A method and system for on-line measurement of no-load and load losses of distribution transformers

Abstract

The present application belongs to the technical field of power system monitoring, and particularly relates to a kind of distribution transformer no-load and load loss on-line measurement method and system, its method includes: obtaining transformer load rate and total loss and storing in buffer;Assess the distribution of load rate to extract distribution scarcity, combine relative measurement accuracy to generate data validity weight by feature fusion;According to the sign and amplitude of load rate change rate, construct asymmetric amplification coefficient;The coefficient, validity weight and dimensionless normalization residual feature fusion based on prediction residual are fused to calculate adaptive forgetting factor;It is embedded in the covariance update process of recursive least squares to realize online identification.The present application comprehensively considers data distribution, measurement accuracy and asymmetric law of thermal inertia, which helps to improve the online identification accuracy and dynamic response speed.

Description

Technical Field

[0001] This invention relates to the field of power system monitoring technology. More specifically, this invention relates to a method and system for online measurement of no-load and load losses of distribution transformers. Background Technology

[0002] With the deepening of lean management in power distribution networks, real-time monitoring of the no-load and load loss parameters of distribution transformers is crucial for area energy efficiency assessment and line loss calculation. In actual operation scenarios, the load exhibits significant uneven temporal distribution due to the influence of industrial production or residential life, and the complex thermodynamic processes within distribution transformers cause their loss characteristics to dynamically drift with temperature rise and load fluctuations. Therefore, the industry urgently needs a distribution transformer loss measurement solution that can identify and accurately track changes in operating conditions online in real time and possesses high stability.

[0003] Currently, in order to achieve online acquisition of loss parameters, the most commonly used traditional technical method in the industry is to fit the model using the recursive least squares method with a fixed forgetting factor. This method usually sets a constant close to 1 as the forgetting factor in advance. By collecting the load rate and total loss data of the distribution transformer, the coefficients of the loss model are iteratively updated using the least squares principle, thereby estimating the no-load loss and load loss of the distribution transformer without stopping operation.

[0004] However, existing technologies have significant shortcomings under complex real-world operating conditions, making it difficult for online identification results to accurately reflect the loss status of distribution transformers. Firstly, existing algorithms fail to differentiate between different levels of data scarcity. In actual operation, data is dense under low-load conditions but extremely scarce under high-load conditions. Using fixed weights leads to the key, scarce features in the high-load region being masked by the massive amount of repetitive data in the low-load region, resulting in model data saturation and limited identification accuracy. Secondly, existing methods ignore the dynamic fluctuations in measurement accuracy. When the transformer is operating under low load, the signal strength collected by sensors is close to the noise floor, resulting in relatively large measurement errors and low signal-to-noise ratios. Traditional algorithms treat measurement data from all operating conditions equally, causing dense hardware noise in the low-load region to severely interfere with the overall fitting stability. Third, traditional methods do not take into account the physical asymmetry of the thermodynamic process of transformers. The windings of distribution transformers heat up quickly when the load increases, but cool down slowly due to heat dissipation limitations when the load drops sharply. Existing methods use symmetrical constant response weights to deal with load increases and decreases, which fail to match this difference in thermal inertia. This results in the identification process lagging behind when the load increases and exhibiting serious overshoot and oscillation phenomena when the load decreases. Summary of the Invention

[0005] To address the technical challenges of uneven data distribution, noise interference, and sudden changes in operating conditions in existing methods, this invention proposes an online measurement method and system for the no-load and load losses of distribution transformers. This system can feature-fuse data density and measurement accuracy, and combine dynamic normalized residual feedback and adaptive adjustment of historical data weights based on the load change rate to achieve highly dynamic and high-precision online identification of distribution transformer losses.

[0006] In a first aspect, the present invention provides an online measurement method for no-load and load losses of a distribution transformer, comprising: acquiring the current load rate, the previous load rate, and the total loss of the distribution transformer, and storing the current load rate in a historical data buffer; evaluating the distribution of the current load rate in the historical data buffer and extracting the data distribution scarcity; calculating the relative measurement accuracy based on the current load rate; performing feature fusion on the relative measurement accuracy and the data distribution scarcity to generate a data validity weight for characterizing the reference value of the current operating condition; obtaining the prediction residual based on the current load rate, calculating the difference between the current load rate and the previous load rate to obtain the directional change rate, constructing an asymmetric amplification coefficient characterizing the thermal inertia difference based on the sign and amplitude of the directional change rate, and calculating an adaptive forgetting factor by combining the data validity weight, the asymmetric amplification coefficient, and the normalized prediction residual feature characterizing the uncertainty of the model state in the recursive least squares method; embedding the adaptive forgetting factor into the covariance update process of the recursive least squares method for online identification, and outputting the no-load loss and load loss of the distribution transformer.

[0007] This invention constructs a multi-dimensional evaluation index covering data distribution, measurement accuracy fluctuations, and load change rates. The evaluation results are mapped to an adaptive forgetting factor and introduced into the covariance update stage of the recursive least squares method. This allows the parameter identification process to dynamically adjust the weight allocation of historical data based on real-time operating conditions. While maintaining the smoothness of the identification results under steady-state conditions, this mechanism improves the model's response performance to sudden load changes, facilitating real-time and effective monitoring of transformer loss characteristics.

