Transformer full life cycle management system and method based on big data analysis

Abstract

The invention relates to the technical field of transformers, and particularly discloses a transformer full life cycle management system and method based on big data analysis, and the method comprises the following steps: periodically collecting health parameters of a transformer in a sub-period, and carrying out the Z-score standardization processing of the health parameters to obtain target parameters; constructing a covariance matrix of the same type of target parameters, obtaining statistical energy of the corresponding type of target parameters, and obtaining a fitting curve based on the statistical energy; deriving the fitted curve, recording a derivative symbol change position, determining a subinterval, calculating the curvature of the fitted curve in the subinterval, and taking an absolute value of the curvature to obtain a target value; and calculating a dynamic weight on each time node in the full life cycle of the transformer based on the target value, and calculating a health score of the transformer based on the target parameter of the current time and the dynamic weight. The reliability of health score calculation is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of transformers, and in particular to a transformer full life cycle management system and method based on big data analysis. Background Art

[0002] Condition assessment is a key component of transformer lifecycle management. During the design and manufacturing phase, simulation and testing are used to control materials and processes to mitigate initial risks. During the initial installation and commissioning phase, installation accuracy and initial degradation are monitored to establish baseline data. During stable operation, dynamic monitoring integrating online and offline data is used to predict faults. During maintenance and overhaul, effectiveness is verified and strategies are optimized. Finally, safety and recovery value are assessed before decommissioning.

[0003] A Chinese patent application with publication number CN107992678A discloses a transformer health status evaluation method based on the entire life cycle. Its scheme mainly includes calculating the transformer health status evaluation score based on the transformer's initial state parameters, aging parameters, and operating value parameters; and generating the transformer's health status based on the health status evaluation score. When calculating the health status assessment score, the weights of the three parameters in the health status assessment model are fixed, ignoring the varying degrees of impact of these three parameters on the health status score over time. This leads to distorted health status scores. This is because, in the early stages of operation, the vast majority of potential faults arise from design and manufacturing defects. Initial state parameters, such as winding geometry deviation and insulation breakdown voltage, have the greatest explanatory power for health status. In the middle stages of operation, chemical aging and thermal-electric coupled fatigue accumulate. The fluctuation amplitude and degradation rate of aging parameters (such as the degree of polymerization of insulation paper and the composition of dissolved gases in oil) directly determine the remaining life. Maintaining the same low weights as in the early stages would reduce perception of the true extent of degradation. As equipment approaches the end of its design life, instantaneous extremes in current-carrying, heat-dissipating, and electric field parameters become the dominant triggers for faults. Immediate anomalies in operating parameters (such as hotspot temperature, load surge current, and number of partial discharge pulses) are most indicative of impending faults. However, fixed-weight models cannot amplify these high-frequency signals within a short window, resulting in a lack of sensitivity to impending failures and impacting decisions based on health status scores. Summary of the Invention

[0004] The purpose of the present invention is to provide a transformer full life cycle management system and method based on big data analysis to solve the above technical problems.

[0005] The purpose of the present invention can be achieved through the following technical solutions:

[0006] The transformer life cycle management method based on big data analysis includes the following steps:

[0007] The transformer's entire life cycle is divided into several sub-periods. Within each sub-period, the transformer's health parameters are periodically collected. The health parameters include initial state parameters, aging parameters, and operating value parameters. The health parameters are then normalized using Z-scores to obtain target parameters.

[0008] Construct the covariance matrix of the same target parameters, calculate the trace of the covariance matrix as the statistical energy of the target parameters of the corresponding type, and generate the coordinate points (t i , X i ), t i represents the i-th sub-period, X i represents the statistical energy corresponding to the i-th sub-period, and the fitting curve is obtained by fitting the coordinate points;

[0009] Derivate the fitting curve, record the position where the derivative sign changes, take the time period between two adjacent reciprocal sign changes as a subinterval, calculate the curvature of the fitting curve in the subinterval, and take the absolute value of the curvature to obtain the target value;

[0010] The dynamic weights of the transformer at each time point in its entire life cycle are calculated based on the target values, and the health score of the transformer is calculated based on the target parameters and dynamic weights at the current time.

