Droop control grid-connected inverter island detection method and system based on deep learning

Abstract

The invention discloses an island detection method and system for a droop control grid-connected inverter based on deep learning, and relates to the technical field of micro-grid detection, and the method comprises the steps: obtaining the voltage of a common coupling point of the droop control grid-connected inverter and a power grid and an RMS signal of the output current of the inverter, extracting the total harmonic distortion (THD) of the voltage RMS signal and the current RMS signal in different scenes to construct a data set; and training and evaluating the constructed LSTM classifier by using the data set, and carrying out island detection on the THD parameters of the input voltage and current by using the LSTM classifier with the optimal learning parameter, so as to classify the input data as an island or non-island scene. The method can replace a traditional threshold value method, does not cause interference to the operation of the inverter, and has the advantages of high accuracy, high reliability and short detection time.

Description

Technical Field

[0001] This invention relates to the field of microgrid detection technology, and in particular to a method and system for detecting islanding in droop-controlled grid-connected inverters based on deep learning. Background Technology

[0002] Islanding detection in droop-controlled grid-forming (GFM) inverters faces challenges due to voltage source characteristics. Traditional islanding detection methods (IDM) primarily target grid-following inverters (GFLs) and can be categorized into active and passive methods. Active methods amplify voltage / frequency offsets by injecting harmonic currents or introducing feedback loops (such as positive feedback disturbances) to achieve detection. However, such disturbances fundamentally contradict the core function of GFMs in maintaining grid stability, often leading to reduced inverter stability in islanding mode. For example, feedback loops in reactive power control branches may degrade grid-connected power quality. Passive methods rely on characteristic thresholds such as Total Harmonic Distortion (THD) and frequency offset, or manually set thresholds, to determine islanding. However, their non-detection zone (NDZ) is susceptible to droop coefficients (M). p M q The system is affected by both load parameters and dynamic changes in system parameters. Conventional methods (such as wavelet transform DWT) fail to meet the standard requirements for NDZ in GFM. In addition, although the hybrid detection scheme combines the characteristics of active and passive methods, it requires instantaneous disturbance injection, which may lead to transient instability of the system.

[0003] Given the challenges of islanding detection in GFM inverters, it is necessary to propose some intelligent techniques to perform accurate islanding detection without interfering with GFM functionality, and to maintain robustness even under droop gain variations. Summary of the Invention

[0004] To address the aforementioned problems and technical requirements, the inventors propose a deep learning-based method and system for islanding detection in droop-controlled grid-connected inverters. Typically, inverters generate harmonics due to the combination of multiple controllers, non-ideal switching of power electronic devices, and pulse width modulation (PWM). Therefore, in the IDM proposed in this application, the voltage and current THD are extracted from the effective values ​​of voltage and current at the point of common coupling (PCC). Then, the prepared dataset is fed into an LSTM classifier for training. Ultimately, the LSTM classifier can classify the input data into islanded or non-islanded scenarios. The technical solution of this invention is as follows:

[0005] In a first aspect, this application provides a deep learning-based method for detecting islanding in droop-controlled grid-connected inverters, comprising the following steps:

[0006] The RMS signals of the common coupling point voltage and inverter output current of the droop-controlled grid-connected inverter and the grid are obtained, and the total harmonic distortion (THD) of the voltage RMS signal and current RMS signal under different scenarios is extracted to construct a dataset.

[0007] The constructed LSTM classifier is trained and evaluated using the dataset. The LSTM classifier with the best learning parameters is then used to perform island detection on the THD parameters of the input voltage and current to obtain the detection results.

[0008] A further technical solution involves extracting the total harmonic distortion (THD) of voltage RMS and current RMS signals under different scenarios, including:

[0009] The Fourier transform algorithm is used to perform harmonic analysis on the RMS signals of the common coupling point voltage and the inverter output current to obtain the voltage harmonics and current harmonics.

[0010] The sum of the squares of the effective values ​​of all voltage harmonic components is calculated as the ratio of the effective value of the fundamental voltage component, which is used as the THD parameter of the voltage RMS signal.

[0011] The sum of the squares of the effective values ​​of all current harmonic components is calculated as the ratio of the effective value of the fundamental current component, which is used as the THD parameter of the current RMS signal.

[0012] Its further technical solution is that the LSTM classifier includes:

[0013] The input layer is used to receive time series of cascaded THD parameters of voltage and current.

[0014] The LSTM layer is used to calculate the current hidden state and update the current storage cell state based on the current input and the previous hidden state.

[0015] In the LSTM classifier, the hidden state is connected to the softmax layer through a fully connected layer. The softmax layer acts as a classification layer to classify the output data into island or non-island scenarios.

