A method and related equipment for identifying transient power quality problems

Abstract

This application relates to the field of power quality technology, and provides a method and related equipment for identifying transient power quality problems. The identification method includes: obtaining low-level and high-level features using a lightweight network model based on all power quality signals; fusing the low-level and high-level features using a feature fusion model to obtain final features; obtaining identification results using a prediction model based on all final features and all high-level features; constructing a loss function based on all identification results, and optimizing the lightweight network model, feature fusion model, and prediction model using the loss function; acquiring the current power quality signal of the distribution network to be identified, and identifying the current power quality signal using the optimized lightweight network model, optimized feature fusion model, and optimized prediction model to obtain the final identification result of the distribution network to be identified. The method of this application can improve the comprehensiveness of transient power quality problem identification.

Description

Technical Field

[0001] This application relates to the field of power quality technology, and in particular to a method and related equipment for identifying transient power quality problems. Background Technology

[0002] With the continuous construction of new power systems, more and more distributed renewable energy sources are being integrated into the power system on a large scale. This process has led to increasingly complex and unpredictable power quality problems in the power system. Frequent power quality problems not only affect the user experience but also adversely impact the stable operation and lifespan of electronic power equipment, and in severe cases, can even cause huge economic losses. Therefore, it is necessary to accurately identify power quality problems occurring in the power system and provide timely information support to the upper-level control center to assess whether early warning and mitigation measures are needed to prevent further significant impacts and ensure the reliability and stability of the overall power supply system.

[0003] Transient power quality problems refer to power quality issues that occur within a very short period of time. Essentially, they are transient manifestations of voltage quality and are often closely related to internal system faults, lightning strikes, and other events.

[0004] Currently, commonly used power quality problem identification systems deploy edge devices in distribution substations and user areas to collect raw power quality data. However, limited by the computing power and storage resources of these edge devices, they transmit the collected power quality data to the control center or cloud via 4G, Ethernet, or other communication methods for further power quality problem monitoring and analysis. With the large-scale integration of distributed renewable energy sources, the number of monitored nodes will increase, and the amount of raw data received by the control center or cloud will double. Faced with this massive increase in data, the control center and cloud will experience enormous pressure on their computing and storage performance. Furthermore, this centralized processing method is inevitably affected by communication bandwidth and latency, making it difficult for existing power quality problem identification systems to meet the urgent needs of new power systems for real-time monitoring of power quality issues.

[0005] To address the aforementioned issues, assigning more computational tasks to edge devices closer to the data source, and completing the identification and monitoring of power quality problems locally after collecting raw data, has become a new approach for power quality problem identification systems. In recent years, deep learning methods, with their powerful learning and mapping capabilities, have gradually become a research hotspot in new power quality problem identification systems. Deploying high-performance deep learning models to edge devices for power quality problem identification and monitoring provides a novel approach to meeting the dual requirements of accuracy and real-time performance in power quality problem identification for modern power systems, and possesses enormous development potential.

[0006] Existing deep learning-based methods for identifying power quality problems possess powerful and flexible feature extraction capabilities, effectively overcoming the shortcomings of traditional methods such as redundancy and slow extraction speed in manual feature extraction. This significantly improves the accuracy of identifying various power quality problems. However, most current deep learning methods applied to power quality problem identification only utilize deep learning algorithms to classify data, distinguishing which type of power quality problem the current data belongs to, resulting in relatively limited functionality. Therefore, current power quality problem identification methods suffer from poor comprehensiveness in identifying transient power quality problems. Summary of the Invention

[0007] This application provides a method and related equipment for identifying transient power quality problems, which can solve the problem of poor comprehensiveness in identifying transient power quality problems.

[0008] In a first aspect, embodiments of this application provide a method for identifying transient power quality problems, the method comprising:

[0009] Acquire multiple power quality signal sequences of the target distribution network; each power quality signal sequence includes power quality signals from multiple historical moments, and the power quality signals are used to describe the power quality of the target distribution network;

[0010] Based on all power quality signals in each power quality signal sequence, a lightweight network model is used to obtain the low-level and high-level features of each power quality signal sequence. The low-level features are used to describe the location information of all power quality signals in the power quality signal sequence, and the high-level features are used to describe the semantic information of all power quality signals in the power quality signal sequence.

