Machine learning based inverter resonance detection method, apparatus and device

Abstract

This invention discloses a machine learning-based inverter resonance detection method, apparatus, and device. The machine learning-based inverter resonance detection method collects real-time operating parameters of the inverter's DC and AC sides, extracts feature vectors containing multi-dimensional features after preprocessing, and performs internal logic verification based on these multi-dimensional feature vectors using a pre-trained classification model. It then outputs the inverter's resonance state result, triggering an alarm and issuing targeted protection and control commands. This solution can not only accurately identify resonance but also effectively distinguish between DC and AC side resonance, significantly reducing the false alarm rate caused by varying operating conditions. It solves the technical problems of existing technologies, such as difficulty in accurately distinguishing resonance types and poor adaptability and high false alarm rates due to reliance on fixed thresholds.

Description

Technical Field

[0001] This invention relates to the field of power electronics control technology, and in particular to a method, apparatus and equipment for inverter resonance detection based on machine learning. Background Technology

[0002] In new energy power generation systems, the inverter, as the core power electronic device for converting DC to AC energy, directly affects the safety and efficiency of the entire grid-connected system due to its operational stability. With the continuous growth of installed capacity of new energy sources such as photovoltaics and wind power, inverters operate in complex and ever-changing grid environments, making them highly susceptible to resonance phenomena in output voltage and current due to factors such as system impedance mismatch, parallel coupling of multiple inverters, or filter parameter resonance, in situations such as weak grids, long-distance power transmission, and nonlinear load connections.

[0003] Traditional inverter resonance detection methods mainly include spectrum analysis based on Fast Fourier Transform (FFT), time-frequency analysis based on wavelet transform, waveform distortion judgment based on fixed thresholds, and analysis methods based on system impedance models. However, these methods face significant technical bottlenecks in practical applications. This results in existing inverter resonance detection methods having difficulty accurately distinguishing between DC and AC side resonances in the inverter, and relying on fixed thresholds leading to high false alarm rates. This technical problem has become a key challenge restricting the reliable grid connection of new energy systems. Summary of the Invention

[0004] The embodiments of the present invention provide an inverter resonance detection method, apparatus and equipment based on machine learning, which aims to solve the technical problems of the prior art, such as the difficulty in accurately distinguishing the DC side and AC side resonance of the inverter, the reliance on fixed thresholds leading to poor adaptability to changing operating conditions and high false alarm rate.

[0005] In a first aspect, embodiments of the present invention provide a machine learning-based inverter resonance detection method, applied to an inverter monitoring system. The system includes an inverter, a data acquisition unit, and an alarm execution unit. The method includes: acquiring DC-side and AC-side operating parameters of the inverter in real time through the data acquisition unit; performing data preprocessing on the DC-side and AC-side operating parameters to obtain preprocessed data; extracting a multi-dimensional feature vector characterizing the resonance state based on the preprocessed data; performing logical verification based on the multi-dimensional feature vector using a pre-trained classification model to output the resonance state result of the inverter; and executing a preset corresponding alarm operation through the alarm execution unit according to the resonance state result, and issuing a corresponding type of resonance protection control command to the inverter.

[0006] Secondly, embodiments of the present invention also provide a machine learning-based inverter resonance detection device, the device comprising: a data acquisition unit for real-time acquisition of DC-side and AC-side operating parameters of the inverter during operation; a data preprocessing unit for preprocessing the DC-side and AC-side operating parameters to obtain preprocessed data; a feature extraction unit for extracting a multi-dimensional feature vector characterizing the resonance state based on the preprocessed data; a resonance identification unit for performing logical verification based on the multi-dimensional feature vector using a pre-trained classification model and outputting the resonance state result of the inverter; and an alarm and control unit for executing a preset corresponding alarm operation through the alarm execution unit according to the resonance state result and issuing a corresponding type of resonance protection control command to the inverter.

[0007] Thirdly, embodiments of the present invention also provide a computer device, the computer device including a memory and a processor connected to the memory; the memory is used to store a computer program; the processor is used to run the computer program stored in the memory to perform the steps of the above-described machine learning-based inverter resonance detection method.

[0008] Compared with the prior art, the beneficial effects of the present invention are: In the technical solution of this invention, the inverter resonance detection method based on machine learning collects the DC-side and AC-side operating parameters of the inverter in real time. After preprocessing, it extracts feature vectors containing multi-dimensional features. A pre-trained classification model then performs internal logic verification based on these multi-dimensional feature vectors, outputting the inverter's resonance state result. This triggers an alarm and issues targeted protection and control commands. Thus, this solution can not only accurately identify resonance but also effectively distinguish between DC-side and AC-side resonance, significantly reducing the false alarm rate caused by varying operating conditions. It solves the technical problems of existing technologies, such as difficulty in accurately distinguishing resonance types and poor adaptability and high false alarm rates due to reliance on fixed thresholds. Attached Figure Description

[0009] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 A flowchart of the inverter resonance detection method based on machine learning provided by the present invention; Figure 2 The first sub-flowchart of the machine learning-based inverter resonance detection method provided by the present invention; Figure 3 This is a sub-flowchart of the second sub-flowchart of the machine learning-based inverter resonance detection method provided by the present invention; Figure 4 The third sub-flowchart of the inverter resonance detection method based on machine learning provided by the present invention; Figure 5 The fourth sub-flowchart of the machine learning-based inverter resonance detection method provided by the present invention; Figure 6 The fifth sub-flowchart of the machine learning-based inverter resonance detection method provided by the present invention; Figure 7 The sixth sub-flowchart of the machine learning-based inverter resonance detection method provided by the present invention; Figure 8 The seventh sub-flowchart of the machine learning-based inverter resonance detection method provided by the present invention; Figure 9 A flowchart illustrating the training process of the classification model for the machine learning-based inverter resonance detection method provided by this invention. Figure 10 A schematic block diagram of the unit of the machine learning-based inverter resonance detection device provided by the present invention; Figure 11 A schematic block diagram of a computer device provided for an embodiment of the present invention. Detailed Implementation

[0011] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0012] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0013] It should also be understood that the terminology used in this specification is for the purpose of describing embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0014] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0015] Based on existing inverter resonance detection methods, the following technical problems and their causes are discussed: First, the FFT method is limited by spectral leakage and picket fence effect, making it difficult to capture non-stationary, instantaneous resonances quickly and accurately. Furthermore, its single frequency domain information is insufficient to pinpoint whether the resonance occurs on the "DC side" (DC / DC level) or the "AC side" (DC / AC level) of the system topology.

