A photovoltaic grid-connected inverter LCL filter network health parameter online evaluation method and system

Abstract

The application discloses a kind of photovoltaic grid-connected inverter LCL filter network health parameter online evaluation method and system, this method is based on dq coordinate system discrete state space model and sampling window operation data, through the evaluation model synchronous output health parameter estimation, identifiable quantification and uncertainty quantification result;According to the identifiable quantification result, the width modulation of basic parameter interval is carried out to construct parameter check interval, and the joint decision of accepting update, rejecting update or degraded output is executed in combination with physical consistency error;When accepting update, write reliable parameter value, when rejecting update or degraded output, keep last reliable parameter value;In the condition of continuous low information, trigger active detection when meeting safety constraint, and reevaluate according to detection data. The application improves the accuracy and engineering usability of LCL filter network health parameter online evaluation under low information condition.

Description

Technical Field

[0001] This invention relates to the field of power electronic equipment condition monitoring and health assessment technology, and in particular to an online assessment method and system for the health parameters of the LCL filter network of a photovoltaic grid-connected inverter. Background Technology

[0002] The photovoltaic grid-connected inverter is a key interface device between the solar power generation system and the power grid. It connects to the grid through an LCL filter network to achieve high-quality power conversion and grid-connected current control. The LCL filter network consists of inverter-side inductors, filter capacitors, and grid-side inductors, along with their equivalent parasitic parameters. During actual operation, the inverter is subjected to the combined effects of electrothermal stress, environmental stress, and high-frequency switching stress over a long period. This causes a slow drift in the electrical parameters of components such as inductors and capacitors in the LCL filter network. This parameter drift directly affects the resonant characteristics, damping capacity, and filtering effect of the filter network, leading to a decrease in grid-connected current quality, resonant frequency shift, and even system instability. Therefore, online assessment of the key health parameters of the LCL filter network is a crucial technical foundation for achieving risk warning, condition-based maintenance, and reliable operation and maintenance of photovoltaic inverter systems.

[0003] Existing methods for online parameter estimation of LCL filter networks can be broadly categorized into three types: analytical model-based, data-driven, and physically constrained. Analytical model-based methods typically rely on recursive least squares, Kalman filters, or state observers. They establish a mathematical model of the inverter system and combine it with real-time sampled voltage and current data to achieve online recursive estimation of the filter inductor and capacitor parameters. Data-driven methods utilize machine learning models such as artificial neural networks and support vector machines. They establish input-output mapping relationships through extensive offline training data, allowing for direct mapping from sampled data to parameter estimates during the online phase. Physically constrained methods introduce the system's physical laws as constraints into the optimization process. They construct an objective function and iteratively solve for the optimal parameters using optimization algorithms such as particle swarm optimization and gradient descent. Furthermore, some schemes pre-screen identifiable parameter combinations through sensitivity analysis or actively inject disturbance signals to improve parameter identification capabilities under specific operating conditions.

[0004] However, all the aforementioned existing technologies have inherent limitations. In the actual operation of grid-connected photovoltaic inverters, the system spends a significant amount of time in steady-state operation, weak disturbance operation, or strong closed-loop suppression conditions—i.e., low-information conditions with insufficient information. In these situations, different parameter combinations may produce approximately identical external electrical responses, meaning that although the algorithm can output parameter estimates, these estimates are not suitable for directly updating the online parameter trajectory. Existing methods use identifiability analysis as an offline analysis tool or a pre-identification screening tool, lacking the ability to directly answer online decisions regarding whether an update should be accepted at the sampling window level. While existing uncertainty quantification methods can characterize the degree of estimation dispersion, they do not incorporate the sufficiency of information in the current sampling window into a unified criterion. Existing active disturbance methods also lack an effective connection between reliable parameter maintenance mechanisms and online update decisions. Therefore, it is necessary to design an online assessment method and system for the health parameters of the LCL filter network of a grid-connected photovoltaic inverter to address these problems. Summary of the Invention

[0005] The technical problem to be solved by this invention is to provide an online evaluation method and system for the health parameters of the LCL filter network of a photovoltaic grid-connected inverter. It aims to overcome the problems of false updates and false alarms, as well as the lack of online update decision and parameter maintenance mechanisms, in the existing online evaluation of the health parameters of the LCL filter network of photovoltaic grid-connected inverters under low information conditions. It features the characteristics of introducing identifiability quantification into the parameter verification interval construction and forming a three-level joint decision with physical consistency error, maintaining reliable parameter values ​​when refusing to update, and restoring updateability through limited active detection under continuous low information conditions.

[0006] To achieve the above technical solution, the technical solution adopted by the present invention is as follows: A method for online evaluation of health parameters of LCL filter network in a photovoltaic grid-connected inverter includes the following steps: S1, Data collection from the grid-connected photovoltaic inverter at the... sampling window The internal operating data, and based on the grid-connected photovoltaic inverter and its LCL filter network, - A discrete state-space model in a synchronously rotating coordinate system is used to uniformly represent the operational data. S2, Sample window The data input is pre-built into the evaluation model, and the estimated health parameters corresponding to the current sampling window are output synchronously. Quantitative results of identifiability for each target health parameter and uncertainty quantification results ; S3, for the first Each target health parameter is estimated based on the health parameter estimation results. Uncertainty quantification results and the first, determined in advance based on the calibration sample set and according to the target coverage. The calibration quantile value corresponding to each target health parameter Constructing the basic parameter range ; And according to the first The quantification results of the identifiability of each target health parameter Using information about Monotonically decreasing interval modulation function The parameter calibration interval is obtained by performing width modulation on the basic parameter range. ; in It is a numerically stable term; S4. Calculate the health parameter estimation results. Substituting the discrete state-space model, a rolling reconstruction is performed on the sampling window, and the physical consistency error is calculated based on the difference between the reconstructed output and the actual measured output. ; S5. Execute joint decision-making based on the identifiability quantification results, the verification interval width of each parameter, and the physical consistency error: when all target health parameters simultaneously meet the requirements... , and When any target health parameter meets the requirements, it is determined to accept the update; or ,or In the case of [specific event], the update is rejected; otherwise, the output is downgraded. , , ; S6. When accepting the update, the current health parameter estimate will be displayed. Write reliable parameter values When performing an update rejection or downgrade output, retain the previous reliable parameter value. Unchanged, and will As the monitoring output benchmark for the current sampling window; S7. When the continuous low information condition determination condition is met, active detection is triggered under the premise of meeting the preset safety constraints. After the active detection ends, a new sampling window is formed based on the collected post-detection data, and S1 to S6 are executed again. Among them, the continuous low information condition determination condition includes at least one of the following: continuous The sampling window did not accept the update; the quantification result of the identifiability of the target health parameter was continuously lower than the preset lower limit; the parameter verification interval was continuously wider than the preset upper limit; the risk indicator continuously indicated a downgraded output or rejection of update; and the data collected during the active detection did not directly overturn the rejection of update or downgraded output conclusion of the current sampling window before the active detection was triggered, but was only used to form the subsequent new sampling window after the active detection ended.

[0007] Preferably, the operating data includes one or more of the following: inverter output voltage command, modulation amount, duty cycle, DC bus voltage, grid voltage, inverter-side inductor current, filter capacitor voltage, and grid-side inductor current; the health parameters include inverter-side inductor current. Inverter-side inductor equivalent series resistance Filter capacitor , filter capacitor equivalent dissipation resistance Grid-side inductance Equivalent series resistance of grid-side inductance One or more of them.

[0008] Preferably, the discrete state-space model is derived from the LCL filter network of the grid-connected photovoltaic inverter. - The continuous model is obtained by discretizing the continuous model in the synchronous rotating coordinate system. The continuous model takes the inverter-side inductor current, filter capacitor voltage and grid-side inductor current as state variables, the inverter output voltage and grid voltage as input variables, and the target health parameter as model parameters. The discretization method is zero-order hold discretization or bilinear transformation discretization.

