Photovoltaic inverter grid-connected test data automatic ladder lookup and processing method

Abstract

This invention discloses an automated method for finding and processing stepped test data of photovoltaic inverters, belonging to the field of photovoltaic power generation system testing technology. The method includes: collecting and preprocessing time-series data using a grid simulator and power analyzer; inputting the preprocessed data into a pre-trained convolutional neural network model to automatically output the coordinates of the start and end points of the stepped changes; acquiring the trigger signal from the grid simulator and aligning it with multi-channel data through time compensation; extracting the steady-state interval of each stepped platform based on the aligned data and the stepped point coordinates, and calculating the platform value, response time, and overshoot; finally, filling the results into a standard test report template to automatically generate a test report. This invention solves the problems of low efficiency, strong subjectivity, and susceptibility to errors caused by reliance on manual analysis in existing technologies. Through fully automated data processing and machine learning recognition, it significantly improves the accuracy, consistency, and efficiency of testing.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power generation system testing technology, and in particular to an automated ladder search and processing method for photovoltaic inverter grid-connected test data. Background Technology

[0002] As the core equipment of a photovoltaic (PV) power generation system, the performance of the PV inverter directly affects the safe and stable operation of the power grid. To evaluate the response characteristics of PV inverters under abnormal grid conditions, industry standards specify a series of test items, among which the step test is a key one. The step test applies sudden voltage or frequency changes to the PV inverter using a grid simulator to simulate instantaneous disturbances in the grid, and records the inverter's response waveform using a power analyzer. Testers need to accurately identify the start and end times of each step change from these response waveforms, and calculate the inverter's steady-state parameters (such as voltage, current, and power plateau values) and dynamic performance indicators (such as response time and overshoot) accordingly, ultimately generating a test report that conforms to the specifications.

[0003] Currently, such tests and analyses rely heavily on manual operation. The testing process typically involves the following steps: First, an operator controls a power grid simulator to generate a step disturbance, simultaneously starting a power analyzer to record data. Then, the collected time-series data is exported to general data processing software (such as MATLAB or Excel). Finally, experienced engineers manually mark the start and end points of the step changes by visually observing the waveforms, and select seemingly stable intervals for parameter calculations. This traditional manual analysis method reveals numerous problems in practical applications.

[0004] First, manual identification of step change points is inherently subjective. Different engineers may have different judgments about the start and end times of a step in the same waveform, and even the same engineer may make inaccurate judgments about the same set of data at different times. This subjectivity directly leads to poor consistency and repeatability of test results, making it difficult to meet the high requirements of accuracy and objectivity in standardized testing. Second, the entire process is extremely inefficient. Testing photovoltaic inverters often requires completing dozens or even hundreds of step test points, each involving the processing of massive amounts of high-sampling-rate data. Manual analysis is time-consuming and severely restricts the product development and certification testing cycle. Third, manual operation is highly prone to introducing errors. When searching for feature points in complex waveforms that may contain noise, the human eye is prone to fatigue, which may cause small changes to be missed or noise to be misjudged as a valid signal, especially under conditions of low signal-to-noise ratio, making it difficult to guarantee accuracy.

[0005] While some automated testing software exists on the market, their core algorithms often rely on simple fixed thresholds or slope judgments to detect step changes. These methods often lack robustness when faced with complex situations common in real-world testing, such as high-frequency noise introduced by inverter switching frequency, signal oscillations, or non-ideal step edges, easily leading to false positives or false negatives. Furthermore, inherent signal transmission and processing delays exist between the grid simulator and power analyzer in the testing system. Without precise time synchronization compensation, this directly results in systematic errors in subsequent calculations of dynamic parameters such as response time, affecting the accuracy of the test results.

[0006] Therefore, there is an urgent need in this field for an automated solution that can overcome the aforementioned shortcomings. An ideal method should be able to automatically, objectively, and accurately identify step changes, intelligently handle time synchronization issues in the testing system, and efficiently complete the entire process from data to report, thereby significantly improving the reliability, efficiency, and standardization of photovoltaic inverter grid-connected testing. Summary of the Invention

[0007] This invention overcomes the problems of low efficiency, strong subjectivity and easy error caused by relying on manual analysis in the prior art. Through fully automated data processing and machine learning recognition, it can significantly improve the accuracy, consistency and efficiency of testing.

[0008] To achieve the above objectives, the present invention adopts the following solution:

[0009] An automated ladder lookup and processing method for photovoltaic inverter grid-connected test data includes the following steps:

[0010] S1: Collect time-series data of photovoltaic inverter grid connection test through grid simulator and power analyzer. The time-series data includes voltage channel data, current channel data and power channel data. Perform low-pass filtering on the time-series data and unify the timestamps of all channel data to generate pre-processed multi-channel data.

[0011] S2: Input the preprocessed multi-channel data into a pre-trained convolutional neural network model. The convolutional neural network model is trained using historical step test data and outputs the coordinates of the start and end points of the step change.

[0012] S3: Obtain the trigger signal output by the power grid simulator, perform rising edge detection on the trigger signal and record its first timestamp, and at the same time detect the starting point of voltage change in the preprocessed multi-channel data and record the second timestamp, calculate the time difference between the first timestamp and the second timestamp, and perform time compensation and alignment on the trigger signal and the preprocessed multi-channel data based on the time difference.

[0013] S4: Based on the time-compensated and aligned multi-channel data and the coordinates of the start and end points of the step change, extract the steady-state interval of each step plateau, calculate the average voltage, average current and average power of the steady-state interval as the plateau value, and calculate the response time and overshoot based on the time point corresponding to the coordinates of the start point of the step change and the plateau value.

[0014] S5: Enter the response time, overshoot, and platform value into the standard test report template to generate a test report document.

[0015] Preferably, step S1 specifically includes the following steps:

[0016] S101: Apply a step-type grid disturbance test signal to the photovoltaic inverter through the grid simulator, and simultaneously collect the response data of the photovoltaic inverter through the power analyzer at a sampling rate of not less than 10kHz to obtain the original voltage channel data, the original current channel data and the original power channel data. The power analyzer and the grid simulator use hardware trigger signals to realize the synchronous start of data acquisition.

[0017] S102: Perform finite-length unit impulse response low-pass filtering on the original voltage channel data, original current channel data and original power channel data based on the Hanning window. The cutoff frequency of the filter is dynamically adjusted according to the switching frequency of the photovoltaic inverter. Specifically, the cutoff frequency is set to 0.4 times the switching frequency and not less than 2kHz.

[0018] S103: A timestamp synchronization method based on a precision clock source is adopted to uniformly stamp the filtered voltage channel data, current channel data and power channel data with microsecond-level precision timestamps. The precision clock source adopts a global positioning system clock signal or a high-precision temperature-controlled crystal oscillator clock signal.

[0019] S104: The timestamped voltage channel data, current channel data, and power channel data are aligned and arranged according to the time series to generate preprocessed multi-channel data. When aligning and arranging, the nearest neighbor interpolation algorithm based on the timestamp is used to handle the time deviation between the data of each channel that is less than one sampling period.

[0020] Preferably, step S5 specifically includes:

[0021] Read the predefined standard test report template file. The standard test report template file adopts the Extensible Markup Language format and contains fixed fields such as test item number, test time, and equipment model, as well as dynamic data filling position markers. Establish the mapping relationship between response time, overshoot and plateau value and the dynamic data filling position in the template. The response time corresponds to the dynamic response characteristics in the test item, the overshoot corresponds to the transient characteristics, and the plateau value corresponds to the steady-state characteristics.

[0022] A rule-based data validation method was adopted, and the validation conditions included: the response time value to be filled was within the valid range of 0 to 500 milliseconds, the overshoot value was within the valid range of 0 to 20 percent, and the platform value was within the valid range of 50 to 120 percent of the rated value. The data that passed all the above validation conditions were normalized according to the International System of Units (SI), retaining three significant figures and adding the corresponding unit symbols. Finally, the document object model interface was used to fill the normalized data into the corresponding position in the template file, generating a portable document format test report document that includes complete test data, test waveform screenshots, and data tables. The document metadata recorded the generation time, operator identifier, and test environment parameter information.

[0023] Preferably, in step S2...

[0024] The convolutional neural network model employs an encoder-decoder architecture. The encoder consists of five convolutional layers for extracting multi-scale features, while the decoder consists of three transposed convolutional layers for generating a time series output of the same length as the input data. Historical step test data includes voltage, frequency, and power step samples of different amplitudes, with Gaussian white noise and impulse noise added to each sample to simulate the actual test environment. The convolutional neural network model is trained using a weighted loss function combining a focus loss function and a mean squared error loss function, with the focus loss function weight coefficient set to 0.7 and the mean squared error loss weight coefficient set to 0.3. The output of the convolutional neural network model consists of two time series of the same length as the input data, representing the probability distribution of the step start point and the probability distribution of the step end point, respectively. The coordinates of the final step change start and end points are determined by finding peak points in the probability distributions that exceed a threshold of 0.8.

