Photovoltaic grid-connected harmonic current prediction method based on deep learning

Abstract

The application discloses a photovoltaic grid-connected harmonic current prediction method based on deep learning, and relates to the fields of power systems and photovoltaic grid-connected technologies. The method obtains three-phase load current and grid-connected point voltage data in a photovoltaic grid-connected system, converts the three-phase load current into two-phase direct current components by adopting Clark and Park transformation, extracts fundamental wave current through a low-pass filter, and obtains harmonic current components. Time domain and frequency domain features of the grid-connected point voltage data are extracted by using fast Fourier transformation, the correlation between voltage harmonics and load current harmonics is analyzed, and grid-connected point voltage features are extracted. Local features of the harmonic current components are extracted through a convolutional neural network, time sequence features are captured through a bidirectional long short-term memory network, and a harmonic current prediction model is constructed. Finally, preprocessed current and voltage features are inputted to generate harmonic current prediction values. Nonlinear and uncertain problems in the photovoltaic grid-connected system can be effectively handled.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of power systems and photovoltaic grid-connected technology, in particular to a photovoltaic grid-connected harmonic current prediction method based on deep learning. BACKGROUND

[0002] With the rapid development of photovoltaic grid-connected systems, the use of a large number of power electronic devices makes the photovoltaic grid-connected system generate more harmonics in operation, and there are power quality problems. At the same time, with the exponential increase of nonlinear loads, the harmonic problem of the power grid system in operation is aggravated, which may cause the power factor of the system to decrease, energy loss, accelerate the insulation aging of generators and transformers, and even cause the misoperation of relays and cause fire accidents. While photovoltaic power generation is booming, the harmonic problem cannot be ignored.

[0003] Under the vigorous promotion of the "double carbon" policy, new energy is gradually rising, but its nonlinear operation characteristics make the traditional harmonic detection technology show fatigue, and cannot meet the requirements. At present, the traditional harmonic detection technology has shown fatigue and cannot meet the requirements of harmonic analysis and prediction in modern complex power systems. Therefore, how to accurately and effectively predict harmonic current has become the focus of research. SUMMARY

[0004] In view of the deficiencies of the prior art, the present application provides a photovoltaic grid-connected harmonic current prediction method based on deep learning, which solves the problems of the above background technology.

[0005] To achieve the above purpose, the present application is realized by the following technical scheme: a photovoltaic grid-connected harmonic current prediction method based on deep learning, comprising the following steps: S1. obtaining three-phase load current and grid-connected point voltage data in a photovoltaic grid-connected system; S2. converting the three-phase load current into two-phase direct current components through Clark transformation and Park transformation, extracting the fundamental current of the two-phase direct current components through a low-pass filter, and obtaining the harmonic current component by subtracting the fundamental current from the total current of the three-phase load current, and preprocessing the harmonic current component; S3. extracting the time domain and frequency domain features of the grid-connected point voltage data through fast Fourier transform, analyzing the amplitude and frequency distribution of the voltage harmonic component, evaluating the correlation between the voltage harmonic and the load current harmonic, and obtaining the grid-connected point voltage feature; S4. extracting local features of the harmonic current component through a convolutional neural network, capturing time sequence features of the harmonic current component through a bidirectional long short-term memory network, and constructing a harmonic current prediction model; S5. inputting the preprocessed harmonic current component and the grid-connected point voltage feature into the wave current prediction model to generate a harmonic current prediction value.

[0006] Further, the specific process of converting three-phase load current into two-phase DC components through Clark transformation and Park transformation is as follows: three-phase load current is converted into two-phase DC components through Clark transformation, and three-phase AC current is projected into a stationary coordinate system; two-phase DC components are converted into DC components in a rotating coordinate system through Park transformation, obtaining DC components of two-phase AC current; the fundamental current component in the DC component of two-phase AC current is extracted through a low-pass filter, and high-frequency harmonic components are removed.

[0007] Further, the specific process of obtaining harmonic current components by subtracting the total current of three-phase load current from the fundamental current is as follows: according to three-phase load current, the total current of three-phase load current is calculated; the fundamental component extracted from the DC component of two-phase AC current is subtracted from the calculated total current of three-phase load current, and the harmonic current component is obtained.

[0008] Further, the specific process of preprocessing harmonic current components is as follows: denoising processing is performed on harmonic current components, and a filter is used to remove high-frequency noise components; standardization processing and smoothing processing are performed on denoised harmonic current components to remove sudden values and abnormal fluctuations; according to the time series characteristics of harmonic current components, data interpolation is performed to fill in missing values and ensure data integrity; through window sliding method, features in local time period are extracted to generate time series data format required for training.

[0009] Further, the specific process of extracting time domain and frequency domain features of grid-connected point voltage data through fast Fourier transform is as follows: fast Fourier transform is performed on grid-connected point voltage data to convert time domain data into frequency domain data, extract frequency components in grid-connected point voltage signal, and identify main frequency peaks of voltage signal; the amplitude of each frequency component is calculated to evaluate the contribution of different frequencies; the harmonic components of voltage signal, including the amplitude and frequency distribution of fundamental wave and each order harmonic, are analyzed to obtain the harmonic features of voltage signal; according to the frequency domain analysis result, the energy distribution of each frequency band in voltage signal is determined.