[0016] Its further technical solution involves training and evaluating the constructed LSTM classifier using a dataset, including:

[0017] The dataset is divided into training and testing sets according to a certain ratio;

[0018] Initialize the LSTM classifier, train the LSTM classifier using the training set, and adjust the learning parameters of the LSTM classifier as training batches are completed.

[0019] The LSTM classifier trained is evaluated using a test set. If the highest accuracy and lowest loss are obtained, the learned parameters of the LSTM classifier trained in this session are taken as the optimal learned parameters. Otherwise, the LSTM classifier is retrained to update the learned parameters.

[0020] A further technical solution is that the method also includes:

[0021] The LSTM classifier performs backpropagation in an end-to-end manner during training, using cross-entropy loss as the loss function. The training objective is to minimize the cross-entropy loss of all training samples between the target distribution and the predicted response distribution.

[0022] Its further technical solution is that the optimal learning parameters of the LSTM classifier include:

[0023] The maximum number of iterations is 40, the number of neurons is 125, the number of iteration samples is 10, the gradient threshold is 1.0, and the learning rate is 0.01.

[0024] A further technical solution involves the following steps before training and evaluating the constructed LSTM classifier using the dataset:

[0025] The dataset is dimensionality reduced to decrease the computational burden on the LSTM classifier.

[0026] Secondly, this application also provides a deep learning-based droop control grid-connected inverter islanding detection system, which includes:

[0027] The data generation module is used to acquire the RMS signals of the common coupling point voltage and inverter output current of the droop-controlled grid-connected inverter and the grid, and to extract the total harmonic distortion (THD) of the voltage RMS signal and current RMS signal under different scenarios to construct a dataset.

[0028] The Intelligent IDM module is used to train and evaluate the constructed LSTM classifier using the dataset. It uses the LSTM classifier with the best learning parameters to perform island detection on the THD parameters of the input voltage and current, and obtains the detection results.

[0029] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the deep learning-based droop control grid-connected inverter islanding detection method described in the first aspect.

[0030] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the deep learning-based droop control grid-connected inverter islanding detection method described in the first aspect.

[0031] The beneficial technical effects of this invention are:

[0032] This invention provides an island detection method based on an LSTM classifier for droop-controlled GFM inverters. First, the total harmonic distortion (THD) of the voltage and current RMS signals at the PCC point is extracted. Then, a Long Short-Term Memory (LSTM) network is used to automatically learn the THD timing features, thus replacing the traditional thresholding mechanism without interfering with the inverter's operation. Compared with other existing classifiers, the proposed intelligent IDM exhibits higher accuracy, reliability, and faster detection time. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the LSTM-based IDM two-step method provided in this application.

[0034] Figure 2 This is a flowchart of the intelligent IDM provided in this application.

[0035] Figure 3 This is the grid-connected equivalent circuit diagram of the inverter provided in this application.

[0036] Figure 4 This is the architecture diagram of the LSTM model provided in this application.

[0037] Figure 5 This is a structural diagram of the droop control grid-connected inverter islanding detection system provided in this application. Detailed Implementation

[0038] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0039] One embodiment of this application provides a deep learning-based islanding detection method for droop-controlled grid-connected inverters. This method mainly includes a two-stage detection process. First, it acquires the RMS (effective value, also known as root mean square value) signal V of the PCC voltage at the common coupling point (PCC) between the droop-controlled grid-connected inverter and the grid. rms and the RMS signal I of the inverter output current rms And extract the voltage RMS signal V under different scenarios. rms and current RMS signal I rms Total harmonic distortion V THD and I THD To construct a dataset, the constructed LSTM classifier is then trained and evaluated using the dataset. The LSTM classifier with the optimal learning parameters is used to evaluate the THD parameters (VT) of the input voltage and current. THD and I THDIsland detection is performed to obtain the detection results. In this embodiment, different scenarios refer to island scenarios and non-island scenarios. A detection result of 1 corresponds to an island scenario, and a detection result of 0 corresponds to a non-island scenario.

[0040] In this embodiment, a Long Short-Term Memory (LSTM) network is used to automatically learn the temporal features of voltage and current THD parameters. This eliminates the need for traditional threshold selection or human intervention, and only requires high-dimensional data to achieve good performance. Furthermore, since it uses sequential input data, it does not require complex transformations of the input data to accurately classify the scenarios corresponding to the input features.

