rising curve of fluctuation cycle i; The above parameters are initialized as follows: a i =P i -V i ; b i =(P i -V i ) / t; c i is set as the midpoint of [V i , P i ]; d i =V i ; The boundary constraints of the above parameters are established as follows: 0 < a i ≤ 2(P i -V i ); 0 < b i ≤ 1; 0.9×V i ≤ d i ≤ 1.1×P i ; The optimal values of a i , b i , c i , d i are solved by using the Levenberg-Marquardt algorithm, and the following objective function is used in the solving process: wherein, k is the sampling point number of the time sequence in the rising curve of fluctuation cycle i, t k is the time corresponding to the sampling point k in the rising curve of fluctuation cycle i, x i,k is the concentration of the dissolved gas in oil at time tk N is the total number of sampling points in the time series of the rising curve of fluctuation period i.

[0052] After fitting the rising curve, as shown in Figure 5 .

[0053] (3) The specific method for modeling the Hodge process according to the fitting function of the rising curve of all fluctuation periods is as follows: wherein, l(t) is the conditional intensity function for calculating the expected number of times of transformer failure corresponding to the dissolved gas, μ is the inherent gas production rate of the transformer without external failure stimulation, only by natural aging or conventional operation, α is the failure self-excitation coefficient, the failure self-excitation coefficient α represents the instantaneous enhancement intensity of a single historical failure on the current failure rate, β is the failure decay coefficient, the failure decay coefficient β controls the decay speed of the influence of historical failures, w i is the weight of the rising curve of fluctuation period i; The optimal values of μ, α, and β are solved by maximizing the likelihood function L(μ, α, β): .

[0054] The starting point t Vi of each rising section is marked as a potential gas production failure event. As can be seen from Figure 4 , there are 9 rising sections, and the marked gas production event time points are {0, 156, 316, 436, 600, 744, 880, 968, 1016}, in hours.

[0055] The fitted parameters and calculated weights are shown in Table 1.

[0056] Table 1 When solving the optimal values of μ, α, and β by maximizing the likelihood function L(μ, α, β), the EM algorithm or gradient descent method is used. The total duration Ttotal is 1196 hours, and according to the data in the above table, the fitting results are μ=0.01, α=0.50, and β=0.01.

[0057] (4) The specific method for calculating the expected number of times of transformer failure corresponding to the dissolved gas in the future time period according to the conditional intensity function of the expected number of times is as follows: where E[N(Δt)] is the expected number of failures in the future time period [T, T+Δt], T is a future moment, and Δt is the length of the time period.

[0058] When E[N(Δt)] < 1.5, monitoring is maintained without issuing an early warning. When 1.5 ≤ E[N(Δt)] < 2, the monitoring cycle is shortened without issuing an early warning. When E[N(Δt)] ≥ 2, an early warning is issued to initiate offline testing and diagnosis. Further judgment is made based on whether R exceeds the threshold specified in relevant standards.

[0059] Based on the expected number of transformer faults corresponding to the dissolved gas in the future time period and the maximum concentration difference of the dissolved gas in the rising curve of all fluctuation cycles, the specific method for calculating the growth amount of the gas in the transformer oil in the future time period is: R=E[N(Δt)]×a max Where R is the growth amount of the gas in the transformer oil in the future time period, a max It is the maximum concentration difference of the dissolved gas in the rising curve of all fluctuation cycles.

[0060] By integrating the above data, we can predict the number of events in the next 168 hours to be E[N(Δt)]=2.27 times. When E[N(Δt)]≥2, an early warning should be issued to start offline detection and diagnosis. The maximum value of the fault amplitude a in the past 50 days is a max =0.402, R=2.27×0.402=0.91, it is predicted that the acetylene content will increase by about 0.91ppm in the next 168 hours.

[0061] Example 3 A computer-readable medium having a computer program stored thereon, wherein the computer program executes the early warning method for dissolved gas in transformer oil based on Hawkes process in embodiment 2 when running.