[0025] Preferably, after determining and outputting the coordinates of the starting and ending points of the step change, the following processing steps are performed:

[0026] The coordinates of all start and end points output by the convolutional neural network model are sorted in temporal order. Then, the time interval between adjacent start and end points is calculated, and coordinate pairs with a time interval of less than 10 milliseconds are removed. The retained coordinate pairs are verified using a template matching method based on dynamic time warping. The similarity between the data segment corresponding to each coordinate pair and the standard step template is calculated, and coordinate pairs with a similarity of less than 0.85 are removed.

[0027] Time-constrained optimization is applied to the remaining coordinate pairs, pairing each starting point with its nearest subsequent ending point, controlling the time interval between the starting and ending points within the standard duration range of the step test, and outputting the validated and optimized final starting and ending point coordinate sequence.

[0028] Preferably, step S4 specifically includes:

[0029] For each stepped platform, the period from 50 milliseconds after the start point of the step to 20 milliseconds before the end point of the step is defined as the analysis interval. Within the analysis interval, the moving standard deviation of the voltage data is calculated using the sliding window method, with a window length of 10 milliseconds and a step size of 1 millisecond. When five consecutive moving standard deviation values ​​are less than 0.5% of the rated voltage value, the region is determined to have entered a steady state. This region is then extended backward to 20 milliseconds before the end point and forward to 50 milliseconds after the start point, ultimately defining it as the steady-state interval. The arithmetic mean of the voltage channel data, the arithmetic mean of the current channel data, and the arithmetic mean of the power channel data within the steady-state interval are calculated as the voltage average, the arithmetic mean of the current channel data, and the arithmetic mean of the power channel data, respectively.

[0030] Preferably, after calculating the average voltage, average current, and average power over the steady-state range, the validity of the platform data is verified, including:

[0031] Calculate the standard deviation of voltage data within the steady-state range. When the standard deviation of voltage data exceeds 0.8% of the average voltage value, it is determined that the platform data has abnormal fluctuations, and the steady-state range is readjusted. Calculate the correlation coefficients between voltage channel data, current channel data, and power channel data. When the correlation coefficient between voltage and current is lower than 0.95, or the correlation coefficient between voltage and power is lower than 0.98, it is determined that the data consistency between channels is insufficient, and the data re-acquisition process is initiated.

[0032] For the verified platform data, a global extreme value search method is used to calculate the overshoot. The maximum voltage deviation is found within a 200-millisecond time range after the step start point. The percentage of the difference between the maximum deviation and the steady-state voltage average value to the steady-state voltage average value is the overshoot.

[0033] As a preferred method, the response time is calculated based on the time point corresponding to the starting point coordinates of the step change and the plateau value. The specific method is as follows:

[0034] The steady-state error band is defined as ±1% of the average steady-state voltage. Starting from the step start point, the voltage data is scanned backward in 1-millisecond steps to find the moment when the voltage value first enters the steady-state error band as the initial entry time. The scan continues from the initial entry time. If the voltage data remains within the steady-state error band for the next 15 consecutive milliseconds, this initial entry time is determined to be the effective response time. If the voltage value jumps out of the error band during the scan, the scan restarts from the jump point until a time period that satisfies the condition of remaining within the error band for 15 consecutive milliseconds is found. The starting point of this time period is recorded as the final response time. The response time is calculated as the time difference between the final response time and the step start point.

[0035] Preferably, step S3 specifically includes:

[0036] A Schmitt trigger circuit is used to shape the trigger signal output from the power grid simulator to eliminate glitches and oscillations. Then, a digital logic circuit detects the rising edge of the shaped trigger signal and records the precise moment of the rising edge as the first timestamp. Simultaneously, in the voltage channel data of the preprocessed multi-channel data, a sliding window variance detection algorithm is used to calculate the variance value of the data within a window of 20 sampling points before and after each sampling point. When the variance value of 5 consecutive sampling points exceeds a set threshold, it is determined as the starting point of voltage change, and the time corresponding to the starting point is recorded as the second timestamp. The difference between the first timestamp and the second timestamp is calculated to obtain the system transmission delay time. The obtained system transmission delay time is used as a compensation amount for the timestamp correction of all channel data.

[0037] Preferably, after aligning the trigger signal with the preprocessed multi-channel data using time compensation, a time compensation accuracy verification step is also performed:

[0038] Multiple step change points are selected during the test, and the time difference calculation process is repeated to obtain multiple time difference samples. The average and standard deviation of these time differences are calculated. When the standard deviation exceeds 50 microseconds, a dynamic compensation adjustment mechanism is activated. The least squares method is used to fit the changing trend of the time difference and establish a linear correction model for the time compensation amount. The corrected time compensation amount is applied to the timestamp correction of subsequent test data. At the same time, the adjustment amount and adjustment time of each time compensation are recorded to form a time compensation accuracy log.

[0039] The present invention has at least the following beneficial effects: (1) By automatically identifying step points through convolutional neural networks, automatically calculating key parameters through algorithms, and automatically generating reports through templates, testers are completely freed from tedious and repetitive manual data analysis work. This not only greatly shortens the analysis cycle of a single test and improves the overall test efficiency, but more importantly, it eliminates the subjectivity of human judgment and ensures a high degree of consistency and repeatability between the test process and the results; (2) By acquiring data at a high sampling rate and preprocessing it with effective low-pass filtering, and further utilizing a deep learning-based convolutional neural network model for identification, the starting and ending features of step changes can be captured more accurately from complex, interference-containing real signals, effectively overcoming the shortcomings of the traditional fixed threshold method in noisy environments; (3) By accurately defining the steady-state interval through the sliding window standard deviation method and combining it with data validity verification, the reliability of the steady-state platform value is ensured; at the same time, a response time with a continuous hold verification mechanism is adopted. The calculation method of the time interval and the calculation method of the global extreme value search overshoot effectively eliminate transient fluctuation interference, making the measurement results of dynamic performance indicators more accurate and reliable; (4) The system transmission delay is accurately calculated by combining hardware shaping and software detection, and all test data are timestamped to solve the problem of timing asynchrony caused by inherent delay between test devices. Through multi-sample statistical evaluation and dynamic compensation adjustment mechanism, it can adapt to the delay drift that may occur in long-term testing, ensuring the long-term stability and accuracy of the time base throughout the test process; (5) Before generating the report, the validity range of key data is verified and the consistency of data between channels is checked, which plays the role of quality checkpoint and effectively prevents abnormal or invalid data from entering the final report. The test report with a unified format is automatically generated through standardized templates, which not only improves the efficiency of document processing, but also ensures the standardization, completeness and traceability of the report content, and meets the document requirements of high-standard testing. Attached Figure Description

[0040] Figure 1 This is a schematic diagram illustrating the principle of the automated ladder search and processing method provided by the present invention. Detailed Implementation

[0041] The present invention will now be described in further detail with reference to the accompanying drawings, so that those skilled in the art can implement it based on the description.

[0042] like Figure 1 As shown, the automated ladder search and processing method for photovoltaic inverter grid-connected test data provided by the present invention includes the following steps:

[0043] S1: Collect time-series data for grid-connected testing of photovoltaic inverters using a grid simulator and power analyzer. The time-series data includes voltage channel data, current channel data, and power channel data. Perform low-pass filtering on the time-series data and unify the timestamps of all channel data to generate pre-processed multi-channel data.

[0044] A grid simulator is used to apply step-type grid disturbance test signals, such as sudden changes in voltage or frequency, to the photovoltaic inverter to simulate faults or fluctuations in the actual power grid. A power analyzer synchronously acquires the photovoltaic inverter's response data at a high sampling rate (typically no less than 10kHz) to ensure the capture of rapid transient changes. To achieve synchronization at the start of acquisition, the power analyzer and grid simulator use hardware trigger signals, such as sending pulse signals through digital output ports, to ensure that both begin recording data simultaneously. The acquired raw data may contain high-frequency noise and interference, therefore low-pass filtering is required. The low-pass filter can employ a finite-length impulse response design based on a Hanning window, with its cutoff frequency dynamically adjusted according to the photovoltaic inverter's switching frequency, for example, set to 0.4 times the switching frequency and no less than 2kHz, to effectively filter out switching noise and other high-frequency components while preserving the main characteristics of the signal. Subsequently, the data from all channels is timestamped and synchronized using a precision clock source (such as a GPS clock signal or a high-precision temperature-controlled crystal oscillator clock signal) to timestamp each data point with microsecond-level precision, ensuring consistency of time information. Finally, a timestamp-based nearest neighbor interpolation algorithm is used to handle time deviations of less than one sampling period between the data channels, aligning the voltage, current, and power data according to the time sequence to generate preprocessed multi-channel data. This step, through a combination of hardware synchronization and software processing, ensures data temporal consistency and signal quality.

[0045] S2: Input the preprocessed multi-channel data into a pre-trained convolutional neural network model. The convolutional neural network model is trained using historical step test data and outputs the coordinates of the start and end points of the step change.