[0008] Preferably, the step of evaluating the distribution of the current load rate in the historical data buffer and extracting the data distribution scarcity includes: dividing the full range of the load rate into multiple sub-intervals with equal step sizes; assigning the current load rate to the corresponding sub-interval based on the left-closed-right-open principle to obtain a target interval; counting the load rate data points in the historical data buffer that fall into the target interval to obtain the number of data points; using the ratio of the number of data points to the total number of actual data points in the historical data buffer as the interval data density, and then calculating the data distribution scarcity based on the interval data density.

[0009] This invention measures the scarcity of working conditions by statistically analyzing the proportion of data points falling into the target interval. This helps to identify and suppress the cumulative effect of excessively repeated working conditions on the identification process, and improves the identification decline problem caused by data saturation from the distribution dimension.

[0010] Preferably, the step of calculating the relative measurement accuracy based on the current load rate includes: constructing an accuracy mapping table for the measuring device, wherein the accuracy mapping table records the correspondence between different load rate intervals and basic measurement errors; using the current load rate as an index to query the accuracy mapping table and extract the basic measurement error that matches the current load rate; and using the difference between the constant 1 and the basic measurement error as the relative measurement accuracy.

[0011] This invention achieves an effective characterization of the physical characteristics of measurement equipment, namely, reduced signal-to-noise ratio and increased error under low load conditions, through a mapping relationship that conforms to engineering practice, thus providing an indicator for subsequent evaluation of data reliability.

[0012] Preferably, the step of performing feature fusion on the relative measurement accuracy and the data distribution scarcity to generate a data validity weight for characterizing the reference value of the current working condition includes: obtaining a preset validity upper limit; calculating the numerical proportionality coefficient between the relative measurement accuracy and the data distribution scarcity; comparing the proportionality coefficient and the validity upper limit, and taking the smaller value of the two as the data validity weight.

[0013] This invention achieves synchronous correction of measurement noise and uneven distribution of historical data through dynamic ratio calculation, which helps to improve the anti-interference capability of online identification.

[0014] Preferably, obtaining the adaptive forgetting factor includes: acquiring the trace of the prediction residual and covariance matrix at the previous time step; quotienting the square of the prediction residual with the trace of the covariance matrix to generate a dimensionless normalized residual feature; determining the orientation adjustment coefficient based on the sign of the orientation change rate combined with a nonlinear mapping function; calculating the product of the orientation adjustment coefficient and the absolute value of the orientation change rate, and summing it with a constant 1 as the asymmetric amplification coefficient; calculating the product of the data validity weight, the normalized residual feature, and the asymmetric amplification coefficient, and using the ratio of this product to the reference uncertainty level as the exponential independent variable; and constructing a negative exponential decay function with the exponential independent variable to obtain the adaptive forgetting factor.

[0015] This invention precisely matches the differences in physical characteristics during the load increase and decrease process by using an asymmetric amplification factor, and constructs a dimensionless control variable using the normalized residual ratio, thereby mitigating the risk of dimensional conflict and improving the smoothness of results under steady-state conditions.

[0016] Preferably, the direction adjustment coefficient is determined based on the sign of the directional change rate combined with a nonlinear mapping function, including: if the directional change rate is greater than or equal to 0, calculating the load increase coefficient as the direction adjustment coefficient based on the load increase temperature rise curve mapping function; if the directional change rate is less than 0, calculating the load decrease coefficient as the direction adjustment coefficient based on the load decrease cooling curve mapping function, wherein the value of the load increase coefficient is greater than the load decrease coefficient; constructing a negative exponential decay function with an exponential independent variable to calculate the adaptive forgetting factor includes: obtaining a preset upper limit and lower limit of the forgetting factor; calculating the difference between the upper limit and lower limit of the forgetting factor; multiplying the difference by the negative exponential term calculated based on the exponential independent variable, and adding the product of the difference and the negative exponential term to the lower limit of the forgetting factor to obtain the adaptive forgetting factor.

[0017] This invention uses a nonlinear mapping function to fit the rate difference between temperature rise and cooling processes, which helps to eliminate overshoot and hysteresis in the identification results caused by thermal inertia.

[0018] Preferably, the adaptive forgetting factor is embedded in the covariance update process of the recursive least squares method for online identification, and the no-load loss and load loss of the distribution transformer are output, including: constructing a regression vector composed of a constant 1 and the square of the load rate at the current time; extracting the parameter estimate value at the previous time, calculating the product of the transpose of the regression vector and the parameter estimate value to obtain the predicted total loss; and calculating the difference between the actual obtained total loss and the predicted total loss to obtain the prediction residual.

[0019] Preferably, after obtaining the predicted residual, the method further includes: extracting the covariance matrix of the previous time step; calculating the continuous product of the transpose of the regression vector, the covariance matrix, and the regression vector to obtain a first intermediate variable; calculating the sum of the adaptive forgetting factor and the first intermediate variable as the gain denominator; calculating the product of the covariance matrix and the regression vector to obtain a second intermediate variable; and using the ratio of the second intermediate variable to the gain denominator as the gain vector.

