A micro-grid island detection method and system based on multi-modal feature cross fusion
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HUAZHONG UNIV OF SCI & TECH
- Filing Date
- 2026-04-13
- Publication Date
- 2026-06-12
Smart Images
Figure CN122196836A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of power system online protection technology, and more specifically, relates to a microgrid islanding detection method and system based on multimodal feature cross-fusion. Background Technology
[0002] While the widespread application of distributed generation and the large-scale deployment of microgrids have significantly promoted decarbonization and improved energy efficiency, they have also brought serious security challenges, with the islanding effect remaining a major safety hazard. The islanding effect refers to the phenomenon where, when the main power grid is disconnected due to a fault or outage, distributed generation in a microgrid fails to disconnect from the grid in time and continues to supply power to the load, forming an uncontrolled, self-sufficient power supply island system. This poses a serious threat to the safety of maintenance personnel, power quality, and equipment lifespan.
[0003] Existing islanding detection methods can be broadly categorized into three types. The first type employs passive detection methods, such as over / under voltage detection and frequency change rate detection. While simple to implement, these methods suffer from significant non-detectable zones (NDZ) under conditions of low power matching, making them prone to failure. The second type uses active detection methods, such as frequency offset injection. While reducing NDZ, these methods introduce additional harmonic pollution, impacting power quality. The third type utilizes artificial intelligence-based detection methods, such as those employing convolutional neural networks and long short-term memory networks. These methods achieve end-to-end automatic feature extraction, reducing NDZ without introducing additional harmonic pollution, resulting in good performance. However, they also suffer from limitations such as single-mode feature extraction, making it difficult to capture various subtle features, while the islanding effect is a condition that changes both in time and frequency. Furthermore, most of these methods only achieve good detection results for specific operating conditions. When power quality disturbances such as capacitor switching, short-circuit faults, or load switching occur, they are easily misjudged as islanding, leading to unnecessary downtime and exhibiting poor robustness. Summary of the Invention
[0004] In view of the above-mentioned defects or improvement needs of existing technologies, the present invention provides a microgrid island detection method and system based on multimodal feature cross-fusion, which is used to solve the technical problem of poor robustness of existing artificial intelligence-based detection methods.
[0005] To achieve the above objectives, in a first aspect, the present invention provides a microgrid islanding detection method based on multimodal feature cross-fusion, comprising: acquiring voltage and current signals at the common coupling point of the microgrid, and performing the following islanding detection operation: The voltage signal and the current signal are sampled separately to obtain multiple voltage time-series data samples and multiple current time-series data samples; Time-frequency transformation is performed on each voltage time-series data sample to obtain the corresponding two-dimensional time-frequency voltage image; time-frequency transformation is performed on each current time-series data sample to obtain the corresponding two-dimensional time-frequency current image. Each voltage time series data sample is processed into a non-overlapping slice to obtain the corresponding time series voltage slice sequence; each current time series data sample is processed into a non-overlapping slice to obtain the corresponding time series current slice sequence. Feature extraction was performed on all two-dimensional time-frequency voltage and two-dimensional time-frequency current images to obtain local features containing local texture information; Feature extraction is performed on all time-series voltage slice sequences and time-series current slice sequences to obtain global features containing global dependency information; Based on the cross-attention mechanism, local features and global features are fused to obtain local-global fused features; By inputting the local and global fusion features into the classifier, the operating status detection results of the microgrid are obtained, so as to realize island detection.
[0006] More preferably, the MSRCN network is used to extract features from all two-dimensional time-frequency voltage images and two-dimensional time-frequency current images to obtain local features containing local texture information; The MSRCN network comprises: cascaded multi-scale depthwise separable convolutional layers, splicing layers, channel attention modules, coordinate attention modules, residual modules, and fully connected layers; the output of the channel attention modules is also connected to the input of the fully connected layers. Multi-scale depth-separable convolutional layers are used to extract multi-scale features from all two-dimensional time-frequency voltage images and two-dimensional time-frequency current images, resulting in feature images at multiple scales. The stitching layer is used to stitch feature images of multiple scales together along the channel dimension to obtain a stitched feature map; The channel attention module is used to extract features from the spliced feature map based on the channel attention mechanism to obtain the channel fusion feature map; The coordinate attention module is used to extract features from the channel fusion feature map based on the coordinate attention mechanism, and obtain the pixel fusion feature map; The residual module is used to perform residual processing on the pixel fusion feature map; The fully connected layer is used to flatten and project the pixel-by-pixel feature map and the feature map after residual processing, resulting in local features containing local texture information.