[0046] The convolutional neural network model employs an encoder-decoder architecture. The encoder, composed of multiple convolutional layers, extracts multi-scale features from the input data, capturing pattern changes at different time resolutions. The decoder, consisting of transposed convolutional layers, maps the extracted features back to a time-series output of the same length as the input data. The model is trained using historical step test data, including voltage, frequency, and power step samples of varying amplitudes. Gaussian white noise and impulse noise are added to simulate various disturbances in real-world testing environments, thereby improving the model's generalization ability. During training, a weighted loss function combining focus loss and mean squared error loss is used. The focus loss function addresses class imbalance issues, such as the difference in the number of step points compared to background points, while the mean squared error loss ensures the smoothness of the output probability distribution. The weighting coefficients can be set to prioritize focus loss (e.g., 0.7) while using mean squared error loss as a secondary factor (e.g., 0.3). The model outputs two time series of the same length as the input data, representing the probability distributions of the start and end points of the steps, respectively. The final step change coordinates are determined by identifying peak points in the probability distributions that exceed a set threshold (e.g., 0.8). To further improve detection accuracy, the output coordinates can be post-processed, such as by sorting them chronologically to remove outliers, or by using a template matching method based on dynamic time warping to verify the reasonableness of the coordinate pairs. This step utilizes deep learning technology to replace traditional threshold-based or manual rule-based methods, enabling more adaptive handling of complex noise and variation patterns, significantly improving the accuracy and efficiency of step point detection.

[0047] S3: Obtain the trigger signal output by the power grid simulator, perform rising edge detection on the trigger signal and record its first timestamp, and at the same time detect the starting point of voltage change in the preprocessed multi-channel data and record the second timestamp, calculate the time difference between the first timestamp and the second timestamp, and perform time compensation alignment on the trigger signal and multi-channel data based on the time difference.

[0048] The trigger signal is typically a digital pulse representing the start of a step test. However, due to equipment transmission delays and signal processing lags, there may be a time deviation between the trigger signal and the actual data change. First, the trigger signal's rising edge is detected, for example, by using a Schmitt trigger circuit to shape the signal, eliminating glitches and oscillations. Then, the precise moment of the rising edge's occurrence is recorded as the first timestamp using digital logic circuitry. Simultaneously, in the pre-processed voltage channel data, the starting point of the voltage change is detected. For example, a sliding window variance detection algorithm is used to calculate the variance of the data within a certain window (e.g., 20 sampling points) before and after each sampling point. When the variance of multiple consecutive sampling points (e.g., 5) exceeds a set threshold, it is determined as the starting point of the voltage change, and the corresponding moment is recorded as the second timestamp. The time difference between the first and second timestamps is calculated; this difference reflects the system transmission delay from the trigger signal's issuance to the detection of the actual voltage change. Then, based on this time difference, the timestamps of all channel data are compensated and corrected, for example, by uniformly shifting the data timestamps forward or backward by the difference amount, achieving precise alignment between the trigger signal and the multi-channel data. To ensure the stability of the compensation, the accuracy of the time compensation can be verified. For example, the statistical characteristics of the time difference can be calculated through multiple step change points. If the standard deviation is too large (e.g., exceeding 50 microseconds), a dynamic compensation adjustment mechanism is activated. The least squares method is used to fit the changing trend of the time difference, establish a linear correction model, and apply it to subsequent data correction. This step, through a time synchronization mechanism combining hardware and software, solves the time drift problem in multi-device collaborative work and ensures the accuracy of subsequent parameter calculations.

[0049] S4: Based on the time-compensated and aligned multi-channel data and the coordinates of the start and end points of the step change, extract the steady-state interval of each step plateau, calculate the average voltage, average current and average power of the steady-state interval as the plateau value, and calculate the response time and overshoot based on the time point corresponding to the coordinates of the start point of the step change and the plateau value.

[0050] For each stepped plateau, the steady-state interval is typically defined as the period from a certain time (e.g., 50 milliseconds) after the start of the plateau to a certain time (e.g., 20 milliseconds) before the end of the plateau, to avoid transient processes. To accurately determine the steady-state interval, a sliding window method is used to calculate the moving standard deviation of the voltage data. The window length is, for example, 10 milliseconds, with a step size of 1 millisecond. When the standard deviation values ​​of multiple consecutive windows (e.g., 5) are all less than 0.5% of the voltage rating, the region is considered to have entered steady state, and this is extended to the entire plateau interval. Within the steady-state interval, the arithmetic mean of the voltage channel data is calculated as the voltage average. The current average and power average are calculated in the same way; these plateau values ​​represent the electrical parameters under steady-state conditions. Then, the response time and overshoot are calculated based on the time point corresponding to the coordinates of the starting point of the stepped change and the plateau values. Response time refers to the time from the start point of the step to the voltage first entering and stabilizing within the steady-state error band (e.g., within ±1% of the average steady-state voltage). Specifically, the voltage data is scanned backward in 1-millisecond steps from the start point, finding the moment the voltage value first enters the error band as the initial entry time, and verifying whether subsequent data continuously remains within the error band (e.g., 15 milliseconds) to determine the final response time. Overshoot is calculated by searching for the maximum voltage deviation within a certain time range (e.g., 200 milliseconds) after the start point of the step and calculating the percentage difference between it and the average steady-state voltage. Furthermore, platform data validity is verified, such as checking whether the standard deviation of the voltage data is too large (e.g., exceeding 0.8% of the average voltage) or whether the correlation coefficient between channels is too low (e.g., the voltage-current correlation coefficient is below 0.95), to ensure consistent data quality. This step automates steady-state parameter extraction and dynamic characteristic calculation through algorithms, reducing the arbitrariness of human judgment and improving the repeatability and reliability of test results.

[0051] S5: Enter the response time, overshoot, and platform value into the standard test report template to generate a test report document.

[0052] The standard test report template uses Extensible Markup Language (EXPLAIN) format and includes fixed fields such as test item number, test time, and equipment model, as well as dynamic data filling location markers. For example, response time corresponds to the dynamic response characteristic in the test item, overshoot corresponds to transient characteristics, and plateau value corresponds to steady-state characteristics. Before data entry, data validation is performed to check whether the response time, overshoot, and plateau value are within the valid range (e.g., 0 to 500 milliseconds), whether they are within the valid range (e.g., 0 to 20%), and whether they are within the valid range of the rated value (e.g., 50% to 120%). These ranges can be adjusted according to specific test standards. Validated data is standardized according to the International System of Units (SI), for example, retaining three significant figures and adding unit symbols to ensure the standardization of data expression. Then, the Document Object Model (DOM) interface is used to fill the data into the corresponding locations in the template file, generating a portable document-formatted test report, including complete test data, test waveform screenshots, and data tables. The report metadata records the generation time, operator identifier, and test environment parameter information for easy traceability and auditing. This step, through templated and automated report generation, achieves closed-loop management of the testing process, improving document processing efficiency and standardization.

[0053] Traditional methods typically rely on manual identification of step change points and parameter calculation, a tedious process prone to subjective errors. This new method, however, achieves efficient and accurate data analysis through automated data acquisition, intelligent signal processing, machine learning model detection, and multi-channel time synchronization. Specifically, a convolutional neural network model is used to automatically detect step change points, improving detection accuracy and robustness, especially in noisy environments. A time compensation alignment mechanism effectively solves synchronization problems between devices, ensuring the accuracy of time measurements. Steady-state interval extraction and parameter calculation are automated through algorithms, reducing human intervention and improving test repeatability and consistency. The report generation module standardizes the output format, improving document quality and processing efficiency. Overall, this method significantly shortens the testing cycle, reduces labor costs, and improves the reliability and traceability of test results, providing strong technical support for the grid-connected performance evaluation of photovoltaic inverters.

[0054] In another technical solution, step S1 specifically includes the following steps:

[0055] S101: Apply a step-type grid disturbance test signal to the photovoltaic inverter through the grid simulator, and simultaneously collect the response data of the photovoltaic inverter through the power analyzer at a sampling rate of not less than 10kHz to obtain the original voltage channel data, the original current channel data and the original power channel data. The power analyzer and the grid simulator use hardware trigger signals to realize the synchronous start of data acquisition.

[0056] S102: Perform finite-length unit impulse response low-pass filtering on the original voltage channel data, original current channel data and original power channel data based on the Hanning window. The cutoff frequency of the filter is dynamically adjusted according to the switching frequency of the photovoltaic inverter. Specifically, the cutoff frequency is set to 0.4 times the switching frequency and not less than 2kHz.

[0057] S103: A timestamp synchronization method based on a precision clock source is adopted to uniformly stamp the filtered voltage channel data, current channel data and power channel data with microsecond-level precision timestamps. The precision clock source adopts a global positioning system clock signal or a high-precision temperature-controlled crystal oscillator clock signal.