[0041] In one possible implementation, the proposed intelligent IDM encompasses the entire process from real-time data measurement using RTDS (Real-Time Digital Simulator), feature selection, dataset preparation, training, and testing of the LSTM classifier. For example... Figure 2 As shown, the implementation of the intelligent IDM program includes three stages: data generation stage, LSTM architecture design stage, and model parameter evaluation stage. The specific implementation content of each stage is described in detail below.

[0042] (1) Data generation stage: At the PCC of the droop-controlled grid-connected inverter and the grid, the RMS signals V of voltage and current are collected. rms and I rms And extract V in different scenarios rms and I rms V THD and I THD Features. In this embodiment, V is extracted. THD and I THD The principle of the feature is analyzed as follows:

[0043] Figure 3 The diagram shows the grid-connected equivalent circuit of the inverter. Common coupling point V. PCC and inverter current signal I IBR h-th harmonic component V PCC(h) and I IBR(h) The expression (1) in the grid-connected mode is as follows:

[0044]

[0045] Among them, Z g Z1 and Z2 represent the grid impedance and bus load, respectively.

[0046] When an island occurs, I IBR(h) It will flow along the high impedance path of the RLC load, instead of flowing into the grid in the manner of equation (2).

[0047]

[0048] Due to the high impedance path, V PCC(h) and I IBR(h) This generates huge harmonics, causing the THD value to increase sharply during islanding. Therefore, the THD parameters of voltage and current can be used to detect islanding scenarios.

[0049] And extract V THD and I THD The specific methods of feature analysis include: firstly, using the Fourier transform algorithm to perform harmonic analysis on the RMS signals of the PCC point voltage and the inverter output current to obtain the voltage harmonics and current harmonics; then, calculating the ratio of the sum of the squares of the effective values ​​of all voltage harmonic components to the effective value of the voltage fundamental component, as the THD parameter of the voltage RMS signal, as shown in equation (3); similarly, calculating the ratio of the sum of the squares of the effective values ​​of all current harmonic components to the effective value of the current fundamental component, as the THD parameter of the current RMS signal, as shown in equation (4).

[0050]

[0051] Among them, the harmonic components are represented by V. h I h The fundamental components are represented by V1 and I1.

[0052] In the above implementation, the THD parameters of the PCC point voltage and current signals are extracted, and the characteristic abrupt changes under high-impedance paths are quantified by Fourier transform, which is beneficial to improving the effectiveness of the subsequent LSTM classifier in distinguishing different scenarios. The extracted V THD and I THD Features are formed into a concatenated dataset, serving as a 2*19999 matrix, to detect isolated or non-isolated scenes. Then, the prepared dataset undergoes dimensionality reduction, for example, reducing the final dimension to a 2*8000 matrix composed of pre-event and post-event data for 10 periods, thereby reducing the computational burden on the LSTM classifier. Common dimensionality reduction methods such as Principal Component Analysis (PCA) can be used.

[0053] (2) LSTM Architecture Design Stage: The LSTM classifier belongs to the category of Recurrent Neural Networks (RNNs), using hidden layers as memory units, which is helpful for processing long-term and short-term sequence data. An RNN is a special type of Artificial Neural Network (ANN), composed of internal network loops. In an RNN, the hidden layers are formed where the output of the previous state is feedback to the current state of the hidden unit; therefore, the output at each step depends on the output of the previous time step. A schematic diagram of the LSTM RNN model is shown below. Figure 4As shown, its main elements consist of gates and storage units, used to control the flow of information and remember the temporary states of the neural network, respectively. The main advantage of LSTM is maintaining a constant error flow. An LSTM unit reserves an encoded memory block that can remember long-term time dependencies to ensure a constant error flow. Input gate I t And the Gate of Oblivion F t Controls the selection of information entering the storage unit and the decision of whether to forget or retain information in the unit state. Information moves to the next hidden state H. t The transfer is handled by output gate O t The decision is made. An LSTM cell has a tanh layer, which is part of the cell state and used for updates. The computation flow of LSTM is shown in equations (5)-(10) below:

[0054] I t =σ(V i H t-1 +U i x t +b i (5)

[0055] F t =σ(V f H t-1 +U f x t +b f (6)

[0056] O t =σ(V o H t-1 +U o x t +b o (7)

[0057] C t =F t ⊙C t-1 +I t ⊙C * t (8)

[0058] C * t =tan h(V) c H t-1 +U c x t +b c (9)

[0059] H t =O t ⊙tanh(C t (10)

[0060] When using the tanh function, the cell state changes; however, the gate uses the sigmoid activation function. The forget gate F at each time step... t Primarily through a previous hidden state H t-1 and a new input state x t Obtained. Last storage unit C t-1 Information can be obtained from the forget gate (value 1), and vice versa. Next, from the input gate I... t The new input function is obtained, and the previous hidden state is added to the storage unit C. t Finally, hide the output state H. t Or the output gate will draw from memory cell C t Determine the new hidden state H t The transformations and predictions in neural networks are accomplished using the sigmoid function and vectors with the operator (⊙) followed by basic multiplication. The learned parameters include modeling the intrinsic parameters (V) of the RNN. i V f V o V c ) and (U i U f U o U c The weights and bias vectors (b) i b f b o b c By minimizing the objective function, the LSTM model can update the weights and biases.