[0020] Preferably, after using the ratio of the second intermediate variable to the gain denominator as the gain vector, the method further includes: calculating the product of the gain vector and the prediction residual; adding the product to the parameter estimate at the previous time step to update the parameter estimate at the current time step; analyzing the parameter estimate at the current time step to output the no-load loss and load loss of the distribution transformer; calculating the continuous product of the gain vector, the transpose of the regression vector, and the covariance matrix to obtain the covariance adjustment term; and dividing the difference between the covariance matrix and the covariance adjustment term by the adaptive forgetting factor to update the covariance matrix at the current time step.

[0021] Secondly, the present invention provides an online measurement system for no-load and load losses of a distribution transformer, comprising a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned method for online measurement of no-load and load losses of a distribution transformer is implemented.

[0022] By adopting the above technical solution, a computer program is generated from the above-mentioned method for online measurement of no-load and load losses of distribution transformers, and stored in a memory for loading and execution by a processor. Terminal equipment is then manufactured based on the memory and processor for convenient use.

[0023] The beneficial effects of this invention are as follows:

[0024] This invention addresses the dimensional conflict of exponential variables to some extent by constructing the ratio of the squared predicted residual to the trace of the covariance matrix as a normalized residual feature, thus improving the robustness of the algorithm when switching between different power units or using units of different orders of magnitude. Simultaneously, the introduction of a predicted residual feedback mechanism allows the forgetting factor to be dynamically adjusted according to the model's identification accuracy, helping to mitigate fluctuations in identification results caused by measurement noise. Furthermore, this invention constructs an asymmetric amplification coefficient through a directional change rate and its accompanying nonlinear mapping function, which can better fit the thermal inertia differences during transformer load increases and decreases, improving the tracking lag phenomenon of traditional symmetric response algorithms during operating condition switching. Finally, by fusing features to determine the scarcity of load distribution, it helps balance the contribution weights of high-value scarce features and densely repetitive operating condition features, further improving identification accuracy and robustness across the entire measurement range. Attached Figure Description

[0025] Figure 1 This is a flowchart of an online measurement method for no-load and load losses of a distribution transformer according to the present invention;

[0026] Figure 2 This is a schematic diagram of the distribution and interval density of historical buffer data in this invention;

[0027] Figure 3 This is a comparison chart of the no-load loss identification results under steady-state low load in this invention. Detailed Implementation

[0028] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0029] This invention discloses an online method for measuring the no-load and load losses of a distribution transformer, referring to... Figure 1 This includes steps S1-S6:

[0030] S1. Obtain the current load rate, the previous load rate, and the total loss of the distribution transformer, and store the current load rate in the historical data buffer.

[0031] In an optional embodiment, the real-time operating data of the distribution transformer is first acquired. This is done by measuring the real-time load current of the distribution transformer using a current transformer installed on the transformer side, and then dividing this real-time load current by the transformer's rated current to obtain the current load rate. Simultaneously, the total losses of the distribution transformer are acquired in real-time through a power measurement terminal. Each time the real-time operating data of the distribution transformer is acquired, in addition to obtaining the current load rate, the load rate from the previous sampling point must also be extracted and retained simultaneously. This allows for the determination of the true direction of load change based on the difference between the two. Furthermore, the sampling time interval for acquiring the real-time operating data needs to be determined based on the rate of change of the distribution transformer load. For example, for distribution transformers in industrial parks where load changes rapidly, a sampling time interval of 15 minutes is preferred.

[0032] Furthermore, to provide the necessary foundational data support for subsequent calculations of data distribution characteristics, it is necessary to rationally store and manage the collected real-time operating data of the distribution transformers, thereby forming a closed-loop data flow. Specifically, the current load rate of the distribution transformer is stored in a historical data buffer. This involves first constructing a historical data buffer using a circular queue structure, and then determining whether the total number of actual data points in the historical data buffer reaches the preset historical window length. If the total number of actual data points reaches the preset historical window length, the current load rate overwrites the earliest stored load rate in the historical data buffer. If the total number of actual data points does not reach the preset historical window length, the total number of actual data points is used as the denominator for calculating the data density of the calculation interval.

[0033] It should be noted that the preset historical window length needs to be determined by balancing statistical accuracy with the computational load of the underlying hardware. For example, in this invention, the preset historical window length is set to 100. With a sampling interval of 15 minutes for the distribution transformer, this preset historical window length corresponds to the load rate collected and stored in the historical data buffer within approximately 25 hours, which precisely covers the load cycle of the distribution transformer in a standard day. In the initial stage of online monitoring, since there are relatively few data points of the current load rate stored in the historical data buffer, the total number of actual data points has not yet reached 100. At this time, the total number of currently stored actual data points is directly used as a benchmark for subsequent density statistics. As the running time progresses, once the total number of actual data points reaches 100, the historical data buffer will remain fully loaded. At this point, once a new current load rate arrives, it will automatically replace the earliest stored load rate in the historical data buffer, thus always keeping the preset historical window length constant.