[0016] Secondly, although time-frequency analysis methods such as wavelet transform can provide time resolution, their computational complexity is high, making it difficult to meet the requirements of online real-time performance. Furthermore, the analysis results depend on the selection of basis functions and have poor robustness.

[0017] Furthermore, existing technologies often rely on fixed amplitude or distortion rate thresholds for resonance detection. However, the actual operating conditions of inverters, such as dynamic changes in light intensity, ambient temperature, load size, and grid strength, cause the fluctuation range of normal signals to change accordingly. The preset fixed threshold cannot adapt to these changes, often resulting in false alarms under high fluctuation conditions or missed alarms in the early stages of slight resonance.

[0018] Finally, neither signal threshold-based nor impedance model-based methods possess the comprehensive analytical capability to understand the complex coupling relationships between multidimensional operating parameters of the system. Furthermore, they cannot effectively integrate DC-side and AC-side parameters for coordinated judgment, making it difficult to accurately distinguish resonance types. The causes, propagation paths, and required protection measures for DC-side and AC-side resonances are drastically different. Existing technologies cannot achieve precise location, resulting in crude or even ineffective protection strategies, seriously threatening equipment safety and power grid stability.

[0019] To address the technical problems of existing inverter resonance detection methods, such as the inability to accurately distinguish between DC and AC resonances in an inverter and the high false alarm rate due to reliance on fixed thresholds, this invention discloses a machine learning-based inverter resonance detection method, based on a discussion of the causes of these bottlenecks. This method is applied to an inverter monitoring system, which includes an inverter, a data acquisition unit, and an alarm execution unit.

[0020] Based on this inverter monitoring system, and referring to Figures 1 to 9 The method includes the following steps: S110. The DC-side operating parameters and AC-side operating parameters of the inverter are collected in real time through the data acquisition unit. S120. Perform data preprocessing on DC-side operating parameters and AC-side operating parameters to obtain preprocessed data; S130. Based on the preprocessed data, extract the multidimensional feature vector representing the resonance state; S140. Using a pre-trained classification model, perform logic verification based on multi-dimensional feature vectors and output the resonant state result of the inverter. S150. Based on the resonance state result, execute the preset corresponding alarm operation through the alarm execution unit, and send the corresponding type of resonance protection control command to the inverter.

[0021] During the power-on operation of the inverter system, the data acquisition unit first collects the DC-side and AC-side operating parameters in real time. The DC-side operating parameters include the maximum power point tracking voltage, maximum power point tracking current, and bus voltage. The AC-side operating parameters include reactive power, three-phase AC voltage, and three-phase AC current. The data acquisition unit may include multi-channel voltage sensors and multi-channel current sensors to simultaneously acquire electrical signals from both the DC and AC sides. During the data acquisition phase, the reactive power parameter is not directly acquired but is derived from the three-phase AC voltage and current samples obtained by the data acquisition unit, based on the sinusoidal power calculation formula.

[0022] The data preprocessing module performs data preprocessing operations on the DC-side and AC-side operating parameters acquired by the data acquisition unit to obtain preprocessed data with clearer structure and lower noise levels. The data preprocessing process includes multi-stage filtering. First, high-frequency noise is removed using a digital filter. Then, the signal is denoised to eliminate random interference. Next, standardization is performed to make parameters of different dimensions comparable. Finally, the continuous signal is divided into multiple time windows according to a preset time length, forming basic data suitable for subsequent feature extraction. During standardization, the Z-Score standardization method is applied individually to each signal channel. The signal value at each moment is subtracted from the long-term mean of the signal and then divided by the standard deviation, thus making parameters of different dimensions comparable and facilitating subsequent analysis.

[0023] Based on the preprocessed DC-side and AC-side operating parameters, the feature extraction module extracts a multi-dimensional feature vector characterizing the resonant state. Feature extraction can include time-domain statistical feature extraction and frequency-domain feature extraction. In the time-domain statistical feature calculation, for the reactive power in the DC-side and AC-side operating parameters, the skewness, kurtosis, coefficient of variation, and extreme value difference of the signal within each sliding window can be calculated. The formula for calculating the coefficient of variation is: Where n is the amount of data within the window. This represents each data point within the window. This represents the sample average. For the three-phase AC voltage and three-phase AC current in the AC side operating parameters, the skewness and extreme value difference of the signal within each sliding window can be calculated. For the three-phase AC voltage, the maximum three-phase voltage unbalance is also calculated. This index is obtained by calculating the AB line voltage unbalance VUF_AB, BC line voltage unbalance VUF_BC, and CA line voltage unbalance VUF_CA, and then taking the maximum value among the three as the maximum three-phase voltage unbalance VUF_MAX. Simultaneously, the maximum three-phase voltage fluctuation rate is also calculated by obtaining the AB line voltage dispersion coefficient V_AB_CV, BC line voltage dispersion coefficient V_BC_CV, and CA line voltage dispersion coefficient V_CA_CV, and then taking the maximum value among the three as the maximum three-phase voltage fluctuation rate V_CV_MAX. For the three-phase AC current, the three-phase unbalance and fluctuation rate are also statistically analyzed. In frequency domain feature extraction, wavelet transform is used to decompose the signal to obtain the wavelet packet energy and energy entropy of each decomposition layer. The energy entropy is calculated based on the Shannon entropy, which is the proportion of energy in each frequency band to the total energy, and is used to characterize the energy distribution of the signal in the frequency domain.

[0024] After feature extraction, the classifier module inputs the extracted multidimensional feature vectors into a pre-trained machine learning classification model. This model has internalized logic verification rules during training, enabling it to directly perform internal logic verification based on the multidimensional feature vectors and output the inverter's resonance state result, including whether resonance has occurred and its type (DC-side resonance or AC-side resonance). This classification model was trained using abundant sample data collected from laboratory and real-world application environments, covering scenarios with and without resonance. The model integrates a comprehensive judgment mechanism for both DC-side and AC-side features, directly outputting the final resonance state result without the need for additional verification steps.

[0025] When resonance is confirmed, the alarm execution unit executes the corresponding preset alarm operation based on the resonance type and sends the corresponding resonance protection control command to the inverter. If DC-side resonance is determined, the alarm execution unit generates DC-side resonance alarm information and sends a preset DC-side resonance protection command to the inverter controller. These commands may include, but are not limited to, adjusting maximum power point tracking control parameters, increasing DC bus-side damping, and temporarily reducing system output power. If AC-side resonance is determined, the alarm execution unit generates AC-side resonance alarm information and sends a preset AC-side resonance protection command to the inverter controller. The command content may include modifying the inverter modulation strategy, enabling AC-side filter damping function, and temporarily switching the additional damping resistor. All protection control commands undergo safety verification before being sent back to the inverter to ensure they do not conflict with the inverter's current operating state and to guarantee system safety.