[0009] Preferably, the evaluation model is an identifiable coupled parameter estimation network, which includes a shared feature encoder, a health parameter estimation branch, an identifiable quantization branch, and an uncertainty quantization branch. The shared feature encoder is used to extract the temporal coupling features between the control quantity and the voltage and current responses within the sampling window, and the health parameter estimation branch outputs the health parameter estimation results. The identifiability quantization branch outputs the identifiability quantization result. The uncertainty quantization branch outputs an uncertainty quantization result that satisfies the positive value constraint. .

[0010] Furthermore, the shared feature encoder adopts a combination structure of one-dimensional convolutional layers and gated recurrent units; the number of output channels of the first three one-dimensional convolutional layers are 32, 64 and 128 respectively, the kernel widths are 5, 5 and 3 respectively, the stride is 1, and each convolutional layer is followed by a normalization layer and a non-linear activation layer; the convolutional output enters a gated recurrent unit with a hidden dimension of 128 to extract time-dependent features.

[0011] Preferably, the training labels for the identifiability quantization branch are generated from parameter sensitivity information, which is based on the Jacobian matrix of the sensitivity of the output to the target health parameters within the sampling window. Fisher Information Matrix Crame-Lower World Alternatively, the equivalent information content can be calculated and converted into a normalized teacher identifiability score using a monotonic mapping function. ;in For regularization terms, To measure the noise covariance matrix.

[0012] Furthermore, the monotonic mapping function is ;in As a scale factor, For numerically stable terms, It is a monotonically increasing function.

[0013] Preferably, the evaluation model is trained using the following total loss function during the offline training phase: ; in, For physical consistency loss, This is the estimation consistency loss constructed based on the health parameter estimation results and the uncertainty quantification results. The constraint loss between the identifiability quantification result and the teacher identifiability score. , , These are the weighting coefficients for each loss term.

[0014] Furthermore, the weighting coefficient is taken as... , , .

[0015] Preferably, for the remaining model parameters not included in the current online evaluation set, one or more of the following methods are used to assign values: rated parameter value, offline identification benchmark value, most recent reliable parameter value, or preset slow update value, in order to maintain the consistency of the discrete state space model rolling reconstruction and joint decision calculation.

[0016] Preferably, the interval modulation function satisfy ,and: ; in For the first The modulation coefficients corresponding to each target health parameter It is a numerically stable term.

[0017] Furthermore, the modulation coefficient According to the The nominal value or percentage of the rated range of each target health parameter is determined so that the interval modulation function is within the identifiability score. When it approaches 0, it tends to ,exist It approaches 1 when it approaches 1.

[0018] Preferably, the physical consistency error Based on the current health parameter estimation results Substituting the discrete state-space model, the weighted error between the output obtained by rolling reconstruction of the sampling window and the actual measured output constitutes the following: ; in To measure the noise covariance matrix, The output is obtained by reconstructing based on the discrete state-space model.

[0019] Preferably, the calibration quantile value The acceptance threshold is based on the inconsistency scores on the calibration sample set and a predetermined quantile value for the target coverage. Rejection threshold Interval width acceptance threshold Interval width rejection threshold Physical consistency error acceptance threshold Physical consistency error rejection threshold Modulation coefficient And at least a portion of the active detection trigger thresholds are determined based on a health sample set, a verification sample set, or historical reliable operating data, through one or more of the following methods: quantile statistics, target coverage constraints, false update rate constraints, false alarm rate constraints, and device safety constraints.

[0020] Furthermore, the inconsistent scores are obtained through... Calculation, where The true value is the target coverage. Take down of quantile value as The acceptance and rejection thresholds are determined by grid search optimization of the false alarm rate and pseudo update rate of the verification sample set.

[0021] Preferably, the reliable parameter value It has an initialization and protection period mechanism, including: the initial reliable parameter values ​​at system startup. The parameters are determined by one or more of the following: factory calibration values, offline identification results during installation and commissioning, the most recent reliable parameter values ​​before equipment restart, or rated parameter values; before the system is powered on or resumes operation. A protection period is set within each sampling window. During the protection period, even if the current sampling window outputs a health parameter estimation result, the result is only used as a candidate update value and is not immediately replaced. Only when a sampling window that meets the update acceptance conditions appears for the first time during the protection period will the corresponding health parameter estimation result be written as a new reliable parameter value.

[0022] Furthermore, the number of sampling windows during the protection period Take an integer between 3 and 10.

[0023] Preferably, the degradation output includes at least one or more of the following information: the previous reliable parameter value. Current sampling window risk indicator, parameter verification range Current identifiability quantification results and current physical consistency error .

[0024] Preferably, the active detection includes One or more of the following: shaft current reference perturbation, reactive power step, duty cycle perturbation, pseudo-random binary excitation, and frequency sweep perturbation.

[0025] Preferably, the active detection satisfies one or more of the following security constraints: ; Furthermore, it meets the requirements for peak grid-connected current, total harmonic distortion, reactive power offset, DC bus voltage fluctuation, upper limit of device thermal stress, and repetitive triggering cooling time. and the maximum number of repetitions At least one of the constraints.

[0026] Furthermore, the aforementioned Axis current reference perturbation amplitude The detection duration is no more than 0.05 pu of the rated grid-connected current. The repeating trigger cooldown time is 1 to 2 fundamental frequency cycles. Not less than 10 fundamental frequency cycles, with an upper limit on the number of repetitions. No more than 3.

[0027] Preferably, an online health parameter assessment system for a photovoltaic grid-connected inverter LCL filter network is provided for executing the online health parameter assessment method for the photovoltaic grid-connected inverter LCL filter network. The system includes: The system includes a data acquisition interface, a processor, a memory, and a probe execution interface. When the processor executes the program instructions stored in the memory, it forms the following cooperative functional units: A window generation unit, whose input terminal is connected to the data acquisition interface, is used to acquire at least one type of operating data, including inverter-side inductor current, filter capacitor voltage, grid-side inductor current, inverter output voltage command, and grid voltage, and to perform sampling according to a preset sampling frequency and window length. The tissue serves as the sampling window. ; An evaluation model inference unit, the input of which is connected to the output of the window generation unit, is used to generate the sampling window. Input a pre-trained offline evaluation model and simultaneously output health parameter estimation results. Quantitative results of identifiability and uncertainty quantification results ; An interval construction unit, whose input is connected to the output of the evaluation model inference unit, is used to construct basic parameter intervals based on the health parameter estimation results and the uncertainty quantification results. And utilizes an interval modulation function that monotonically decreases with respect to the quantization result of identifiability. The basic parameter range is width modulated to output the parameter verification range. ; A physical consistency calculation unit, whose input is connected to the output of the evaluation model inference unit and the window generation unit, is used to substitute the health parameter estimation results into the discrete state-space model to perform rolling reconstruction of the sampling window, and calculate the physical consistency error based on the difference between the reconstructed output and the actual measurement output. ; The joint decision-making unit has its input terminals connected to the output terminals of the evaluation model inference unit, the interval construction unit, and the physical consistency calculation unit, respectively. It is used to perform joint decision-making based on the identifiability quantification result, the parameter verification interval width, and the physical consistency error, and output a judgment result of accepting the update, rejecting the update, or downgrading the output. A reliable value memory unit, the input of which is connected to the output of the joint decision-making unit, is used to write the current health parameter estimation result as a reliable parameter value when the joint decision-making unit determines that the update is accepted. And when it is determined that the update or downgrade output should be rejected, the previous reliable parameter value should be retained. constant; An active detection triggering unit, the input of which is connected to the output of the joint decision-making unit, is used to send an active detection command to the inverter controller through the detection execution interface when the continuous low information condition judgment condition is met, and to trigger the window generation unit to generate a new sampling window after the active detection is completed.

[0028] Preferably, a computer device includes a memory and a processor, which are communicatively connected to each other. The memory stores computer instructions, and the processor executes the computer instructions to perform the online evaluation method for the health parameters of the LCL filter network of the photovoltaic grid-connected inverter.

[0029] Preferably, a computer-readable storage medium stores computer instructions that cause a computer to execute the online evaluation method for health parameters of the LCL filter network of the photovoltaic grid-connected inverter.