[0061] The LSTM classifier constructed in this embodiment includes an input layer, an LSTM layer, a fully connected layer, and a classification layer connected in sequence. The input layer receives the time series of cascaded THD parameters of voltage and current. The LSTM layer, based on the aforementioned LSTMRNN model, calculates the current hidden state and updates the current storage cell state based on the current input and the previous hidden state. The LSTM layer automatically extracts island features. Finally, the fully connected layer and the softmax layer are fused. The softmax layer, acting as the classification layer, categorizes the output data into island or non-island scenarios; that is, the softmax layer maps the results to probability values ​​between 0 and 1, and sums all output probability values ​​to 1.

[0062] In one possible implementation, the dimensionality-reduced dataset is divided into training and test sets according to a certain ratio. An LSTM classifier is initialized; that is, the learning parameters are selected in the proposed IDM LSTM model. Then, the LSTM classifier is trained using the training set to achieve island detection. The learning parameters of the LSTM classifier are adjusted with each training batch to obtain maximum accuracy and minimum loss. The learning parameters include the maximum number of iterations, the number of neurons, the number of iteration samples (10), the gradient threshold (1.0), and the learning rate.

[0063] Furthermore, the LSTM classifier performs backpropagation end-to-end during training, using cross-entropy loss as the loss function, and assuming Y is the target distribution value, Y * To predict the response distribution, the training objective is to combine Y and Y * The cross-entropy loss of all training samples is minimized, as shown in Equation (11):

[0064]

[0065] (3) Model Parameter Evaluation Stage: The performance of the trained LSTM classifier is evaluated using the test set, and the LSTM model is experimentally tested to achieve maximum accuracy. If the highest accuracy and lowest loss can be obtained in this step, the learned parameters of the LSTM classifier obtained in this training are taken as the optimal learning parameters; otherwise, the LSTM classifier needs to be retrained to update the learning parameters until the maximum accuracy is achieved. The optimal learning parameters obtained by selecting the values ​​of the training and testing parameters are shown in Table 1.

[0066] Table 1. Values ​​of learning parameters in intelligent IDM

[0067] Learning parameters Optimal parameters combination Maximum number of iterations 40 10-20-40-60 Number of neurons 125 16-64-125 Number of iterative samples 10 10-16-32-64 Gradient threshold 1.0 0.5-1.0-2.0 Learning rate 0.01 0.001-0.1

[0068] Therefore, the optimal LSTM model parameters for distinguishing between isolated and non-isolated events can be obtained.

[0069] Based on the same inventive concept, another embodiment of this application provides a droop control grid-connected inverter islanding detection system based on deep learning. The solution provided by this system is similar to the solution described in the above method. Therefore, the specific limitations of one or more smart IDM system embodiments provided below can be found in the limitations of the smart IDM method above, and will not be repeated here.

[0070] In one possible implementation, such as Figure 5 As shown, the system includes a data generation module and an intelligent IDM module, wherein: the data generation module is used to acquire the RMS signal V of the PCC voltage at the common coupling point between the droop-controlled grid-connected inverter and the grid. rmsand the RMS signal I of the inverter output current rms And extract the voltage RMS signal V under different scenarios. rms and current RMS signal I rms Total harmonic distortion V THD and I THD This is used to construct a dataset. The Intelligent IDM module is used to train and evaluate the constructed LSTM classifier using the dataset, employing the LSTM classifier with optimal learning parameters to evaluate the THD parameters (V) of the input voltage and current. THD and I THD Island detection is performed to obtain the detection results.

[0071] Each module in the aforementioned intelligent IDM system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of a computer device in hardware form or independent of it, or stored in the memory of the computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0072] In one embodiment, a computer device is also provided. This computer device can be a terminal, including a processor, memory, communication interface, display screen, and input device connected via a system bus. The processor of the computer device provides computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The communication interface of the computer device is used for wired or wireless communication with external terminals. Wireless communication can be achieved through Wi-Fi, mobile cellular networks, NFC (Near Field Communication), or other technologies. When the computer program is executed by the processor, it implements a deep learning-based droop control grid-connected inverter islanding detection method. The display screen of the computer device can be a liquid crystal display (LCD) or an e-ink display. The input device of the computer device can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.