[0011] For each power quality signal sequence, a feature fusion model is used to fuse the low-level and high-level features of the power quality signal sequence to obtain the final features;

[0012] Based on all final features and all high-level features, the identification results for each power quality signal sequence are obtained using a prediction model. The identification results include the prediction center point, prediction length, confidence level, and prediction category corresponding to multiple predicted transient power quality problems. The prediction length is the number of all power quality signals corresponding to the duration of the predicted transient power quality problem. The prediction center point is the central power quality signal among all power quality signals corresponding to the duration of the predicted transient power quality problem. The confidence level is used to describe the probability that the predicted transient power quality problem actually occurs. The prediction category is the category of the predicted transient power quality problem.

[0013] A loss function is constructed based on all recognition results, and the lightweight network model, feature fusion model, and prediction model are optimized using the loss function to obtain the optimized lightweight network model, optimized feature fusion model, and optimized prediction model; the loss function is used to describe the accuracy of all recognition results.

[0014] Multiple current power quality signals of the distribution network to be identified are acquired. The optimized lightweight network model, the optimized feature fusion model, and the optimized prediction model are used to identify all current power quality signals to obtain the final identification result of the distribution network to be identified. The final identification result includes the prediction center point, prediction length, confidence level, and prediction category corresponding to multiple final predicted transient power quality problems.

[0015] Optionally, the lightweight network model includes a first extraction sub-model, a second extraction sub-model, a third extraction sub-model, a fourth extraction sub-model, a fifth extraction sub-model, a sixth extraction sub-model, a first one-dimensional convolutional sub-model, and a second one-dimensional convolutional model connected in sequence.

[0016] Each of the first extraction sub-model, the second extraction sub-model, the third extraction sub-model, the fourth extraction sub-model, the fifth extraction sub-model, and the sixth extraction sub-model includes an extraction sub-model;

[0017] The extraction sub-model consists of a one-dimensional convolutional unit and a pooling unit connected in sequence. The input of the one-dimensional convolutional unit is the input of the extraction sub-model, and the output of the pooling unit is the output of the extraction sub-model.

[0018] The input of the first extraction sub-model is the input of the lightweight network model, the output of the one-dimensional convolutional unit in the fifth extraction sub-model is the first output of the lightweight network model, and the output of the second one-dimensional convolutional sub-model is the second output of the lightweight network model.

[0019] Optionally, both the first and second one-dimensional convolutional sub-models include a one-dimensional convolutional unit.

[0020] Optionally, a one-dimensional convolutional unit includes a one-dimensional convolutional layer, a batch normalization layer, an activation layer, and an attention mechanism layer connected in sequence.

[0021] The input of a one-dimensional convolutional layer is the input of a one-dimensional convolutional unit, and the output of the attention mechanism layer is the output of a one-dimensional convolutional unit.

[0022] Optionally, the feature fusion model includes a one-dimensional convolutional sub-model, an upsampling sub-model, and a feature concatenation sub-model connected in sequence;

[0023] The inputs of the one-dimensional convolutional sub-model and the feature splicing sub-model are both inputs of the feature fusion model. The input of the feature splicing sub-model is connected to the first output of the lightweight network model, the input of the one-dimensional convolutional sub-model is connected to the second output of the lightweight network model, and the output of the feature splicing sub-model is the output of the feature fusion model.

[0024] Optionally, the prediction model includes a first prediction sub-model, a second prediction sub-model, and a spliced ​​sub-model;

[0025] The input of the splicing sub-model is connected to the output of the first prediction sub-model and the output of the second prediction sub-model, respectively.

[0026] Optionally, both the first and second prediction sub-models include a prediction sub-model.