[0010] Further, the specific process of evaluating the correlation between voltage harmonics and load current harmonics to obtain grid-connected point voltage features is as follows: by analyzing the harmonic components of voltage signal and the harmonic components of load current, the similarity of voltage signal and load current in frequency domain is compared, and the mutual relationship of both in different frequency bands is evaluated; the correlation between voltage harmonics and load current harmonics is quantified through correlation, and according to the correlation analysis result, the correlation features between voltage signal and load current signal are extracted to obtain the feature vector of grid-connected point voltage.

[0011] Further, the specific process of extracting local features of the harmonic current component by the convolutional neural network is as follows: inputting the harmonic current component time series data into the convolutional neural network, extracting features of the input harmonic current component through the convolutional layer; capturing the local spatial features of the harmonic current component by sliding the convolution kernel in the local area, reducing the data dimension through the pooling layer, and reducing the calculation complexity; gradually extracting different levels of local features of the harmonic current component through the convolutional layer and the pooling layer, and representing the time series feature learning.

[0012] Further, the specific process of capturing the time series features of the harmonic current component by the bidirectional long short-term memory network is as follows: inputting the local features of the harmonic current component extracted by the convolutional neural network into the bidirectional long short-term memory network; the bidirectional long short-term memory network captures the time series features in the time series data of the harmonic current component by modeling the time series data; the bidirectional long short-term memory network learns forward and reverse time series information respectively in the process of forward and reverse propagation, and generates forward and reverse hidden state vectors; according to the forward and reverse hidden state vectors, the hidden state matrix of the bidirectional time series information is generated.

[0013] Further, the specific process of generating the harmonic current prediction value is as follows: inputting the preprocessed harmonic current component and the grid-connected point voltage feature into the harmonic current prediction model, modeling the input harmonic current component and the grid-connected point voltage feature in time series, and fusing the dynamic change features of the current and the voltage to generate the prediction value of the harmonic current.

[0014] The present application has the following beneficial effects:

[0015] (1) The photovoltaic grid-connected harmonic current prediction method based on deep learning provides accurate basic data sources for harmonic current prediction by obtaining three-phase load current and grid-connected point voltage data of the photovoltaic grid-connected system. The three-phase load current is converted into two-phase direct current components through Clark transformation and Park transformation, and the fundamental current is extracted by combining a low-pass filter, which can effectively separate the harmonic current component. By eliminating the interference of the fundamental current, the accuracy of the harmonic current component extraction is improved, providing more pure harmonic feature data for subsequent model training. The time domain and frequency domain features of the grid-connected point voltage are extracted by fast Fourier transform, which not only quantifies the amplitude and frequency distribution of the voltage harmonic, but also evaluates its correlation with the load current harmonic. By comprehensively analyzing the characteristics of the grid-connected point voltage, it helps to reveal the formation mechanism of the harmonic current and its association with the power grid. The identification ability of complex harmonic components is improved, which provides key inputs for the prediction model and improves its adaptability to nonlinearity and time variability.

[0016] (2), the photovoltaic grid-connected harmonic current prediction method based on deep learning, through the convolutional neural network extracts the local features of the harmonic current component, and then combines the bidirectional long short-term memory network to capture the time sequence features, so that the deep fusion of multi-dimensional features can be realized. The convolutional neural network focuses on spatial feature extraction, while the bidirectional long short-term memory network accurately captures the dynamic changes in the time sequence, thereby improving the representation ability of the complex characteristics of the harmonic current. The problem that the traditional prediction method cannot fully utilize the time sequence information is solved, so that the model can better reflect the evolution law of the actual harmonic current, and the accuracy and robustness of the prediction result are improved. The preprocessed harmonic current component and the grid-connected point voltage feature are input into the harmonic current prediction model, and the prediction accuracy is improved by using multi-feature fusion. By combining the current and voltage characteristic information, the model can comprehensively reflect the grid operation state and load characteristics, and dynamically generate accurate harmonic current prediction values. The stability and safety of system operation are significantly improved.

[0017] Of course, implementing any product of the application does not necessarily require achieving all the advantages described above at the same time. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 The flowchart of the photovoltaic grid-connected harmonic current prediction method based on deep learning of the application. DETAILED DESCRIPTION

[0019] The photovoltaic grid-connected harmonic current prediction method based on deep learning in the embodiments of the application solves the problems of low prediction accuracy, poor real-time performance and inability to effectively capture complex time sequence characteristics in existing photovoltaic grid-connected systems.

[0020] The scheme in the embodiments of the application has the following overall idea:

[0021] Obtain three-phase load current and grid-connected point voltage data in a photovoltaic grid-connected system.

[0022] Convert the three-phase load current into two-phase direct current components through Clark transformation and Park transformation, extract the fundamental current of the two-phase direct current components through a low-pass filter, and obtain the harmonic current component by subtracting the fundamental current from the total current of the three-phase load current. The harmonic current component is preprocessed.

[0023] Extract the time domain and frequency domain features of the grid-connected point voltage data through fast Fourier transform, analyze the amplitude and frequency distribution of the voltage harmonic component, evaluate the correlation between the voltage harmonic and the load current harmonic, and obtain the grid-connected point voltage feature.

[0024] Extract local features of the harmonic current component through a convolutional neural network, capture time sequence features of the harmonic current component through a bidirectional long short-term memory network, and construct a harmonic current prediction model.