[0073] In one embodiment, a computer-readable storage medium is also provided, on which a computer program is stored, which, when executed by a processor, implements the steps in the above method embodiments.

[0074] The above descriptions are merely preferred embodiments of this application, and the present invention is not limited to the above embodiments. It is understood that other improvements and variations directly derived or conceived by those skilled in the art without departing from the spirit and concept of the present invention should be considered to be included within the protection scope of the present invention.

Claims

1. A method for detecting islanding in a droop-controlled grid-connected inverter based on deep learning, characterized in that, The method includes: The RMS signals of the common coupling point voltage and inverter output current of the droop-controlled grid-connected inverter and the grid are obtained, and the total harmonic distortion (THD) of the voltage RMS signal and current RMS signal under different scenarios is extracted to construct a dataset. The constructed LSTM classifier is trained and evaluated using the dataset. The LSTM classifier with the best learning parameters is then used to perform island detection on the THD parameters of the input voltage and current to obtain the detection results.

2. The method for detecting islanding in a droop-controlled grid-connected inverter based on deep learning according to claim 1, characterized in that, The extraction of the total harmonic distortion (THD) of voltage RMS and current RMS signals under different scenarios includes: The RMS signals of the common coupling point voltage and the inverter output current are analyzed for harmonics using the Fourier transform algorithm to obtain the voltage harmonics and current harmonics. The sum of the squares of the effective values ​​of all voltage harmonic components is calculated as the ratio of the effective value of the fundamental voltage component, which is used as the THD parameter of the voltage RMS signal. The sum of the squares of the effective values ​​of all current harmonic components is calculated as the ratio of the effective value of the fundamental current component, which is used as the THD parameter of the current RMS signal.

3. The method for detecting islanding in a droop-controlled grid-connected inverter based on deep learning according to claim 1, characterized in that, The LSTM classifier includes: The input layer is used to receive the time series of the cascaded THD parameters of the voltage and current; The LSTM layer is used to calculate the current hidden state and update the current storage cell state based on the current input and the previous hidden state. The hidden states in the LSTM classifier are connected to the softmax layer through a fully connected layer. The softmax layer acts as a classification layer to classify the output data into island or non-island scenarios.

4. The method for detecting islanding in a droop-controlled grid-connected inverter based on deep learning according to claim 1, characterized in that, Training and evaluating the constructed LSTM classifier using the dataset includes: The dataset is divided into a training set and a test set according to a certain ratio; Initialize the LSTM classifier, train the LSTM classifier using the training set, and adjust the learning parameters of the LSTM classifier as training batches are performed; The LSTM classifier trained is evaluated using the test set. If the maximum accuracy and minimum loss are obtained, the learning parameters of the LSTM classifier trained in this session are taken as the optimal learning parameters. Otherwise, the LSTM classifier is retrained to update the learning parameters.

5. The method for detecting islanding in a droop-controlled grid-connected inverter based on deep learning according to claim 4, characterized in that, The method further includes: The LSTM classifier is trained by backpropagation in an end-to-end manner, using cross-entropy loss as the loss function. The training objective is to minimize the cross-entropy loss of all training samples between the target distribution and the predicted response distribution.

6. The method for islanding detection in a droop control grid-connected inverter based on deep learning according to claim 4, characterized in that, The optimal learning parameters for the LSTM classifier include: The maximum number of iterations is 40, the number of neurons is 125, the number of iteration samples is 10, the gradient threshold is 1.0, and the learning rate is 0.

01.

7. The method for detecting islanding in a droop-controlled grid-connected inverter based on deep learning according to claim 1, characterized in that, Before training and evaluating the constructed LSTM classifier using the dataset, the method further includes: The dataset is dimensionality reduced to decrease the computational burden on the LSTM classifier.

8. A droop control grid-connected inverter islanding detection system based on deep learning, characterized in that, The system includes: The data generation module is used to acquire the RMS signals of the common coupling point voltage and inverter output current of the droop-controlled grid-connected inverter and the grid, and to extract the total harmonic distortion (THD) of the voltage RMS signal and current RMS signal under different scenarios to construct a dataset. The intelligent IDM module is used to train and evaluate the constructed LSTM classifier using the dataset. The LSTM classifier with the best learning parameters is used to perform island detection on the THD parameters of the input voltage and current to obtain the detection results.

9. A computer device, characterized in that, The device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the deep learning-based droop control grid-connected inverter islanding detection method as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the deep learning-based droop control grid-connected inverter islanding detection method as described in any one of claims 1-7.

Images