[0007] More preferably, the MSRCN network further includes a convolutional pooling layer set between the channel attention module and the coordinate attention module, which is used to perform dimensionality reduction convolution and pooling operations on the channel fusion feature map and then output it to the coordinate attention module.
[0008] More preferably, the above-mentioned cross-attention mechanism is a bidirectional cross-attention mechanism.
[0009] More preferably, local-global fusion features for:
[0010]
[0011]
[0012]
[0013]
[0014] in, These are learnable fusion weight coefficients; and All are intermediate features; It is a normalization operation; This is a local query matrix; ; Local features; A learnable projection matrix that maps local features onto the query matrix; This represents an intermediate fusion feature from local to global perspectives. This is the global key matrix; ; For global features; A learnable projection matrix that maps global features onto the key matrix; To preset the number of attention points; This is a global value matrix; ; A learnable projection matrix that maps global features onto a value matrix; For global query matrix; ; A learnable projection matrix that maps global features onto the query matrix; This is an intermediate fusion feature from global to local; It is a local bond matrix; ; This is a learnable projection matrix that maps local features onto the key matrix; It is a local value matrix; ; This is a learnable projection matrix that maps local features onto a value matrix.
[0015] More preferably, a Transformer network is used to extract features from all time-series voltage slice sequences and time-series current slice sequences to obtain global features containing global dependency information.
[0016] More preferably, the above classifier is trained in the following manner: Obtain the training set; wherein, the training set includes: training samples and corresponding labels of the microgrid in different operating states; the training samples include: voltage and current signals at the common coupling point of the microgrid; the labels are the corresponding microgrid operating states; the operating state categories include normal operating state, islanded operating state, and non-islanded disturbance state; For each training sample in the training set, an island detection operation is performed to obtain the corresponding running state detection result; the classifier is trained by minimizing the difference loss between the running state detection result and the corresponding label.
[0017] More preferably, while training the classifier by minimizing the difference loss between the running state detection result and the corresponding label, the learnable parameters involved in the local feature extraction, global feature extraction, and local-global fusion feature extraction processes in the island detection operation are also trained.
[0018] More preferably, the above-mentioned islanded operation state categories include: zero power mismatch state, active power surplus state, active power shortage state, reactive power surplus state, reactive power shortage state, frequency drift state, mixed power mismatch state, and slow frequency drift state. The above-mentioned non-islanding disturbance state categories include: three-phase unbalance state, harmonic pollution state, voltage sag state, voltage swell state, short circuit fault state, capacitor switching state, and load change state.
[0019] In a second aspect, the present invention provides a microgrid islanding detection system, comprising: a memory and a processor, wherein the memory stores a computer program, and the processor executes the microgrid islanding detection method provided in the first aspect of the present invention when executing the computer program.
[0020] Thirdly, the invention also provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the microgrid islanding detection method provided in the first aspect of the invention.
[0021] In summary, the above-described technical solutions conceived in this invention can achieve the following beneficial effects: 1. This invention provides a microgrid islanding detection method based on multimodal feature cross-fusion. The voltage and current signals at the common coupling point of the microgrid are processed into corresponding two-dimensional time-frequency images and time-series slice sequences. Feature extraction is performed on the two-dimensional time-frequency images and time-series slice sequences to obtain local features containing local texture information and global features containing global dependency information. Then, based on a cross-attention mechanism, the local and global features are fused to obtain a local-global fused feature. This local-global fused feature retains sensitivity to local texture and the ability to capture long-term dependencies, accurately characterizing the islanding features of the microgrid. It effectively solves the problem of traditional artificial intelligence methods misjudging non-islanding disturbance states with power quality disturbances as islanding operation states, achieving accurate islanding detection. Simultaneously, this invention maintains the accuracy of islanding detection even in scenarios with additional disturbance signals (such as power quality disturbances), exhibiting good robustness and strong noise resistance.
[0022] 2. The microgrid islanding detection method provided by this invention has a small non-detection area and does not introduce additional disturbance signals (such as harmonic pollution).