[0027] The prediction sub-model includes a one-dimensional convolutional unit, a semantic convolutional unit, a semantic prediction unit, a detection convolutional unit, a detection prediction unit, a confidence prediction unit, and a concatenation unit;

[0028] The input of a one-dimensional convolutional unit is the input of the prediction sub-model, and the output of a concatenation unit is the output of the prediction sub-model.

[0029] The output of the one-dimensional convolutional unit is connected to the input of the semantic convolutional unit and the input of the detection convolutional unit, respectively. The output of the semantic convolutional unit is connected to the input of the semantic prediction unit. The output of the detection convolutional unit is connected to the input of the detection prediction unit and the input of the confidence prediction unit, respectively. The input of the concatenation unit is connected to the output of the semantic prediction unit, the output of the detection prediction unit, and the output of the confidence prediction unit, respectively.

[0030] Optionally, the loss function is:

[0031] loss=loss_iou+loss_obj+loss_cls

[0032] Where loss represents the value of the loss function, loss_iou represents the location loss function, loss_obj represents the confidence loss function, and loss_cls represents the classification loss function.

[0033] loss_iou = 1 - IOU 2

[0034]

[0035] Where IOU represents the intersection-over-union ratio (IoU) between all predicted center points and the true center points, and the IoU represents the intersection-over-union ratio between all predicted lengths and the true lengths; y represents the set of true confidence scores for all predicted transient power quality problems. Let represent the set of confidence levels for all predicted transient power quality problems, and let c represent the set of true categories for transient power quality problems in all power quality signal sequences. This represents the set of all prediction categories for predicting transient power quality problems.

[0036] Secondly, embodiments of this application provide a device for identifying transient power quality problems, including:

[0037] The acquisition module acquires multiple power quality signal sequences of the target distribution network; each power quality signal sequence includes power quality signals from multiple historical moments, and the power quality signals are used to describe the power quality of the target distribution network.

[0038] The feature acquisition module uses a lightweight network model to obtain the low-level and high-level features of each power quality signal sequence based on all power quality signals in each power quality signal sequence. The low-level features are used to describe the location information of all power quality signals in the power quality signal sequence, and the high-level features are used to describe the semantic information of all power quality signals in the power quality signal sequence.

[0039] The fusion module, for each power quality signal sequence, uses a feature fusion model to fuse the low-level and high-level features of the power quality signal sequence to obtain the final features;

[0040] The identification result acquisition module, based on all final features and all high-level features, uses a prediction model to obtain the identification results for each power quality signal sequence. The identification results include the prediction center point, prediction length, confidence level, and prediction category corresponding to multiple predicted transient power quality problems. The prediction length is the number of all power quality signals corresponding to the duration of the predicted transient power quality problem. The prediction center point is the center power quality signal among all power quality signals corresponding to the duration of the predicted transient power quality problem. The confidence level is used to describe the probability that the predicted transient power quality problem actually occurs. The prediction category is the category of the predicted transient power quality problem.

[0041] The optimization module constructs a loss function based on all recognition results and uses the loss function to optimize the lightweight network model, feature fusion model, and prediction model, resulting in optimized lightweight network model, optimized feature fusion model, and optimized prediction model; the loss function is used to describe the accuracy of all recognition results.

[0042] The identification module acquires multiple current power quality signals of the distribution network to be identified, and uses an optimized lightweight network model, an optimized feature fusion model, and an optimized prediction model to identify all current power quality signals, thereby obtaining the final identification result of the distribution network to be identified. The final identification result includes the prediction center point, prediction length, confidence level, and prediction category corresponding to multiple final predicted transient power quality problems.

[0043] Thirdly, embodiments of this application provide a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the aforementioned method for identifying transient power quality problems.

[0044] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned method for identifying transient power quality problems.