[0058] S104: The timestamped voltage channel data, current channel data, and power channel data are aligned and arranged according to the time series to generate preprocessed multi-channel data. When aligning and arranging, the nearest neighbor interpolation algorithm based on the timestamp is used to handle the time deviation between the data of each channel that is less than one sampling period.

[0059] The grid simulator is configured to generate step-type grid disturbance test signals, such as a voltage amplitude jump from its rated value to a dip or spike within milliseconds, or a step change in grid frequency, to simulate faults or fluctuations that may occur in a real power grid. The power analyzer synchronously acquires the response data of the photovoltaic inverter at a sampling rate of at least 10kHz. This high sampling rate ensures the capture of the rapidly changing transient voltage, current, and power waveforms of the inverter under step disturbances. To achieve strict synchronization of acquisition, the power analyzer and the grid simulator are linked by a hardware trigger signal. For example, at the instant the grid simulator begins outputting the step signal, it sends a TTL-level pulse signal through its digital I / O port to the external trigger port of the power analyzer, thereby initiating data recording by the power analyzer. This hardware-level synchronization mechanism effectively eliminates random delays that may be introduced by software commands, ensuring consistency from the occurrence of the disturbance to the start of data recording, laying the foundation for subsequent accurate time analysis. The acquired raw voltage, current, and power channel data serve as the raw input for all subsequent processing operations.

[0060] Because photovoltaic inverters operate in a high-frequency switching state, their output response inevitably contains high-frequency noise introduced by the switching frequency and its harmonics. This noise interferes with subsequent judgments of key characteristics such as the step start point and steady-state value. Therefore, this solution uses a finite-length unit impulse response low-pass filter based on the Hanning window to process the data from each channel. The Hanning window function has smooth attenuation characteristics in the time domain, which can effectively reduce the passband ripple and stopband attenuation of the filter. In the frequency domain, it helps to concentrate energy and reduce spectral leakage. The cutoff frequency of the filter is a key parameter, and its setting is dynamically related to the switching frequency of the photovoltaic inverter. For example, it can be set to 0.4 times the switching frequency, but in order to ensure the preservation of the fundamental frequency and the main low-frequency components, the cutoff frequency is usually set to a value of not less than 2kHz. In actual operation, it can be calculated based on the typical switching frequency of the inverter under test (such as 20kHz), and the cutoff frequency can be set to 8kHz. This dynamic adjustment strategy ensures that high-frequency switching noise is effectively filtered out without excessively damaging the useful rapidly changing components in the signal, providing a clean and reliable data source for subsequent step identification and parameter calculation.

[0061] Although the acquisition is hardware-triggered and synchronized, slight time deviations (less than one sampling period) may still occur between different channels during transmission, conditioning, and analog-to-digital conversion. To address this, this solution introduces a timestamp synchronization method based on a precision clock source. This precision clock source can be a highly stable temperature-controlled crystal oscillator tamed by GPS signals, possessing microsecond-level or even higher precision time output capabilities. During data acquisition, each sampling point is marked with a precise timestamp provided by this precision clock source. When generating pre-processed multi-channel data, the system employs a timestamp-based nearest-neighbor interpolation algorithm to handle potential time deviations (less than one sampling period) between channels. Specifically, the system selects a reference time axis. For each data point in a non-reference channel, it finds the nearest time point on the reference time axis based on its timestamp and assigns the value of that data point to the reference time point. In this way, the data from the voltage, current, and power channels are strictly aligned to a unified time series, forming a multi-channel dataset that is completely synchronized in time.

[0062] Through a refined data acquisition and preprocessing process, the quality and reliability of the test data were improved. High sampling rate and hardware-triggered synchronization ensured complete capture of transient characteristics and precise alignment of the start time; dynamic cutoff frequency low-pass filtering based on the Hanning window effectively suppressed switching noise while preserving the key features of the signal to the greatest extent, thus improving the signal-to-noise ratio; and microsecond-level precision timestamps and nearest-neighbor interpolation synchronization algorithms fundamentally eliminated time misalignment between channels, providing a highly consistent time reference for all subsequent analyses.

[0063] In another technical solution, step S5 specifically includes:

[0064] Read the predefined standard test report template file. The standard test report template file adopts the Extensible Markup Language format and contains fixed fields such as test item number, test time, and equipment model, as well as dynamic data filling position markers. Establish the mapping relationship between response time, overshoot and plateau value and the dynamic data filling position in the template. The response time corresponds to the dynamic response characteristics in the test item, the overshoot corresponds to the transient characteristics, and the plateau value corresponds to the steady-state characteristics.

[0065] A rule-based data validation method was adopted, and the validation conditions included: the response time value to be filled was within the valid range of 0 to 500 milliseconds, the overshoot value was within the valid range of 0 to 20 percent, and the platform value was within the valid range of 50 to 120 percent of the rated value. The data that passed all the above validation conditions were normalized according to the International System of Units (SI), retaining three significant figures and adding the corresponding unit symbols. Finally, the document object model interface was used to fill the normalized data into the corresponding position in the template file, generating a portable document format test report document that includes complete test data, test waveform screenshots, and data tables. The document metadata recorded the generation time, operator identifier, and test environment parameter information.

[0066] The standard test report template is defined using Extensible Markup Language (XML), a highly readable and structured file format that facilitates both machine parsing and manual verification. The template file includes not only fixed fields such as test item number, test execution time, and device under test model, but also pre-defined markers for populating dynamic test data. These markers are identified using XML tags. The system internally establishes a clear mapping rule that associates the calculated response time, overshoot, and steady-state plateau values ​​(average voltage, current, and power) with the corresponding dynamic data population locations in the template. For example, response time data is mapped to the test item column describing the inverter's "dynamic response characteristics," overshoot data corresponds to the "transient characteristics" column, and plateau values ​​are filled into the "steady-state characteristics" column. This template-based and mapping mechanism separates variable test data from a fixed report framework, enabling a highly structured and automated report generation process. Data is simply entered according to the mapping relationship, avoiding the tedious work of manually editing the report format after each test and ensuring consistency in format across different test reports.

[0067] To ensure the validity and reasonableness of the final report data, the system employs a rule-based data validation method to screen the data to be entered. These rules define the valid ranges for various data points. For example, response time should be between 0 and 500 milliseconds, overshoot should be between 0% and 20%, and plateau value (taking voltage as an example) should be between 50% and 120% of its rated value. These range thresholds are reference values ​​set based on relevant standards for photovoltaic inverter grid-connected testing and common inverter performance indicators, and can be adaptively adjusted according to specific test specifications in practical applications. Any data exceeding these ranges will be marked as abnormal, and operators can intervene to check or trigger a data re-acquisition process. For data that passes validation, the system will standardize it according to the International System of Units (SI). For example, voltage is standardized to volts, current to amperes, power to watts, and time to milliseconds, and values ​​are usually retained to three significant figures to ensure the standardization and readability of the report data. This series of validation and standardization steps acts as a "gatekeeper" for data quality, effectively preventing invalid or abnormal data from entering the final report, thus improving the rigor and professionalism of the report.

[0068] The Document Object Model (DOM) interface provides a standard way to access and manipulate the structure and content of XML documents, making the population process accurate and efficient. After population, the system converts the entire document into a portable document format, which offers good cross-platform compatibility and a fixed layout, facilitating distribution, printing, and archiving. The generated PDF test report not only contains all the populated test data but also automatically embeds relevant test waveform screenshots and detailed data organized into tables, forming a complete test document. Furthermore, the system records the report's generation timestamp, the operator's identifier, and environmental parameters (such as ambient temperature and humidity) in the PDF file's metadata. This additional information enhances the test report's traceability, providing valuable contextual information for subsequent data analysis, troubleshooting, or auditing.

[0069] This solution completely transforms the traditional, inefficient, and error-prone manual test report generation mechanism through template-based, automated verification, and standardized output. It automates the entire test report generation process, significantly improving work efficiency. Furthermore, by validating and standardizing data, it significantly enhances the accuracy and standardization of report data. The generated standardized PDF reports are uniform in format, complete in content, and traceable, greatly facilitating the review, comparison, and archiving of test results. Overall, this solution frees testers from tedious report documentation, allowing them to focus more on the analysis and optimization of the tests themselves, while also improving the overall professionalism and quality management capabilities of the testing work.

[0070] In another technical solution, in step S2,

[0071] The convolutional neural network model employs an encoder-decoder architecture. The encoder consists of five convolutional layers for extracting multi-scale features, while the decoder consists of three transposed convolutional layers for generating a time series output of the same length as the input data. Historical step test data includes voltage, frequency, and power step samples of different amplitudes, with Gaussian white noise and impulse noise added to each sample to simulate the actual test environment. The convolutional neural network model is trained using a weighted loss function combining a focus loss function and a mean squared error loss function, with the focus loss function weight coefficient set to 0.7 and the mean squared error loss weight coefficient set to 0.3. The output of the convolutional neural network model consists of two time series of the same length as the input data, representing the probability distribution of the step start point and the probability distribution of the step end point, respectively. The coordinates of the final step change start and end points are determined by finding peak points in the probability distributions that exceed a threshold of 0.8.