[0023] 3. Furthermore, the microgrid islanding detection method provided by this invention employs an MSRCN network to extract features from all two-dimensional time-frequency voltage and current images, obtaining local features containing local texture information. The multi-scale depth-separable convolutional layers in the MSRCN network balance the receptive field and resolution of local features while maximizing computational efficiency. In a microgrid, different disturbance events exhibit drastically different texture features on the time-frequency images. For example, transient high-frequency faults appear as sharp, fine edges, while steady-state low-frequency disturbances appear as broad, blurred color block evolutions. Unlike a single convolutional kernel that either loses low-frequency feature patterns or blurs high-frequency edge features, the multi-scale depth-separable convolutional layer uses multiple convolutional kernels with different receptive fields to analyze the power graph. Parallel feature extraction enables the network to accurately perceive power quality disturbances at multiple scales, further improving island detection accuracy. Meanwhile, considering that in a time-frequency image, the horizontal axis represents time and the vertical axis represents frequency, traditional attention mechanisms directly average time and frequency information, causing the network to fail to pinpoint the time and frequency band of the fault. The coordinate attention module in the MSRCN network, based on a coordinate attention mechanism, aggregates features along both horizontal and vertical directions, accurately capturing the energy distribution location of specific frequency bands in a two-dimensional time-frequency image and accurately capturing local texture information in the two-dimensional time-frequency image. Through this dual-axis decoupling mechanism, the network can assign high weights to the coordinate points where the fault occurs and low weights to background noise areas, achieving strong noise resistance and further improving island detection accuracy.
[0024] 4. Furthermore, the microgrid islanding detection method provided by this invention is based on a bidirectional cross-attention mechanism, which fuses local features and global features. Features are only activated when the semantics of local features and global features are consistent, thereby effectively filtering out one-sided noise interference, exhibiting strong noise resistance and maintaining high availability even under strong noise.
[0025] 5. Furthermore, the microgrid islanding detection method provided by this invention introduces learnable fusion weight coefficients when fusing local and global features based on a bidirectional cross-attention mechanism. , It acts like a switch when the fault is a fast, transient fluctuation problem. It can learn higher values with a greater probability, allowing the detection process to focus on local texture features; when the fault is a slow, dynamic change, It can learn smaller values with a higher probability, allowing the detection process to focus on global trends, thus making the detection process more adaptable and further improving detection accuracy.
[0026] 6. Furthermore, the microgrid islanding detection method provided by the present invention trains a classifier using data samples from 16 different operating states of the microgrid. This not only distinguishes between islanded and non-islanded operating conditions, but also differentiates between the 16 specific operating conditions, enabling more granular detection of the microgrid's operating state.
[0027] 7. The microgrid islanding detection method provided by this invention is conducive to quickly identifying fault information, providing a basis for rapid maintenance or fault-tolerant control of distributed generation systems, and improving the reliability of the system. Attached Figure Description
[0028] Figure 1 This is a schematic diagram of the microgrid island detection method based on multimodal feature cross-fusion provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of signal processing for voltage and current signals at the common coupling point of a microgrid, provided in an embodiment of the present invention. Figure 3 This is a schematic diagram of the microgrid islanding detection framework provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the microgrid grid-connected power generation system provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the confusion matrix of the classification results of 16 working conditions by the method provided in this invention; Figure 6 This is a comparison chart of the detection accuracy of the method provided by this invention and existing technologies; Figure 7This is a schematic diagram illustrating the accuracy of the method provided by the present invention under different signal-to-noise ratio noise environments. Detailed Implementation
[0029] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.
[0030] To achieve the above objectives, in a first aspect, the present invention provides a microgrid islanding detection method based on multimodal feature cross-fusion, comprising: acquiring voltage and current signals at the common coupling point of the microgrid, and performing the following islanding detection operation: The voltage signal and the current signal are sampled separately to obtain multiple voltage time-series data samples and multiple current time-series data samples; Time-frequency transformation is performed on each voltage time-series data sample to obtain the corresponding two-dimensional time-frequency voltage image; time-frequency transformation is performed on each current time-series data sample to obtain the corresponding two-dimensional time-frequency current image. Each voltage time series data sample is processed into a non-overlapping slice to obtain the corresponding time series voltage slice sequence; each current time series data sample is processed into a non-overlapping slice to obtain the corresponding time series current slice sequence. Feature extraction was performed on all two-dimensional time-frequency voltage and two-dimensional time-frequency current images to obtain local features containing local texture information; Feature extraction is performed on all time-series voltage slice sequences and time-series current slice sequences to obtain global features containing global dependency information; Based on the cross-attention mechanism, local features and global features are fused to obtain local-global fused features; By inputting the local and global fusion features into the classifier, the operating status detection results of the microgrid are obtained, so as to realize island detection.