[0026] In one embodiment, step S110 includes: S111. The maximum power point tracking voltage and bus voltage on the DC side of the inverter are collected in real time by a voltage sensor. S112. The maximum power point tracking current of the inverter is collected in real time by a current sensor. S113. Real-time acquisition of the three-phase AC voltage and three-phase AC current of the inverter through voltage and current sensors; S114. Calculate the reactive power based on the three-phase AC voltage and three-phase AC current.

[0027] The data acquisition unit may consist of a voltage sensor, a current sensor, and a voltage and current sensor. Each sensor is connected to the corresponding electrical node of the inverter through hardware circuitry to collect the corresponding operating parameters.

[0028] The voltage sensor uses a high-precision isolated voltage sampling chip, which can sample the photovoltaic side input voltage and DC bus voltage in real time to achieve maximum power point tracking control and bus voltage monitoring.

[0029] The current sensor employs a Hall effect current sensor, with its through-hole fitting onto the positive cable of the photovoltaic array's DC output. This allows for non-invasive, real-time detection of the maximum power point tracking current. The sensor features high linearity and a wide bandwidth response, accurately capturing high-frequency current oscillations caused by DC / DC stage resonance. The voltage and current sensors are a three-phase combined sensor unit, comprising three independent voltage measurement channels and three independent current measurement channels. The voltage channels are connected to the three-phase AC output terminals of the inverter (phases A, B, and C) via a high-voltage resistor network to acquire three-phase AC voltage. The current channels utilize three closed-loop Hall current sensors, each fitted onto one of the three-phase output lines, to synchronously acquire three-phase AC current, ensuring strict time alignment of the voltage and current signals. The sampling clock is driven by a unified master clock source.

[0030] After acquiring the three-phase AC voltage and three-phase AC current, the original signal is first sampled and digitally filtered to extract the fundamental component; then, based on the fundamental component, a method based on instantaneous reactive power theory is used. The reactive power is calculated using a method. This calculation process is executed periodically in the embedded processor of the data acquisition unit or the main control computing unit, and the obtained reactive power is used as one of the AC side operating parameters in subsequent feature extraction.

[0031] In one embodiment, step S120 includes: S121. Filter the DC side operating parameters and AC side operating parameters to remove noise and interference signals; S122. Perform noise reduction processing on the filtered DC-side operating parameters and AC-side operating parameters; S123. Standardize the noise-reduced DC-side operating parameters and AC-side operating parameters. S124. The standardized DC-side operating parameters and AC-side operating parameters are divided into a sliding window to obtain the preprocessed data.

[0032] In the data preprocessing stage, the inverter monitoring system performs multi-stage processing on the DC-side and AC-side operating parameters acquired from the data acquisition unit to obtain preprocessed data suitable for subsequent feature extraction. First, the system performs filtering on the DC-side and AC-side operating parameters, employing a two-stage cascaded digital filter approach to remove noise and interference signals. The first stage is a second-order Butterworth low-pass filter used to eliminate high-frequency switching noise and interference; the second stage is an FIR filter used to compensate for phase distortion, smooth the signal, and ensure the integrity of signal edges. The filtering is implemented using a digital filtering algorithm, with each channel having independently set filtering parameters. The filtering parameters for the DC-side channel and the AC-side channel are appropriately adjusted according to their respective signal and noise characteristics.

[0033] After filtering, the DC-side and AC-side operating parameters enter the denoising stage. The denoising process employs a wavelet thresholding algorithm. First, the signal undergoes multi-scale wavelet decomposition, followed by thresholding of the wavelet coefficients at each level using an adaptive threshold formula: in Let be the noise standard deviation, and N be the signal length. A soft thresholding method is used for thresholding, shrinking coefficients above the threshold and setting coefficients below the threshold to zero, balancing denoising effectiveness and signal fidelity. After denoising, clean signal samples are obtained through wavelet reconstruction. This process ensures that no new phase delay is introduced, meeting the real-time requirements of resonance detection.

[0034] The denoised DC-side and AC-side operating parameters then undergo standardization. Standardization uses the Z-Score method: for each signal channel, the mean and standard deviation within a sliding time window are calculated, and then the mean is subtracted from each data point within the window, and the result is divided by the standard deviation. The standardization formula is: Where x is the original signal value, μ is the mean within the sliding window, and σ is the standard deviation within the sliding window. This processing makes signals of different dimensions comparable while eliminating the influence of signal amplitude variations with operating conditions. In particular, for three-phase AC voltage and current, additional verification is performed using three-phase balance conditions to eliminate signal values ​​exceeding reasonable ranges, preventing abnormal signals caused by single-phase faults from affecting the overall standardization effect.

[0035] Finally, the standardized DC-side and AC-side operating parameters are segmented into preprocessed data using a sliding window. During the sliding window segmentation, an overlapping window technique is employed to effectively avoid feature loss due to window boundaries. For the data within each window, a Hanning window is added to reduce spectral leakage, and the window data is encapsulated into a standard data structure, including window ID, start timestamp, sampling interval, and point-by-point data values, facilitating accurate processing by the subsequent feature extraction module.

[0036] In practical implementation, the data preprocessing stage can be deployed on edge computing devices, using open-source stream processing frameworks such as Apache Flink to build the processing pipeline. For the specific application scenario of photovoltaic inverters in this case, the sliding window length can be appropriately shortened during periods of significant light variation to enhance the response to rapid changes, while the time window can be appropriately extended during the daytime when the grid voltage is stable to improve the stability of feature extraction.

[0037] The entire data preprocessing stage is a key guarantee for the performance of resonance detection. It provides a high-quality data foundation for subsequent feature extraction and machine learning model judgment, effectively solves the problem of misjudgment caused by poor data quality in traditional resonance detection schemes, and significantly improves the accuracy and robustness of resonance detection.

[0038] In one embodiment, the AC side operating parameters include first AC side operating parameters and second AC side operating parameters, and step S130 includes: S131. Calculate the time-domain statistical characteristics of the DC side operating parameters and the first AC side operating parameters, including skewness, kurtosis, coefficient of variation, and extreme value difference. S132. Calculate the time-domain statistical characteristics of the operating parameters of the second AC side, including skewness, extreme value difference, three-phase imbalance, and volatility. S133. Form a multidimensional feature vector based on time-domain statistical features.