[0034] In this way, by establishing a historical data buffer with a fixed length and supporting dynamic updates, not only is reliable recent basic data support provided for subsequent calculation of the unevenness of the load rate distribution of distribution transformers, but the risk of memory overflow caused by storing massive amounts of data indefinitely is also avoided, thereby ensuring that the data statistics process can accurately reflect the real load distribution characteristics of distribution transformers in the near term.

[0035] S2. Evaluate the distribution of the current load rate in the historical data buffer and extract the data distribution scarcity.

[0036] In an optional embodiment, firstly, the target interval to which the current load rate of the distribution transformer belongs in the historical data buffer is determined, and the statistical method for the number of data points within the target interval is determined based on the distribution of the load rate in the historical data buffer. Secondly, the full range of the load rate is divided into multiple sub-intervals with equal step sizes; then, the number of data points in the historical data buffer that fall into the target interval is counted; and the interval data density is obtained by dividing the number of data points in the historical data buffer that fall into the target interval by the total number of actual data points.

[0037] Here, "full range" refers to the complete range that the load rate of the distribution transformer may cover. Based on the overload capacity of the distribution transformer, the value range of the full range is set to 0 to 1.2. Simultaneously, the full range is equally divided according to a preset step size to achieve refined density statistics. The step size setting directly affects the smoothness of the statistics. For example, in this invention, the preset step size is set to 0.05. By calculating the quotient of the current load rate of the distribution transformer relative to the preset step size and performing a rounding down operation, the specific sub-interval to which the current load rate of the distribution transformer belongs, i.e., the target interval, can be quickly located. Subsequently, the historical data buffer is traversed, and the load rate data points falling into the target interval are counted, thereby obtaining the number of data points in the historical data buffer that fall into the target interval.

[0038] Furthermore, to measure the density of data points falling into the target interval within the historical data buffer, an interval data density was constructed, and its calculation method is as follows:

[0039]

[0040] in, This represents the interval data density of the target interval to which the load rate of the distribution transformer belongs at the current moment; This represents the number of data points in the historical data buffer that fall within the target range. This represents the total number of actual data points within the historical data buffer. Its value is determined based on the statistical results mentioned above and is not greater than the preset historical window length of 100.

[0041] It should be noted that the interval data density reflects the duration and frequency of the distribution transformer's stay within a specific load interval. A high interval data density indicates that the distribution transformer's operating point has been sufficiently sampled from the actual data points in the historical data buffer. Continuing to assign a high weight to this operating point may lead to local overfitting in the identification model or algorithm convergence stagnation. Conversely, a low interval data density indicates that the operating point is a real data point with an extremely low frequency of occurrence and deviates from the common distribution range in the historical data buffer. Therefore, the density is obtained by calculating the proportion of data points falling into the target interval in the historical data buffer to the total number of actual data points. Its reciprocal essentially reflects the scarcity of data distribution, and can more objectively measure the contribution weight of the currently collected load rate to the subsequent online identification of loss parameters.

[0042] Reference Figure 2It can be seen that the load rate distribution of the distribution transformer exhibits significant unevenness, with most data concentrated in the low-load range, while data points in the high-load range (over 70%) are extremely sparse. By adjusting the interval data density, these high-load data points can be accurately identified. In subsequent calculations, the validity weight of high-load data is increased by reducing the denominator of the interval data density, thus ensuring that scarce, high-value operating condition characteristics are not masked by dense low-load noise.

[0043] In this way, by conducting real-time statistics on the spatial distribution of the load rate of distribution transformers, the scarcity of the current operating conditions of distribution transformers can be assessed, providing a weight basis for the distribution dimension for the effectiveness of subsequent dynamic adjustment of data, and effectively avoiding the problem of decreased parameter identification accuracy caused by excessive accumulation of data under a single operating condition.

[0044] S3. Calculate the relative measurement accuracy based on the load rate at the current moment.

[0045] In an optional embodiment, due to the objective physical differences in the measurement characteristics of current transformers and power measuring devices installed on the distribution transformer side under different loads, the signal-to-noise ratio of these measuring devices decreases under extremely low load conditions, and the measurement error increases accordingly. To characterize this characteristic, the present invention employs a lookup table method that conforms to engineering practice.

[0046] Specifically, based on existing national measurement equipment accuracy standards, such as those for 0.2S class current transformers, a precision mapping table for measurement equipment is pre-constructed. This precision mapping table divides the load rate into multiple discrete lookup intervals, such as 1%–5%, 5%–20%, 20%–100%, and 100%–120%, and directly assigns a fixed standard basic measurement error.

[0047] The current load rate is then directly input as an index into the accuracy mapping table for matching, and the corresponding basic measurement error is extracted. Finally, the basic measurement error is subtracted from the constant 1 to directly output the dimensionless relative measurement accuracy.

[0048] Thus, by adopting a lookup table method that conforms to engineering practice, the objective physical characteristics of reduced signal-to-noise ratio and increased measurement error of measuring equipment under extremely low load conditions are effectively characterized, thereby directly outputting dimensionless relative measurement accuracy.