[0025] The preprocessed harmonic current component and the grid-connected point voltage feature are input into the wave current prediction model to generate a harmonic current prediction value.

[0026] Please refer to Figure 1 The embodiment of the present application provides a technical scheme: a photovoltaic grid-connected harmonic current prediction method based on deep learning, comprising the following steps: S1. obtaining three-phase load current and grid-connected point voltage data in a photovoltaic grid-connected system; S2. converting the three-phase load current into two-phase direct current components through Clark transformation and Park transformation, extracting the fundamental current of the two-phase direct current components through a low-pass filter, obtaining the harmonic current component by subtracting the fundamental current from the total current of the three-phase load current, and preprocessing the harmonic current component; S3. extracting the time domain and frequency domain features of the grid-connected point voltage data through fast Fourier transform, analyzing the amplitude and frequency distribution of the voltage harmonic component, evaluating the correlation between the voltage harmonic and the load current harmonic, and obtaining the grid-connected point voltage feature; S4. extracting local features of the harmonic current component through a convolutional neural network, capturing time sequence features of the harmonic current component through a bidirectional long short-term memory network, and constructing a harmonic current prediction model; S5. inputting the preprocessed harmonic current component and the grid-connected point voltage feature into the wave current prediction model to generate a harmonic current prediction value.

[0027] In this embodiment, step S1: By collecting three-phase load current and grid-connected point voltage data in the photovoltaic grid-connected system, the core input data of the system is obtained. The three-phase load current reflects the dynamic changes of the load in the grid-connected system, while the grid-connected point voltage data reflects the overall characteristics of the grid operation. By obtaining these data, it provides a basic guarantee for the subsequent extraction and feature analysis of harmonic current components. Step S2: Through Clark transformation and Park transformation, the three-phase load current is converted into two-phase direct current components. This processing method can simplify the characteristic expression of three-phase signals, making it easier for subsequent calculations. At the same time, through a low-pass filter, the fundamental current of the two-phase direct current components is extracted, and the total current of the three-phase load current is subtracted from the fundamental current, thereby accurately separating the harmonic current component. Step S3: Using fast Fourier transform (FFT) to extract time and frequency domain features from grid-connected point voltage data can comprehensively analyze the harmonic components in the voltage signal. Specifically, through FFT, the amplitude and frequency distribution of harmonics can be quantified, and the correlation between voltage harmonics and load current harmonics can be further evaluated. These features provide important information reflecting the operating state of the grid, improving the globality and accuracy of the prediction. Step S4: Using convolutional neural network (CNN) to extract local features of harmonic current components, CNN can identify the spatial distribution characteristics of harmonic signals. Combined with bidirectional long short-term memory network (Bi-LSTM), the time sequence characteristics of harmonic current components are captured, and Bi-LSTM learns the dynamic characteristics of time series through forward and backward propagation. Step S5: The preprocessed harmonic current components and grid-connected point voltage features are input into the constructed harmonic current prediction model. The model generates a predicted value of the harmonic current through multi-feature fusion. This predicted value provides an important basis for subsequent harmonic control and system optimization, helping to detect harmonic problems in advance and take effective measures. Clark transformation is a mathematical method used to convert three-phase alternating current signals (such as current or voltage) from a three-dimensional coordinate system to a two-dimensional plane coordinate system. This transformation simplifies the complex three-phase signal into two components, making it easier for subsequent calculations and analysis. In power systems, Clark transformation is widely used to process symmetric and asymmetric three-phase signals to more intuitively analyze signal characteristics. Park transformation is a method developed on the basis of Clark transformation, which is used to further convert signals in a two-dimensional plane coordinate system to a rotating coordinate system. Through this transformation, the changes of alternating current signals can be described as steady-state direct current signals, which is particularly suitable for extracting the fundamental component of signals. This method is very important in grid and motor control fields, which can effectively distinguish harmonic components and fundamental components. Fast Fourier transform is an efficient calculation method used to convert time-domain signals to frequency-domain signals. Through this transformation, the frequency composition of the signal can be analyzed, and the fundamental and harmonic components can be identified.The fast Fourier transform is very important in harmonic analysis, which can reveal the frequency distribution and amplitude characteristics of the signal, so as to evaluate the harmonic characteristics of voltage and current in the power grid. Convolutional neural network is a deep learning model that can automatically extract features from data. In the present invention, CNN is used to process the harmonic current component and extract local features. Compared with traditional feature extraction methods, CNN does not need artificial design of features, but automatically captures patterns and details in the signal through learning, making the processing process more efficient and accurate. Bidirectional long short-term memory network is a deep learning model for processing time series data, which can capture the forward and reverse characteristics of the signal in time. Its advantage is that it can fully utilize historical information and future information to model the dynamic changes of the signal. In harmonic current prediction, Bi-LSTM can accurately capture the time sequence characteristics of the signal, which helps to improve the prediction accuracy.

[0028] Specifically, the specific process of converting three-phase load current into two-phase DC component through Clark transformation and Park transformation and extracting the fundamental current of the two-phase DC component through a low-pass filter is as follows: converting three-phase load current into two-phase DC component through Clark transformation, projecting three-phase alternating current into a stationary coordinate system; converting two-phase DC component into DC component in a rotating coordinate system through Park transformation to obtain DC component of two-phase alternating current; extracting fundamental current component in DC component of two-phase alternating current through a low-pass filter to remove high-frequency harmonic components.