[0030] The beneficial effects of this invention are as follows: 1. This invention unifies health parameter estimation, identifiability quantification, and uncertainty quantification into a single evaluation process based on sampling windows. This enables the evaluation model to synchronously output the estimated values ​​of each target health parameter, identifiability score, and uncertainty value. This solves the problems in the prior art where parameter estimation results are separated from online update decisions and it is difficult to determine the reliability of the current estimate under low-information conditions. It achieves refined evaluation for each sampling window, improving the relevance and information utilization of online evaluation.

[0031] 2. This invention constructs a parameter verification interval modulated by identifiability and directly introduces the identifiability score into the width modulation process of the parameter verification interval. This allows the parameter verification interval to simultaneously reflect the degree of estimation dispersion and the information sufficiency of the current sampling window. This solves the defects in the prior art where the representation of parameter estimation uncertainty does not include the information sufficiency dimension and the update risk under low information conditions cannot be explicitly quantified. It realizes the transformation of the identifiability quantification result from a bypass analysis quantity to an online judgment participation quantity, and explicitly maps low information risk into a more conservative parameter judgment boundary.

[0032] 3. This invention employs a joint decision-making mechanism based on parameter verification interval width, physical consistency error, and identifiability score to perform a three-level judgment process of accepting updates, rejecting updates, and downgrading outputs. When rejecting updates or downgrading outputs, it maintains the previous reliable parameter value. Under continuous low-information conditions, it restores the updability of subsequent sampling windows through limited active detection. This solves the problems of existing technologies that rely solely on single-point estimation or single residuals for parameter updates, are prone to false updates and false alarms under low-information conditions, and lack online parameter maintenance and recovery mechanisms. It achieves a balance between the continuity of online monitoring and the robustness of evaluation. Attached Figure Description

[0033] Figure 1 This is the overall flowchart in the embodiments of the present invention; Figure 2 This is a schematic diagram of the diagnostic model of a three-level NPC grid-connected inverter in the dq synchronous rotating coordinate system in an embodiment of the present invention. Figure 3 This is a schematic diagram of the health reference observer and the perturbation subspace projection module in an embodiment of the present invention; Figure 4 This is a schematic diagram illustrating the determination of the excitation capability of the conduction path and the generation of the fault visibility mask in an embodiment of the present invention. Figure 5 This is a comparison chart showing the changes over time of the estimated health parameters, the basic parameter range, and the parameter verification range modulated by identifiability in an embodiment of the present invention. Figure 6 This is a graph showing the changes in the joint decision-making state, reliable parameter retention value, and physical consistency error before and after the parameter drift event in this embodiment of the invention. Figure 7 This is a structural diagram of a computer device in an embodiment of the present invention. Detailed Implementation

[0034] Example 1: A method for online evaluation of health parameters of LCL filter network in a photovoltaic grid-connected inverter includes the following steps: S1, Data collection from the grid-connected photovoltaic inverter at the... sampling window The internal operating data, and based on the grid-connected photovoltaic inverter and its LCL filter network, - A discrete state-space model in a synchronously rotating coordinate system is used to uniformly represent the operational data. S2, Sample window The data input is pre-built into the evaluation model, and the estimated health parameters corresponding to the current sampling window are output synchronously. Quantitative results of identifiability for each target health parameter and uncertainty quantification results ; S3, for the first Each target health parameter is estimated based on the health parameter estimation results. Uncertainty quantification results and the first, determined in advance based on the calibration sample set and according to the target coverage. The calibration quantile value corresponding to each target health parameter Constructing the basic parameter range ; And according to the first The quantification results of the identifiability of each target health parameter Using information about Monotonically decreasing interval modulation function The parameter calibration interval is obtained by performing width modulation on the basic parameter range. ; in It is a numerically stable term; S4. Calculate the health parameter estimation results. Substituting the discrete state-space model, a rolling reconstruction is performed on the sampling window, and the physical consistency error is calculated based on the difference between the reconstructed output and the actual measured output. ; S5. Execute joint decision-making based on the identifiability quantification results, the verification interval width of each parameter, and the physical consistency error: when all target health parameters simultaneously meet the requirements... , and When any target health parameter meets the requirements, it is determined to accept the update; or ,or In the case of [specific event], the update is rejected; otherwise, the output is downgraded. , , ; S6. When accepting the update, the current health parameter estimate will be displayed. Write reliable parameter values When performing an update rejection or downgrade output, retain the previous reliable parameter value. Unchanged, and will As the monitoring output benchmark for the current sampling window; S7. When the continuous low information condition determination condition is met, active detection is triggered under the premise of meeting the preset safety constraints. After the active detection ends, a new sampling window is formed based on the collected post-detection data, and S1 to S6 are executed again. Among them, the continuous low information condition determination condition includes at least one of the following: continuous The sampling window did not accept the update; the quantification result of the identifiability of the target health parameter was continuously lower than the preset lower limit; the parameter verification interval was continuously wider than the preset upper limit; the risk indicator continuously indicated a downgraded output or rejection of update; and the data collected during the active detection did not directly overturn the rejection of update or downgraded output conclusion of the current sampling window before the active detection was triggered, but was only used to form the subsequent new sampling window after the active detection ended.

[0035] Preferably, the operating data includes one or more of the following: inverter output voltage command, modulation amount, duty cycle, DC bus voltage, grid voltage, inverter-side inductor current, filter capacitor voltage, and grid-side inductor current; the health parameters include inverter-side inductor current. Inverter-side inductor equivalent series resistance Filter capacitor , filter capacitor equivalent dissipation resistance Grid-side inductance Equivalent series resistance of grid-side inductance One or more of them.

[0036] Preferably, the discrete state-space model is derived from the LCL filter network of the grid-connected photovoltaic inverter. - The continuous model is obtained by discretizing the continuous model in the synchronous rotating coordinate system. The continuous model takes the inverter-side inductor current, filter capacitor voltage and grid-side inductor current as state variables, the inverter output voltage and grid voltage as input variables, and the target health parameter as model parameters. The discretization method is zero-order hold discretization or bilinear transformation discretization.

[0037] Preferably, the evaluation model is an identifiable coupled parameter estimation network, which includes a shared feature encoder, a health parameter estimation branch, an identifiable quantization branch, and an uncertainty quantization branch. The shared feature encoder is used to extract the temporal coupling features between the control quantity and the voltage and current responses within the sampling window, and the health parameter estimation branch outputs the health parameter estimation results. The identifiability quantization branch outputs the identifiability quantization result. The uncertainty quantization branch outputs an uncertainty quantization result that satisfies the positive value constraint. .

[0038] Furthermore, the shared feature encoder adopts a combination structure of one-dimensional convolutional layers and gated recurrent units; the number of output channels of the first three one-dimensional convolutional layers are 32, 64 and 128 respectively, the kernel widths are 5, 5 and 3 respectively, the stride is 1, and each convolutional layer is followed by a normalization layer and a non-linear activation layer; the convolutional output enters a gated recurrent unit with a hidden dimension of 128 to extract time-dependent features.

[0039] Preferably, the training labels for the identifiability quantization branch are generated from parameter sensitivity information, which is based on the Jacobian matrix of the sensitivity of the output to the target health parameters within the sampling window. Fisher Information Matrix Crame-Lower World Alternatively, the equivalent information content can be calculated and converted into a normalized teacher identifiability score using a monotonic mapping function. ;in For regularization terms, To measure the noise covariance matrix.

[0040] Furthermore, the monotonic mapping function is ;in As a scale factor, For numerically stable terms, It is a monotonically increasing function.

[0041] Preferably, the evaluation model is trained using the following total loss function during the offline training phase: ; in, For physical consistency loss, This is the estimation consistency loss constructed based on the health parameter estimation results and the uncertainty quantification results. The constraint loss between the identifiability quantification result and the teacher identifiability score. , , These are the weighting coefficients for each loss term.

[0042] Furthermore, the weighting coefficient is taken as... , , .

[0043] Preferably, for the remaining model parameters not included in the current online evaluation set, one or more of the following methods are used to assign values: rated parameter value, offline identification benchmark value, most recent reliable parameter value, or preset slow update value, in order to maintain the consistency of the discrete state space model rolling reconstruction and joint decision calculation.