[0049] S4. Perform feature fusion on relative measurement accuracy and data distribution scarcity to generate data validity weights that characterize the reference value of the current working condition.

[0050] In an optional embodiment, the contribution of the load rate acquired at the current moment to parameter identification is evaluated by comprehensively measuring the sensor's measurement reliability and the distribution characteristics of the current operating condition within the historical data buffer. Higher sensor measurement reliability and a scarcer distribution of the current operating condition indicate a greater amount of effective information contained in the actual data points sampled at the current moment.

[0051] Furthermore, firstly, a preset validity upper limit is obtained. Then, a multiplicative fusion index is calculated, which is the ratio of relative measurement accuracy to the scarcity of data distribution. It should be noted that this ratio is the reciprocal of density. Finally, this fusion index is compared with the preset validity upper limit, and the smaller value is used as the data validity weight. Data validity is thus obtained by calculating the ratio of relative measurement accuracy to interval data density.

[0052] The preset validity upper limit is a pre-set threshold constant used to prevent data validity from overflowing due to excessively low interval data density under extreme operating conditions. If the interval data density is extremely low, for example, if the current load rate falls into a completely new target interval that has never been recorded before, the ratio of relative measurement accuracy to interval data density will become extremely large, which may lead to instability in the step size of the subsequent identification process. By introducing a preset validity upper limit for truncation, the robustness of the identification process can be ensured. For example, in this invention, the preset validity upper limit is set to 20. Based on the above logic, data validity is constructed, and its calculation method is as follows:

[0053]

[0054] in, Weights for data validity; For relative measurement accuracy; The interval data density is used as the denominator to reflect the fusion of scarcity. The upper limit of validity is preset, and in this invention, it is set to 20.

[0055] The above formula establishes a dual weighting mechanism: when the distribution transformer operates in a high-load range, the relative measurement accuracy is higher, and if the proportion of actual data points in this high-load range within the historical data buffer is low, the data density of the range is lower, significantly increasing the validity of the final calculated data. This means that the identification process assigns greater weight to operating conditions with high quality and scarcity. Conversely, if the operating conditions are extremely widespread in a low-load range, the data validity will remain at a low level. Thus, through this dynamic ratio calculation, adaptive grading of input data quality is achieved.

[0056] In this way, by calculating the ratio of relative measurement accuracy to interval data density and combining it with the upper limit truncation mechanism, synchronous correction of measurement noise and uneven distribution of actual data points in different load rate intervals within the historical data buffer is achieved. This ensures that only load rates with high reliability and representativeness of new operating conditions can have a significant impact on the identification model, thereby improving the anti-interference capability of online measurement from the source.

[0057] S5. Based on the current load rate, obtain the prediction residual, calculate the difference between the current load rate and the previous load rate, obtain the directional change rate, construct an asymmetric amplification coefficient representing the difference in thermal inertia based on the sign and magnitude of the directional change rate, and calculate the adaptive forgetting factor by combining the data validity weight, the asymmetric amplification coefficient, and the normalized prediction residual characteristics representing the uncertainty of the model state in the recursive least squares method.

[0058] In an optional embodiment, to accurately track the evolution of loss characteristics of the distribution transformer during load fluctuations, especially to compensate for the thermal inertia caused by the inconsistency between the oil temperature and winding temperature changes of the distribution transformer, it is necessary to dynamically guide the convergence speed of the identification process through the load change trend. Specifically, firstly, the load change characteristics are characterized, that is, the directional change rate is obtained by calculating the difference between the load rate at the current moment and the load rate at the previous moment. The calculation method is as follows:

[0059]

[0060] in, For directional change rate; The load rate at the current moment; The load rate is the load rate data retained from the previous sampling period.

[0061] It should be noted that the directional change rate not only reflects the drastic nature of load changes, but its sign also indicates the direction of load change. When the directional change rate is positive, it means that the distribution transformer is under load increase conditions, at which time the windings heat up rapidly and the loss characteristics change quickly; when the directional change rate is negative, it means that the distribution transformer is under load decrease conditions, at which time the loss characteristics change relatively slowly due to the limitation of the heat dissipation rate of the distribution transformer.

[0062] Furthermore, the trace of the prediction residual and covariance matrix of the current identification period is obtained; then, the dynamic direction adjustment coefficient is constructed by looking up the table based on the sign and magnitude of the directional change rate, and the asymmetric amplification coefficient is calculated; then, the exponential independent variable is constructed by continuous multiplication based on the data validity weight, the asymmetric amplification coefficient, and the dimensionless normalization feature composed of the prediction residual.

[0063] The reference uncertainty level is used to set the basic sensitivity of the identification process to parameter fluctuations. Its specific value is determined based on the noise benchmark of the distribution transformer during rated operation. For example, in this invention, the specific value of the reference uncertainty level is preferably 0.05. The directional adjustment coefficient is used to distinguish different physical feedback rates during load increases and decreases. The process of determining the directional adjustment coefficient based on the sign of the directional change rate is as follows: if the directional change rate is greater than or equal to 0, the preset load increase coefficient is used as the directional adjustment coefficient; if the directional change rate is less than 0, the preset load decrease coefficient is used as the directional adjustment coefficient.