[0029] In the present embodiment, Clark transformation: the purpose of Clark transformation is to project three-phase alternating current (such as I a , I b , I c ) from a symmetrical three-phase coordinate system into an α-β stationary coordinate system. Through projection, complex three-phase signals can be converted into two-phase signals, which is convenient for subsequent calculation and analysis. The formula is represented as: Formula parameter explanation: I a , I b , I c : three-phase current signal, representing the instantaneous current of the load in the photovoltaic grid-connected system. I α , I β : two-phase current component in the stationary coordinate system, representing the projection on the α-axis and β-axis respectively. Projection matrix: mapping three-phase system to two-dimensional plane (stationary coordinate system). This transformation is essentially projecting three-phase signals from a rotating space to a stationary coordinate system, obtaining I α , I βThe signal is still an alternating signal, but has a simplified mathematical property, which is convenient for subsequent processing. Park transformation: the role of Park transformation is to convert the α-β signal in the stationary coordinate system into the d-q signal in the rotating coordinate system, wherein the d-axis signal is a direct current component, and the q-axis signal reflects the rotating component. The formula is expressed as: Formula parameter explanation: θ: the phase angle at the current moment, which is synchronized with the angular frequency of the grid voltage, and is used to define the direction of the rotating coordinate system I d : the component consistent with the direction of θ in the rotating coordinate system, corresponding to the direct current fundamental component. I q : the component perpendicular to θ in the rotating coordinate system, mainly reflecting the oscillating part of the alternating current. Through this transformation, the grid fundamental signal is extracted to the d-axis component, which is expressed as an approximately constant direct current value, while the high-order harmonic and noise components are still reflected as oscillating signals, which are concentrated in the q-axis component.

[0030] Specifically, the specific process of obtaining the harmonic current component by subtracting the total current of the three-phase load current from the fundamental current is as follows: according to the three-phase load current, the total current of the three-phase load current is calculated; the fundamental component is extracted from the direct current component of the two-phase alternating current, and the calculated total current of the three-phase load current is subtracted from the extracted fundamental current to obtain the harmonic current component.

[0031] In this embodiment, according to the three-phase load current I a ,I b ,I c , the total current of the three-phase load current can be calculated by simple vector sum operation. Assuming I total represents the total current of the three-phase current, the calculation method is as follows: Formula parameter explanation: I a ,I b ,I c : respectively represent the instantaneous value of the three-phase current. I total : represents the total current of the three-phase current. The purpose of this step is to integrate the size of the three-phase current to provide a reference value for subsequent calculation. Through methods such as Clark transformation and Park transformation, the three-phase current is converted into two-phase direct current component, and then the fundamental current component is extracted through a low-pass filter. The fundamental current component represents the most important frequency component in the current signal, and the fundamental current extracted synchronously with the grid frequency is usually I base , which can be obtained from the I d component, and the I d component after low-pass filtering represents the fundamental direct current. The total current I total of the three-phase load current and the extracted fundamental current component I base are subtracted to obtain the harmonic current component. This process can be represented by the following formula: Iharmonic = I total - I base ; formula parameter explanation: I total : total current of three-phase load current, calculated synthetic current. I base : fundamental current extracted from the direct current component, the fundamental component after low-pass filter processing. I harmonic : the final obtained harmonic current component, indicating the part after removing the fundamental component in the total current.

[0032] Specifically, the specific process of preprocessing the harmonic current component is as follows: denoising the harmonic current component, removing high-frequency noise components by using a filter; normalizing and smoothing the denoised harmonic current component to remove sudden values and abnormal fluctuations; according to the time series characteristics of the harmonic current component, data interpolation is performed to handle missing values and ensure data integrity; through the window sliding method, the features in the local time period are extracted to generate the time series data format required for training.

[0033] In this embodiment, the harmonic current component may be disturbed by high-frequency noise. In order to ensure the accuracy of subsequent analysis, it is necessary to remove these noise components. At this time, a filter is used for denoising. The function of the filter is to set a frequency range, only keep the meaningful signal components, and remove the high-frequency noise components. The filter can be designed as a low-pass, high-pass or band-pass filter according to the needs to eliminate the unwanted high-frequency noise. The harmonic current component after denoising often has sudden values or abnormal fluctuations, which may affect the stability of subsequent analysis. Therefore, it needs to be standardized to ensure that the mean value of the data is 0 and the standard deviation is 1, so that it is more suitable for the training of machine learning models. At the same time, smoothing processing is used to remove high-frequency jitter in the signal to avoid affecting the accuracy of the prediction model. The smoothing method can be a simple moving average or a more complex weighted average method, aiming to reduce the fluctuations of the data. In the actual acquisition process, the harmonic current data may have missing values. Missing data will affect the accuracy of subsequent analysis, so it is necessary to interpolate and fill in the missing values. The interpolation method predicts the missing values according to the time series characteristics of the harmonic current component, and the commonly used interpolation methods include linear interpolation, spline interpolation, etc. Through interpolation processing, the integrity of the time series data is ensured, so that subsequent analysis can be carried out on continuous data. The characteristics of the harmonic current component not only reflect the trend of the overall data, but also the characteristics in the local time period. In order to better capture the details of the time series, a window sliding method is used to extract the characteristics in the local time period. Specifically, the entire time series is divided into multiple small time windows, each window contains a certain number of data points. In each window, extract features related to the prediction target, such as mean, standard deviation, maximum value, etc. These local features are crucial for model training and help improve the model's ability to capture time series changes.