[0044] Preferably, the interval modulation function satisfy ,and: ; in For the first The modulation coefficients corresponding to each target health parameter It is a numerically stable term.

[0045] Furthermore, the modulation coefficient According to the The nominal value or percentage of the rated range of each target health parameter is determined so that the interval modulation function is within the identifiability score. When it approaches 0, it tends to ,exist It approaches 1 when it approaches 1.

[0046] Preferably, the physical consistency error Based on the current health parameter estimation results Substituting the discrete state-space model, the weighted error between the output obtained by rolling reconstruction of the sampling window and the actual measured output constitutes the following: ; in To measure the noise covariance matrix, The output is obtained by reconstructing based on the discrete state-space model.

[0047] Preferably, the calibration quantile value The acceptance threshold is based on the inconsistency scores on the calibration sample set and combined with a pre-determined quantile value for the target coverage. Rejection threshold Interval width acceptance threshold Interval width rejection threshold Physical consistency error acceptance threshold Physical consistency error rejection threshold Modulation coefficient And at least a portion of the active detection trigger thresholds are determined based on a health sample set, a verification sample set, or historical reliable operating data, through one or more of the following methods: quantile statistics, target coverage constraints, false update rate constraints, false alarm rate constraints, and device safety constraints.

[0048] Furthermore, the inconsistent scores are obtained through... Calculation, where The true value is the target coverage. Take down of quantile value as The acceptance and rejection thresholds are determined by grid search optimization of the false alarm rate and pseudo update rate of the verification sample set.

[0049] Preferably, the reliable parameter value It has an initialization and protection period mechanism, including: the initial reliable parameter values ​​at system startup. The parameters are determined by one or more of the following: factory calibration values, offline identification results during installation and commissioning, the most recent reliable parameter values ​​before equipment restart, or rated parameter values; before the system is powered on or resumes operation. A protection period is set within each sampling window. During the protection period, even if the current sampling window outputs a health parameter estimation result, the result is only used as a candidate update value and is not immediately replaced. Only when a sampling window that meets the update acceptance conditions appears for the first time during the protection period will the corresponding health parameter estimation result be written as a new reliable parameter value.

[0050] Furthermore, the number of sampling windows during the protection period Take an integer between 3 and 10.

[0051] Preferably, the degradation output includes at least one or more of the following information: the previous reliable parameter value. Current sampling window risk indicator, parameter verification range Current identifiability quantification results and current physical consistency error .

[0052] Preferably, the active detection includes One or more of the following: shaft current reference perturbation, reactive power step, duty cycle perturbation, pseudo-random binary excitation, and frequency sweep perturbation.

[0053] Preferably, the active detection satisfies one or more of the following security constraints: ; Furthermore, it meets the requirements for peak grid-connected current, total harmonic distortion, reactive power offset, DC bus voltage fluctuation, upper limit of device thermal stress, and repetitive triggering cooling time. and the maximum number of repetitions At least one of the constraints.

[0054] Furthermore, the aforementioned Axis current reference perturbation amplitude The detection duration is no more than 0.05 pu of the rated grid-connected current. The repeating trigger cooldown time is 1 to 2 fundamental frequency cycles. Not less than 10 fundamental frequency cycles, with an upper limit on the number of repetitions. No more than 3.

[0055] Preferably, an online health parameter assessment system for a photovoltaic grid-connected inverter LCL filter network is provided for executing the online health parameter assessment method for the photovoltaic grid-connected inverter LCL filter network. The system includes: The system includes a data acquisition interface, a processor, a memory, and a probe execution interface. When the processor executes the program instructions stored in the memory, it forms the following cooperative functional units: A window generation unit, whose input terminal is connected to the data acquisition interface, is used to acquire at least one type of operating data, including inverter-side inductor current, filter capacitor voltage, grid-side inductor current, inverter output voltage command, and grid voltage, and to perform sampling according to a preset sampling frequency and window length. The tissue serves as the sampling window. ; An evaluation model inference unit, the input of which is connected to the output of the window generation unit, is used to generate the sampling window. Input a pre-trained offline evaluation model and simultaneously output health parameter estimation results. Quantitative results of identifiability and uncertainty quantification results ; An interval construction unit, whose input is connected to the output of the evaluation model inference unit, is used to construct basic parameter intervals based on the health parameter estimation results and the uncertainty quantification results. And utilizes an interval modulation function that monotonically decreases with respect to the quantization result of identifiability. The basic parameter range is width modulated to output the parameter verification range. ; A physical consistency calculation unit, whose input is connected to the output of the evaluation model inference unit and the window generation unit, is used to substitute the health parameter estimation results into the discrete state-space model to perform rolling reconstruction of the sampling window, and calculate the physical consistency error based on the difference between the reconstructed output and the actual measurement output. ; The joint decision-making unit has its input terminals connected to the output terminals of the evaluation model inference unit, the interval construction unit, and the physical consistency calculation unit, respectively. It is used to perform joint decision-making based on the identifiability quantification result, the parameter verification interval width, and the physical consistency error, and output a judgment result of accepting the update, rejecting the update, or downgrading the output. A reliable value memory unit, the input of which is connected to the output of the joint decision-making unit, is used to write the current health parameter estimation result as a reliable parameter value when the joint decision-making unit determines that the update is accepted. And when it is determined that the update or downgrade output should be rejected, the previous reliable parameter value should be retained. constant; An active detection triggering unit, the input of which is connected to the output of the joint decision-making unit, is used to send an active detection command to the inverter controller through the detection execution interface when the continuous low information condition judgment condition is met, and to trigger the window generation unit to generate a new sampling window after the active detection is completed.

[0056] Preferably, a computer device includes a memory and a processor, which are communicatively connected to each other. The memory stores computer instructions, and the processor executes the computer instructions to perform the online evaluation method for the health parameters of the LCL filter network of the photovoltaic grid-connected inverter.

[0057] Preferably, a computer-readable storage medium stores computer instructions that cause a computer to execute the online evaluation method for health parameters of the LCL filter network of the photovoltaic grid-connected inverter.

[0058] Example 2: like Figure 1 As shown, this embodiment uses a grid-connected photovoltaic inverter and its LCL filter network as the technical object. Let the target health parameter vector be: ; in, For the inverter side inductor, The equivalent series resistance of the inverter-side inductance. For filtering capacitors, The equivalent dissipation resistance is connected in parallel with the filter capacitor. For grid-side inductance, It is the equivalent series resistance of the grid-side inductance.

[0059] To reduce the impact of three-phase time-varying coupling, the original abc quantities are mapped to a variable that rotates synchronously with the grid voltage. - Coordinate system. Define the state vector as... ; - In a synchronously rotating coordinate system, the continuous-time model can be written as: ; ; ; in, To synchronize the rotational angular frequency, It is a two-dimensional rotation matrix.

[0060] The above continuous model can be uniformly written as: ; ; in, This indicates an unmodeled perturbation. This represents the measurement noise. Zero-order hold discretization is preferred for obtaining this value. ; ; in, The sampling period.

[0061] The system operates according to a preset sampling frequency and window length. Collect continuously running data to form the first sampling window .

[0062] Preferably, the sampling window duration covers at least one fundamental frequency cycle of the power grid and includes multiple control cycles to balance dynamic information content with real-time online calculation. In one implementable example, the control sampling frequency is selected as follows: Window length The corresponding window covers two The fundamental frequency period; the window step size can be set to 100 sampling points to enable rolling updates. The operating data may include one or more of the following: inverter output voltage command, grid voltage, modulation amount, duty cycle, DC bus voltage, grid-connected current reference value, inverter-side inductor current, filter capacitor voltage, and grid-side inductor current.

[0063] like Figure 2 As shown, for the remaining model parameters not included in the current online evaluation set, one or more of the following methods can be used to assign values: rated parameter value, offline identification benchmark value, most recent reliable parameter value, or preset slow update value, in order to maintain consistency in subsequent rolling reconstruction and joint decision-making.

[0064] Sampling window Input a pre-built evaluation model ,get ; in, For the health parameter estimation results, To quantify the results of identifiability, This is the result of quantifying uncertainty.