[0011] Preferably, calculating the health score of the transformer includes:

[0012] Several time nodes are set at preset time intervals throughout the life cycle of the transformer. The curvature corresponding to each target parameter is obtained at a single time node, and the dynamic weight Yk=Qk / Q' of the kth target parameter is calculated. Qk and Q' represent the target values ​​of the kth target parameter and the sum of the target values ​​of all target parameters at that time node, respectively.

[0013] Get the target parameters and corresponding dynamic weights at the current time, and perform weighted summation of the real-time parameters and dynamic weights to obtain the health score of the transformer at the current time.

[0014] Preferably, obtaining the health score also includes:

[0015] Generate transformer capacity expansion and renovation plans based on health scores and grid load forecasts.

[0016] Preferably, if the curvature in a subinterval is zero, a preset value is assigned to all target parameters in the subinterval.

[0017] Preferably, the method for fitting the coordinate points includes cubic smoothing spline or local weighted regression.

[0018] Preferably, in computing the trace of the covariance matrix:

[0019] When the initial state parameters of the transformer are constant or approximately constant in the time dimension, if the covariance matrix trace value tends to zero, the lateral variance of the initial state parameters of the transformers in the same batch is introduced as an alternative trace to avoid the statistical energy being constant to zero.

[0020] Preferably, when the health score is less than a preset health score threshold, an early warning message is sent to a preset manager.

[0021] The transformer life cycle management system based on big data analysis is characterized by including:

[0022] Acquisition module: The entire life cycle of the transformer is divided into several sub-periods. The health parameters of the transformer are periodically collected in each sub-period. The health parameters include initial state parameters, aging parameters, and operating value parameters. The health parameters are normalized using Z scores to obtain target parameters.

[0023] Analysis module: Construct the covariance matrix of the same target parameters, calculate the trace of the covariance matrix as the statistical energy of the target parameters of the corresponding type, and generate the coordinate points (t i , X i ), t i represents the i-th sub-period, X i represents the statistical energy corresponding to the i-th sub-period, and the fitting curve is obtained by fitting the coordinate points;

[0024] Optimization module: Derivate the fitting curve, record the position where the derivative sign changes, take the time period between two adjacent reciprocal sign changes as a subinterval, calculate the curvature of the fitting curve in the subinterval, and take the absolute value of the curvature to obtain the target value;

[0025] Scoring module: Calculates the dynamic weights at each time point in the transformer's entire life cycle based on the target value, and calculates the transformer's health score based on the target parameters and dynamic weights at the current time.

[0026] Beneficial effects of the present invention:

[0027] 1. The present invention obtains the curvature of three types of target parameters at continuous time nodes throughout the life cycle and converts the absolute value of the curvature into a dynamic weight that changes with time. This highlights the influence of initial state parameters in the early stage and gradually increases the proportion of aging parameters and operating value parameters in the middle and late stages. The health status evaluation score can keep pace with the actual degradation rhythm of the transformer, avoiding the deviation caused by the fixed weight model ignoring time series differences, and thus making the scoring result closer to the actual operating conditions.

[0028] 2. The present invention first performs Z-score normalization on the collected data, then uses the covariance matrix trace value to characterize the statistical energy, and introduces the horizontal variance of the same batch as a substitute when the trace value approaches zero. Subsequently, smoothing splines or local weighted regression are used to complete the fitting, and the numerical continuity is maintained through the absolute value of curvature and zero zone compensation. In this way, a complete set of robust data processing links is formed in terms of dimensional unification, noise suppression and zero division protection, thereby improving the reliability and comparability of the health status assessment model. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] The present invention will be further described below with reference to the accompanying drawings.

[0030] Figure 1 It is a flow chart of the transformer life cycle management method based on big data analysis of the present invention. DETAILED DESCRIPTION

[0031] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0032] See also Figure 1 As shown, the present invention is a transformer life cycle management method based on big data analysis, which includes the following steps:

[0033] Step 1: Divide the entire life cycle of the transformer into several sub-periods, periodically collect the health parameters of the transformer in each sub-period, and perform Z-score normalization on the health parameters to obtain the target parameters.