[0034] Specifically, the specific process of extracting time domain and frequency domain features of grid-connected point voltage data through fast Fourier transform, analyzing the amplitude and frequency distribution of voltage harmonic components is as follows: performing fast Fourier transform on the grid-connected point voltage data, converting time domain data into frequency domain data, extracting frequency components in the grid-connected point voltage signal, identifying the main frequency peak value of the voltage signal; calculating the amplitude of each frequency component to evaluate the contribution of different frequencies; analyzing the harmonic components of the voltage signal, including the amplitude and frequency distribution of the fundamental wave and each harmonic, obtaining the harmonic characteristics of the voltage signal; according to the frequency domain analysis result, determine the energy distribution of each frequency band in the voltage signal.

[0035] In this embodiment, the grid-connected point voltage data is converted from the time domain to the frequency domain. This is achieved through the Fast Fourier Transform (FFT). FFT is an efficient algorithm that converts time-domain signals into frequency-domain signals, allowing us to observe the contribution of different frequency components in the signal. In simple terms, FFT reveals the amplitude and phase information of the input voltage signal at different frequencies by performing a mathematical transformation on it. After Fourier transformation, the resulting frequency-domain data represents the strength of the voltage signal at each frequency. In the frequency domain, the size (amplitude) of each frequency component corresponds to the contribution of that frequency to the voltage signal. In this way, the main frequency peaks in the voltage signal, i.e. the most significant frequency components in the signal, can be identified. These frequency components may represent the fundamental frequency in the power system (such as the fundamental frequency) and harmonic signals. For each frequency component in the frequency-domain data, we calculate its amplitude, i.e. the energy size of that frequency in the voltage signal. The larger the amplitude, the more significant the frequency component's contribution to the overall voltage signal. The process of calculating the amplitude is done by taking the modulus of the complex frequency-domain representation obtained through the Fourier transform. Amplitude information is crucial for subsequent harmonic analysis, revealing the frequency spectrum structure of the voltage signal. Harmonic components in the voltage signal refer to high-frequency components other than the fundamental frequency (i.e. the power frequency), which are caused by common nonlinear factors in the power system. Through Fourier transform analysis, the voltage signal can be decomposed into fundamental and harmonic components. The fundamental is usually the main frequency of the system, while the harmonics are integer multiples of the fundamental frequency. These harmonic components affect the operation of power equipment, so accurately identifying and analyzing the amplitude and frequency distribution of each order of harmonic is crucial for system optimization and fault prevention. According to the frequency-domain data obtained from the Fourier transform, the energy distribution of the voltage signal in different frequency bands can be further evaluated. This means we can identify the energy concentration of different frequency bands in the voltage signal. Analysis of energy distribution helps us understand which frequency bands have a greater impact on voltage quality, especially in cases of severe harmonic pollution, providing a basis for system optimization.

[0036] Specifically, the correlation between voltage harmonics and load current harmonics is evaluated, and the specific process of obtaining the grid-connected point voltage features is as follows: by analyzing the harmonic components of the voltage signal and the harmonic components of the load current, comparing the similarity of the voltage signal and the load current in the frequency domain, and evaluating the mutual relationship between the two in different frequency bands; the correlation between voltage harmonics and load current harmonics is quantified, and according to the correlation analysis results, the correlation features between the voltage signal and the load current signal are extracted, and the feature vector of the grid-connected point voltage is obtained.

[0037] In this embodiment, the harmonic components of the voltage signal and the load current signal are extracted by performing Fourier transforms on them separately. The voltage signal and the load current signal typically contain fundamental and harmonic components. In this step, we focus on the amplitude and frequency distribution of each frequency component (including the fundamental and harmonics) in the voltage and load current. Then, by comparing these harmonic components, we analyze the similarity in the frequency domain between the voltage signal and the load current signal, finding similarities and differences in different frequency bands. By comparing the harmonic components of the voltage and load current signals in the frequency domain, we can understand the relationship between the two in different frequency ranges. For example, some frequency bands may show significant synchronization, while others may have significant differences. The comparison process not only focuses on the alignment of frequencies, but also includes the matching of amplitudes. This analysis helps to identify the relationship between the voltage signal and the load current, and to understand how changes in the load current affect the quality of the grid voltage, especially when the harmonic pollution is severe. To quantitatively evaluate the correlation between the voltage harmonics and the load current harmonics, the correlation coefficient is a commonly used indicator to measure the similarity between two signals, and the formula is as follows: where ρ represents the correlation coefficient between the voltage and the load current; x i and y i are the amplitudes of the voltage signal and the load current signal in a certain frequency band, respectively; and are the mean values of the voltage signal and the load current signal in that frequency band, respectively; n is the number of data points. This formula calculates the correlation between the voltage and the load current signals in each frequency band to determine whether there is a significant linear relationship between them in the frequency domain. Its value range is usually between -1 and +1, where +1 indicates a complete positive correlation, -1 indicates a complete negative correlation, and 0 indicates no linear correlation. Here, we calculate the correlation coefficient of each harmonic component of the voltage signal and the load current signal to quantify the correlation between them and evaluate the mutual influence between the voltage and load current harmonics in different frequency bands. According to the correlation coefficient analysis results, we can extract the correlation features between the voltage signal and the load current signal. For example, in some frequency bands, the voltage signal and the load current signal may have very similar trends, while in other frequency bands, the correlation is low. By extracting these features, we can construct a feature vector that describes the relationship between the voltage and the load current. This feature vector contains the correlation information between the voltage harmonics and the load current harmonics and is one of the key inputs for the subsequent harmonic current prediction model. Based on the above correlation analysis, the correlation features between the voltage signal and the load current signal are obtained, and the feature vector of the grid-connected point voltage is extracted. This feature vector contains the characteristics of the voltage signal in the frequency domain, especially the relationship with the harmonic components of the load current. This feature vector will be used as input for the subsequent harmonic current prediction model, helping to improve the accuracy of the prediction and the responsiveness of the system.