[0065] In a preferred embodiment, the evaluation model is implemented using a identifiability-coupled parameter estimation network. This network includes a shared feature encoder, a health parameter estimation branch, an identifiability quantization branch, and an uncertainty quantization branch. The shared feature encoder extracts features from the multi-channel temporal signals within a window, preferably using a combination of one-dimensional convolutional layers and gated recurrent units: the first three one-dimensional convolutional layers have output channels of 32, 64, and 128 respectively, kernel widths of 5, 5, and 3 respectively, and a stride of 1. Each convolutional layer is followed by a normalization layer and a non-linear activation layer. The convolutional output enters a gated recurrent unit with a hidden dimension of 128 for further extraction of time-dependent features. Subsequently, the shared features enter the three output branches: the health parameter estimation branch outputs through two fully connected layers. ; Identifiability quantification branch The function outputs the corresponding values ​​for each target parameter. Uncertainty Quantification Branch The function outputs the positive values ​​corresponding to each target parameter. .

[0066] like Figure 3 As shown, to enhance the model's generalization ability across different operating ranges, the input quantities can be normalized according to rated values, per-unit values, or historical reliable statistics. In one example, voltage signals are normalized to rated voltage or bus voltage per unit, current signals are normalized to rated current per unit, and modulation and duty cycles are normalized to their physical upper limits. These are then combined into a multi-channel input tensor in chronological order.

[0067] Identifiable teacher labels can be constructed based on the output's sensitivity to target health parameters using a Jacobian matrix. For the first... A sampling point, defined as: ; And based on this, the Fisher information matrix is ​​constructed: ; Further construct the Cramer-Loh lower bound: ; Then through the monotonic mapping function Convert it into a teacher identifiability score ; in, For regularization terms, As a scale factor, It is a numerically stable term.

[0068] During the offline training phase, the following total loss function is preferred: ; in, Used to constrain the consistency between the health parameter estimation results and the discrete state-space model; Used to constrain the coupling output between the health parameter estimation branch and the uncertainty quantification branch; This is used to constrain the consistency between the network's output identifiability quantification and the teacher's identifiability score. A feasible training parameter setting is: , , The optimizer adopts The initial learning rate is The batch size is 64, and the number of training rounds is 100. The above values ​​are for illustrative purposes only and can be adjusted appropriately according to the equipment size, sample size, and real-time requirements in actual applications.

[0069] Training samples can consist of one or more of digital simulation samples, bench test samples, and historical operating samples. In a specific example, for a rated power of... A three-phase grid-connected photovoltaic inverter is constructed with the following training range: , and Around the nominal value Internal disturbances , and Around the nominal value Internal disturbances; grid voltage amplitude is to The power factor covers the variation between these values. 0.95 to The training samples are superimposed with steady-state, power ramp-up, and small-disturbance operating conditions. This training sample setup is only for illustrating how to construct an implementable dataset and does not constitute a limitation on the scope of this invention.

[0070] Obtaining only the health parameter estimation results is insufficient to directly update the online parameter trajectory; therefore, this invention further constructs a parameter verification interval. For the... Each target health parameter is estimated based on the health parameter estimation results. Uncertainty quantification results and calibration quantile values Constructing the basic parameter range: ; Wherein, the calibration quantile value The quantile can be obtained by taking the corresponding quantile value from the inconsistency scores on the calibration sample set under the target coverage constraint. For example, in one embodiment, the corresponding The above-mentioned non-consistent scores can be taken as follows: Quantities at the confidence level.

[0071] Subsequently, based on the quantification results of identifiability The basic parameter range is subjected to width modulation. Preferably, the range modulation function is: ; Thus, the parameter verification interval is obtained: ; Based on the above design, the identifiability quantification results It is no longer a bypass analysis quantity separate from the updated decision, but directly participates in interval construction. When there is insufficient effective information within the sampling window, the corresponding parameter verification interval will widen as the identifiability decreases, thus explicitly mapping low information risk to a more conservative parameter verification boundary.

[0072] Physical consistency error is defined as: ; in, To estimate the current health parameters The output is obtained by substituting the sample window into the discrete state-space model and then reconstructing it using a rolling process.

[0073] Let the identifiability acceptance threshold and rejection threshold be respectively... and ,and The interval width acceptance threshold and rejection threshold are respectively and ,and The physical consistency error acceptance threshold and rejection threshold are respectively and ,and The preferred joint decision-making rule is: When all target health parameters are simultaneously satisfied At that time, execute the update acceptance; When any target health parameter satisfies or At that time, the update will be rejected. In all other cases, a downgraded output is performed.

[0074] Define the reliability parameter value as follows: When accepting the update, let When executing a rejection update or degradation output, make ,in, This represents the health parameter value written when the update acceptance conditions were most recently met. During degradation output, it is preferable to output the previous reliable parameter value to the upper-level system. Current risk indicator, parameter verification range Quantitative results of identifiability and physical consistency error At least one of them.

[0075] To enhance the robustness of the system during startup, initial reliability parameter values ​​are set. The preferred methods are to use the equipment's factory calibration values, offline identification results during the installation and commissioning phase, the most recent reliable parameter values ​​before the equipment restarts, or the rated parameter values. (The last part, "after system startup," appears to be an unrelated fragment and is omitted from the translation.) Each sampling window has a protection period. During this period, the health parameter estimates output by the current sampling window are recorded only as candidate update values ​​and are not immediately replaced. The corresponding estimation result is written as a new reliable parameter value only when a sampling window that meets the acceptance update criteria appears for the first time during the protection period. In one example, it can be... Set to 5 windows.

[0076] Threshold parameters can be determined using a healthy sample set, a calibration sample set, a validation sample set, or historical reliable operational data. In one example, the threshold parameter can be... The value is set to 0.65. Set it to 0.35; and Take the corresponding nominal values ​​of the parameters respectively. and ; and Take the physical consistency error statistics of the verification sample set respectively. quantiles and Quantile values. The above values ​​are preferred examples for illustrative purposes only, and can be recalibrated based on false alarm rate, false update rate, and device security requirements during actual deployment.

[0077] When multiple consecutive sampling windows fail to meet the conditions for accepting updates, the system determines that it is currently in a continuous low-information condition. This condition can be determined by at least one or more of the following conditions: (1) Continuous None of the sampling windows performed the update acceptance action; (2) Quantitative results of the identifiability of at least one target health parameter Continuously below the preset lower limit; (3) The width of the parameter verification interval for at least one target health parameter Continuously wider than the preset upper limit; (4) Maintain the previous reliable parameter value At the same time, the risk indicator continuously indicates a downgraded output or a refusal to update.

[0078] When any of the above conditions are met, and the current grid-connected power margin, reactive power regulation margin, power quality constraints, and equipment thermal stress constraints allow, the system enters active detection mode. Active detection can employ one or more of the following methods: (1) Axis current reference perturbation; (2) Reactive power step; (3) Duty cycle perturbation; (4) Pseudo-random binary excitation; (5) Frequency sweeping perturbation.

[0079] Active detection should meet one or more of the following security constraints: ; In addition, it can further meet the requirements of grid-connected current peak, total harmonic distortion, reactive power offset, DC bus voltage fluctuation, upper limit of device thermal stress, and repetitive trigger cooling time. and the maximum number of repetitions Equal constraints. In an implementable example, The current can be no greater than the rated grid-connected current. , One to two fundamental frequency cycles can be selected to repeatedly trigger the cooldown time. Not less than 10 fundamental frequency cycles, with an upper limit on the number of repetitions. No more than 3.

[0080] like Figure 4 As shown, in this invention, the data collected during active probing is not directly used to overturn the rejection or downgrade output conclusion of the current sampling window before the active probing was triggered. Instead, it forms one or more subsequent new sampling windows after the active probing ends. , The system then re-executes health parameter estimation, identifiability quantification, uncertainty quantification, parameter verification interval construction, and joint decision-making based on the subsequent new sampling window. If the subsequent new sampling window after the initial active probe still does not meet the conditions for accepting updates, active probes are allowed to be executed again, provided that the cooldown time, the maximum number of repetitions, and the safety boundary are met. When the maximum number of repetitions is reached, the persistent low-information state is lifted, or the system enters a grid-connected operating condition where probes are not allowed, repeated probes are paused, and the previous reliable parameter value is maintained. .