[0062] Example 4 An electronic device, characterized in that it includes: a memory and a processor, the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the Hawkes process-based dissolved gas early warning method in transformer oil in Example 2 by executing the computer instructions.

[0063] Any material not described in detail in this specification is prior art known to those skilled in the art. Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take 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.) containing computer-usable program code.

[0064] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, 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 processes in the flowcharts and / or block diagrams. 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.

[0065] 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 1 The function specified in one or more boxes.

[0066] 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.

[0067] It should be pointed out finally that the above examples are only used for illustrating the technical solutions of the present application but not for limiting the protection scope thereof, and although the present application has been described in detail with reference to the above examples, it should be understood by those skilled in the art that the specific embodiments of the present application can be changed, modified or replaced equivalently by those skilled in the art after reading the present application, but these changes, modifications or equivalent replacements are all within the protection scope of the claims to be filed of the present application.

Claims

1. A transformer oil dissolved gas early warning system based on Hawkes process, characterized in that: include: The historical data fitting module is used to divide the historical fluctuation curve of the concentration of a certain dissolved gas in the transformer oil over time into fluctuation periods. In each divided fluctuation period, the rising curve of the fluctuation period is selected with the concentration valley of the dissolved gas as the starting point and the concentration peak of the dissolved gas as the end point, and the rising curves of all fluctuation periods are fitted to obtain the fitting function of the rising curves of all fluctuation periods. A Hawkes process modeling module is used to perform Hawkes process modeling based on the fitting function of the rising curve of all fluctuation cycles to obtain a conditional intensity function for calculating the expected number of transformer faults corresponding to the dissolved gas; The dissolved gas warning module is used to calculate the expected number of transformer faults corresponding to the dissolved gas in a future time period based on the conditional intensity function of the expected number of times, compare the expected number of transformer faults corresponding to the dissolved gas in the future time period with a set expected number threshold, and determine whether to issue a warning based on the comparison result.

2. The transformer oil dissolved gas early warning system based on Hawkes process according to claim 1, characterized in that: The historical data fitting module is further used to: Set the minimum peak-to-valley difference threshold, and record the dissolved gas concentration range in the rising curve of all fluctuation cycles with a peak-to-valley difference greater than the minimum peak-to-valley difference threshold as [V i ,P i ],V i is the valley value of the dissolved gas concentration in the rising curve of fluctuation period i, P i It is the peak value of the dissolved gas concentration in the rising curve of the fluctuation period i.

3. The transformer oil dissolved gas early warning system based on Hawkes process according to claim 2, characterized in that: The historical data fitting module is further used to: The Logistic function is used to fit the rising curves of all fluctuation cycles. The formula is as follows: Among them, f i (t) is the fitting function of the rising curve of the fluctuation period i, t is time, a i is the peak-to-valley difference of the dissolved gas concentration in the rising curve of fluctuation period i, b i is the growth rate of the dissolved gas concentration in the rising curve of the fluctuation period i, c i is the time point when the concentration of the dissolved gas increases fastest in the rising curve of the fluctuation period i, d i is the initial dissolved concentration of the gas in the rising curve of fluctuation period i; Initialize the above parameters: a i =P i -V i ; b i =(P i -V i ) / t;c i Set to [V i ,P i ] midpoint; d i =V i ; Establish bounds for the above parameters: 0 i ≤2(P i -V i );0 i ≤1;0.9×V i ≤d i ≤1.1×P i ;​​ Use the Levenberg-Marquardt algorithm to solve a i 、b i 、c i d i The optimal value of , the following objective function is used in the solution process: Among them, k is the sampling point number of the time series in the rising curve of fluctuation period i, t k is the time corresponding to sampling point k in the rising curve of fluctuation period i, x i,k is the rising curve of the fluctuation period i at time t k is the dissolved gas concentration, and N is the total number of sampling points in the time series of the rising curve of fluctuation period i.