[0031] It should be noted that feature extraction for all two-dimensional time-frequency voltage and current images can be performed using any existing feature extraction model, such as CNN, VGG, YOLO, autoencoders, etc., and no limitation is made here.
[0032] Preferably, in one optional implementation, the MSRCN network is used to extract features from all two-dimensional time-frequency voltage images and two-dimensional time-frequency current images to obtain local features containing local texture information; The MSRCN network comprises: cascaded multi-scale depthwise separable convolutional layers, splicing layers, channel attention modules, coordinate attention modules, residual modules, and fully connected layers; the output of the channel attention modules is also connected to the input of the fully connected layers. Multi-scale depth-separable convolutional layers are used to extract multi-scale features from all two-dimensional time-frequency voltage images and two-dimensional time-frequency current images, resulting in feature images at multiple scales. The stitching layer is used to stitch feature images of multiple scales together along the channel dimension to obtain a stitched feature map; The channel attention module is used to extract features from the spliced feature map based on the channel attention mechanism to obtain the channel fusion feature map; The coordinate attention module is used to extract features from the channel fusion feature map based on the coordinate attention mechanism, and obtain the pixel fusion feature map; The residual module is used to perform residual processing on the pixel fusion feature map to prevent gradient vanishing; The fully connected layer is used to flatten and project the pixel-by-pixel feature map and the feature map after residual processing, resulting in local features containing local texture information.
[0033] Preferably, the MSRCN network further includes a convolutional pooling layer between the channel attention module and the coordinate attention module. This layer is used to perform dimensionality reduction convolution and pooling operations on the channel fusion feature map obtained by the channel attention module, and then input it into the coordinate attention module, so as to solve the problem of loss of position information in global pooling in traditional methods.
[0034] It should be noted that the above-mentioned cross-attention mechanism can be a unidirectional (standard) cross-attention mechanism or a bidirectional cross-attention mechanism, and no limitation is made here. Preferably, in an optional implementation, the above-mentioned cross-attention mechanism is a bidirectional cross-attention mechanism.
[0035] Preferably, in one optional implementation, local-global fusion features for:
[0036]
[0037]
[0038]
[0039]
[0040] in, These are learnable fusion weight coefficients; and All are intermediate features; It is a normalization operation; This is a local query matrix; ; Local features; A learnable projection matrix that maps local features onto the query matrix; This represents an intermediate fusion feature from local to global perspectives. This is the global key matrix; ; For global features; A learnable projection matrix that maps global features onto the key matrix; To preset the number of attention points; This is a global value matrix; ; A learnable projection matrix that maps global features onto a value matrix; For global query matrix; ; A learnable projection matrix that maps global features onto the query matrix; This is an intermediate fusion feature from global to local; It is a local bond matrix; ; This is a learnable projection matrix that maps local features onto the key matrix; It is a local value matrix; ; This is a learnable projection matrix that maps local features onto a value matrix.
[0041] Compared to conventional bidirectional cross-attention mechanisms, this implementation uses learnable fusion weight coefficients to dynamically balance the weight contributions of local and global features to fault feature information. During the fusion process... It acts like a switch; when the fault is a fast transient fluctuation problem, the parameters... It is highly likely that the model will learn higher values, therefore emphasizing that the model focuses on local texture features; when the fault is a slow, dynamic change, the parameters... It is highly likely that the model will learn smaller values, therefore the emphasis is on the model focusing on the global trend.
[0042] It should be noted that feature extraction for all time-series voltage slice sequences and time-series current slice sequences can be performed using any existing feature extraction model, such as Transformer, LSTM, RNN, etc., and is not limited here. Preferably, in one optional implementation, a Transformer network is used to extract features from all time-series voltage slice sequences and time-series current slice sequences to obtain global features containing global dependency information.
[0043] It should be noted that the classifier, as well as the models used for local feature extraction, global feature extraction, and local-global fusion feature extraction in this invention, can be existing pre-trained models, or they can be obtained through step-by-step training or end-to-end training. No limitation is made here.
[0044] In one alternative implementation, the classifier described above is trained in the following manner: Obtain the training set; wherein, the training set includes: training samples and corresponding labels of the microgrid in different operating states; the training samples include: voltage and current signals at the common coupling point of the microgrid; the labels are the corresponding microgrid operating states; the operating state categories include normal operating state, islanded operating state, and non-islanded disturbance state; For each training sample in the training set, an island detection operation is performed to obtain the corresponding running state detection result; the classifier is trained by minimizing the difference loss between the running state detection result and the corresponding label.