[0081] The following is a specific example of an implementable method to illustrate the closed-loop online evaluation process of the present invention.

[0082] A unit with a rated power of The three-phase grid-connected photovoltaic inverter uses an LCL filter network for grid connection, and the nominal parameters can be: , , , , , Control the sampling frequency. Window length is taken The rolling step size is set to 100. The system collects the inverter output voltage command, grid voltage, grid-connected current reference value, inverter-side inductor current, filter capacitor voltage, and grid-side inductor current in real time, and... - Complete window construction in the coordinate system.

[0083] In the offline phase, the evaluation model is jointly trained using digital simulation samples and historical bench samples. In the online phase, the... sampling window After inputting the evaluation model, estimates of six target health parameters are obtained. Identifiability score and uncertainty quantification results Then, according to the pre-defined... Constructing the basic parameter range and combined with identifiability score Calculate the parameter verification interval Next, using the currently estimated parameters... Substituting the discrete state-space model, the output within the window is reconstructed using a scrolling method, and the physical consistency error is calculated. If all target health parameters meet the acceptance update criteria, then the estimated values ​​will be updated. Write reliable parameter values If an update rejection condition is encountered, the previous reliable parameter value will be retained. If the result remains unchanged, and the result falls between the two, then a downgrade result and a risk indicator will be output.

[0084] If the system fails to accept updates for three consecutive sampling windows, and the current grid-connected power margin and power quality constraints allow, the trigger amplitude will not exceed [a certain value]. of The axis current reference perturbation lasts for one fundamental frequency cycle. After active detection ends, the detected data is reorganized into a new sampling window, and joint decision-making is performed again. This method improves the identifiability of subsequent windows without violating grid connection safety boundaries, thereby restoring the ability to update parameters online.

[0085] Example 3: This embodiment uses a 50kW three-phase grid-connected photovoltaic inverter as the application object. The inverter is connected to the grid via an LCL filter network, and the online health parameter evaluation system of this invention is deployed on the photovoltaic power station monitoring server. The experimental platform uses an RT-LAB real-time simulator to simulate the physical behavior of the inverter main circuit and the LCL filter network. The sampled data is transmitted to the industrial control computer where the evaluation algorithm is located via TCP / IP, and the algorithm output results and decision status are recorded in real time to the local database.

[0086] The nominal parameters of the LCL filter network are set as follows: inverter-side inductance. The equivalent series resistance of the inverter-side inductor is 1.8mH. The filter capacitor has a resistance of 0.12Ω. The equivalent dissipation resistance of the filter capacitor in parallel is 18μF. The inductance on the grid side is 15kΩ. The equivalent series resistance of the grid-side inductor is 1.2mH. The impedance is 0.08Ω. The sampling frequency is set to 10kHz, and the sampling window length is... The value is 400, the rolling step size is 100 sampling points, and 2000 sampling windows are continuously collected. At sampling window 700, the simulation model is injected... The parameter abrupt change event, from a step drift of 1.8mH to 1.71mH, simulates a fault scenario of inductor aging or loose connection. The measurement noise is zero-mean Gaussian white noise, and the signal-to-noise ratio is set to 45dB.

[0087] The evaluation model employs a identifiability-coupled parameter estimation network. The network structure consists of three one-dimensional convolutional layers with 32, 64, and 128 output channels respectively, followed by gated recurrent units with a hidden dimension of 128. Key parameter values: identifiability acceptance threshold. The rejection threshold is 0.65. The value is 0.35; the acceptable threshold for interval width is... 5% of the nominal value, rejection threshold 15% of the nominal value; Physical consistency error acceptance threshold The 50th percentile of the validation sample set is 0.08, which is the rejection threshold. Take the 90th percentile value of 0.20; modulation coefficient The weight of the loss function is 0.5. For 1.0, 0.5 The value is 0.3; the number of protection windows. The value is 5; among the active detection parameters It is 0.05 pu. It takes 20ms. It takes 200ms. The value is 3.

[0088] The comparative test uses two existing methods under the same conditions. Method 1 is an online parameter estimation method based on recursive least squares. This method refreshes the online parameter trajectory with the latest estimate in each sampling window, but it lacks identifiability judgment, joint decision-making, and a reliable parameter preservation mechanism. Method 2 is a parameter estimation method based on Kalman filtering. This method outputs the parameter estimation result but lacks uncertainty quantization and identifiability quantization as accompanying outputs. Both comparative methods use the same sampling window data as this method as input, as shown in Tables 1 and 2 below.

[0089] Table 1: Online evaluation performance of the method of the present invention under different working conditions;

[0090] Table 2: Comparison of false update rate and evaluation accuracy between the method of the present invention and the comparative method;

[0091] As can be seen from the data in Table 1, the method of this invention exhibits differentiated evaluation and decision-making characteristics under different operating conditions. During the high-information dynamic operating period from window 150 to 300, the identifiability score... The parameter verification interval width is only 3.8% to 4.1% of the nominal value, reaching 0.78 to 0.82, and the physical consistency error is... Around 0.04, the update acceptance condition was met simultaneously, and the reliable parameter value was accurately updated to the estimated result consistent with the true value. During the steady-state low-information period from window 350 to 600, the identifiability score dropped to 0.22 to 0.31, and the interval width increased to 14.2% to 18.9%. The system automatically determined a downgraded output, and the reliable parameter value remained unchanged at 1.801mH, avoiding false updates caused by low information. During the parameter drift period from window 680 to 710, the physical consistency error surged to 0.225, exceeding the rejection threshold of 0.20. The system immediately determined a rejection of the update, and the reliable parameter value remained at 1.801mH, effectively preventing transient abnormal data from being written into the online parameter trajectory. During the detection and recovery period after window 800... The value rose to 0.56 to 0.71, the interval width narrowed to 4.5% to 7.2%, the system re-determined to accept the update, and the reliable parameter value was correctly tracked to 1.711mH.

[0092] As shown in Table 2, the pseudo-update rate of the method in this invention is 3.2% for the full window and only 4.8% for the low-information window. In contrast, the corresponding rates for recursive least squares are 28.6% and 41.5%, respectively, and for Kalman filtering, they are 21.4% and 33.7%. This method demonstrates a significant effect in suppressing pseudo-updates under low-information conditions. Regarding accuracy, this method… The mean deviation of reliable values ​​is only 0.42%, far lower than the 1.93% and 1.52% of recursive least squares and Kalman filtering, respectively. This is due to the reliable parameter preservation mechanism, which eliminates the interference of unreliable estimates during low-information and abnormal transient phases. Furthermore, this method is the only one with active detection triggering and degradation output risk labeling functions, achieving a degradation output risk label matching rate of 91.2%, providing clear and actionable information for the upper-level operation and maintenance system. The above data collectively demonstrate that this method significantly outperforms existing technologies in terms of parameter evaluation accuracy, false update suppression, and engineering usability.

[0093] Figure 5 This paper demonstrates the modulation effect of the parameter verification interval proposed in this invention compared to the traditional basic parameter interval under low-information conditions. The gray band in the figure represents the traditional method's quantization result based solely on uncertainty. Constructing the basic parameter range Its width depends only on the estimated discreteness, and varies only slightly between sampling windows with different levels of information sufficiency. The blue band represents the identifiable quantization result of this invention. The parameter verification interval obtained after modulation The bottom sub-image simultaneously provides a recognition score. The curve showing the change over time. From the steady-state low-information period between window 400 and 550, it can be seen that when... When the value drops below 0.35, the width of the parameter verification interval significantly increases, while the width of the basic parameter interval remains unchanged. This indicates that the present invention explicitly transforms the update risk under low-information conditions into a more conservative parameter judgment boundary, effectively avoiding false updates caused by fluctuations in estimated values ​​during low-information periods. During the dynamic high-information period within a window of 300 to 350, Above 0.65, the parameter verification interval nearly matches the basic parameter interval, indicating that the modulation effect automatically weakens under high information conditions without sacrificing parameter tracking sensitivity. This figure visually demonstrates the technical effect of identifiable participation in interval modulation, a feature not found in existing parameter identification methods.