4. The transformer oil dissolved gas early warning system based on Hawkes process according to claim 3 is characterized in that: In the Hawkes process modeling module, the specific formula of the conditional intensity function is as follows: Where l(t) is the conditional intensity function used to calculate the expected number of transformer faults corresponding to the dissolved gas, μ is the inherent gas production rate of the transformer without external fault stimulation, only caused by natural aging or normal operation, α is the fault self-excitation coefficient, which represents the instantaneous enhancement intensity of a single historical fault on the current fault occurrence rate, β is the fault attenuation coefficient, which controls the attenuation rate of the impact of historical faults, and w i is the weight of the rising curve of fluctuation period i; By maximizing the likelihood function L(μ,α,β), we can find the optimal values ​​of μ, α, and β: 。 5. The transformer oil dissolved gas early warning system based on Hawkes process according to claim 4 is characterized in that: In the dissolved gas warning module, the formula for the expected number of transformer faults corresponding to the dissolved gas in the future time period is as follows: Where E[N(Δt)] is the expected number of failures in the future period [T, T+Δt], T is a certain time in the future, and Δt is the length of the period; When E[N(Δt)]<1.5, monitoring is maintained and no warning is issued; when 1.5≤E[N(Δt)]<2, the monitoring cycle is shortened and no warning is issued; when E[N(Δt)]≥2, a warning is issued to start offline detection and diagnosis.

6. The transformer oil dissolved gas early warning system based on Hawkes process according to claim 5, characterized in that: The dissolved gas warning module is further configured to calculate the growth amount of the gas in the transformer oil within the future time period based on the expected number of transformer faults corresponding to the dissolved gas within the future time period and the maximum concentration difference of the dissolved gas in the rising curve of all fluctuation cycles. The growth amount calculation formula is as follows: R=E[N(Δt)]×a max Where R is the growth amount of the gas in the transformer oil in the future time period, a max It is the maximum concentration difference of the dissolved gas in the rising curve of all fluctuation cycles.

7. A method for warning of dissolved gas in transformer oil based on Hawkes process, characterized in that: include: The historical fluctuation curve of the concentration of a certain dissolved gas in transformer oil over time is divided according to the fluctuation period. In each divided fluctuation period, the rising curve of the fluctuation period is selected with the concentration valley of the dissolved gas as the starting point and the concentration peak of the dissolved gas as the end point. The rising curves of all fluctuation periods are fitted to obtain the fitting function of the rising curves of all fluctuation periods. Hawkes process modeling is performed based on the fitting function of the rising curve of all fluctuation cycles to obtain a conditional intensity function for calculating the expected number of transformer faults corresponding to the dissolved gas; Based on the conditional intensity function of the expected number of times, the expected number of transformer faults corresponding to the dissolved gas in the future time period is calculated. The expected number of transformer faults corresponding to the dissolved gas in the future time period is compared with the set expected number threshold, and a judgment is made based on the comparison result whether to issue an early warning.

8. The method for early warning of dissolved gas in transformer oil based on Hawkes process according to claim 7, characterized in that: The specific method of selecting the rising curve of each fluctuation period with the concentration valley of the dissolved gas as the starting point and the concentration peak of the dissolved gas as the end point is as follows: Set the minimum peak-to-valley difference threshold, and record the dissolved gas concentration range in the rising curve of all fluctuation cycles with a peak-to-valley difference greater than the minimum peak-to-valley difference threshold as [V i ,P i ],V i is the valley value of the dissolved gas concentration in the rising curve of fluctuation period i, P i It is the peak value of the dissolved gas concentration in the rising curve of the fluctuation period i.