[0038] Specifically, the specific process of extracting local features of the harmonic current component by the convolutional neural network is as follows: the time series data of the harmonic current component is input into the convolutional neural network, the input harmonic current component is subjected to feature extraction by a convolutional layer; the convolution kernel slides in the local area to capture the local spatial features of the harmonic current component, and the data dimension is reduced by a pooling layer to reduce the calculation complexity; the local features of the harmonic current component at different levels are gradually extracted by the convolutional layer and the pooling layer, and the time series feature learning is represented.

[0039] In the embodiment, the time series data of the harmonic current component is input into the convolutional neural network. The time series data is usually composed of a plurality of continuous current values, which change over time. In this process, the harmonic current value at each time point is taken as an input feature for the neural network to learn and analyze. The core of the convolutional neural network is the convolutional layer. The convolutional layer extracts local features in the data by using a plurality of convolution kernels (also known as filters) to perform convolution operations on the input data. Specifically, the convolution kernel slides on the input data (i.e., the time series of the harmonic current component) and calculates the weighted sum in each local area. In this way, the convolutional layer can capture the local spatial features of the input data, such as the change pattern of the current waveform or the behavior of certain frequency bands. The output of the convolutional layer is processed by the pooling layer, which reduces the dimension of the features extracted by the convolutional layer to reduce the calculation complexity and avoid overfitting. The pooling operation usually has two forms: max pooling and average pooling. Max pooling selects the maximum value in the local area, while average pooling selects the average value in the area. Through the pooling layer, the dimension of the data can be effectively reduced while retaining the main features of the input data, maintaining the calculation efficiency. In the convolutional neural network, there is usually a structure composed of multiple convolutional layers and pooling layers stacked together. Each convolutional layer and pooling layer extracts features at different levels from the data. In the first convolutional layer, the network usually extracts relatively simple local features (such as edges, waveform changes, etc.); in deeper convolutional layers, the network can capture more complex features (such as periodic changes or complex patterns of harmonic components). Through this layer-by-layer feature extraction, the convolutional neural network can extract more and more abstract and recognizable features from the time series data. After processing by multiple convolutional layers and pooling layers, the network will extract rich local features from the original harmonic current data, which not only contain the regularity in time series, but also contain the frequency domain characteristics of the signal. Through such feature learning, the convolutional neural network can identify the key patterns in the time series data, which can be used for harmonic current prediction in subsequent steps. These extracted time series features can effectively reflect the periodic changes and abnormal behaviors of the current waveform.

[0040] Specifically, the specific process of capturing the time sequence characteristics of the harmonic current component by the bidirectional long short-term memory network is as follows: inputting the local features of the harmonic current component extracted by the convolutional neural network into the bidirectional long short-term memory network; the bidirectional long short-term memory network models the time sequence data of the harmonic current component to capture the time sequence characteristics in the time sequence; the bidirectional long short-term memory network learns the forward and reverse time sequence information respectively in the process of forward and reverse propagation to generate forward and reverse hidden state vectors; and the hidden state matrix of the bidirectional time sequence information is generated according to the forward and reverse hidden state vectors.

[0041] In the embodiment, the local features of the harmonic current component are extracted by the convolutional neural network. These features extracted by the convolutional network are inputted into the bidirectional long short-term memory network. The local features here contain the local change pattern of the current component in time, which can help the bidirectional long short-term memory network better understand the time sequence characteristics of the harmonic current component. The bidirectional long short-term memory network (BiLSTM) is a neural network for processing sequence data, which combines the advantages of traditional LSTM (Long Short-Term Memory Network) and bidirectional structure. LSTM can effectively handle long-term dependencies in time series, while the bidirectional structure further enhances the ability to capture time sequence information by modeling the sequence data forward and backward. In BiLSTM, each input of the time sequence passes through two LSTM layers: a forward LSTM layer (from front to back) and a reverse LSTM layer (from back to front). In this way, BiLSTM can capture both past and future time sequence information, thereby improving the learning ability of time sequence characteristics. In the training process of the BiLSTM network, the information in the time sequence data is learned through forward propagation and reverse propagation: forward propagation (forward propagation): the forward LSTM layer processes the input time sequence, starting from the first time step of the sequence and propagating forward one by one. The output of each time step depends on the current input and the hidden state of the previous time step, so it can capture the forward time sequence characteristics of the time sequence. Reverse propagation: the reverse LSTM layer propagates from the last time step of the sequence to the front, modeling the information of the time sequence in reverse. This reverse propagation can capture the dependency between the current time step and the future time steps. Through the two propagation methods of forward and reverse, BiLSTM can learn the information in the time sequence data from two directions at the same time. The output of the bidirectional long short-term memory network is the combination of the forward hidden state and the reverse hidden state, forming a bidirectional hidden state matrix. This matrix will contain both past and future time sequence information, thereby providing more abundant context for further prediction or classification tasks. The hidden state matrix (H t ): The bidirectional hidden state matrix concatenates or combines the forward and reverse hidden state vectors. Assuming that the dimensions of the forward and reverse hidden states are d, the final hidden state matrix has a dimension of 2d. wherein, is the forward-propagated hidden state, is the backward-propagated hidden state, and finally generates a hidden state matrix H containing forward and backward timing information t . Formula and parameter explanation: hidden state h t : represents the hidden state at time step t, which is the memory of the previous and subsequent timing information. The hidden states obtained by forward and backward propagation are and weight matrix W f ,W b : In LSTM, the weight matrix controls how the input information affects the update of the hidden state. In the forward propagation process, W f controls the update of forward propagation, while W b controls the update of backward propagation.