[0072] The core of a convolutional neural network (CNN) model is an encoder-decoder structure, where the encoder consists of five consecutive convolutional layers. Each convolutional layer slides its internal kernel across the multi-channel temporal data input to extract multi-scale features, from local to global and from subtle to abstract. For example, shallow convolutional layers can capture sharp transitions in the signal, while deeper convolutional layers can identify complex patterns associated with the transitions of the entire stepped platform. Next, the decoder consists of three transposed convolutional layers, also known as deconvolutional layers. Their function is the opposite of the encoder; they are responsible for progressively upsampling and reconstructing the highly abstracted and compressed feature maps extracted by the encoder into a time-series output with the exact same length as the original input data. This "compress encoding then expand decoding" structure allows the network to fully understand the global context of the input signal and accurately locate events at each time point in the output, making it ideal for tasks requiring precise temporal localization, such as identifying the start and end points of a stepped platform.

[0073] To ensure the trained model possesses strong generalization ability and can handle various complex situations in real-world testing environments, the historical step test data used needs to be sufficiently diverse. This includes voltage step samples of different amplitudes, such as voltage surges or drops of different percentages; frequency step samples of different sizes; and power step samples. More importantly, before using these ideal step samples for training, noise simulating common interference in real-world testing environments is artificially added, such as Gaussian white noise covering a wide frequency range and impulse noise simulating transient impulse interference. This allows the model to be exposed to and learn how to suppress these interferences during the training phase, thus maintaining robustness in real-world noisy data. During training, the loss function used for model optimization is a weighted combination of the focus loss function and the mean squared error loss. The focus loss function addresses the problem of a severe imbalance between the number of step change points (positive samples) and stable background points (negative samples). By adjusting the weights, it allows the model to focus more on learning the minority of samples that are difficult to classify correctly. The mean squared error loss, on the other hand, focuses on ensuring the overall smoothness and continuity of the probability distribution curve of the network output. By setting the weight coefficient of the focus loss function to a relatively high value (e.g., 0.7) and the weight coefficient of the mean squared error loss to an auxiliary value (e.g., 0.3), this weighted loss function guides the model to remain highly sensitive to key step point events while keeping the output curve smooth.

[0074] The final output of the convolutional neural network model is not a direct coordinate value, but rather two one-dimensional time series of the same length as the input data. One series represents the probability that each time point is identified as the "starting point of the ladder," and the other series represents the probability that each time point is identified as the "ending point of the ladder." This probabilistic output contains rich location information. To determine precise coordinates from these probability distribution curves, the system searches for peak points whose probability values ​​exceed a preset threshold (e.g., 0.8). This threshold setting aims to balance recall and precision; a higher threshold results in more reliable detections but may miss some subtle changes, while a lower threshold may introduce more false positives. By scanning the entire probability series, all local maxima exceeding the threshold are identified, and the times corresponding to these points are preliminarily determined as the starting and ending coordinates of the ladder change. This method based on deep learning probabilistic output, compared to the traditional fixed threshold method, can better adapt to changes in signal amplitude and fluctuations in background noise, providing a more flexible and reliable detection mechanism.

[0075] By employing an encoder-decoder network structure, a carefully constructed noisy training dataset, and a training strategy incorporating focus loss, the proposed step detection model demonstrates significant superiority. It can accurately and robustly identify the start and end positions of step changes from test data containing complex noise and variations, overcoming the inherent limitations of traditional methods that are prone to missed or false detections when the signal-to-noise ratio is low or the step shape is irregular. This model possesses excellent perception capabilities for both subtle changes and macroscopic patterns in the signal, thus elevating the automation and reliability of step detection to a new level.

[0076] After determining and outputting the coordinates of the starting and ending points of the step change, the following processing steps are performed:

[0077] The coordinates of all start and end points output by the convolutional neural network model are sorted in temporal order. Then, the time interval between adjacent start and end points is calculated, and coordinate pairs with a time interval of less than 10 milliseconds are removed. The retained coordinate pairs are verified using a template matching method based on dynamic time warping. The similarity between the data segment corresponding to each coordinate pair and the standard step template is calculated, and coordinate pairs with a similarity of less than 0.85 are removed.

[0078] Time-constrained optimization is applied to the remaining coordinate pairs, pairing each starting point with its nearest subsequent ending point, controlling the time interval between the starting and ending points within the standard duration range of the step test, and outputting the validated and optimized final starting and ending point coordinate sequence.

[0079] The initial set of start and end point coordinates output by the convolutional neural network model may contain isolated points or obviously unreasonable points caused by noise interference or momentary model misjudgment. First, the system sorts all detected start and end points chronologically to establish a basic timeline. Next, it calculates the time interval between each start point and its nearest neighboring end point; this time interval theoretically represents the duration of a step platform. Based on the physical characteristics of photovoltaic inverter step testing, an effective step platform should not have too short a duration, otherwise a stable measurement interval cannot be formed. Therefore, the system sets a minimum duration threshold, such as 10 milliseconds, and then compares all calculated time intervals with this threshold, decisively eliminating start-end point coordinate pairs with time intervals shorter than this threshold. This step effectively filters out spurious step events caused by brief pulse interference or model jitter, retaining candidate platforms with reasonable durations.

[0080] Even if a stepped platform meets the duration requirements, its specific voltage or power change waveform may differ significantly from the standard step shape. This could stem from specific disturbances during testing or abnormal inverter responses. To assess this morphological similarity, this scheme employs a template matching method based on dynamic time warping. Dynamic time warping is a powerful algorithm capable of aligning and calculating the similarity between two time series of potentially different lengths. Here, the system matches each data segment (e.g., voltage channel data) defined by the coordinates of the candidate start and end points with a predefined, ideal standard step template, resulting in a similarity score between 0 and 1. The system sets a similarity threshold, such as 0.85; only when the similarity between the candidate data segment and the standard template reaches or exceeds this threshold is the coordinate pair considered valid. This step performs deep screening at the waveform morphology level, ensuring that each identified stepped platform not only has a reasonable duration but also conforms to the basic characteristics of a step, further improving the reliability of the detection results.

[0081] The starting and ending points retained after the first two rounds of screening need to be correctly paired to ensure the temporal coherence of the entire sequence. This step employs a strategy combining nearest neighbor pairing and duration range verification. Specifically, the system traverses each starting point in chronological order and then searches for the nearest ending point within the time range following that starting point, pairing them to form a complete step event. This pairing method follows the natural logic that "a start must be immediately followed by an end." After pairing, the system re-checks the time interval between each pair of starting and ending points to ensure it falls within the standard duration range expected by the step test. This range can be set according to specific test criteria, for example, from several hundred milliseconds to several seconds. Through this optimization, the final output sequence of starting and ending point coordinates is a high-quality detection result that is strictly ordered in time, has reasonable pairing relationships, and meets the expected duration, providing a precise time interval definition for subsequent steady-state parameter calculations.

[0082] By introducing a multi-layered, multi-faceted post-processing optimization process after the initial detection by the neural network, this scheme significantly improves the final output quality of the staircase point coordinate sequence. It not only eliminates obviously unreasonable and transient false detections but also verifies the authenticity of each staircase event through waveform morphology and ensures the integrity of the event sequence through rigorous time logic pairing. This combined strategy enables the entire detection system to exhibit extremely strong anti-interference and error correction capabilities when facing complex real-world data, resulting in highly reliable and accurate output of the staircase point coordinates.

[0083] In another technical solution, step S4 specifically includes:

[0084] For each stepped platform, the period from 50 milliseconds after the start point of the step to 20 milliseconds before the end point of the step is defined as the analysis interval. Within the analysis interval, the moving standard deviation of the voltage data is calculated using the sliding window method, with a window length of 10 milliseconds and a step size of 1 millisecond. When five consecutive moving standard deviation values ​​are less than 0.5% of the rated voltage value, the region is determined to have entered a steady state. This region is then extended backward to 20 milliseconds before the end point and forward to 50 milliseconds after the start point, ultimately defining it as the steady-state interval. The arithmetic mean of the voltage channel data, the arithmetic mean of the current channel data, and the arithmetic mean of the power channel data within the steady-state interval are calculated as the voltage average, the arithmetic mean of the current channel data, and the arithmetic mean of the power channel data, respectively.

[0085] After successfully identifying the coordinates of the start and end points of the step change, it is necessary to avoid the transition process within a step plateau period and select a data segment that can represent the steady-state characteristics. This scheme defines an initial "analytical interval," which begins at a specific time after the start point of the step change, such as 50 milliseconds, and ends at a specific time before the end point of the step change, such as 20 milliseconds. Delaying the start point by 50 milliseconds aims to skip the unavoidable violent oscillations and adjustment processes at the initial stage of the step change, ensuring that the analysis is not contaminated by transient responses. Advancing the end point by 20 milliseconds avoids new transient processes that may be triggered by the impending arrival of the next step event or the pullback after the end of the current step. This analytical interval serves as a preliminary, conservative search range, providing a foundation for subsequently and accurately locating the steady-state region.