[0039] Once the inverter monitoring system acquires standardized, preprocessed data divided into sliding windows, the feature extraction module first calculates the time-domain statistical characteristics of the DC-side operating parameters. These DC-side operating parameters include the maximum power point tracking voltage, maximum power point tracking current, and bus voltage signal. For each signal window, the module calculates the skewness (representing waveform asymmetry), kurtosis (representing waveform sharpness), coefficient of variation (representing relative signal volatility), and extreme value difference (representing maximum waveform oscillation amplitude). The formula for calculating the coefficient of variation is: Where n represents the amount of data within the window. This represents the value of each data point within the window. This represents the average value of all data points within the window. For DC-side signals, the coefficient of variation calculated by this formula can sensitively reflect the degree of amplification of DC voltage or current fluctuations when resonance occurs.

[0040] The first AC-side operating parameters may include reactive power, and the second AC-side operating parameters may include three-phase AC voltage and three-phase AC current. When performing characteristic calculations on the AC-side operating parameters, for reactive power, skewness, kurtosis, coefficient of variation, and extreme value difference are calculated; for three-phase AC voltage and three-phase AC current, in addition to calculating skewness and extreme value difference, three-phase unbalance and volatility are also calculated. For three-phase AC voltage and three-phase AC current, the feature extraction module first determines the line voltages, calculates the AB line voltage unbalance VUF_AB, BC line voltage unbalance VUF_BC, and CA line voltage unbalance VUF_CA, and then takes the maximum value among the three as the maximum three-phase voltage unbalance VUF_MAX. This index is highly sensitive to grid asymmetry and resonance. Similarly, the AB line voltage dispersion coefficient V_AB_CV, BC line voltage dispersion coefficient V_BC_CV, and CA line voltage dispersion coefficient V_CA_CV are calculated, and then the maximum value is taken as the maximum three-phase voltage volatility V_CV_MAX. The three-phase current imbalance and fluctuation rate are calculated using similar methods, respectively reflecting load imbalance and current oscillation. These additional calculations enable the system to accurately distinguish the types of resonance specific to the AC side.

[0041] Multidimensional feature vectors can be formed by combining time-domain statistical features and frequency-domain features in a preset order. After the time-domain feature calculation is completed, the feature extraction module can also perform frequency-domain decomposition on the DC-side operating parameters and the AC-side operating parameters respectively. Frequency-domain decomposition is performed separately on the DC-side operating parameters using wavelet transform or empirical mode decomposition, and the same applies to the AC-side operating parameters, so that each signal source obtains its corresponding decomposed signal. After decomposition, frequency-domain features are calculated for each decomposed signal, including wavelet packet energy and energy entropy. The wavelet packet energy is calculated as the sum of squares of the coefficients of each frequency component within the decomposed signal, representing the energy magnitude of the corresponding frequency band. The energy entropy is calculated based on the ratio of energy in each frequency band to the total energy using the Shannon entropy formula: in This represents the proportion of energy in each sub-band to the total energy. The energy imbalance between specific frequency bands obtained through wavelet packet decomposition constitutes an important criterion for identifying resonance, because resonance causes signal energy to be significantly concentrated in a specific frequency band.

[0042] After completing the time-domain and frequency-domain feature calculations, the feature extraction module combines the two sets of features in a preset order. The feature vector organization follows the principle of DC-side features first, followed by AC-side features, and is sorted according to feature type. DC-side features include the skewness, kurtosis, coefficient of variation, and extreme value difference of each DC-side operating parameter, while AC-side features include first AC-side features and second AC-side features. The first AC-side features include the skewness, kurtosis, coefficient of variation, and extreme value difference of reactive power, while the second AC-side features include the skewness, extreme value difference, three-phase unbalance, and fluctuation rate of three-phase AC voltage and three-phase AC current. Frequency-domain features are arranged from low to high frequency, recording the energy value and entropy value of each frequency band. The entire feature vector typically includes features in multiple dimensions, but the system can be configured with a feature filtering mechanism to retain a predetermined number of features in the top ranks after sorting by importance, thereby reducing computational complexity while maintaining sufficient recognition capability.

[0043] In practice, feature computation is implemented using an efficient mathematical library, with each operation executed immediately upon arrival of the data window. The system evaluates the actual discriminative power of each feature using pre-defined feature selection algorithms, such as using a random forest algorithm to calculate feature importance and periodically adjusting feature selection and ranking to ensure that the features input into the machine learning model always have the highest discriminative value. To enhance system adaptability, when the system detects a new type of resonance event, it can automatically analyze the salient features in the new event, incorporating high-potential new features into the feature set, thus achieving self-evolution of the feature extraction function.

[0044] In one embodiment, the multidimensional feature vector includes DC-side features extracted based on DC-side operating parameters and first AC-side features extracted based on first AC-side operating parameters; step S140 includes: S141. Input the multidimensional feature vector into the pre-trained classification model to instruct the classification model to perform internal logic verification by checking whether the DC side feature and the first AC side feature are abnormal. If the classification model verifies that the DC side feature and the first AC side feature are abnormal, the classification model outputs the resonance state result of the inverter experiencing DC side resonance. Wherein, abnormal DC side feature and abnormal first AC side feature mean that the DC side feature and the first AC side feature exceed the dynamic range determined during the pre-training of the classification model, respectively.

[0045] Specifically, DC-side operating parameters may include maximum power point tracking voltage, maximum power point tracking current, and bus voltage. DC-side anomalies may be caused by at least one of the following DC-side operating parameters—skewness, kurtosis, coefficient of dispersion, or extreme value difference—exceeding the dynamic range determined during the pre-training of the classification model.

[0046] The first AC side operating parameters may include reactive power. Anomalies in the first AC side characteristics may be that at least one of the following: skewness, kurtosis, coefficient of dispersion, or extreme value difference of the first AC side operating parameters exceeds the dynamic range determined during the pre-training of the classification model.

[0047] Furthermore, the multidimensional feature vector includes AC-side features extracted based on AC-side operating parameters, and step S140 further includes: S142. After the multidimensional feature vector is input into the pre-trained classification model, the classification model performs internal logic verification by checking whether the AC side features and DC side features are in an abnormal state. If both the AC side features and DC side features are abnormal, the classification model outputs the resonance state result of the inverter experiencing AC side resonance. Here, AC side feature abnormality means that the AC side features exceed the dynamic range determined during the pre-training of the classification model.

[0048] Specifically, the second AC side operating parameters may include three-phase AC voltage and three-phase AC current. AC side characteristic anomalies include first AC side characteristic anomalies and second AC side characteristic anomalies. The second AC side characteristic anomaly may be that at least one of the following—skewness, extreme value difference, three-phase imbalance, and volatility—of the second AC side operating parameters exceeds the dynamic range determined during the pre-training of the classification model.