[0045] The above-mentioned solution in this application has the following beneficial effects:

[0046] In some embodiments of this application, multiple power quality signal sequences of the target distribution network are acquired. Then, based on all power quality signals in each power quality signal sequence, a lightweight network model is used to obtain the low-level and high-level features of each power quality signal sequence. For each power quality signal sequence, a feature fusion model is used to fuse the low-level and high-level features to obtain the final features. Then, based on all the final features and all the high-level features, a prediction model is used to obtain the identification result of each power quality signal sequence. Then, a loss function is constructed based on all the identification results, and the lightweight network model, feature fusion model, and prediction model are optimized using the loss function to obtain the optimized lightweight network model, optimized feature fusion model, and optimized prediction model. Finally, multiple current power quality signals of the distribution network to be identified are acquired, and the optimized lightweight network model, optimized feature fusion model, and optimized prediction model are used to identify all the current power quality signals to obtain the final identification result of the distribution network to be identified. By acquiring low-level and high-level features, the information of the power quality signal sequence can be represented at two levels, increasing the comprehensiveness of the information. The low-level and high-level features are fused to obtain the final features, which contain information from both low-level and high-level features, thus improving the information richness of the final features. Based on the information-rich final and high-level features, the identification results can be obtained, which can improve the accuracy of the identification results. At the same time, the identification results include four aspects of data: predicted center point, predicted length, confidence level, and predicted category, which improves the comprehensiveness of power quality problem identification.

[0047] Other beneficial effects of this application will be described in detail in the following detailed description section. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 A flowchart illustrating a method for identifying transient power quality problems provided in an embodiment of this application;

[0050] Figure 2 This is a schematic diagram of the structure of a lightweight network model provided in an embodiment of this application;

[0051] Figure 3 This is a schematic diagram of the structure of a one-dimensional convolutional unit provided in an embodiment of this application;

[0052] Figure 4 This is a schematic diagram of the structure of a feature fusion model provided in an embodiment of this application;

[0053] Figure 5 This is a schematic diagram of the structure of a prediction model provided in an embodiment of this application;

[0054] Figure 6 A loss function curve provided for one embodiment of this application;

[0055] Figure 7 A schematic diagram of the structure of a transient power quality problem identification device provided in an embodiment of this application;

[0056] Figure 8 This is a schematic diagram of the structure of a terminal device provided in an embodiment of this application. Detailed Implementation

[0057] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0058] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0059] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0060] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0061] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0062] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0063] To address the issue of poor comprehensiveness in existing transient power quality problem identification methods, this application provides a method for identifying transient power quality problems. This method acquires multiple power quality signal sequences of a target distribution network. Then, based on all power quality signals in each sequence, a lightweight network model is used to obtain the low-level and high-level features of each power quality signal sequence. Next, for each power quality signal sequence, a feature fusion model is used to fuse the low-level and high-level features to obtain the final features. Based on all the final features and all the high-level features, a prediction model is used to obtain the identification result for each power quality signal sequence. A loss function is then constructed based on all the identification results, and the lightweight network model, feature fusion model, and prediction model are optimized using the loss function to obtain optimized lightweight network model, optimized feature fusion model, and optimized prediction model. Finally, multiple current power quality signals of the distribution network to be identified are acquired, and the optimized lightweight network model, optimized feature fusion model, and optimized prediction model are used to identify all current power quality signals to obtain the final identification result of the distribution network to be identified. By acquiring low-level and high-level features, the information of the power quality signal sequence can be represented at two levels, increasing the comprehensiveness of the information. The low-level and high-level features are fused to obtain the final features, which contain information from both low-level and high-level features, thus improving the information richness of the final features. Based on the information-rich final and high-level features, the identification results can be obtained, which can improve the accuracy of the identification results. At the same time, the identification results include four aspects of data: predicted center point, predicted length, confidence level, and predicted category, which improves the comprehensiveness of power quality problem identification.

[0064] Step 11: Obtain multiple power quality signal sequences of the target distribution network.

[0065] Each of the above power quality signal sequences includes power quality signals from multiple historical moments. These power quality signals are used to describe the power quality of the target distribution network, such as voltage amplitude, voltage frequency, or voltage waveform quality. All power quality signals contain the same type of information, such as signals describing voltage amplitude.