[0094] Figure 6 This demonstrates the operational effectiveness of the three-level joint decision-making mechanism and reliable parameter preservation strategy of this invention under parameter drift events. The upper subplot of the figure shows the parameters. Real-time estimates Reliable parameter values ​​of joint decision control The middle sub-figure indicates the decision state of each sampling window: green squares indicate accepting the update, yellow squares indicate downgrading the output, and red squares indicate rejecting the update. The lower sub-figure shows the physical consistency error. and its acceptance threshold and rejection threshold The simulation results show that at the instant when the true parameter value undergoes a step drift near window 700, the physical consistency error... Rising sharply to The joint decision-making process immediately determines that the update should be rejected, and the parameter value is reliable. By keeping the reliable values ​​from before the drift unchanged, unreliable estimates during transient disturbances are avoided from being written into the online parameter trajectory. During the subsequent transition period from window 710 to 760, Falling back to and If the identifiability has not been fully restored, the joint decision-making process results in a downgraded output, while the reliable parameter values ​​remain unchanged. Only after the window reaches 800 seconds and all indicators simultaneously meet the acceptance criteria does the decision switch to accepting updates, and the reliable parameter values ​​begin to track the true values ​​again. This figure fully demonstrates the superiority of the joint decision-making mechanism of this invention compared to existing technologies that rely solely on point estimation for parameter updates.

[0095] Example 4: like Figure 7As shown, this embodiment of the invention also provides a computer device, which includes one or more processors 10, a memory 20, and interfaces for connecting the various components, including high-speed interfaces and low-speed interfaces. The various components are interconnected via different buses and can be mounted on a common motherboard or otherwise installed as needed. The processors can process instructions executed within the computer device, including instructions stored in or on memory for displaying graphical information of a GUI on external input / output devices, such as display devices coupled to the interfaces. In some alternative embodiments, multiple processors and / or multiple buses can be used with multiple memories and multiple memory modules, if desired. Similarly, multiple computer devices can be connected, each providing some of the necessary operations, for example, as a server array, a group of blade servers, or a multiprocessor system.

[0096] Processor 10 may be a central processing unit, a network processor, or a combination thereof. Processor 10 may further include a hardware chip. The hardware chip may be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The programmable logic device may be a complex programmable logic device (CAMP), a field-programmable gate array (FPGA), a general-purpose array logic (GDA), or any combination thereof.

[0097] The memory 20 stores instructions executable by at least one processor 10 to cause the at least one processor 10 to perform the method shown in the above embodiments.

[0098] The memory 20 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the computer device. Furthermore, the memory 20 may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some alternative embodiments, the memory 20 may optionally include memory remotely located relative to the processor 10, and these remote memories may be connected to the computer device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0099] The memory 20 may include volatile memory, such as random access memory; the memory may also include non-volatile memory, such as flash memory, hard disk or solid-state drive; the memory 20 may also include a combination of the above types of memory.

[0100] The computer device also includes a communication interface 30 for communicating with other devices or communication networks.

[0101] This invention also provides a computer-readable storage medium. The methods described above according to embodiments of the invention can be implemented in hardware or firmware, or implemented as computer code that can be recorded on a storage medium, or implemented as computer code downloaded via a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and then stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code, which, when accessed and executed by the computer, processor, or hardware, implements the methods shown in the above embodiments.

[0102] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A method for online evaluation of health parameters of LCL filter network in a photovoltaic grid-connected inverter, characterized in that, Includes the following steps: S1, Data collection from the grid-connected photovoltaic inverter at the... sampling window The internal operating data, and based on the grid-connected photovoltaic inverter and its LCL filter network, - A discrete state-space model in a synchronously rotating coordinate system is used to uniformly represent the operational data. S2, Sampling window The data input is pre-built into the evaluation model, and the estimated health parameters corresponding to the current sampling window are output synchronously. Quantitative results of identifiability for each target health parameter and uncertainty quantification results ; S3, for the first Each target health parameter is estimated based on the health parameter estimation results. Uncertainty quantification results and the first, determined in advance based on the calibration sample set and according to the target coverage. The calibration quantile value corresponding to each target health parameter Constructing the basic parameter range ； And according to the first The quantification results of the identifiability of each target health parameter Using information about Monotonically decreasing interval modulation function The parameter calibration interval is obtained by performing width modulation on the basic parameter range. ； in It is a numerically stable term; S4. Calculate the health parameter estimation results. Substituting the discrete state-space model, a rolling reconstruction is performed on the sampling window, and the physical consistency error is calculated based on the difference between the reconstructed output and the actual measured output. ; S5. Execute joint decision-making based on the identifiability quantification results, the verification interval width of each parameter, and the physical consistency error: when all target health parameters simultaneously meet the requirements... , and When any target health parameter meets the requirements, it is determined to accept the update; or ,or In the case of [specific event], the update is rejected; otherwise, the output is downgraded. , , ; S6. When accepting the update, the current health parameter estimate will be displayed. Write reliable parameter values When performing an update rejection or downgrade output, retain the previous reliable parameter value. Unchanged, and will This serves as the monitoring output benchmark for the current sampling window; S7. When the continuous low information condition determination condition is met, active detection is triggered under the premise of meeting the preset safety constraints. After the active detection ends, a new sampling window is formed based on the collected post-detection data, and S1 to S6 are executed again. Among them, the continuous low information condition determination condition includes at least one of the following: continuous The sampling window did not accept the update; the quantification result of the identifiability of the target health parameter was continuously lower than the preset lower limit; the parameter verification interval was continuously wider than the preset upper limit; the risk indicator continuously indicated a downgraded output or rejection of update; and the data collected during the active detection did not directly overturn the rejection of update or downgraded output conclusion of the current sampling window before the active detection was triggered, but was only used to form the subsequent new sampling window after the active detection ended.

2. The method for online evaluation of health parameters of LCL filter network in a photovoltaic grid-connected inverter according to claim 1, characterized in that, The operating data includes one or more of the following: inverter output voltage command, modulation amount, duty cycle, DC bus voltage, grid voltage, inverter-side inductor current, filter capacitor voltage, and grid-side inductor current; the health parameters include inverter-side inductor current. Inverter-side inductor equivalent series resistance Filter capacitor , filter capacitor equivalent dissipation resistance Grid-side inductance Equivalent series resistance of grid-side inductance One or more of them.

3. The method for online evaluation of health parameters of LCL filter network in a photovoltaic grid-connected inverter according to claim 1, characterized in that, The discrete state-space model is derived from the LCL filter network of the grid-connected photovoltaic inverter. - The continuous model is obtained by discretizing the continuous model in the synchronous rotating coordinate system. The continuous model takes the inverter-side inductor current, filter capacitor voltage and grid-side inductor current as state variables, the inverter output voltage and grid voltage as input variables, and the target health parameter as model parameters. The discretization method is zero-order hold discretization or bilinear transformation discretization.

4. The method for online evaluation of health parameters of LCL filter network in a photovoltaic grid-connected inverter according to claim 1, characterized in that, The evaluation model is an identifiability coupled parameter estimation network, which includes a shared feature encoder, a health parameter estimation branch, an identifiability quantization branch, and an uncertainty quantization branch. The shared feature encoder is used to extract the timing coupling features between the control quantity and the voltage and current response within the sampling window, and the health parameter estimation branch outputs the health parameter estimation result. The identifiability quantization branch outputs the identifiability quantization result. The uncertainty quantization branch outputs an uncertainty quantization result that satisfies the positive value constraint. . Furthermore, the shared feature encoder adopts a combination structure of one-dimensional convolutional layers and gated recurrent units; the number of output channels of the first three one-dimensional convolutional layers are 32, 64 and 128 respectively, the kernel widths are 5, 5 and 3 respectively, the stride is 1, and each convolutional layer is followed by a normalization layer and a non-linear activation layer; the convolutional output enters a gated recurrent unit with a hidden dimension of 128 to extract time-dependent features.