[0086] Steady state is essentially characterized by small, random fluctuations in electrical parameters around a set average value, statistically manifested by a very small standard deviation. Based on this principle, the system applies a sliding window method to calculate the moving standard deviation of voltage channel data (typically the primary criterion for determining steady state) within the analysis interval. Specifically, a fixed-length window (e.g., 10 milliseconds) slides along the time axis in small steps (e.g., 1 millisecond). Each slide calculates the standard deviation of all voltage data within the current window. When the system detects that the standard deviations calculated for several consecutive windows (e.g., 5 windows) are all less than a set relative threshold (e.g., 0.5% of the rated voltage), it determines that the region has entered a steady state. This continuous judgment mechanism effectively prevents misjudgments caused by accidental fluctuations in a single data point, ensuring the robustness of the steady-state criterion.

[0087] Once the starting position satisfying the steady-state criterion is found using the sliding window method, the system formally defines this region as the steady-state interval. To fully utilize the stable data, this interval is extended forward to 50 milliseconds after the start point of the step (i.e., the starting boundary of the interval to be analyzed) and backward to 20 milliseconds before the end point of the step (i.e., the ending boundary of the interval to be analyzed), thus forming a final steady-state interval covering most of the stable data during the plateau period. Within this defined steady-state interval, the system calculates the arithmetic mean of all data points in the voltage channel as the voltage average of the plateau, and similarly calculates the arithmetic mean of the current channel data as the current average, and the arithmetic mean of the power channel data as the power average. Using the arithmetic mean is the most direct and effective way to characterize steady-state performance, reflecting the central trend of electrical parameters during the plateau period.

[0088] By precisely defining the analysis interval and combining it with a dynamic detection strategy that uses a sliding window to move the standard deviation, this method can accurately and reliably extract the pure steady-state interval from the stepped platform data, effectively avoiding the interference of transient processes on the calculation of steady-state parameters. Compared with methods that simply fix the interval or rely on a single threshold, this method is more adaptable to different degrees of step response and fluctuations, thus enabling the calculated average values ​​of voltage, current, and power platforms to more realistically and consistently reflect the operating characteristics of the photovoltaic inverter under steady-state conditions, significantly improving the accuracy and repeatability of steady-state parameter measurements.

[0089] After calculating the average voltage, average current, and average power over the steady-state range, the validity of the platform data is verified, including:

[0090] Calculate the standard deviation of voltage data within the steady-state range. When the standard deviation of voltage data exceeds 0.8% of the average voltage value, it is determined that the platform data has abnormal fluctuations, and the steady-state range is readjusted. Calculate the correlation coefficients between voltage channel data, current channel data, and power channel data. When the correlation coefficient between voltage and current is lower than 0.95, or the correlation coefficient between voltage and power is lower than 0.98, it is determined that the data consistency between channels is insufficient, and the data re-acquisition process is initiated.

[0091] For the verified platform data, a global extreme value search method is used to calculate the overshoot. The maximum voltage deviation is found within a 200-millisecond time range after the step start point. The percentage of the difference between the maximum deviation and the steady-state voltage average value to the steady-state voltage average value is the overshoot.

[0092] Even within a range deemed steady-state, abnormal fluctuations may occur due to measurement interference or slight instability of the equipment itself. These fluctuations can distort the calculated platform average. Therefore, after calculating the voltage average within the steady-state range, the system further calculates the standard deviation of the voltage data over that entire range. This standard deviation quantifies the dispersion of the data around the average. The system sets a relatively lenient threshold, such as 0.8% of the voltage average. If the calculated standard deviation of the voltage data exceeds this threshold, it is determined that the platform data exhibits significant abnormal fluctuations, and its average may not be sufficient to represent stable performance. In this case, the system will not directly use this data but will trigger a process to readjust the steady-state range. For example, it might slightly narrow the range of the steady-state range, avoid edge data with large fluctuations, and recalculate the average and standard deviation until the stability requirements are met or prompt the operator to check.

[0093] Under ideal steady-state conditions, the data changes of the voltage, current, and power channels should be highly coordinated. If the data in one channel is abnormal while the other channels are normal, or if the channels lose synchronization due to some fault, the mathematical relationship between them will deviate from expectations. This coordination is quantified by calculating the correlation coefficients between the channels. Specifically, the system calculates the correlation coefficients between the voltage channel data and the current channel data, as well as the correlation coefficients between the voltage channel data and the power channel data within the steady-state range. According to basic circuit principles, voltage and current should exhibit a strong positive correlation under resistive loads, and power is also closely related to the square of voltage under purely resistive conditions. Therefore, the system sets a relatively high correlation threshold; for example, the correlation coefficient between voltage and current should not be lower than 0.95, and the correlation coefficient between voltage and power should not be lower than 0.98. If any set of correlation coefficients falls below its threshold, the data consistency between channels is deemed insufficient, indicating that the test system may have channel misalignment, sensor failure, or other unknown interference. In this case, the system will initiate a data re-acquisition process to ensure that a completely reliable dataset is obtained.

[0094] Overshoot is a crucial indicator of the maximum deviation of the response from the final steady-state value during a transient process. To ensure the capture of the true global extremum, the search range is defined as a sufficiently long time window, such as 200 milliseconds, following the start point of the step disturbance. This time range needs to cover the moment when most photovoltaic inverters might experience maximum overshoot under step disturbances. Within this time range, the system performs a global scan of the voltage channel data, searching for its maximum value (for upward overshoot) or minimum value (for downward overshoot). This extremum is called the maximum voltage deviation. The overshoot is calculated by dividing the difference between this maximum deviation and the previously determined steady-state voltage average value by the steady-state voltage average value, and finally expressed as a percentage. This global extremum search method is simple and direct, and can reliably capture peak phenomena during transient processes.

[0095] By introducing two key quality control steps—internal fluctuation verification of platform data and inter-channel correlation verification—the reliability and credibility of test data results are greatly enhanced. This effectively identifies and eliminates abnormal data caused by measurement noise, equipment instability, or channel misalignment, preventing problematic data from entering the final report and thus reducing the risk of misjudgment. Simultaneously, a clearly defined overshoot calculation method ensures the consistency and accuracy of transient characteristic assessments. This verification mechanism acts as an automated quality filter, significantly improving the robustness of the entire testing system and the professionalism of the output results.

[0096] The response time is calculated based on the time point corresponding to the starting point coordinates of the step change and the plateau value. The specific method is as follows:

[0097] The steady-state error band is defined as ±1% of the average steady-state voltage. Starting from the step start point, the voltage data is scanned backward in 1-millisecond steps to find the moment when the voltage value first enters the steady-state error band as the initial entry time. The scan continues from the initial entry time. If the voltage data remains within the steady-state error band for the next 15 consecutive milliseconds, this initial entry time is determined to be the effective response time. If the voltage value jumps out of the error band during the scan, the scan restarts from the jump point until a time period that satisfies the condition of remaining within the error band for 15 consecutive milliseconds is found. The starting point of this time period is recorded as the final response time. The response time is calculated as the time difference between the final response time and the step start point.

[0098] Response time measures the time required for a system to move from a step input to a point where its output stabilizes within an acceptable range near its steady-state value. This "acceptable range" is called the steady-state error band, defined as the interval between plus and minus 1% of the average steady-state voltage. For example, if the average steady-state voltage is 220 volts, then the error band is from 217.8 volts to 222.2 volts. This 1% threshold is a common engineering practice reference value, taking into account both measurement uncertainties and reasonable requirements for inverter performance. This error band provides a clear, quantified target boundary for subsequent scanning and judgment.

[0099] At the start of the calculation, the system sequentially examines the voltage data of each sample point at a high time resolution (e.g., 1 millisecond step size), beginning from a precisely known step start point. The system searches for the moment when the voltage value first enters the error band from outside the error band; this moment is recorded as the "initial entry time." However, a single entry into the error band may be merely a transient coincidence or fluctuation, insufficient to prove stability. Therefore, starting from this initial entry time, the system continues scanning and initiates a continuous verification: it needs to confirm that all voltage data points remain within the steady-state error band for a subsequent continuous period (e.g., 15 milliseconds). This continuous hold time requirement is crucial for confirming "stability," ensuring that the response is not accidentally touching the boundary but has truly entered the steady-state region.

[0100] In actual response, the voltage may initially enter the error band but then exit it again due to system adjustments or interference. If the system detects that the voltage value has exited the error band during the verification of continuous hold time, the previous "initial entry time" is deemed invalid. At this point, the system does not stop but restarts the scanning and verification process from the point of exit: it continues to search for the next moment when the voltage first enters the error band and verifies again whether it can remain within the band for 15 milliseconds. This process continues until a complete time period satisfying both the "first entry" and "continuous hold" conditions is found. The starting point of this successful time period (i.e., the valid first entry time) is recorded as the "final response time." The final response time calculation result is the time difference between this final response time and the step start point. This algorithm with anomaly handling and reconfirmation mechanisms greatly enhances the anti-interference capability and accuracy of response time calculation.