[0066] In some embodiments of this application, all power quality signals of the target distribution network over a historical period can be acquired, and then all power quality signals can be divided into multiple consecutive power quality signal sequences. For each power quality signal sequence, a start power quality signal and an end power quality signal for a transient power quality problem are randomly generated from all power quality signals in the sequence. The start power quality signal is the power quality signal corresponding to the time the transient power quality problem occurs, and the end power quality signal is the power quality signal corresponding to the time the transient power quality problem ends. The true category of the transient power quality problem (e.g., voltage sag, voltage swell, brief interruption, etc.) is set, and the true center point and true length of the transient power quality problem are calculated based on the start and end power quality signals using the following formula:

[0067]

[0068] Calculate the index x of the true center point c And the actual length w.

[0069] Where t1 represents the sequence number of the starting power quality signal, t2 represents the sequence number of the ending power quality signal, and K represents the number of power quality signals in the dataset. The power quality signal corresponding to the sequence number of the true centroid is taken as the true centroid.

[0070] Thus, we obtain the transient power quality problem for each power quality signal sequence, described by the true center point, true length, and true category.

[0071] Step 12: Based on all power quality signals in each power quality signal sequence, use a lightweight network model to obtain the low-level and high-level features of each power quality signal sequence.

[0072] The aforementioned low-level features are used to describe the positional information of all power quality signals in the power quality signal sequence (such as the relative position of each power quality signal in the power quality signal sequence), while the high-level features are used to describe the semantic information of all power quality signals in the power quality signal sequence (such as the information contained in each power quality signal).

[0073] like Figure 2 As shown, the lightweight network model includes a first extraction sub-model, a second extraction sub-model, a third extraction sub-model, a fourth extraction sub-model, a fifth extraction sub-model, a sixth extraction sub-model, a first one-dimensional convolution sub-model, and a second one-dimensional convolution sub-model, which are connected in sequence.

[0074] Each of the first, second, third, fourth, fifth, and sixth extraction sub-models includes one extraction sub-model.

[0075] The extraction sub-model consists of a one-dimensional convolutional unit and a pooling unit connected in sequence. The input of the one-dimensional convolutional unit is the input of the extraction sub-model, and the output of the pooling unit is the output of the extraction sub-model.

[0076] The input of the first extraction sub-model is the input of the lightweight network model, the output of the one-dimensional convolutional unit in the fifth extraction sub-model is the first output of the lightweight network model, and the output of the second one-dimensional convolutional sub-model is the second output of the lightweight network model.

[0077] It should be noted that the first output of the lightweight network model outputs low-level features, and the second output outputs high-level features. Both the first and second one-dimensional convolutional sub-models include a one-dimensional convolutional unit. For example... Figure 3 As shown, a one-dimensional convolutional unit comprises a one-dimensional convolutional layer, a batch normalization layer, an activation layer, and an attention mechanism layer connected in sequence. The input of the one-dimensional convolutional layer is the input of the one-dimensional convolutional unit, and the output of the attention mechanism layer is the output of the one-dimensional convolutional unit. The one-dimensional convolutional layer is used to extract features, the batch normalization layer is used for normalization, the activation function layer (i.e., the activation layer) is used to introduce non-linear features, and the attention mechanism layer is used for further feature extraction. The pooling unit performs pooling operations, used for dimensionality reduction and to prevent overfitting.

[0078] In some embodiments of this application, for all power quality signals corresponding to each power quality signal sequence, all power quality signals are divided into multiple sets on an average basis. The number of sets is the same as the number of input channels of the first extraction sub-model, and each set corresponds one-to-one with an input channel. Then, the sets are input into the corresponding input channels. The operation of the one-dimensional convolutional layer in the one-dimensional convolutional unit of the lightweight network can be represented as:

[0079]