[0034] In a specific embodiment, the entire operation history is continuously divided into multiple interconnected time periods according to the expected service life of the transformer and the characteristics of the operation stage. Each time period is called a sub-period, which is used to define a complete sampling window.

[0035] In each sub-period, three types of data collection are carried out in sequence at a uniform time interval: initial state parameters, aging parameters, and operating value parameters. After completing the collection of the three types of original data, similar data are assembled into column vectors in a sub-period, and then similar historical samples are extracted from a large database accumulated across years. The overall mean and overall variance of this type of samples are calculated as the normalization benchmark, and then the standardized target parameters are generated by subtracting the mean from each original observation value and dividing it by the square root of the variance. Through this transformation, measurements of different dimensions can present a zero-centered, unit-variance distribution within the same statistical scale, thereby laying a unified baseline for subsequent covariance calculations and time series comparisons.

[0036] Step 2: Construct the covariance matrix of the same target parameters, calculate the trace of the covariance matrix as the statistical energy of the target parameters of the corresponding type, and generate the coordinate points (t i , X i ), t i represents the i-th sub-period, X i It represents the statistical energy corresponding to the ith sub-period, and the fitting curve is obtained by fitting the coordinate points.

[0037] In a specific embodiment, taking the target parameter corresponding to the aging parameter (aging target parameter) as an example, the values ​​of the aging target parameter continuously obtained at a fixed sampling interval within a sub-period are arranged in chronological order as a single-column vector. Each component in this vector corresponds to an independent observation, so there are natural differences between the components caused by fluctuations in equipment operation.

[0038] Taking the single-column vector as input, the vector is outer-producted with itself according to the mathematical definition of the covariance matrix, and then the expectation of the outer-product result is taken. At this time, the covariance matrix obtained has only one diagonal and the order is equal to one. Because the matrix elements degenerate into the average of the squares of all the observations of the vector deviating from their arithmetic mean, the matrix trace is the same as the only non-zero element in the matrix, which is essentially equivalent to the variance of the vector.

[0039] This variance is regarded as the statistical energy of the aging target parameter in the current sub-period. The start time of the sub-period is mapped as the horizontal coordinate and the variance is mapped as the vertical coordinate to form a coordinate pair.

[0040] The coordinate pairs formed by multiple consecutive sub-periods are classified into the same sequence in chronological order. In order to avoid excessive traction of accidental peaks on the overall trend, a smoothing step based on the weighted average of adjacent coordinate pairs is introduced when processing the sequence. The temporal proximity is used to reduce the impact of short-term sharp fluctuations while retaining the overall form of the statistical energy changing over time. After this continuous smoothing process, a fitting curve describing the change of the statistical energy of the aging target parameter over the life cycle is obtained.

[0041] It should be noted that the methods for fitting the coordinate points include cubic smoothing splines or local weighted regression.

[0042] In a preferred embodiment of the present invention, in the process of calculating the trace of the covariance matrix:

[0043] When calculating the trace of the covariance matrix for the initial state parameter sequence, all observed values ​​of the parameter of the same transformer in a sub-period are combined into a single-column vector along the time axis. Then, according to the statistical definition, the outer product of this vector and itself is performed and the expectation is taken. If it is found that the obtained variance tends to zero, it means that the parameter is almost unchanged in the time series of this device. At this time, directly using this variance as statistical energy will cause the initial state information to be ignored in subsequent steps. In order to retain the quality characteristics brought about by differences in the manufacturing links, it is necessary to switch the perspective from the vertical time dimension to the horizontal batch dimension.