[0045] Preferably, in one optional implementation, while training the classifier by minimizing the difference loss between the running state detection result and the corresponding label, the learnable parameters involved in the local feature extraction, global feature extraction, and local-global fusion feature extraction processes in the island detection operation are also trained.
[0046] Preferably, in one optional implementation, the above-mentioned islanded operation state categories include: zero power mismatch state, active power surplus state, active power shortage state, reactive power surplus state, reactive power shortage state, frequency drift state, mixed power mismatch state, and slow frequency drift state. The above-mentioned non-islanding disturbance state categories include: three-phase unbalance state, harmonic pollution state, voltage sag state, voltage swell state, short circuit fault state, capacitor switching state, and load change state.
[0047] Under the above implementation method, it is not only possible to accurately distinguish between islanded and non-islanded systems, but also to identify operating conditions under power quality disturbances.
[0048] It should be noted that the classifier in this invention can be any type of classifier, such as SVM, softmax, MLP, fully connected layer, etc., and there is no limitation here.
[0049] To further illustrate the microgrid islanding detection method based on multimodal feature cross-fusion provided by this invention, a specific embodiment is described in detail below: like Figure 1 The diagram shown is a flowchart of a microgrid islanding detection method based on multimodal feature cross-fusion provided in this embodiment. The following section will discuss the method in conjunction with... Figure 2 and Figure 3The steps of this method are explained in detail: S1. Obtain the voltage and current signals at the common coupling point of the microgrid. , , , , , Then, the voltage signal and the current signal are sampled respectively to obtain multiple voltage time series data samples and multiple current time series data samples. Specifically, in this embodiment, for voltage / current signals Sliding window sampling is performed using the following formula:
[0050] in, It is a time-series signal segment after sliding window sampling. The sampling window size, For the sampling step size, For time steps, For the first One sample segment.
[0051] S2. Perform time-frequency transformation on each voltage time-series data sample to obtain the corresponding two-dimensional time-frequency voltage image; perform time-frequency transformation on each current time-series data sample to obtain the corresponding two-dimensional time-frequency current image; Each voltage time series data sample is processed into a non-overlapping slice to obtain the corresponding time series voltage slice sequence; each current time series data sample is processed into a non-overlapping slice to obtain the corresponding time series current slice sequence. Specifically, in this embodiment, in the obtained voltage / current timing data samples The 2D time-frequency plot and patch segments are obtained using the following formula:
[0052]
[0053]
[0054] in, It is the Morlet wavelet transform function; Scale factor; The translation factor; Indicates complex conjugation. The bilinear interpolation operator fixes the size of the resulting image. This is a two-dimensional time-frequency feature map with dimensions of . ; It is a time-series slice sequence. The length of the patch. Indicates the index range for the time dimension.
[0055] S3. This embodiment constructs a two-stream feature extraction network, such as... Figure 3 As shown, it includes a parallel first branch network MSRCN (i.e., multi-scale residual coordinate network) and a second branch network Transformer; the MSRCN network is used to extract features from all two-dimensional time-frequency voltage images and two-dimensional time-frequency current images to obtain local features containing local texture information; the Transformer network is used to extract features from all time-series voltage slice sequences and time-series current slice sequences to obtain global features containing global dependency information. The MSRCN network includes a multi-scale depthwise separable convolutional module to balance the receptive field and resolution of local features while maximizing computational efficiency. It also includes a coordinate attention module to aggregate features along both horizontal and vertical directions to accurately capture the energy distribution of specific frequency bands in the time-frequency map. Through multiple experiments, this embodiment has been verified to employ… , , When using a combination of multi-scale convolutional kernels, the model achieves optimal capture of high-frequency transient noise in microgrids.
[0056] Specifically, all two-dimensional time-frequency voltage images and two-dimensional time-frequency current images are batch-input into the MSRCN network. After passing through multi-scale depth-separable convolutional layers and splicing layers, the resulting features are:
[0057] in, A batch image package for all two-dimensional time-frequency voltage images and two-dimensional time-frequency current images; These represent three depthwise separable convolutional kernels. It is a non-linear activation function. This is a feature splicing operation.
[0058] Subsequently, to address the issue of lost positional information during global pooling in traditional methods, this embodiment performs simultaneous pooling on both the time and frequency axes using a coordinate attention module after dimensionality reduction convolution and pooling of the concatenated features.