[0080] Among them, y n [i] represents the output of the nth output channel of the one-dimensional convolutional layer, where n represents the number of output channels, m represents the mth input channel, k represents the length of the convolutional kernel, s represents the convolutional stride, and x... m w represents the input of the m-th input channel. n,m Let represent the kernel weights between the nth output channel and the mth input channel, where i represents the i-th index of the nth output channel, i ∈ l', and l' represents the sequence length of the output channels. The final output of the one-dimensional convolutional layer is y = [n, l'], which is then input into the batch normalization layer for computation. The expression is:

[0081]

[0082] Where y' represents the output of the batch normalization layer, γn Represents the scaling parameter, μ n This represents the mean of the nth channel. Let ε represent the variance of the nth channel, β represent the adjustment parameter, and β represent the variance of the nth channel. n This represents the offset parameter, which is then input into the activation function layer for calculation. The expression is:

[0083] y” = max(0, y')

[0084] Where y” represents the output of the activation function layer, this formula simply sets all negative values ​​in the input to 0, while keeping positive values ​​unchanged, thus introducing non-linearity. Finally, it is input into the attention mechanism layer for calculation, and the expression is:

[0085]

[0086] f n =σ(s n )

[0087]

[0088] in, f represents the output of the attention mechanism layer. n s represents the channel attention coefficients for n channels. n s represents the channel attention weights for n channels. j Let s represent the channel attention weight of the j-th channel. j ∈s n w i Z represents the convolution kernel weights. n Let y' represent the global feature vector of the nth channel. n,i This represents the output of the i-th index of the n-th channel in y. The sequence length obtained by the above formula is n-k+1. We fill it with 0s evenly on the left and right to make the sequence length n.

[0089] It should be noted that in the above expression, the shape of the output of each layer is the same, i.e., [n, l'], with a total of n*l' data points. A padding operation is performed on the output of each one-dimensional convolutional unit, uniformly padding the left and right sides of the output sequence with 0s, making the length of the output channel sequence the same as the input. Dimensionality reduction is then performed only through subsequent pooling units.

[0090] In some embodiments of this application, the feature representation is insufficient in lower layers or convolutional layers with fewer channels, and the effect of the attention mechanism on such features is not obvious. Therefore, in some embodiments, the attention mechanism layer can be added only in one-dimensional convolutional units with more than 64 output channels.

[0091] It is worth mentioning that by acquiring both low-level and high-level features, the information of the power quality signal sequence can be represented at two levels, increasing the comprehensiveness of the information. A lightweight network model mainly composed of one-dimensional convolutional units and pooling units was constructed, enabling accurate feature extraction from the input data.

[0092] Meanwhile, the lightweight network model constructed in this application has a simple structure, effectively reducing computational and storage resources.

[0093] Step 13: For each power quality signal sequence, the feature fusion model is used to fuse the low-level and high-level features of the power quality signal sequence to obtain the final features.

[0094] like Figure 4 As shown, the feature fusion model includes a one-dimensional convolutional sub-model, an upsampling sub-model, and a feature splicing sub-model connected in sequence. Figure 4 Input 1 corresponds to the low-level feature, and input 2 corresponds to the high-level feature.

[0095] The inputs of the one-dimensional convolutional sub-model and the feature splicing sub-model are both inputs of the feature fusion model. The input of the feature splicing sub-model is connected to the first output of the lightweight network model, the input of the one-dimensional convolutional sub-model is connected to the second output of the lightweight network model, and the output of the feature splicing sub-model is the output of the feature fusion model.

[0096] In some embodiments of this application, the input end of the one-dimensional convolutional sub-model is high-level features, the input end of the feature splicing sub-model is low-level features, and the one-dimensional convolutional sub-model includes a one-dimensional convolutional unit.

[0097] It should be noted that the upsampling module is a nearest neighbor interpolation calculation used to interpolate the channel sequence in the output of the one-dimensional convolutional sub-model so that the sequence length of its channel sequence is the same as the sequence length of the underlying features. Then, it is concatenated with the underlying features in the feature concatenation sub-model to obtain the final features.

[0098] It is worth mentioning that fusing low-level features and high-level features to obtain the final feature means that the final feature contains information from both low-level and high-level features, thereby improving the information richness of the final feature.