[0064] In an optional embodiment, the typical load rate of the distribution transformer stored in the historical data buffer during the calibration period is first obtained, and the typical load rate is divided into a first interval where the load increases continuously and a second interval where the load decreases continuously. Finally, parameter identification iteration is performed in the first interval and the second interval respectively, and parameter optimization is performed with the goal of minimizing the sum of squared predicted residuals in the identification process, so as to obtain the load increase coefficient and the load decrease coefficient.

[0065] Specifically, in order to achieve the optimal setting of the load increase and load decrease factors, an objective function for evaluating the identification effect was constructed, namely, the sum of squared predicted residuals satisfying the following relationship:

[0066]

[0067] in, To predict the sum of squared residuals; The total loss measured at time t; For the direction adjustment coefficient is The predicted total loss is identified at that time; T is the total number of actual data points contained in the first or second interval.

[0068] It should be noted that, in order to accurately reflect the complex nonlinear temperature rise curve inside the transformer, this invention obtains typical load rate rise and fall process data through pre-calibration periodic calibration, and constructs a continuous nonlinear thermodynamic mapping function. Based on the directional change rate... The symbols are dynamically retrieved from the load increase or decrease curve mapping table to achieve accurate fitting of thermal inertia differences in complex industrial scenarios.

[0069] Furthermore, the method for calculating the asymmetric magnification factor is as follows:

[0070]

[0071] in, It is an asymmetric magnification factor; This is the directional adjustment coefficient, which is based on the rate of change of orientation. The sign and amplitude are dynamically mapped from the temperature rise or cooling nonlinear curve function of the transformer.

[0072] After obtaining the asymmetric amplification factor, the exponential independent variable is calculated by combining data quality and the current convergence state of the identification process. This eliminates the use of units like the square root of power in traditional algorithms. or By incorporating the dimensional collapse and scaling distortion caused by the exponential function, and introducing closed-loop residual feedback to ensure the noise-resistant stability of the system in steady state, this invention constructs dimensionless normalized residual characteristics by predicting the trace quotient of the squared residuals and the covariance matrix. The calculation method for the exponential independent variable is as follows:

[0073]

[0074] in, It is the independent variable of the index; Weights for data validity; For reference purposes, the uncertainty benchmark coefficient is set to 0.05 in this invention; This is to normalize the predicted residual characteristics.

[0075] It should be noted that the normalized prediction residual characteristics are calculated as follows:

[0076]

[0077] in The current time prediction residual is calculated in advance based on the parameter estimates of the previous time step, which is the difference between the actual total loss and the predicted total loss. The trace of the covariance matrix in the recursive least squares method reflects the total uncertainty of the current parameter estimation. This is due to the squared prediction residuals. With the trace of the covariance matrix Having the same physical dimension, namely the square of power, the ratio of the two naturally becomes a dimensionless control variable. This rigorous mathematical construction prevents logical breakdowns caused by switching power units, for example... to Meanwhile, when the steady-state model converges and the prediction residual becomes extremely small, the normalized residual feature approaches zero, effectively preventing drastic jumps in the forgetting factor caused by minute noise.

[0078] Finally, the final adaptive forgetting factor is output using a negative exponential decay function and then subjected to amplitude limiting. Therefore, the adaptive forgetting factor is calculated as follows:

[0079]

[0080] in, An adaptive forgetting factor; This is the upper limit of the forgetting factor, which is set to 0.99 in this invention; The lower limit of the forgetting factor is 0.95 in this invention; It is an exponential independent variable.

[0081] It should be noted that when data validity is high, identification uncertainty is significant, and the load is in a phase of rapid increase, the exponential independent variable will increase rapidly, causing the value of the negative exponential decay function to approach 0. This makes the adaptive forgetting factor approach its lower limit of 0.95, meaning that the identification process will accelerate the forgetting of old data, improving the ability to track sudden changes in the current distribution transformer loss. Conversely, under steady-state conditions, the adaptive forgetting factor will rise back to 0.99 to enhance the stability of the identification results. This constructs a highly nonlinear dynamic adjustment mechanism.

[0082] Thus, by introducing an asymmetric tracking mechanism based on the direction of load change and a nonlinear decay model based on data quality, an adaptive balance between identification sensitivity and stability is achieved, effectively solving the problem of inconsistent identification accuracy during load increase and decrease caused by thermal inertia in distribution transformers.

[0083] S6. Embed the adaptive forgetting factor into the covariance update process of the recursive least squares method for online identification, and output the no-load loss and load loss of the distribution transformer.

[0084] In an optional embodiment, after obtaining the adaptive forgetting factor, an improved recursive least squares method is needed for parameter identification in order to achieve dynamic updates of the physical model parameters of the distribution transformer. By introducing the adaptive forgetting factor, which reflects the characteristics of the load conditions, into the iterative logic, the identification process can have differentiated memory lengths at different load stages.