[0044] The specific approach is: retrieve all equipment belonging to the same production batch as the target transformer from the original production records, extract the initial state parameters of the same name obtained by these equipment in the corresponding factory inspection link, arrange these horizontal observations into sets and calculate their variances. This variance comes from slight differences in raw material purity, winding geometric tolerance, insulation dryness, etc. under the same production conditions, which can reflect the innate discreteness of the parameter caused by manufacturing process fluctuations. Therefore, this horizontal variance is written into the trace position of the covariance matrix instead of the time variance. Using this replacement value to continue the subsequent statistical energy process can avoid the initial state information being completely diluted due to zero variance, while still maintaining consistency with the statistical energy calculation caliber of other sub-periods.

[0045] Step 3: Derivate the fitting curve, record the position where the derivative sign changes, take the time period between two adjacent reciprocal sign changes as a subinterval, calculate the curvature of the fitting curve in the subinterval, and take the absolute value of the curvature to obtain the target value.

[0046] In a specific embodiment, after obtaining a fitting curve describing the change of statistical energy over time, the mathematical expression of the curve is first written in a form expanded according to the time variable, and then the expression is differentiated once. The slope information of the curve at any time can be obtained by differentiation. A positive slope value indicates that the curve changes upward at this moment, while a negative slope value indicates a downward change.

[0047] The slope sign is calculated point by point along the time axis by sampling at equal intervals, and the time coordinates of each change from positive to negative or from negative to positive are recorded sequentially. These time coordinates are the slope zero points. The zero points are adjacent to each other to define several continuous time periods, and each continuous time period is defined as a subinterval.

[0048] In each subinterval, the fitting curve is further differentiated twice. The second derivative gives the instantaneous rate change of the curvature of the curve. In a geometric sense, the curvature formula can be used to link the second derivative with the first derivative to obtain the curvature expression. The curvature expression is then integrated and averaged or densely sampled to obtain the average curvature in the subinterval. The curvature measures the strength of the curve in this interval. In order to avoid positive and negative cancellation caused by the sign, the absolute value of the average curvature is taken and recorded as the target value of the subinterval.

[0049] For example, when the fitting curve is convex in a certain period of time, the second derivative is positive and the absolute value of the curvature is large. If it is close to a straight line in another period of time, the second derivative approaches zero and the absolute value of the curvature is small. The absolute value of the curvature can be used to compare the differences in the statistical energy change rates in different sub-intervals.

[0050] It should be noted that after performing curvature extraction, if the quadratic derivatives of the three types of fitting curves in a certain subinterval are all zero at all sampling points in the entire interval, it can be determined that the statistical energy of the interval maintains a linear or approximately linear change over time, and the calculation results of the absolute value of the curvature are all zero.

[0051] After detecting this situation, the system first queries a preset non-zero constant, which is derived from the lower limit of the absolute value of the curvature in the transformer's historical samples and has the same dimension as the curvature.

[0052] The constant is written into the curvature positions corresponding to the initial state target parameter, aging target parameter and running value target parameter in the sub-interval respectively to form a new set of curvature absolute values; thereby raising the three curvatures from zero to the same minimum positive value, maintaining the relative balance of the curvature set, and avoiding numerical pauses caused by the dynamic weight denominator being zero in the next step.

[0053] Step 4: Calculate the dynamic weights at each time point in the transformer's entire life cycle based on the target values, and calculate the transformer's health score based on the target parameters and dynamic weights at the current time.

[0054] In a preferred embodiment of the present invention, calculating the health score of the transformer includes:

[0055] A fixed time interval is set according to the planned maintenance rhythm or the computing load of the monitoring platform, and a series of equally spaced time nodes are marked along the time axis throughout the entire life cycle. At each time node, the absolute value of the curvature of the initial state target parameter, aging target parameter, and operating value target parameter previously calculated within the subinterval to which the node belongs is read. The three absolute values ​​of curvature are denoted as Q1, Q2, and Q3 respectively, and the sum of the three is denoted as Q'. Subsequently, the absolute value of each curvature is divided by Q' using the proportional relationship to obtain the dynamic weights Y1, Y2, and Y3.