[0049] During implementation, the multi-dimensional feature vector generated in the previous step is first input into a pre-trained classification model. This model has been trained using a large amount of labeled historical operating data, including normal operating conditions, DC-side resonance, and AC-side resonance scenarios, through supervised learning algorithms before deployment. It has also been validated and optimized, and possesses the ability to classify resonance states into multiple categories. After receiving the feature vector, the embedded logic verification mechanism is immediately activated, directly outputting the resonance state result.

[0050] If the classification model identifies a need to check for DC-side resonance, it uses an internal logic verification mechanism to check whether the DC-side features and the first AC-side feature in the multi-dimensional feature vector are in an abnormal state. An abnormal DC-side feature means that at least one of the time-domain statistical features corresponding to the DC-side operating parameters in the multi-dimensional feature vector—namely, skewness, kurtosis, coefficient of variation, or extreme value difference—exceeds the dynamic range determined during the pre-training of the classification model. An abnormal first AC-side feature means that at least one of the time-domain statistical features corresponding to the reactive power in the multi-dimensional feature vector—namely, skewness, kurtosis, coefficient of variation, or extreme value difference—exceeds the dynamic range determined during the pre-training of the classification model. If both the DC-side feature and the first AC-side feature are in an abnormal state, the classification model directly outputs the resonance state result indicating that the inverter has experienced DC-side resonance.

[0051] For AC-side resonance, the classification model uses an internal logic verification mechanism to check whether both the AC-side and DC-side features in the multi-dimensional feature vector are in an abnormal state. AC-side feature anomalies include a first AC-side feature anomaly and a second AC-side feature anomaly. The second AC-side feature anomaly refers to at least one of the time-domain statistical features corresponding to the second AC-side operating parameters in the multi-dimensional feature vector—namely, skewness, extreme value difference, three-phase imbalance, and volatility—exceeding the dynamic range determined during the pre-training of the classification model. If both the AC-side and DC-side features are in an abnormal state, the classification model directly outputs the resonance state result indicating that the inverter has experienced AC-side resonance.

[0052] This method combines data-driven intelligent recognition with logic rules based on physical laws. It utilizes the powerful nonlinear pattern recognition capabilities of machine learning models to cope with changing operating conditions, and effectively suppresses false alarms through interpretable physical rule verification. It significantly improves the ability to accurately distinguish between "DC-side resonance" and "AC-side resonance" faults, and solves the problems of poor adaptability and ambiguous positioning caused by the reliance on fixed thresholds in traditional methods.

[0053] In one embodiment, step S150 includes: S151. If the resonance state result is DC-side resonance, then the alarm execution unit generates DC-side resonance alarm information and sends a preset DC-side resonance protection command to the inverter controller. S152. If the resonance state result is AC side resonance, then the alarm execution unit generates AC side resonance alarm information and sends a preset AC side resonance protection command to the inverter controller.

[0054] When the system ultimately determines that a "DC-side resonance" has occurred, the alarm execution unit immediately initiates a preset response procedure: First, alarm information at the "DC-side resonance" level is generated simultaneously on the local monitoring terminal and the remote centralized control platform. This information includes the fault type, occurrence time, relevant characteristic parameters, and severity level. It is then communicated to maintenance personnel via audible and visual alarms, pop-up notifications, and remote push notifications such as SMS, email, or system alarms to ensure timely fault detection. Simultaneously, the system sends a "preset DC-side resonance protection command" to the inverter's main controller through a standard communication interface. This command may include: reducing the duty cycle of the DC / DC boost stage to lower input voltage stress; activating the DC-side active damping control algorithm to suppress the resonance by injecting a reverse resonance component into the control loop; or triggering soft-dip power logic to smoothly reduce the inverter's output power to a safe level to avoid the resonance operating point, thus protecting power devices while maximizing grid-connected operation.

[0055] If the system ultimately determines that it is an "AC-side resonance," the alarm execution unit generates an "AC-side resonance" alarm message and notifies the user using the same multi-path alarm mechanism as the DC side. Simultaneously, it sends a "preset AC-side resonance protection command" to the inverter controller. This command, tailored to the characteristics of AC-side resonance, employs a different control strategy than the DC side. Specifically, it may include: activating the active damping function of the LCL filter to dissipate resonance energy by adjusting control loop parameters; dynamically adjusting the phase-locked loop (PLL) bandwidth to enhance the system's adaptability to grid voltage disturbances; or generating and deploying an adaptive notch filter online, with its notch frequency dynamically set according to the identified resonance frequency range to precisely suppress resonance components in specific frequency bands. In cases of abnormally severe resonance and a risk of equipment damage, the command can be upgraded to perform a brief grid disconnection and resynchronization operation to achieve fault isolation. This implementation method, by issuing differentiated and targeted protection commands based on the resonance type, avoids the power generation loss caused by traditional single-protection methods, such as direct shutdown, significantly improving the system's operational continuity and safety.

[0056] In one embodiment, the machine learning-based inverter resonance detection method of the present invention further includes: S160. Collect and store the identified resonance state results as resonance samples, and update the training dataset periodically based on the resonance samples to perform incremental learning and updating of the classification model.

[0057] During system operation, all logically verified "DC-side resonance" or "AC-side resonance" status results, along with their original signal data, preprocessed data, multi-dimensional feature vectors, decision logs, and timestamps, are completely collected and stored as new resonance samples in a local database or cloud data warehouse. The resonance type label for each sample is automatically generated by a hybrid decision-making mechanism and can be reviewed and corrected by maintenance personnel through a human-computer interaction interface to ensure label accuracy. The system automatically triggers an update process weekly, monthly, or when the cumulative number of new samples reaches a set quantity.

[0058] During updates, the new resonance samples are merged with the historical training dataset to form an updated training dataset, and the original classification model is incrementally learned based on this dataset. Frameworks supporting continuous learning, such as H2O, Spark MLlib, or PyTorch, can be used to fine-tune the model parameters for retraining, preserving the original generalization ability while quickly absorbing new knowledge. After training, the new model is evaluated on an independent validation set, with metrics including classification accuracy, recall, F1 score, and the ability to identify novel resonances.

[0059] The system automatically deploys the new model and archives the old model only when the new model outperforms the current online model; otherwise, it keeps the old model running and issues an alarm to prompt technicians to investigate the problem. This mechanism enables the resonance detection system to continuously learn and adapt to unknown resonance modes brought about by inverter equipment aging, component parameter drift, grid structure evolution, and new grid connection conditions, achieving self-evolution of detection capabilities.