[0085] Specifically, firstly, a regression vector is constructed based on the current load rate, and the predicted total loss is calculated based on the parameter estimates and regression vector from the previous time step. Then, the difference between the actual total loss and the predicted total loss is used as the prediction residual. Next, a gain vector is calculated based on the covariance matrix, regression vector, and adaptive forgetting factor from the previous time step. Then, the parameter estimates are updated based on the gain vector and prediction residual to obtain the no-load loss and load loss at the current time step. Finally, the covariance matrix at the current time step is updated based on the adaptive forgetting factor, gain vector, regression vector, and the covariance matrix from the previous time step. Here, the parameter estimate is a vector containing the physical parameters to be identified; in this invention, it specifically represents the no-load loss and load loss of the distribution transformer. To clarify the specific mapping relationship of the online identification process in the distribution transformer physical loss model, the regression vector consists of a constant 1 and the square of the current load rate.

[0086] Furthermore, the total loss is predicted as follows:

[0087]

[0088] in, To predict total losses; The no-load loss identified at the previous moment; The load loss identified at the previous moment; The load factor is the current value. The above formula characterizes the basic loss physical characteristics of a distribution transformer, namely, the total loss consists of constant no-load loss and load loss that varies with the square of the load factor.

[0089] Next, the difference between the actual total loss and the predicted total loss is calculated to obtain the prediction residual. Then, the gain vector is calculated based on the covariance matrix, regression vector, and adaptive forgetting factor from the previous time step. The calculation method is as follows:

[0090]

[0091] in, It is the gain vector; Let be the covariance matrix of the previous time step; For the regression vector, by constitute; This is the adaptive forgetting factor. When the load undergoes a drastic change in orientation and the data validity is high, the value of the adaptive forgetting factor decreases, resulting in an increase in the magnitude of the gain vector. This means that the identification process will assign higher weight to the current prediction residual, thereby accelerating the convergence of parameters to the true value.

[0092] After obtaining the gain vector, iterative updates of the parameter estimates and covariance matrix are performed. The product of the gain vector and the prediction residual is calculated, and this product is added to the parameter estimates from the previous time step to update the current parameter estimates. The current parameter estimates are then analyzed to output the no-load loss and load loss of the distribution transformer. Simultaneously, the current covariance matrix is ​​updated based on the adaptive forgetting factor, gain vector, regression vector, and the covariance matrix from the previous time step. The current covariance matrix satisfies the following relationship:

[0093]

[0094] in, Let the covariance matrix be the current time step. An adaptive forgetting factor; It is the identity matrix; It is the gain vector; This is the transpose of the regression vector; Let be the covariance matrix of the previous time step.

[0095] It should be noted that the covariance matrix is ​​used to characterize the uncertainty of parameter estimation. By embedding the adaptive forgetting factor into the covariance matrix, when the value of the adaptive forgetting factor decreases, the value of the covariance matrix will be magnified proportionally, thereby mathematically restarting the search capability of the identification process and preventing the identification process from falling into a state of convergence stagnation due to excessive accumulation of previously obtained load rates and total losses.

[0096] Reference Figure 3 A comparison of the proposed solution with the traditional method using a fixed forgetting factor for identifying no-load losses reveals that under low-load conditions, the signal-to-noise ratio drops significantly due to sensor measurement errors, leading to drastic fluctuations in the identification results of traditional methods. The proposed solution, however, introduces data validity, objectively evaluating the measurement reliability under different load levels. Results show that the identification curve corresponding to the proposed solution is smoother and consistently remains within an extremely narrow confidence interval, effectively suppressing interference from inherent hardware noise and significantly improving the accuracy and stability of loss measurement under steady-state conditions.

[0097] Thus, by embedding an adaptive forgetting factor that reflects physical thermal inertia and load rate distribution characteristics into the core iterative link of the recursive least squares method, high dynamic and high-precision online identification of key loss parameters of distribution transformers is achieved. This can keenly capture parameter transients caused by load fluctuations and effectively solve the problems of recognition lag and slow convergence of traditional fixed forgetting factors under complex working conditions.

[0098] This invention also discloses an online measurement system for no-load and load losses of a distribution transformer, including a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement an online measurement method for no-load and load losses of a distribution transformer according to the present invention.