[0056] The absolute value of curvature corresponds to the degree of curvature of the fitting curve in a geometric sense. The higher the degree of curvature, the faster the statistical energy of the target parameter changes at this stage, reflecting its greater contribution to the time-varying sensitivity of the health status. Therefore, using the absolute value of curvature as the numerator can allow the weight to be adjusted in real time with the contribution, while introducing the sum of the absolute values ​​of curvature in the denominator can constrain the three weights to maintain the normalized property within the same comparable proportional framework.

[0057] After the weight calculation is completed, the real-time initial state target parameters, real-time aging target parameters and real-time operation value target parameters just stored in the monitoring system are synchronously called, and these three real-time target parameters are arranged in the same order as a column vector P. Then the three dynamic weights are arranged in the same order as a column vector Y. The corresponding components of vector P and vector Y are multiplied and summed to obtain the health score of the current time node. This summation step is equivalent to weighting the contribution of the three real-time target parameters within the same statistical scale. Therefore, the current status of the three types of parameters can be condensed into a comprehensive indicator of a single scale for subsequent decision logic calls.

[0058] Another preferred embodiment of the present invention further includes:

[0059] Generate transformer capacity expansion and renovation plans based on health scores and grid load forecasts.

[0060] In a specific embodiment, the health scores are arranged into a curve in chronological order. By observing the downward trend of the curve in several future prediction windows, the load-bearing range of the remaining life of the equipment can be determined. Then, the power grid load prediction curve is mapped to the same time coordinate. The peak-to-valley difference of the load curve shows the fluctuation range of future operating stress.

[0061] In each prediction window, the vertical distance between the lowest point of the health score curve and the highest point of the load curve is calculated, and this distance is compared with the safety margin of the transformer's current rated capacity. The smaller the vertical distance, the closer the health margin is to the upcoming high load point, indicating that the rated capacity can no longer meet subsequent growth needs.

[0062] When the comparison results trigger a capacity shortage judgment, the thermal-electrical-mechanical bidirectional expansion coefficients reserved for this model during the factory design stage are retrieved from the equipment production archive. These coefficients reflect the upper limit potential of the winding conductor cross-section, core magnetic density, and oil channel cooling capacity under design redundancy. Combined with the on-site layout space constraints and the possible additional modification range of the heat dissipation circuit, the scalable conductor cross-section increase, core magnetic density increase value, and oil pump flow increase ratio are compiled into a list of alternative expansion plans.

[0063] In order to determine the optimal capacity expansion plan, the peak duration of the load forecast curve is statistically analyzed. A peak with a longer duration means that higher heat dissipation stability is required after the transformation, while a shorter duration allows the use of a local enhanced cooling solution. Based on this, an expansion plan that can meet the thermal stability requirements and can be implemented within the operational space is selected from the list, and construction steps such as winding replacement process, rearrangement of the insulating oil circulation path, and additional location of the cooler group are listed according to the plan. Finally, a capacity expansion and transformation implementation plan is formed, which includes technical routes, construction period and raw material configuration.

[0064] Transformer full life cycle management system based on big data analysis, including:

[0065] Acquisition module: The entire life cycle of the transformer is divided into several sub-periods. The health parameters of the transformer are periodically collected in each sub-period. The health parameters include initial state parameters, aging parameters, and operating value parameters. The health parameters are normalized by Z-score to obtain the target parameters.

[0066] Analysis module: Construct the covariance matrix of the same target parameters, calculate the trace of the covariance matrix as the statistical energy of the target parameters of the corresponding type, and generate the coordinate points (t i , X i ), t i represents the i-th sub-period, X i It represents the statistical energy corresponding to the ith sub-period, and the fitting curve is obtained by fitting the coordinate points.

[0067] Optimization module: Derivate the fitting curve, record the position where the derivative sign changes, take the time period between two adjacent reciprocal sign changes as a subinterval, calculate the curvature of the fitting curve in the subinterval, and take the absolute value of the curvature to obtain the target value.

[0068] Scoring module: Calculates the dynamic weights at each time point in the transformer's entire life cycle based on the target value, and calculates the transformer's health score based on the target parameters and dynamic weights at the current time.