[0042] Specifically, the specific process of generating the harmonic current prediction value is as follows: input the preprocessed harmonic current component and grid-connected point voltage feature into the wave current prediction model, model the input harmonic current component and grid-connected point voltage feature in time sequence, and fuse the dynamic change characteristics of the current and voltage to generate the prediction value of the harmonic current.

[0043] In this embodiment, the preprocessed data is input into the wave current prediction model for time sequence modeling. By simultaneously inputting the harmonic current component and grid-connected point voltage feature into the model, the model can simultaneously learn the dynamic change characteristics of the current and voltage over time. This fusion process helps the model capture the interaction and common change patterns between the current and voltage, which is crucial for accurately predicting the change of the harmonic current. On the basis of time sequence modeling and feature fusion, the model transmits the final timing features to the output layer through forward propagation to generate the prediction value. This prediction value is the harmonic current value at a future time. Output layer: The output layer of the model is usually a regression layer, whose output is a continuous numerical value representing the predicted harmonic current value. The prediction value reflects the possible change trend of the harmonic current at a future time.

[0044] In summary, the present application has at least the following effects:

[0045] The photovoltaic grid-connected harmonic current prediction method based on deep learning can effectively capture the time sequence characteristics and dynamic changes of harmonic current components and grid-connected point voltages, significantly improve the prediction accuracy of harmonic currents, and reduce the prediction error. The dynamic change characteristics are used for joint modeling, so that the model can comprehensively consider the mutual relationship between the current and the voltage, thereby improving the sensitivity and prediction ability of the system to the change of the harmonic current. Through the deep learning method for time sequence modeling, the automatic harmonic current prediction can be realized, the running state of the power grid can be reflected in real time, the data support can be provided for the dynamic monitoring and optimization of the power system, and the real-time response ability can be improved. The bidirectional long short-term memory network can capture the long-term and short-term dynamic changes of the current and the voltage, has strong adaptability, and can accurately predict the harmonic current under different load conditions and power grid environments. The convolutional neural network reduces the data dimension in the feature extraction stage, thereby reducing the computational complexity of the subsequent time sequence modeling, and making the model training and prediction process more efficient.

[0046] Those skilled in the art will appreciate that embodiments of the application can be provided as methods, systems, or computer program products. Accordingly, the application can be embodied in the form of complete hardware embodiments, complete software embodiments, or embodiments combining software and hardware aspects. Moreover, the application can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage, etc.) having computer-usable program code embodied therein.

[0047] The present application is described with reference to flowcharts and / or block diagrams of systems, apparatuses (systems), and computer program products according to embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, as well as combinations of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing apparatus to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing apparatus produce an apparatus that implements the flow Figure 1 The functions specified in a flow or multiple flows and / or blocks Figure 1 The functions specified in a flow or multiple flows and / or blocks

[0048] These computer program instructions can also be stored in a computer-readable memory that can direct the computer or other programmable data processing apparatus to work in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including instruction apparatus, which implements the flow Figure 1 The functions specified in a flow or multiple flows and / or blocks Figure 1 The functions specified in a flow or multiple flows and / or blocks

[0049] These computer program instructions can also be loaded into computer or other programmable data processing devices, so that a series of operations steps are performed on the computer or other programmable data processing devices to generate computer-implemented processes, thus the instructions executed on the computer or other programmable data processing devices provide the function of realizing the processes specified in the flowcharts Figure 1 one flowchart or multiple flowcharts and / or blocks Figure 1 one block or multiple blocks.

[0050] Although the preferred embodiments of the application have been described, those skilled in the art will be able to make additional modifications and variations to these embodiments without departing from the spirit and scope of the application. Accordingly, it is intended that the appended claims be construed to include all such modifications and variations as fall within the scope of the application.

[0051] Obviously, various modifications and changes are possible in the present application without departing from the spirit and scope of the application. It is to be understood that the application includes any such modifications and changes only insofar as they come within the scope of the appended claims and their equivalents.