[0099] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0100] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for online measurement of no-load and load losses of a distribution transformer, characterized in that, include: The current load rate, previous load rate, and total loss of the distribution transformer are obtained, and the current load rate is stored in the historical data buffer. The distribution of the current load rate in the historical data buffer is evaluated, and the data distribution scarcity is extracted. The relative measurement accuracy is calculated based on the current load rate. The relative measurement accuracy and the data distribution scarcity are fused to generate a data validity weight that characterizes the reference value of the current operating condition. Based on the current load rate, the prediction residual is obtained. The difference between the current load rate and the previous load rate is calculated to obtain the directional change rate. An asymmetric amplification coefficient representing the thermal inertia difference is constructed based on the sign and amplitude of the directional change rate. The adaptive forgetting factor is calculated by combining the data validity weight, the asymmetric amplification coefficient, and the normalized prediction residual feature representing the uncertainty of the model state in the recursive least squares method. The adaptive forgetting factor is embedded in the covariance update process of the recursive least squares method for online identification, and the no-load loss and load loss of the distribution transformer are output. The process of evaluating the distribution of the current load rate in the historical data buffer and extracting the data distribution scarcity includes: dividing the full range of the load rate into multiple sub-intervals with equal step sizes; assigning the current load rate to the corresponding sub-interval based on the left-closed-right-open principle to obtain the target interval; counting the load rate data points in the historical data buffer that fall into the target interval to obtain the number of data points; using the ratio of the number of data points to the total number of actual data points in the historical data buffer as the interval data density, and then calculating the data distribution scarcity based on the interval data density. Calculating relative measurement accuracy based on the current load rate includes: constructing an accuracy mapping table for the measuring equipment, wherein the accuracy mapping table records the correspondence between different load rate intervals and basic measurement errors; using the current load rate as an index, querying the accuracy mapping table, and extracting the basic measurement error that matches the current load rate; and using the difference between a constant 1 and the basic measurement error as the relative measurement accuracy. The relative measurement accuracy and the data distribution scarcity are fused to generate a data validity weight that characterizes the reference value of the current working condition. This includes: obtaining a preset validity upper limit; calculating the product of the relative measurement accuracy and a numerical proportionality coefficient reflecting the data distribution scarcity; comparing the product with the validity upper limit, and taking the smaller value as the data validity weight. The adaptive forgetting factor is obtained by: acquiring the trace of the prediction residual and covariance matrix at the previous time step; quotienting the square of the prediction residual with the trace of the covariance matrix to generate a dimensionless normalized residual feature; determining the orientation adjustment coefficient based on the sign of the orientation change rate combined with a nonlinear mapping function; calculating the product of the orientation adjustment coefficient and the absolute value of the orientation change rate, and summing it with a constant 1 as the asymmetric amplification coefficient; calculating the product of the data validity weight, the normalized residual feature, and the asymmetric amplification coefficient, and using the ratio of this product to the reference uncertainty level as the exponential independent variable; and constructing a negative exponential decay function with the exponential independent variable to obtain the adaptive forgetting factor.

2. The method for online measurement of no-load and load losses of a distribution transformer according to claim 1, characterized in that, The directional adjustment coefficient is determined based on the sign of the directional change rate and the nonlinear mapping function, including: if the directional change rate is greater than or equal to 0, the load increase coefficient is calculated as the directional adjustment coefficient based on the load increase temperature rise curve mapping function; if the directional change rate is less than 0, the load decrease coefficient is calculated as the directional adjustment coefficient based on the load decrease cooling curve mapping function, wherein the value of the load increase coefficient is greater than the load decrease coefficient. Constructing a negative exponential decay function with an exponential independent variable to calculate an adaptive forgetting factor includes: obtaining a preset upper limit and lower limit of the forgetting factor; calculating the difference between the upper limit and lower limit of the forgetting factor; and adding the product of the difference and the negative exponential term calculated based on the exponential independent variable to the lower limit of the forgetting factor to obtain the adaptive forgetting factor.

3. The method for online measurement of no-load and load losses of a distribution transformer according to claim 1, characterized in that, The adaptive forgetting factor is embedded in the covariance update process of the recursive least squares method for online identification, and the no-load loss and load loss of the distribution transformer are output. This includes: constructing a regression vector consisting of a constant 1 and the square of the load rate at the current moment; extracting the parameter estimate value at the previous moment, calculating the product of the transpose of the regression vector and the parameter estimate value to obtain the predicted total loss; and calculating the difference between the actual obtained total loss and the predicted total loss to obtain the prediction residual.

4. The method for online measurement of no-load and load losses of a distribution transformer according to claim 3, characterized in that, After obtaining the predicted residual, the process further includes: extracting the covariance matrix from the previous time step; calculating the transpose of the regression vector, the continuous product of the covariance matrix and the regression vector to obtain a first intermediate variable; calculating the sum of the adaptive forgetting factor and the first intermediate variable as the gain denominator; calculating the product of the covariance matrix and the regression vector to obtain a second intermediate variable; and using the ratio of the second intermediate variable to the gain denominator as the gain vector.

5. The method for online measurement of no-load and load losses of a distribution transformer according to claim 4, characterized in that, After using the ratio of the second intermediate variable to the gain denominator as the gain vector, the method further includes: calculating the product of the gain vector and the prediction residual; adding the product to the parameter estimate at the previous time step to update the parameter estimate at the current time step; analyzing the parameter estimate at the current time step to output the no-load loss and load loss of the distribution transformer; calculating the continuous product of the gain vector, the transpose of the regression vector, and the covariance matrix to obtain the covariance adjustment term; and subtracting the covariance adjustment term from the covariance matrix and dividing the result by the adaptive forgetting factor to update the covariance matrix at the current time step.

6. An online measurement system for no-load and load losses of a distribution transformer, characterized in that, include: A processor and a memory, wherein the memory stores computer program instructions that, when executed by the processor, implement the online measurement method for no-load and load losses of a distribution transformer according to any one of claims 1-5.

Images