[0059]
[0060] in, This indicates that for a height of The first Average pooling is performed on each channel. This indicates that the width is The first Average pooling is performed on each channel.
[0061] The above operations yielded features in two directions. The two feature maps were then concatenated and transformed to obtain the attention weights.
[0062] in, Indicates splicing along spatial dimensions. It is a non-linear activation function. It is the sigmoid activation function. and These are the attention weights in the two directions.
[0063] Finally, the attention weights are applied to the original feature map to obtain a local feature output with orientation awareness. :
[0064] This MSRCN network can accurately pinpoint the specific local texture of fault features on the time-frequency map.
[0065] In this embodiment, the temporal slice sequence is mapped to the embedding vector sequence in the Transformer network, and then the long-range dependencies of global features in the temporal data sample are captured by a multi-head self-attention encoder.
[0066] Specifically, for all time-series voltage slice sequences and time-series current slice sequences, the input signals of the Transformer encoder are obtained through linear projection mapping and position encoding. :
[0067] in, For linear projection layers, For all time-series voltage slice sequences and time-series current slice sequences, the first... A time-series slice sequence For position encoding.
[0068] For the input sequence First, it requires three linear transformation matrices. , , Generate query, key, and value matrix:
[0069] This invention employs a multi-head attention mechanism, utilizing multiple single-head attention... Used to simultaneously monitor comprehensive information such as frequency, amplitude, and phase:
[0070] Finally, after stacking multiple Transformer encoders, the output sequence is... Including context-aware features for each patch, to further integrate and generate the final global feature vector, the sequence is flattened and then linearly projected onto the same feature dimension as the MSRCN part:
[0071] This Transformer network can be used to determine the energy distribution pattern of fault characteristics in the global voltage and current signals.
[0072] S4. Using a bidirectional cross-attention mechanism, global features query related information in local features, and local features query related information in global features, to obtain a fused feature vector; Specifically, for the obtained local features and global features To avoid information loss caused by simple feature concatenation, local and global features are first projected into a fusion space of the same dimension, and simultaneously made into sequences of length 1:
[0073]
[0074] in , , , , and These are the learnable projection matrices that map local and global features onto the query, key, and value matrices, respectively.
[0075] by As a query vector and As a key-value vector, it can realize the function of querying global features from local features, confirming whether the high-frequency texture detected on the image has a corresponding trend support in the global temporal waveform. If there is no support, the weight of the local feature is suppressed.
[0076] by As a query vector and As a key-value vector, it can realize the function of querying local features from global features, and confirm whether the detected frequency micro-changes in time series have corresponding structured texture features in the time-frequency map:
[0077] The features obtained at this point have achieved semantic alignment. To prevent information loss during signal transmission as the network depth increases, residual connections are added to obtain the final semantically enhanced feature vector. and :
[0078] To dynamically balance their direct weighted contributions to fault feature information, a learnable gating parameter is used to adaptively weight the outputs of both, ultimately yielding the mutually verified fused features:
[0079]
[0080] During the integration process It's like a switch, considering that if the network is allowed to learn directly... It is easy for this weight parameter to exceed the range of 0 to 1 during updates. Therefore, in order to make this weight parameter absolutely stable, an adaptive adjustment parameter is introduced. As an independent variable within the network, it has no boundary restrictions and will freely update and adjust itself based on the feedback from the loss function during backpropagation. Finally, it is smoothly normalized to obtain the required gating parameters. When the fault is a fast transient fluctuation problem, the parameters It is highly likely that the model will learn higher values, therefore emphasizing that the model focuses on local texture features; when the fault is slowly and dynamically changing, the parameters... It is highly likely that the model will learn smaller values, therefore the emphasis is on the model focusing on the global trend.
[0081] S5. Input the fused feature vector into the classifier and output the current microgrid operating status category, which includes normal operation status, islanded operation status, and non-islanded disturbance status.
[0082] Specifically, such as Figure 3 As shown, the fused features are input into a fully connected layer classifier to achieve fault classification for island detection, and the current operating status of the microgrid is output. Figure 5 As shown on the coordinate axes, this embodiment defines 16 specific working conditions, including: (1) Normal operation (2) Non-islanding disturbances, including: three-phase imbalance, harmonic pollution, voltage sag / boost, short circuit fault, capacitor switching and load change; (3) Island operation, including: zero power mismatch, excess or deficiency of active / reactive power, frequency drift and mixed power mismatch.