5. The method for online evaluation of health parameters of LCL filter network in a photovoltaic grid-connected inverter according to claim 4, characterized in that, The training labels for the identifiability quantization branch are generated from parameter sensitivity information, which is based on the Jacobian matrix of the sensitivity of the output to the target health parameters within the sampling window. Fisher Information Matrix Crame-Lower World Alternatively, the equivalent information content can be calculated and converted into a normalized teacher identifiability score using a monotonic mapping function. ;in For regularization terms, To measure the noise covariance matrix. Furthermore, the monotonic mapping function is ;in As a scale factor, For numerically stable terms, It is a monotonically increasing function.

6. The method for online evaluation of health parameters of LCL filter network in a photovoltaic grid-connected inverter according to claim 4, characterized in that, The evaluation model is trained using the following total loss function during the offline training phase: ； in, For physical consistency loss, This is the estimation consistency loss constructed based on the health parameter estimation results and the uncertainty quantification results. The constraint loss between the identifiability quantification result and the teacher identifiability score. , , These are the weighting coefficients for each loss term. Furthermore, the weighting coefficient is taken as... , , .

7. The method for online evaluation of health parameters of LCL filter network in a photovoltaic grid-connected inverter according to claim 1, characterized in that, For the remaining model parameters not included in the current online evaluation set, one or more of the following methods are used to assign values: rated parameter value, offline identification benchmark value, most recent reliable parameter value, or preset slow update value, in order to maintain the consistency of the discrete state space model rolling reconstruction and joint decision calculation.

8. The method for online evaluation of health parameters of LCL filter network in a photovoltaic grid-connected inverter according to claim 1, characterized in that, The interval modulation function satisfy ,and: ； in For the first The modulation coefficients corresponding to each target health parameter It is a numerically stable term. Furthermore, the modulation coefficient According to the The nominal value or percentage of the rated range of each target health parameter is determined so that the interval modulation function is within the identifiability score. When it approaches 0, it tends to ,exist It approaches 1 when it approaches 1.

9. The method for online evaluation of health parameters of LCL filter network in a photovoltaic grid-connected inverter according to claim 1, characterized in that, The physical consistency error Based on the current health parameter estimation results Substituting the discrete state-space model, the weighted error between the output obtained by rolling reconstruction of the sampling window and the actual measured output constitutes the following: ； in To measure the noise covariance matrix, The output is obtained by reconstructing based on the discrete state-space model.

10. The method for online evaluation of health parameters of LCL filter network in a photovoltaic grid-connected inverter according to claim 1, characterized in that, The calibration quantile The acceptance threshold is based on the inconsistency scores on the calibration sample set and combined with a pre-determined quantile value for the target coverage. Rejection threshold Interval width acceptance threshold Interval width rejection threshold Physical consistency error acceptance threshold Physical consistency error rejection threshold Modulation coefficient And at least a portion of the active detection trigger thresholds are determined based on a health sample set, a verification sample set, or historical reliable operating data, through one or more of the following methods: quantile statistics, target coverage constraints, false update rate constraints, false alarm rate constraints, and device safety constraints. Furthermore, the inconsistent scores are obtained through... Calculation, where The true value is the target coverage. Take down of quantile value as The acceptance and rejection thresholds are determined by grid search optimization of the false alarm rate and pseudo update rate of the verification sample set.

11. The method for online evaluation of health parameters of LCL filter network in a photovoltaic grid-connected inverter according to claim 1, characterized in that, The reliable parameter value It has an initialization and protection period mechanism, including: the initial reliable parameter values ​​at system startup. The parameters are determined by one or more of the following: factory calibration values, offline identification results during installation and commissioning, the most recent reliable parameter values ​​before equipment restart, or rated parameter values; before the system is powered on or resumes operation. A protection period is set within each sampling window. During the protection period, even if the current sampling window outputs a health parameter estimation result, the result is only used as a candidate update value and is not immediately replaced. Only when a sampling window that meets the update acceptance conditions appears for the first time during the protection period will the corresponding health parameter estimation result be written as a new reliable parameter value. Furthermore, the number of sampling windows during the protection period Take an integer between 3 and 10.

12. The method for online evaluation of health parameters of LCL filter network in a photovoltaic grid-connected inverter according to claim 1, characterized in that, The degradation output includes at least one or more of the following information: the previous reliable parameter value. Current sampling window risk indicator, parameter verification range Current identifiability quantification results and current physical consistency error .

13. The method for online evaluation of health parameters of LCL filter network in a photovoltaic grid-connected inverter according to claim 1, characterized in that, The active detection includes One or more of the following: shaft current reference perturbation, reactive power step, duty cycle perturbation, pseudo-random binary excitation, and frequency sweep perturbation.

14. The method for online evaluation of health parameters of LCL filter network in a photovoltaic grid-connected inverter according to claim 13, characterized in that, The active detection satisfies one or more of the following security constraints: ； Furthermore, it meets the requirements for peak grid-connected current, total harmonic distortion, reactive power offset, DC bus voltage fluctuation, upper limit of device thermal stress, and repetitive triggering cooling time. and the maximum number of repetitions At least one of the constraints. Furthermore, the aforementioned Axis current reference perturbation amplitude The detection duration is no more than 0.05 pu of the rated grid-connected current. The repeating trigger cooldown time is 1 to 2 fundamental frequency cycles. Not less than 10 fundamental frequency cycles, with an upper limit on the number of repetitions. No more than 3.

15. An online health parameter evaluation system for a photovoltaic grid-connected inverter LCL filter network, used to execute the online health parameter evaluation method for a photovoltaic grid-connected inverter LCL filter network according to any one of claims 1-14, characterized in that, The system includes: The system includes a data acquisition interface, a processor, a memory, and a probe execution interface. When the processor executes the program instructions stored in the memory, it forms the following cooperative functional units: A window generation unit, whose input terminal is connected to the data acquisition interface, is used to acquire at least one type of operating data, including inverter-side inductor current, filter capacitor voltage, grid-side inductor current, inverter output voltage command, and grid voltage, and to perform sampling according to a preset sampling frequency and window length. The tissue serves as the sampling window. ; An evaluation model inference unit, the input of which is connected to the output of the window generation unit, is used to generate the sampling window. Input a pre-trained offline evaluation model and simultaneously output health parameter estimation results. Quantitative results of identifiability and uncertainty quantification results ; An interval construction unit, whose input is connected to the output of the evaluation model inference unit, is used to construct basic parameter intervals based on the health parameter estimation results and the uncertainty quantification results. And utilizes an interval modulation function that monotonically decreases with respect to the quantization result of identifiability. The basic parameter range is width modulated to output the parameter verification range. ; A physical consistency calculation unit, whose input is connected to the output of the evaluation model inference unit and the window generation unit, is used to substitute the health parameter estimation results into the discrete state-space model to perform rolling reconstruction of the sampling window, and calculate the physical consistency error based on the difference between the reconstructed output and the actual measurement output. ; The joint decision-making unit has its input terminals connected to the output terminals of the evaluation model inference unit, the interval construction unit, and the physical consistency calculation unit, respectively. It is used to perform joint decision-making based on the identifiability quantification result, the parameter verification interval width, and the physical consistency error, and output a judgment result of accepting the update, rejecting the update, or downgrading the output. A reliable value memory unit, the input of which is connected to the output of the joint decision-making unit, is used to write the current health parameter estimation result as a reliable parameter value when the joint decision-making unit determines that the update is accepted. And when it is determined that the update or downgrade output should be rejected, the previous reliable parameter value should be retained. constant; An active detection triggering unit, the input of which is connected to the output of the joint decision-making unit, is used to send an active detection command to the inverter controller through the detection execution interface when the continuous low information condition judgment condition is met, and to trigger the window generation unit to generate a new sampling window after the active detection is completed.

16. A computer device, characterized in that, It includes a memory and a processor, which are interconnected and communicate with each other. The memory stores computer instructions, and the processor executes the computer instructions to perform the online evaluation method for health parameters of the LCL filter network of the photovoltaic grid-connected inverter as described in any one of claims 1 to 14.

17. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to execute the online evaluation method for health parameters of the LCL filter network of the photovoltaic grid-connected inverter as described in any one of claims 1 to 14.

Images