[0060] In one embodiment, step S160 includes: S161. The sample data that is determined to be in a resonant state is cached as a resonant sample, and the running parameters and multi-dimensional feature vectors are recorded at the same time. S162. Cluster analysis of cached resonance samples is performed using an unsupervised learning algorithm to identify novel resonance samples; S163. Incorporate the identified novel resonance samples and manually labeled samples into the updated training dataset; S164. Incrementally train the original classification model using the updated training dataset, and evaluate the performance of the new model using the validation set.

[0061] Specifically, all sample data that have undergone a complete identification process, including preliminary machine learning judgment and logical verification, and are ultimately determined to be "DC-side resonance" or "AC-side resonance," are cached in real time to an edge server or cloud buffer. Each cached resonance sample not only includes its classification result, but also fully records the original operating parameters that triggered the result and the multi-dimensional feature vector calculated from the preprocessed data, ensuring the traceability and reproducibility of the sample information.

[0062] To further uncover potential patterns in new samples, the system periodically initiates a clustering analysis process. Specifically, unsupervised learning algorithms, such as Isolation Forest or K-Means clustering, are used to process the multidimensional feature vectors corresponding to all cached resonant samples. The Isolation Forest algorithm assesses the anomaly level of samples by constructing random split trees, identifying novel anomalous samples that deviate from known resonant patterns. The K-Means algorithm, on the other hand, divides samples into clusters based on the similarity of their feature spaces. When a new sample clusters into a completely new cluster that is far from the centers of all historical clusters, it is labeled as a "novel resonant sample." The purpose of this step is to discover potential, unknown resonant patterns from a large amount of automatically labeled data that have not appeared in the training set, such as broadband oscillations caused by new grid-connected equipment or composite resonances under specific operating conditions.

[0063] Subsequently, when generating the updated training dataset, the system incorporates these identified "novel resonance samples" along with manually labeled samples that have been reviewed and confirmed by operations personnel into the original historical training dataset. Manual labeling serves as the final quality control step, ensuring the accuracy of the labels for the newly added samples. Finally, this updated training dataset, which integrates novel patterns and high-quality labeling, is used to incrementally train the original classification model. The training process prioritizes maintaining the core structure of the model, only fine-tuning parameters in the last few layers or key nodes to efficiently absorb new knowledge. After training, the new model retains the ability to recognize old resonance patterns while enhancing its ability to recognize novel resonances. The system conducts a comprehensive performance evaluation of the new model using an independent validation set containing both known and newly discovered resonance cases. Only when all metrics, such as overall accuracy and novel sample recall, reach preset thresholds is the new model approved to replace the online model, thus completing a safe and reliable closed-loop self-evolution. This implementation significantly improves the system's foresight and intelligence in dealing with unknown faults.

[0064] In one embodiment, the classification model of the present invention is trained through the following steps: S171. Collect inverter operating data, including normal operating status and various resonance scenarios; S172. Extract training feature vectors from the running data and label the resonance type of the training feature vectors; S173. Use the labeled training feature vectors as the training dataset to train the classification model to identify resonance states; S174. Use a subset of data labeled with resonance types to train a classification model to distinguish resonance types.

[0065] The training begins with the collection of a large amount of inverter operation data, which comes from historical operation records of actual power plants, laboratory simulation platforms, and hardware-in-the-loop simulation systems. The data covers the normal operating conditions of inverters under different lighting, temperature, load, and grid strength conditions, as well as various resonance scenarios that are artificially injected or occur naturally, including DC-side boost stage LC resonance and subsynchronous oscillation, and AC-side LCL filter resonance and broadband oscillation caused by grid impedance mismatch, to ensure the comprehensiveness and representativeness of the dataset.

[0066] After data preprocessing, training feature vectors are extracted from the operational data using the same algorithm as the online dataset. These vectors include time-domain statistical features, volatility, imbalance, and wavelet packet energy and energy entropy in the frequency domain for both DC and AC sides. Each extracted training feature vector is then labeled at two levels. First, experienced power electronics engineers, combining the original waveform, spectrum, and operating condition information, label the overall resonance state, such as "normal" or "resonance." Second, samples labeled as "resonance" are further labeled with their resonance type, either "DC-side resonance" or "AC-side resonance," forming precise classification labels.

[0067] During the model training phase, the fully labeled training feature vectors are used as the training dataset. A supervised learning algorithm is used to train the classification model, enabling it to learn the mapping relationship from feature vectors to "resonance states," thus acquiring preliminary recognition capabilities. Building upon this, a subset of data labeled with resonance types is further used for targeted optimization training or fine-tuning of the model. This step can be achieved by setting a multi-task loss function, allowing the model to simultaneously output "whether resonance exists" and "resonance type," or by using a hierarchical classification strategy: first determining whether resonance exists, then training a sub-model on the resonance samples to distinguish resonance types. Ultimately, this yields an ensemble classification model capable of end-to-end recognition of resonance states and accurate differentiation of resonance types. After training, the model's performance is validated on an independent test set. Only after confirming high accuracy and strong generalization ability can it be deployed as a pre-trained classification model to an online system.

[0068] Figure 10 This is a schematic block diagram of an inverter resonance detection device 600 based on machine learning provided in an embodiment of the present invention. Figure 10 As shown, corresponding to the above-described machine learning-based inverter resonance detection method, the present invention also provides a machine learning-based inverter resonance detection device 600. This machine learning-based inverter resonance detection device 600 includes a unit for executing the above-described machine learning-based inverter resonance detection method, and the device can be configured in terminals such as inverters, inverter monitoring systems, power generation equipment, desktop computers, tablet computers, and smartphones.

[0069] Specifically, please refer to Figure 10 The machine learning-based inverter resonance detection device 600 includes: The data acquisition unit 610 is used to acquire the DC side operating parameters and AC side operating parameters of the inverter in real time during operation. The data preprocessing unit 620 is used to preprocess the DC side operating parameters and AC side operating parameters to obtain preprocessed data. The feature extraction unit 630 is used to extract a multi-dimensional feature vector representing the resonance state based on the preprocessed data. The resonance identification unit 640 is used to perform logic verification based on multi-dimensional feature vectors through a pre-trained classification model and output the resonance state result of the inverter. The alarm and control unit 650 is used to execute the preset corresponding alarm operation through the alarm execution unit according to the resonance state result, and to send the corresponding type of resonance protection control command to the inverter.