[0101] By employing a clearly defined steady-state error band, a scanning strategy combining initial entry and continuous hold verification, and a reconfirmation mechanism to handle out-of-state anomalies, the resulting response time calculation method exhibits high accuracy and robustness. It effectively eliminates interference from common jitter and secondary adjustments in the response curve, objectively and consistently determining the inverter's true steady-state moment, avoiding significant errors that may arise from single-point judgments or simple averaging in traditional methods. This makes the measurement results of response time, a key dynamic indicator, more reliable, providing a data foundation for evaluating the rapid response capabilities of photovoltaic inverters.

[0102] In another technical solution, step S3 specifically includes:

[0103] A Schmitt trigger circuit is used to shape the trigger signal output from the power grid simulator to eliminate glitches and oscillations. Then, a digital logic circuit detects the rising edge of the shaped trigger signal and records the precise moment of the rising edge as the first timestamp. Simultaneously, in the voltage channel data of the preprocessed multi-channel data, a sliding window variance detection algorithm is used to calculate the variance value of the data within a window of 20 sampling points before and after each sampling point. When the variance value of 5 consecutive sampling points exceeds a set threshold, it is determined as the starting point of voltage change, and the time corresponding to the starting point is recorded as the second timestamp. The difference between the first timestamp and the second timestamp is calculated to obtain the system transmission delay time. The obtained system transmission delay time is used as a compensation amount for the timestamp correction of all channel data.

[0104] When the power grid simulator begins outputting the step test signal, it simultaneously generates a digital trigger signal, typically manifested as a voltage jump. However, this raw signal may introduce ringing, glitches, or insufficiently steep edges during transmission, making direct detection prone to errors. Therefore, a Schmitt trigger circuit is first used to shape the signal. The Schmitt trigger has inherent hysteresis characteristics, effectively suppressing small noise and oscillations in the signal. It is only considered high when the input voltage exceeds a higher upper threshold level, and switches back to low when the voltage falls below a lower lower threshold level, thus outputting a clean, steeply edged rectangular wave. Subsequently, this shaped signal is fed into digital logic circuits or detected by software algorithms for rising edge detection. The detection logic continuously monitors the signal level, and once a low-to-high transition is detected, it immediately records the time value of the precision clock source corresponding to that moment. This moment is precisely recorded as the first timestamp, representing the theoretical start time of the step test command.

[0105] Although the trigger signal has been issued, there will be a lag in observing actual changes in the voltage channel data due to factors such as the response delay of the power analyzer and the signal transmission path. To accurately capture this actual starting point, this scheme employs a variance-based sliding window detection algorithm on the voltage channel data. The algorithm works as follows: during a stable signal phase, the variance is small; however, when a step change occurs, data volatility increases sharply, and the variance rises significantly. The system slides across the voltage data sequence with a fixed window length (e.g., covering 20 sampling points), calculating the variance of the data within the window point by point. An empirical variance threshold is set; when the algorithm detects that the variance calculated from multiple consecutive sampling points (e.g., 5 points) exceeds this threshold, it determines that a significant change has occurred in that region, and records the time corresponding to the first sampling point that consecutively exceeds the threshold as a second timestamp. This method is highly sensitive to signal abrupt changes and effectively resists random noise interference, thus reliably identifying the moment when the voltage actually begins to change.

[0106] The system calculates the difference between the first and second timestamps, and the resulting difference represents the total system transmission delay from the issuance of the trigger command to the actual start of voltage change. This delay includes various inherent lags in the hardware and software processing chain. Subsequently, the system uses this calculated delay as a fixed compensation amount, uniformly applying it to the timestamp correction of all channel data (including voltage, current, and power). For example, subtracting this delay from the timestamps of all data points shifts the data time axis forward, aligning the detected starting point in the voltage data with the theoretical issuance time of the trigger signal on the time axis. Through this operation, the trigger signal and multi-channel data, which were originally time-biased due to system delays, are corrected to a unified time reference, laying the foundation for all subsequent calculations based on precise time relationships (such as response time).

[0107] By combining hardware signal shaping with intelligent software detection, the inherent system delay can be determined with extremely high precision, and high-precision time alignment between trigger events and response data can be achieved. This method effectively eliminates the time asynchrony errors introduced by differences in response speed and transmission paths between different devices in the test system. This significantly improves the accuracy and reliability of the time base for subsequent calculations of various performance indicators based on time parameters, providing a crucial guarantee for obtaining test results that truly reflect the dynamic characteristics of the inverter.

[0108] After aligning the trigger signal with the preprocessed multi-channel data using time compensation, a time compensation accuracy verification step is also performed:

[0109] Multiple step change points are selected during the test, and the time difference calculation process is repeated to obtain multiple time difference samples. The average and standard deviation of these time differences are calculated. When the standard deviation exceeds 50 microseconds, a dynamic compensation adjustment mechanism is activated. The least squares method is used to fit the changing trend of the time difference and establish a linear correction model for the time compensation amount. The corrected time compensation amount is applied to the timestamp correction of subsequent test data. At the same time, the adjustment amount and adjustment time of each time compensation are recorded to form a time compensation accuracy log.

[0110] During testing, the system does not rely solely on the time difference calculated from a single step event. Instead, it actively selects multiple step change points (e.g., 5 to 10 consecutive step events) and repeatedly executes the time difference calculation process to obtain a time difference sequence composed of multiple samples. The system then calculates the mean and standard deviation of this sequence. The mean reflects the typical magnitude of the system latency, while the standard deviation quantifies the degree of fluctuation in these latency times, i.e., the stability of time synchronization. The system sets a stability threshold, such as 50 microseconds. When the calculated standard deviation does not exceed this threshold, it indicates that the system latency is relatively stable, and the current compensation amount can be used directly. This step elevates the time synchronization performance from a single judgment to a statistical confidence level, providing a scientific basis for determining whether dynamic adjustments are needed.

[0111] When the standard deviation of the time difference exceeds a set stability threshold (e.g., 50 microseconds), it indicates that the system latency exhibits significant volatility or trend changes, potentially stemming from equipment temperature drift, clock drift, or other time-varying factors. In this case, the system uses the least squares method to fit multiple calculated time difference samples. The least squares method can find an optimal straight line (or curve) to describe the trend of the time difference over time. Through fitting, a linear correction model can be established, whose output is a compensation correction value that varies with time. This model can predict possible changes in system latency during subsequent testing, thus upgrading the compensation from a static, fixed value to a dynamic, predictable adjustment value.

[0112] The system applies the corrected time compensation calculated by the linear correction model to the timestamp correction of subsequently collected test data in real time. This means that time compensation is no longer fixed but can be dynamically fine-tuned as testing progresses, continuously combating system latency drift. Simultaneously, the system records the specific value and time of each compensation adjustment, forming a complete time compensation accuracy log. This log not only helps in real-time monitoring of synchronization status but also provides data support for subsequent analysis of test data reliability and diagnosis of potential system problems.

[0113] By introducing multi-sample statistical evaluation and trend-based dynamic adjustment mechanisms on top of single-component compensation, the robustness and stability of the time synchronization system under long-term operation are greatly improved. It can proactively detect and compensate for slow-varying drift in system latency, effectively preventing cumulative errors in subsequent test results caused by gradual degradation of synchronization accuracy, and ensuring the long-term consistency and accuracy of the time base throughout the testing process. Simultaneously, detailed logging enhances the traceability and transparency of the testing process, providing a deeper level of assurance for obtaining high-quality test data.

[0114] It should be noted that although the steps are described in a specific order above, this does not mean that they must be performed in that order. In fact, some of these steps can be executed concurrently, or even in a different order, as long as the required functionality is achieved. The number of devices and processing scale described herein are for simplification of the invention; applications, modifications, and variations of this invention will be readily apparent to those skilled in the art.

[0115] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and illustrations shown and described herein.