9. The method for early warning of dissolved gas in transformer oil based on Hawkes process according to claim 8, characterized in that: The specific method of fitting the rising curves of all fluctuation periods to obtain the fitting function of the rising curves of all fluctuation periods is: The Logistic function is used to fit the rising curves of all fluctuation cycles. The formula is as follows: Among them, f i (t) is the fitting function of the rising curve of the fluctuation period i, t is time, a i is the peak-to-valley difference of the dissolved gas concentration in the rising curve of fluctuation period i, b i is the growth rate of the dissolved gas concentration in the rising curve of the fluctuation period i, c i is the time point when the concentration of the dissolved gas increases fastest in the rising curve of the fluctuation period i, d i is the initial dissolved concentration of the gas in the rising curve of fluctuation period i; Initialize the above parameters: a i =P i -V i ; b i =(P i -V i ) / t;c i Set to [V i ,P i ] midpoint; d i =V i ; Establish bounds for the above parameters: 0 i ≤2(P i -V i );0 i ≤1;0.9×V i ≤d i ≤1.1×P i ;​​ Use the Levenberg-Marquardt algorithm to solve a i 、b i 、c i d i The optimal value of , the following objective function is used in the solution process: Among them, k is the sampling point number of the time series in the rising curve of fluctuation period i, t k is the time corresponding to sampling point k in the rising curve of fluctuation period i, x i,k is the rising curve of the fluctuation period i at time t k is the dissolved gas concentration, and N is the total number of sampling points in the time series of the rising curve of fluctuation period i.

10. The method for early warning of dissolved gas in transformer oil based on Hawkes process according to claim 9, characterized in that: The specific method of performing Hawkes process modeling based on the fitting function of the rising curve of all fluctuation cycles to obtain the conditional intensity function for calculating the expected number of transformer faults corresponding to the dissolved gas is: Where l(t) is the conditional intensity function used to calculate the expected number of transformer faults corresponding to the dissolved gas, μ is the inherent gas production rate of the transformer without external fault stimulation, only caused by natural aging or normal operation, α is the fault self-excitation coefficient, which represents the instantaneous enhancement intensity of a single historical fault on the current fault occurrence rate, β is the fault attenuation coefficient, which controls the attenuation rate of the impact of historical faults, and w i is the weight of the rising curve of fluctuation period i; By maximizing the likelihood function L(μ,α,β), we can find the optimal values ​​of μ, α, and β: 。 11. The method for early warning of dissolved gas in transformer oil based on Hawkes process according to claim 10, characterized in that: The specific method of calculating the expected number of transformer faults corresponding to the dissolved gas in a future time period based on the conditional intensity function of the expected number of times, comparing the expected number of transformer faults corresponding to the dissolved gas in the future time period with a set expected number threshold, and determining whether to issue an early warning based on the comparison result is as follows: Where E[N(Δt)] is the expected number of failures in the future period [T, T+Δt], T is a certain time in the future, and Δt is the length of the period; When E[N(Δt)]<1.5, monitoring is maintained and no warning is issued; when 1.5≤E[N(Δt)]<2, the monitoring cycle is shortened and no warning is issued; when E[N(Δt)]≥2, a warning is issued to start offline detection and diagnosis.

12. The method for early warning of dissolved gas in transformer oil based on Hawkes process according to claim 11, characterized in that: Based on the expected number of transformer faults corresponding to the dissolved gas in the future time period and the maximum concentration difference of the dissolved gas in the rising curve of all fluctuation cycles, the growth amount of the gas in the transformer oil in the future time period is calculated. The specific method is: R=E[N(Δt)]×a max Where R is the growth amount of the gas in the transformer oil in the future time period, a max It is the maximum concentration difference of the dissolved gas in the rising curve of all fluctuation cycles.

13. A computer-readable medium, characterized in that The computer-readable medium stores a computer program / instruction, which, when run, executes the early warning method for dissolved gas in transformer oil based on the Hawkes process according to any one of claims 7 to 12.

14. An electronic device, characterized in that: include: A memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores a computer program / instruction, and the processor executes the Hawkes process-based early warning method for dissolved gas in transformer oil according to any one of claims 7 to 12 by executing the computer program / instruction.

Images