[0070] In one embodiment, the data acquisition unit 610 includes: The voltage signal acquisition unit is used to acquire the maximum power point tracking voltage and bus voltage on the DC side of the inverter in real time through a voltage sensor. The DC current acquisition unit is used to acquire the inverter's maximum power point tracking current in real time through a current sensor; The AC electrical quantity acquisition unit is used to acquire the three-phase AC voltage and three-phase AC current of the inverter in real time through voltage and current sensors. The reactive power calculation unit is used to calculate and obtain reactive power based on the three-phase AC voltage and three-phase AC current.

[0071] In one embodiment, the data preprocessing unit 620 includes: The signal filtering unit is used to filter the DC-side operating parameters and AC-side operating parameters to remove noise and interference signals; The signal denoising unit is used to denoise the filtered DC-side operating parameters and AC-side operating parameters. The data standardization unit is used to standardize the denoised DC-side operating parameters and AC-side operating parameters. The sliding window segmentation unit is used to segment the standardized DC-side operating parameters and AC-side operating parameters into a sliding window to obtain preprocessed data.

[0072] In one embodiment, the feature extraction unit 630 includes: The first time-domain feature calculation unit is used to calculate the time-domain statistical characteristics of the DC side operating parameters and the first AC side operating parameters, including skewness, kurtosis, coefficient of dispersion, and extreme value difference. The second time-domain feature calculation unit is used to calculate the time-domain statistical characteristics of the second AC side operating parameters, including skewness, extreme value difference, three-phase imbalance, and volatility. The feature vector forming unit is used to form multidimensional feature vectors based on time-domain statistical features.

[0073] In one embodiment, the multidimensional feature vector includes DC-side features extracted based on DC-side operating parameters and first AC-side features extracted based on first AC-side operating parameters; the resonance identification unit 640 includes: The first logic verification unit is used to input the multi-dimensional feature vector into the pre-trained classification model to instruct the classification model to perform internal logic verification by checking whether the DC-side feature and the first AC-side feature are abnormal. If the classification model verifies that the DC-side feature and the first AC-side feature are abnormal, the classification model outputs the resonance state result of the inverter experiencing DC-side resonance. Here, the abnormality of the DC-side feature and the abnormality of the first AC-side feature are respectively the DC-side feature and the first AC-side feature exceeding the dynamic range determined during the pre-training of the classification model.

[0074] In one embodiment, the multidimensional feature vector includes AC-side features extracted based on AC-side operating parameters, and the resonance identification unit 640 further includes: The second logic verification unit is used to perform internal logic verification by checking whether the AC side features and DC side features are in an abnormal state after the multi-dimensional feature vector is input into the pre-trained classification model. If both the AC side features and DC side features are abnormal, the classification model outputs the resonance state result of the inverter experiencing AC side resonance. Here, AC side feature abnormality means that the AC side features exceed the dynamic range determined during the pre-training of the classification model.

[0075] In one embodiment, the alarm and control unit 650 includes: The DC-side alarm and control unit is used to generate DC-side resonance alarm information through the alarm execution unit and send a preset DC-side resonance protection command to the inverter controller if the status result is DC-side resonance. The AC side alarm and control unit is used to generate AC side resonance alarm information through the alarm execution unit and send a preset AC side resonance protection command to the inverter controller if the status result is AC side resonance.

[0076] In one embodiment, the machine learning-based inverter resonance detection device 600 further includes: The incremental learning unit is used to collect and store the identified resonance state results as resonance samples, and to periodically update the training dataset based on the resonance samples to perform incremental learning and updating of the classification model.

[0077] In one embodiment, the model incremental learning unit includes: The resonant sample caching unit is used to cache the sample data that has been determined to be in a resonant state as a resonant sample, and at the same time record the running parameters and multi-dimensional feature vectors. A novel resonance identification unit is used to perform cluster analysis on cached resonance samples using an unsupervised learning algorithm to identify novel resonance samples. The training set update unit is used to incorporate the newly identified resonant samples and manually labeled samples into the updated training dataset. The model retraining and evaluation unit is used to incrementally train the original classification model using the updated training dataset and evaluate the performance of the new model using the validation set.

[0078] In one embodiment, the classification model in the machine learning-based inverter resonance detection device 600 is trained using the following units: The operation data collection unit is used to collect the inverter's operating data, including normal operating status and various resonance scenarios; The feature extraction and labeling unit is used to extract training feature vectors from the running data and label the resonance type of the training feature vectors. The classification model training unit is used to train a classification model to identify resonance states using labeled training feature vectors as the training dataset. Type-discriminating training units are used to train a classification model to distinguish resonance types using a subset of data labeled with resonance types.

[0079] The aforementioned machine learning-based inverter resonance detection device 600 can be implemented as a computer program, which can, for example... Figure 11 It runs on the computer device shown.

[0080] Please see Figure 11 , Figure 11 This is a schematic block diagram of a computer device 500 provided in an embodiment of this application. The computer device 500 can be a terminal or a server. The terminal can be an inverter, inverter monitoring system, power generation equipment, desktop computer, tablet computer, smartphone, or other electronic device with communication functions. The server can be a standalone server or a server cluster composed of multiple servers.

[0081] See Figure 11 The computer device 500 includes a processor 502, a memory, and a network interface 505 connected via a system bus 501. The memory may include a non-volatile storage medium 503 and internal memory 504.

[0082] The non-volatile storage medium 503 may store an operating system 5031 and a computer program 5032. The computer program 5032 includes program instructions that, when executed, cause the processor 502 to perform a machine learning-based inverter resonance detection method.

[0083] The processor 502 provides computing and control capabilities to support the operation of the entire computer device 500.

[0084] The internal memory 504 provides an environment for the operation of the computer program 5032 in the non-volatile storage medium 503. When the computer program 5032 is executed by the processor 502, the processor 502 can execute a machine learning-based inverter resonance detection method.

[0085] This network interface 505 is used for network communication with other devices. Those skilled in the art will understand that... Figure 11 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device 500 to which the present application is applied. The specific computer device 500 may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0086] The processor 502 is used to run a computer program 5032 stored in a memory to implement the steps of the above method.

[0087] It should be understood that in the embodiments of this application, the processor 502 may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0088] It will be understood by those skilled in the art that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program includes program instructions and can be stored in a storage medium, which is a computer-readable storage medium. The program instructions are executed by at least one processor in the computer system to implement the process steps of the embodiments of the above methods.

[0089] Therefore, the present invention also provides a storage medium. This storage medium can be a computer-readable storage medium. The storage medium stores a computer program, wherein the computer program includes program instructions. When executed by a processor, the program instructions cause the processor to perform the steps of the above-described method.