[0069] The above is a detailed description of an embodiment of the present invention. However, the content is only a preferred embodiment of the present invention and should not be considered to limit the scope of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the scope of the present invention.

Claims

1. A transformer life cycle management method based on big data analysis, characterized in that: The following steps are involved: The transformer's entire life cycle is divided into several sub-periods. Within each sub-period, the transformer's health parameters are periodically collected. The health parameters include initial state parameters, aging parameters, and operating value parameters. The health parameters are then normalized using Z-scores to obtain target parameters. Construct the covariance matrix of the same target parameters, calculate the trace of the covariance matrix as the statistical energy of the target parameters of the corresponding type, and generate the coordinate points (t i , X i ), t i represents the i-th sub-period, X i represents the statistical energy corresponding to the i-th sub-period, and the fitting curve is obtained by fitting the coordinate points; Derivate the fitting curve, record the position where the derivative sign changes, take the time period between two adjacent reciprocal sign changes as a subinterval, calculate the curvature of the fitting curve in the subinterval, and take the absolute value of the curvature to obtain the target value; The dynamic weights of the transformer at each time point in its entire life cycle are calculated based on the target values, and the health score of the transformer is calculated based on the target parameters and dynamic weights at the current time.

2. The transformer life cycle management method based on big data analysis according to claim 1 is characterized in that: The calculation of the transformer health score includes: Several time nodes are set at preset time intervals throughout the life cycle of the transformer. The curvature corresponding to each target parameter is obtained at a single time node, and the dynamic weight Yk=Qk / Q' of the kth target parameter is calculated. Qk and Q' represent the target values ​​of the kth target parameter and the sum of the target values ​​of all target parameters at that time node, respectively. Get the target parameters and corresponding dynamic weights at the current time, and perform weighted summation of the real-time parameters and dynamic weights to obtain the health score of the transformer at the current time.

3. The transformer life cycle management method based on big data analysis according to claim 1 is characterized in that: After obtaining the health score, the following steps are also included: Generate transformer capacity expansion and renovation plans based on health scores and grid load forecasts.

4. The transformer life cycle management method based on big data analysis according to claim 1 is characterized in that: If the curvature is zero in a subinterval, a preset value is assigned to all the target parameters in the subinterval.

5. The transformer life cycle management method based on big data analysis according to claim 1 is characterized in that: Methods for fitting the coordinate points include cubic smoothing splines or locally weighted regression.

6. The transformer life cycle management method based on big data analysis according to claim 1 is characterized in that: In the process of calculating the trace of the covariance matrix: When the initial state parameters of the transformer are constant or approximately constant in the time dimension, if the covariance matrix trace value tends to zero, the lateral variance of the initial state parameters of the transformers in the same batch is introduced as an alternative trace to avoid the statistical energy being constant to zero.

7. The transformer life cycle management method based on big data analysis according to claim 1 is characterized in that: When the health score is lower than the preset health score threshold, an early warning message is sent to the preset manager.

8. The transformer life cycle management system based on big data analysis is characterized by: include: Acquisition module: The entire life cycle of the transformer is divided into several sub-periods. The health parameters of the transformer are periodically collected in each sub-period. The health parameters include initial state parameters, aging parameters, and operating value parameters. The health parameters are normalized using Z scores to obtain target parameters. Analysis module: Construct the covariance matrix of the same target parameters, calculate the trace of the covariance matrix as the statistical energy of the target parameters of the corresponding type, and generate the coordinate points (t i , X i ), t i represents the i-th sub-period, X i represents the statistical energy corresponding to the i-th sub-period, and the fitting curve is obtained by fitting the coordinate points; Optimization module: Derivate the fitting curve, record the position where the derivative sign changes, take the time period between two adjacent reciprocal sign changes as a subinterval, calculate the curvature of the fitting curve in the subinterval, and take the absolute value of the curvature to obtain the target value; Scoring module: Calculates the dynamic weights at each time point in the transformer's entire life cycle based on the target value, and calculates the transformer's health score based on the target parameters and dynamic weights at the current time.

Images