Claims

1. An automated ladder search and processing method for grid-connected test data of photovoltaic inverters, characterized in that, Includes the following steps: S1: Collect time-series data of photovoltaic inverter grid connection test through grid simulator and power analyzer. The time-series data includes voltage channel data, current channel data and power channel data. Perform low-pass filtering on the time-series data and unify the timestamps of all channel data to generate pre-processed multi-channel data. S2: Input the preprocessed multi-channel data into a pre-trained convolutional neural network model. The convolutional neural network model is trained using historical step test data and outputs the coordinates of the start and end points of the step change. S3: Obtain the trigger signal output by the power grid simulator, perform rising edge detection on the trigger signal and record its first timestamp, and at the same time detect the starting point of voltage change in the preprocessed multi-channel data and record the second timestamp, calculate the time difference between the first timestamp and the second timestamp, and perform time compensation and alignment on the trigger signal and the preprocessed multi-channel data based on the time difference. S4: Based on the time-compensated and aligned multi-channel data and the coordinates of the start and end points of the step change, extract the steady-state interval of each step plateau, calculate the average voltage, average current and average power of the steady-state interval as the plateau value, and calculate the response time and overshoot based on the time point corresponding to the coordinates of the start point of the step change and the plateau value. S5: Fill in the response time, overshoot, and platform value into the standard test report template to generate a test report document; In step S4, after calculating the average voltage, average current, and average power over the steady-state range, the validity of the platform data is verified, including: Calculate the standard deviation of voltage data within the steady-state range. When the standard deviation of voltage data exceeds 0.8% of the average voltage value, it is determined that the platform data has abnormal fluctuations, and the steady-state range is readjusted. Calculate the correlation coefficients between voltage channel data, current channel data, and power channel data. When the correlation coefficient between voltage and current is lower than 0.95, or the correlation coefficient between voltage and power is lower than 0.98, it is determined that the data consistency between channels is insufficient, and the data re-acquisition process is initiated. For the verified platform data, the global extreme value search method is used to calculate the overshoot. The maximum voltage deviation is found within 200 milliseconds after the step start point. The percentage of the difference between the maximum deviation and the steady-state voltage average value to the steady-state voltage average value is the overshoot. The specific method for calculating the response time based on the time point corresponding to the starting point coordinates of the step change and the plateau value is as follows: The steady-state error band is defined as ±1% of the average steady-state voltage. Starting from the step start point, the voltage data is scanned backward in 1-millisecond steps to find the moment when the voltage value first enters the steady-state error band as the initial entry time. The scan continues from the initial entry time. If the voltage data remains within the steady-state error band for the next 15 consecutive milliseconds, this initial entry time is determined to be the effective response time. If the voltage value jumps out of the error band during the scan, the scan restarts from the jump point until a time period that satisfies the condition of remaining within the error band for 15 consecutive milliseconds is found. The starting point of this time period is recorded as the final response time. The response time is calculated as the time difference between the final response time and the step start point.

2. The automated ladder search and processing method for photovoltaic inverter grid-connected test data according to claim 1, characterized in that, Step S1 specifically includes the following steps: S101: Apply a step-type grid disturbance test signal to the photovoltaic inverter through the grid simulator, and simultaneously collect the response data of the photovoltaic inverter through the power analyzer at a sampling rate of not less than 10kHz to obtain the original voltage channel data, the original current channel data and the original power channel data. The power analyzer and the grid simulator use hardware trigger signals to realize the synchronous start of data acquisition. S102: Perform finite-length unit impulse response low-pass filtering on the original voltage channel data, original current channel data and original power channel data based on the Hanning window. The cutoff frequency of the filter is dynamically adjusted according to the switching frequency of the photovoltaic inverter. Specifically, the cutoff frequency is set to 0.4 times the switching frequency and not less than 2kHz. S103: A timestamp synchronization method based on a precision clock source is adopted to uniformly stamp the filtered voltage channel data, current channel data and power channel data with microsecond-level precision timestamps. The precision clock source adopts a global positioning system clock signal or a high-precision temperature-controlled crystal oscillator clock signal. S104: The timestamped voltage channel data, current channel data, and power channel data are aligned and arranged according to the time series to generate preprocessed multi-channel data. When aligning and arranging, the nearest neighbor interpolation algorithm based on the timestamp is used to handle the time deviation between the data of each channel that is less than one sampling period.

3. The automated ladder search and processing method for photovoltaic inverter grid-connected test data according to claim 1, characterized in that, Step S5 specifically includes: Read the predefined standard test report template file. The standard test report template file adopts the Extensible Markup Language format and contains fixed fields such as test item number, test time, and equipment model, as well as dynamic data filling position markers. Establish the mapping relationship between response time, overshoot and plateau value and the dynamic data filling position in the template. The response time corresponds to the dynamic response characteristics in the test item, the overshoot corresponds to the transient characteristics, and the plateau value corresponds to the steady-state characteristics. A rule-based data validation method was adopted, and the validation conditions included: the response time value to be filled was within the valid range of 0 to 500 milliseconds, the overshoot value was within the valid range of 0 to 20 percent, and the platform value was within the valid range of 50 to 120 percent of the rated value. The data that passed all the above validation conditions were normalized according to the International System of Units (SI), retaining three significant figures and adding the corresponding unit symbols. Finally, the document object model interface was used to fill the normalized data into the corresponding position in the template file, generating a portable document format test report document that includes complete test data, test waveform screenshots, and data tables. The document metadata recorded the generation time, operator identifier, and test environment parameter information.

4. The automated ladder search and processing method for photovoltaic inverter grid-connected test data according to claim 1, characterized in that, In step S2, The convolutional neural network model employs an encoder-decoder architecture. The encoder consists of five convolutional layers for extracting multi-scale features, while the decoder consists of three transposed convolutional layers for generating a time series output of the same length as the input data. Historical step test data includes voltage, frequency, and power step samples of different amplitudes, with Gaussian white noise and impulse noise added to each sample to simulate the actual test environment. The convolutional neural network model is trained using a weighted loss function combining a focus loss function and a mean squared error loss function, with the focus loss function weight coefficient set to 0.7 and the mean squared error loss weight coefficient set to 0.

3. The output of the convolutional neural network model consists of two time series of the same length as the input data, representing the probability distribution of the step start point and the probability distribution of the step end point, respectively. The coordinates of the final step change start and end points are determined by finding peak points in the probability distributions that exceed a threshold of 0.

8.

5. The automated ladder search and processing method for photovoltaic inverter grid-connected test data according to claim 4, characterized in that, After determining and outputting the coordinates of the starting and ending points of the step change, the following processing steps are performed: The coordinates of all start and end points output by the convolutional neural network model are sorted in temporal order. Then, the time interval between adjacent start and end points is calculated, and coordinate pairs with a time interval of less than 10 milliseconds are removed. The retained coordinate pairs are verified using a template matching method based on dynamic time warping. The similarity between the data segment corresponding to each coordinate pair and the standard step template is calculated, and coordinate pairs with a similarity of less than 0.85 are removed. Time-constrained optimization is applied to the remaining coordinate pairs, pairing each starting point with its nearest subsequent ending point, controlling the time interval between the starting and ending points within the standard duration range of the step test, and outputting the validated and optimized final starting and ending point coordinate sequence.

6. The automated ladder search and processing method for photovoltaic inverter grid-connected test data according to claim 1, characterized in that, Step S4 specifically includes: For each stepped platform, the period from 50 milliseconds after the start point of the step to 20 milliseconds before the end point of the step is defined as the analysis interval. Within the analysis interval, the moving standard deviation of the voltage data is calculated using the sliding window method, with a window length of 10 milliseconds and a step size of 1 millisecond. When five consecutive moving standard deviations are all less than 0.5% of the rated voltage value, the area covered by these five consecutive windows is determined to have entered a steady state. This area is then extended backward to 20 milliseconds before the end point and forward to 50 milliseconds after the start point, finally determining it as the steady-state interval. The arithmetic mean of the voltage channel data, the arithmetic mean of the current channel data, and the arithmetic mean of the power channel data within the steady-state interval are calculated as the voltage average, the arithmetic mean of the current channel data, and the arithmetic mean of the power channel data, respectively.

7. The automated ladder search and processing method for photovoltaic inverter grid-connected test data according to claim 1, characterized in that, Step S3 specifically includes: A Schmitt trigger circuit is used to shape the trigger signal output from the power grid simulator to eliminate glitches and oscillations. Then, a digital logic circuit detects the rising edge of the shaped trigger signal and records the precise moment of the rising edge as the first timestamp. Simultaneously, in the voltage channel data of the preprocessed multi-channel data, a sliding window variance detection algorithm is used to calculate the variance value of the data within a window of 20 sampling points before and after each sampling point. When the variance value of 5 consecutive sampling points exceeds a set threshold, it is determined as the starting point of voltage change, and the time corresponding to the starting point is recorded as the second timestamp. The difference between the first timestamp and the second timestamp is calculated to obtain the system transmission delay time. The obtained system transmission delay time is used as a compensation amount for the timestamp correction of all channel data.

8. The automated ladder search and processing method for photovoltaic inverter grid-connected test data according to claim 7, characterized in that, After aligning the trigger signal with the multi-channel data using time compensation, a time compensation accuracy verification step is also performed: Multiple step change points were selected during the test, and the time difference calculation process was repeated to obtain multiple time difference samples. Calculate the average and standard deviation of these time differences. When the standard deviation exceeds 50 microseconds, activate the dynamic compensation adjustment mechanism. Use the least squares method to fit the changing trend of the time difference and establish a linear correction model for the time compensation amount. Apply the corrected time compensation amount to the timestamp correction of subsequent test data. At the same time, record the adjustment amount and adjustment time of each time compensation to form a time compensation accuracy log.

Images