[0090] The storage medium can be any computer-readable storage medium capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory (ROM), magnetic disk, or optical disk.

[0091] The present invention also provides a computer program product, including a computer program that, when executed by a processor, causes the processor to perform the steps of the above-described method.

[0092] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0093] In the several embodiments provided by this invention, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For example, the division of each unit is only a logical functional division, and there may be other division methods in actual implementation. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0094] The steps in the method of this invention can be adjusted, merged, or reduced in order according to actual needs. The units in the device of this invention can be merged, divided, or reduced according to actual needs. Furthermore, the functional units in the various embodiments of this invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0095] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a terminal, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0096] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A machine learning-based inverter resonance detection method, characterized in that, An inverter monitoring system, comprising an inverter, a data acquisition unit, and an alarm execution unit, is used; the method includes: The data acquisition unit collects the DC-side and AC-side operating parameters of the inverter in real time during operation. The DC-side operating parameters and the AC-side operating parameters are preprocessed to obtain preprocessed data. Based on the preprocessed data, a multidimensional feature vector characterizing the resonance state is extracted. The inverter's resonant state result is output by performing logical verification based on the multidimensional feature vector using a pre-trained classification model. Based on the resonance state result, the alarm execution unit executes the preset corresponding alarm operation and sends the corresponding type of resonance protection control command to the inverter.

2. The inverter resonance detection method based on machine learning according to claim 1, characterized in that, The AC side operating parameters include first AC side operating parameters and second AC side operating parameters. The step of extracting a multidimensional feature vector characterizing the resonant state based on the preprocessed data includes: Calculate the time-domain statistical characteristics of the DC-side operating parameters and the first AC-side operating parameters, including skewness, kurtosis, coefficient of variation, and extreme value difference; Calculate the time-domain statistical characteristics of the second AC side operating parameters, including skewness, extreme value difference, three-phase imbalance, and volatility; A multidimensional feature vector is formed based on the aforementioned time-domain statistical features.

3. The inverter resonance detection method based on machine learning according to claim 2, characterized in that, The multidimensional feature vector includes DC-side features extracted based on the DC-side operating parameters and first AC-side features extracted based on the first AC-side operating parameters. The step of performing logical verification based on the multi-dimensional feature vector using a pre-trained classification model and outputting the resonant state result of the inverter includes: The multidimensional feature vector is input into the pre-trained classification model to instruct the classification model to perform internal logic verification by checking whether the DC side feature and the first AC side feature are abnormal. If the classification model verifies that the DC side feature and the first AC side feature are abnormal, the classification model outputs the resonance state result of the inverter experiencing DC side resonance. Wherein, the DC-side feature anomaly and the first AC-side feature anomaly refer to the DC-side feature and the first AC-side feature respectively exceeding the dynamic range determined during the pre-training of the classification model.

4. The inverter resonance detection method based on machine learning according to claim 2, characterized in that, The multidimensional feature vector includes AC-side features extracted based on the AC-side operating parameters. The step of performing logical verification based on the multidimensional feature vector using a pre-trained classification model and outputting the resonant state result of the inverter includes: After the multidimensional feature vector is input into the pre-trained classification model, the classification model performs internal logic verification by checking whether the AC side feature and the DC side feature are in an abnormal state. If the AC side feature is abnormal and the DC side feature is abnormal, the classification model outputs the resonance state result of the inverter experiencing AC side resonance. The anomaly in the communication side features refers to the communication side features exceeding the dynamic range determined during the pre-training of the classification model.

5. The inverter resonance detection method based on machine learning according to claim 1, characterized in that, The data acquisition unit includes a voltage sensor, a current sensor, and a voltage-current sensor. The step of acquiring the DC-side operating parameters and AC-side operating parameters of the inverter in real time through the data acquisition unit includes: The maximum power point tracking voltage and bus voltage on the DC side of the inverter are collected in real time by a voltage sensor. The maximum power point tracking current of the inverter is collected in real time by a current sensor. The three-phase AC voltage and three-phase AC current of the inverter are collected in real time by voltage and current sensors. The reactive power is calculated based on the three-phase AC voltage and the three-phase AC current.

6. The inverter resonance detection method based on machine learning according to claim 1, characterized in that, The step of preprocessing the DC-side operating parameters and the AC-side operating parameters to obtain preprocessed data includes: The DC-side operating parameters and the AC-side operating parameters are filtered to remove noise and interference signals. The filtered DC-side operating parameters and AC-side operating parameters are then subjected to noise reduction processing. The denoised DC-side operating parameters and AC-side operating parameters are standardized. The standardized DC-side operating parameters and the AC-side operating parameters are divided into a sliding window to obtain the preprocessed data.

7. The inverter resonance detection method based on machine learning according to claim 1, characterized in that, The method further includes: The sample data that is determined to be the result of the resonance state is cached as the resonance sample, and the running parameters and the multidimensional feature vector are recorded at the same time. Cluster analysis of cached resonance samples is performed using an unsupervised learning algorithm to identify novel resonance samples; The identified novel resonance samples and manually labeled samples are incorporated into the updated training dataset; The original classification model is incrementally trained using the updated training dataset, and the performance of the new model is evaluated using the validation set.

8. The inverter resonance detection method based on machine learning according to claim 1, characterized in that, The classification model is trained through the following steps: Collect the inverter's operating data, including normal operating status and various resonance scenarios; Extract training feature vectors from the running data and label the resonance type of the training feature vectors; The classification model is trained to identify resonance states using the labeled training feature vectors as the training dataset. The classification model is trained using a subset of data labeled with resonance types to distinguish resonance types.

9. A machine learning-based inverter resonance detection device, characterized in that, The device includes: The data acquisition unit is used to acquire the DC-side operating parameters and AC-side operating parameters of the inverter in real time during operation. A data preprocessing unit is used to preprocess the DC-side operating parameters and the AC-side operating parameters to obtain preprocessed data. The feature extraction unit is used to extract a multidimensional feature vector characterizing the resonance state based on the preprocessed data. The resonance identification unit is used to perform logical verification based on the multi-dimensional feature vector using a pre-trained classification model, and output the resonance state result of the inverter. The alarm and control unit is used to execute a preset alarm operation through the alarm execution unit based on the resonance state result, and to send a corresponding type of resonance protection control command to the inverter.

10. A computer device, characterized in that, The computer device includes a memory and a processor connected to the memory; the memory is used to store a computer program; the processor is used to run the computer program stored in the memory to perform the steps of the method as described in any one of claims 1 to 8.

Images