Claims

1. A deep learning-based photovoltaic grid-connected harmonic current prediction method, characterized in that, The method comprises the following steps: S1. Obtain three-phase load current and point-of-connection voltage data in a photovoltaic grid-connected system; S2. Convert the three-phase load current into two-phase DC components through Clark transformation and Park transformation, extract the fundamental current of the two-phase DC components through a low-pass filter, and obtain the harmonic current component by subtracting the fundamental current from the total current of the three-phase load current, and pre-process the harmonic current component; S3. Extract the time-domain and frequency-domain features of the point-of-connection voltage data through fast Fourier transform, analyze the amplitude and frequency distribution of the voltage harmonic component, evaluate the correlation between the voltage harmonic and the load current harmonic, and obtain the point-of-connection voltage characteristics; the specific process of analyzing the amplitude and frequency distribution of the voltage harmonic component through fast Fourier transform is as follows: Perform fast Fourier transform on the point-of-connection voltage data to convert the time-domain data into frequency-domain data, extract the frequency components in the point-of-connection voltage signal, and identify the main frequency peaks of the voltage signal; Calculate the amplitude of each frequency component to evaluate the contribution of different frequencies; Analyze the harmonic components of the voltage signal, including the amplitude and frequency distribution of the fundamental wave and each harmonic, and obtain the harmonic characteristics of the voltage signal; According to the frequency domain analysis result, determine the energy distribution of each frequency band in the voltage signal; S4. Extract local features of the harmonic current component through a convolutional neural network, capture time series features of the harmonic current component through a bidirectional long short-term memory network, and build a harmonic current prediction model; S5. Input the pre-processed harmonic current component and the point-of-connection voltage characteristics into the harmonic current prediction model to generate a harmonic current prediction value.

2. The deep learning-based photovoltaic grid-connected harmonic current prediction method according to claim 1, characterized in that: The specific process of converting the three-phase load current into two-phase DC components through Clark transformation and Park transformation is as follows: Convert the three-phase load current into two-phase DC components through Clark transformation, and project the three-phase alternating current into a stationary coordinate system; Convert the two-phase DC components into DC components in a rotating coordinate system through Park transformation to obtain the DC components of the two-phase alternating current; Extract the fundamental current component in the DC component of the two-phase alternating current through a low-pass filter to remove high-frequency harmonic components.

3. The deep learning-based photovoltaic grid-connected harmonic current prediction method according to claim 2, characterized in that: The specific process of obtaining the harmonic current component by subtracting the fundamental current from the total current of the three-phase load current is as follows: According to the three-phase load current, calculate the total current of the three-phase load current; From the fundamental component extracted from the DC component of the two-phase alternating current, perform difference operation on the calculated total current of the three-phase load current and the extracted fundamental current to obtain the harmonic current component.

4. The deep learning-based photovoltaic grid-connected harmonic current prediction method according to claim 3, characterized in that: The specific process of pre-processing the harmonic current component is as follows: Denoising the harmonic current component to remove high-frequency noise components; Standardize and smooth the denoised harmonic current component to remove sudden changes and abnormal fluctuations; According to the time series characteristics of the harmonic current component, perform data interpolation and filling to handle missing values and ensure data integrity; Extract features in a local time period through window sliding method to generate time series data format required for training.

5. The deep learning-based photovoltaic grid-connected harmonic current prediction method according to claim 4, characterized in that: The specific process of evaluating the correlation between voltage harmonics and load current harmonics and obtaining the characteristics of the grid-connected point voltage is as follows: By analyzing the harmonic components of the voltage signal and the harmonic components of the load current, the similarity of the voltage signal and the load current in the frequency domain is compared, and the mutual relationship between the two in different frequency bands is evaluated. The correlation between the voltage harmonics and the load current harmonics is quantified, and according to the correlation analysis results, the correlation features between the voltage signal and the load current signal are extracted, and the characteristic vector of the grid-connected point voltage is obtained.

6. The deep learning-based photovoltaic grid-connected harmonic current prediction method according to claim 5, characterized in that: The specific process of extracting local features of harmonic current components by convolutional neural network is as follows: The harmonic current component time series data is input into the convolutional neural network, and the harmonic current component is extracted by the convolutional layer. Through the sliding of the convolution kernel in the local area, the local spatial features of the harmonic current component are captured, and the data dimension is reduced by the pooling layer to reduce the computational complexity. Through the convolutional layer and the pooling layer, the different levels of local features of the harmonic current component are gradually extracted, and the time series feature learning is represented.

7. The deep learning-based photovoltaic grid-connected harmonic current prediction method according to claim 6, characterized in that: The specific process of capturing the time series features of the harmonic current component by the bidirectional long short-term memory network is as follows: The local features of the harmonic current component extracted by the convolutional neural network are input into the bidirectional long short-term memory network. The bidirectional long short-term memory network models the time series data of the harmonic current component and captures the time series features in the time series. In the process of forward and backward propagation, the bidirectional long short-term memory network learns the forward and backward time series information respectively, and generates the forward hidden state vector and the backward hidden state vector. According to the forward hidden state vector and the backward hidden state vector, the hidden state matrix of the bidirectional time series information is generated.

8. The deep learning-based photovoltaic grid-connected harmonic current prediction method according to claim 7, characterized in that: The specific process of generating the harmonic current prediction value is as follows: The preprocessed harmonic current component and the grid-connected point voltage characteristics are input into the harmonic current prediction model, the time series modeling of the input harmonic current component and the grid-connected point voltage characteristics is performed, and the dynamic change characteristics of the current and voltage are fused to generate the prediction value of the harmonic current.

Images