[0083] To verify the effectiveness of the method of the present invention, examples are applied. Figure 4 As shown, this microgrid grid-connected power generation system mainly consists of distributed generation units (including DC-side capacitors and inverter bridge power transistors), RLC local loads, the main grid, a controller, and an islanding detection module equipped with the algorithm of this invention. The islanding detection module uses the three-phase voltage and current signals of the shared control system, avoiding the need for additional sensors and significantly reducing the implementation cost and deployment complexity of the diagnostic system. The diagnostic results are fed back to the controller in real time, achieving integration of detection and control. Subsequently, the detection results are as follows... Figure 5 , Figure 6 and Figure 7 As shown.
[0084] Figure 5 This diagram illustrates the confusion matrix of the microgrid islanding detection method provided by this invention for 16 different operating conditions. C1 represents normal operation; C2 represents non-islanded - imbalance; C3 represents non-islanded - harmonics; C4 represents non-islanded - sag; C5 represents non-islanded - sag; C6 represents non-islanded - short circuit; C7 represents non-islanded - capacitor; C8 represents non-islanded - load; C9 represents islanded - zero power; C10 represents islanded - excess active power; C11 represents islanded - insufficient active power; C12 represents islanded - excess reactive power; C13 represents islanded - insufficient reactive power; C14 represents islanded - frequency drift; C15 represents islanded - mixed mismatch; and C16 represents islanded - slow frequency drift. Figure 5 It can be seen that the microgrid islanding detection method provided by this invention can basically achieve error-free classification, and at the same time, it can accurately distinguish between normal operation (category 0) and zero power mismatch (category 8), which are easily confused by existing detection methods.
[0085] Figure 6 A comparison chart showing the accuracy of the microgrid islanding detection method provided by this invention with existing technologies is presented. From... Figure 6 As can be seen, the accuracy of the microgrid island detection method provided by this invention is significantly improved compared to existing advanced solutions such as CNN-Transformer (93.06%) and ResNet-LSTM (92.53%), ultimately achieving an accuracy of 99.22%. This result demonstrates that the multimodal feature cross-fusion method implemented in this invention significantly improves the overall detection performance of the model.
[0086] Figure 7This paper illustrates the accuracy of the microgrid island detection method provided by this invention under different signal-to-noise ratio (SNR) noise environments. The method is compared with the two best-performing existing models, CNN-Transformer and ResNet-LSTM, in noise-free conditions. The four bars of each model represent the performance under noise-free, 30dB, 20dB, and 10dB conditions, respectively. The accuracy of the two existing techniques shows a significant stepwise decrease with increasing noise. Particularly under strong noise conditions of 10dB, the accuracy of the comparative methods drops sharply to 80.73% and 77.26%, respectively, which is insufficient for industrial applications. In contrast, the microgrid island detection method provided by this invention (leftmost group) exhibits a smooth change in the height of the four bars, demonstrating excellent robustness. Even under strong noise of 10dB, the accuracy remains at 93.92%. This fully demonstrates that the multimodal feature cross-fusion method implemented in this invention can filter out random noise interference in single-mode environments.
[0087] In a second aspect, the present invention provides a microgrid islanding detection system, comprising: a memory and a processor, wherein the memory stores a computer program, and the processor executes the microgrid islanding detection method provided in the first aspect of the present invention when executing the computer program.
[0088] The related technical solutions are the same as the microgrid islanding detection method provided in the first aspect of this invention, and are not limited here.
[0089] Thirdly, the invention also provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the microgrid islanding detection method provided in the first aspect of the invention.
[0090] The related technical solutions are the same as the microgrid islanding detection method provided in the first aspect of this invention, and are not limited here.
[0091] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A microgrid islanding detection method based on multimodal feature cross-fusion, characterized in that, include: Acquire the voltage and current signals at the microgrid's common coupling point and perform the following islanding detection operation: The voltage signal and the current signal are sampled separately to obtain multiple voltage time-series data samples and multiple current time-series data samples; Time-frequency transformation is performed on each voltage time-series data sample to obtain the corresponding two-dimensional time-frequency voltage image; time-frequency transformation is performed on each current time-series data sample to obtain the corresponding two-dimensional time-frequency current image. Each voltage time series data sample is processed into a non-overlapping slice to obtain the corresponding time series voltage slice sequence. Each current time series data sample is processed into non-overlapping slices to obtain the corresponding time series current slice sequence. Feature extraction was performed on all two-dimensional time-frequency voltage and two-dimensional time-frequency current images to obtain local features containing local texture information; Feature extraction is performed on all time-series voltage slice sequences and time-series current slice sequences to obtain global features containing global dependency information; Based on the cross-attention mechanism, the local features and the global features are fused to obtain local-global fused features; The local-to-global fusion features are input into the classifier to obtain the operating status detection results of the microgrid, thereby achieving island detection.
2. The microgrid islanding detection method according to claim 1, characterized in that, The MSRCN network was used to extract features from all two-dimensional time-frequency voltage and two-dimensional time-frequency current images to obtain local features containing local texture information. The MSRCN network comprises: cascaded multi-scale depth-separable convolutional layers, splicing layers, channel attention modules, coordinate attention modules, residual modules, and fully connected layers; the output of the channel attention modules is also connected to the input of the fully connected layers. The multi-scale depth-separable convolutional layer is used to extract multi-scale features from all two-dimensional time-frequency voltage images and two-dimensional time-frequency current images to obtain feature images of multiple scales. The stitching layer is used to stitch feature images of multiple scales together in the channel dimension to obtain a stitched feature map; The channel attention module is used to extract features from the spliced feature map based on the channel attention mechanism to obtain the channel fusion feature map; The coordinate attention module is used to extract features from the channel fusion feature map based on the coordinate attention mechanism to obtain the pixel fusion feature map; The residual module is used to perform residual processing on the pixel fusion feature map; The fully connected layer is used to flatten and project the pixel-by-pixel feature map and the feature map after residual processing, to obtain local features containing local texture information.
3. The microgrid islanding detection method according to claim 2, characterized in that, The MSRCN network further includes a convolutional pooling layer between the channel attention module and the coordinate attention module, which is used to perform dimensionality reduction convolution and pooling operations on the channel fusion feature map and then output it to the coordinate attention module.
4. The microgrid islanding detection method according to claim 1, characterized in that, The cross-attention mechanism is a bidirectional cross-attention mechanism.
5. The microgrid islanding detection method according to claim 4, characterized in that, The local-global fusion feature for: in, These are learnable fusion weight coefficients; and All are intermediate features; It is a normalization operation; This is a local query matrix; ; Local features; A learnable projection matrix that maps local features onto the query matrix; This is an intermediate fusion feature from local to global; This is the global key matrix; ; For global features; A learnable projection matrix that maps global features onto the key matrix; To preset the number of attention points; This is a global value matrix; ; A learnable projection matrix that maps global features onto a value matrix; This is a global query matrix; ; A learnable projection matrix that maps global features onto the query matrix; This is an intermediate fusion feature from global to local; It is a local bond matrix; ; This is a learnable projection matrix that maps local features onto the key matrix; It is a local value matrix; ; This is a learnable projection matrix that maps local features onto a value matrix.
6. The microgrid islanding detection method according to any one of claims 1-5, characterized in that, A Transformer network is used to extract features from all time-series voltage slice sequences and time-series current slice sequences to obtain global features containing global dependency information.
7. The microgrid islanding detection method according to any one of claims 1-5, characterized in that, The classifier is trained in the following way: Obtain the training set; The training set includes: training samples and corresponding labels for the microgrid in different operating states; the training samples include: voltage and current signals at the common coupling point of the microgrid; the labels are the corresponding microgrid operating states; the operating state categories include normal operating state, islanded operating state, and non-islanded disturbance state. An island detection operation is performed on each training sample in the training set to obtain the corresponding running state detection result; the classifier is trained by minimizing the difference loss between the running state detection result and the corresponding label.
8. The microgrid islanding detection method according to claim 7, characterized in that, While training the classifier by minimizing the difference loss between the running state detection results and the corresponding labels, the learnable parameters involved in the local feature extraction, global feature extraction, and local-global fusion feature extraction processes in the island detection operation are also trained.
9. The microgrid islanding detection method according to claim 7, characterized in that, The islanded operation status categories include: zero power mismatch, active power surplus, active power shortage, reactive power surplus, reactive power shortage, frequency drift, mixed power mismatch, and slow frequency drift. Non-islanded disturbance states include: three-phase unbalanced state, harmonic pollution state, voltage sag state, voltage swell state, short-circuit fault state, capacitor switching state, and load change state.
10. A microgrid islanding detection system, characterized in that, include: A memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the microgrid islanding detection method according to any one of claims 1-9.