[0099] Step 14: Based on all final features and all high-level features, use the prediction model to obtain the identification results of each power quality signal sequence.

[0100] The above identification results include multiple predicted transient power quality problems, including the prediction center point, prediction length, confidence level, and prediction category. The prediction length is the number of all power quality signals corresponding to the duration of the predicted transient power quality problem. The prediction center point is the central power quality signal among all power quality signals corresponding to the duration of the predicted transient power quality problem. The confidence level describes the probability that the predicted transient power quality problem will actually occur. The prediction category is the type of predicted transient power quality problem (such as voltage swell, voltage drop, or brief interruption).

[0101] like Figure 5 As shown, the above prediction model includes a first prediction sub-model, a second prediction sub-model, and a splicing sub-model. Figure 5 Input 1 is the input to the first prediction sub-model, and input 2 is the input to the second prediction sub-model.

[0102] The input of the splicing sub-model is connected to the output of the first prediction sub-model and the output of the second prediction sub-model, respectively.

[0103] Both the first and second predictive sub-models include a predictive sub-model.

[0104] The prediction sub-model includes one-dimensional convolutional units (such as...) Figure 5 One-dimensional convolution in semantic convolution units (such as...) Figure 5 Cls convolution in semantic prediction units (such as...) Figure 5 Cls prediction in the detection convolutional unit (such as Cls prediction) Figure 5 Reg convolution in (e.g., detection prediction unit) Figure 5 Reg prediction in (e.g., confidence prediction unit) Figure 5 Obj prediction and splicing unit.

[0105] The input of a one-dimensional convolutional unit is the input of the prediction sub-model, and the output of a concatenation unit is the output of the prediction sub-model.

[0106] The output of the one-dimensional convolutional unit is connected to the input of the semantic convolutional unit and the input of the detection convolutional unit, respectively. The output of the semantic convolutional unit is connected to the input of the semantic prediction unit. The output of the detection convolutional unit is connected to the input of the detection prediction unit and the input of the confidence prediction unit, respectively. The input of the concatenation unit is connected to the output of the semantic prediction unit, the output of the detection prediction unit, and the output of the confidence prediction unit, respectively.

[0107] In some embodiments of this application, the input terminal of the first prediction sub-model is the final feature, and the input terminal of the second prediction sub-model is the high-level feature. Both the semantic convolution unit and the detection convolution unit perform convolution operations. Taking the first prediction sub-model as an example, the detection prediction unit outputs multiple detection boxes (i.e., the prediction center points and prediction lengths of multiple predicted transient power quality problems, with each detection box corresponding to a specific predicted transient power quality problem). The semantic prediction unit outputs the prediction category corresponding to each detection box (e.g., voltage sag, voltage swell, brief interruption, etc.), and the confidence prediction unit outputs the confidence level of each detection box. The detection prediction unit is an object detection algorithm (such as the anchor-free algorithm). The semantic prediction unit (cls unit) and the confidence prediction unit (obj unit) are both simple one-dimensional convolutional layers, and there is only one layer. The inputs of these two prediction units correspond to the semantic convolutional unit and the detection convolutional unit, respectively (the architecture and output size of these two convolutional units are the same). The difference between the two prediction units is that the output channel size of the semantic prediction unit is equal to the predicted class size, while the confidence prediction unit only outputs one channel. In addition, the length of each channel output by these two prediction units is the same (equal to the number of predicted boxes).

[0108] The first prediction sub-model outputs the prediction center point, prediction length, confidence level, and prediction category corresponding to multiple predicted transient power quality problems obtained from the final features. The second prediction sub-model outputs the prediction center point, prediction length, confidence level, and prediction category corresponding to multiple predicted transient power quality problems obtained from the high-level features. Then, the outputs of the first and second prediction sub-models are concatenated using a splicing sub-model to obtain the recognition result of the power quality signal sequence. The recognition result includes the outputs of the first and second prediction sub-models.

[0109] It is worth mentioning that the identification results are obtained based on the information-rich final features and high-level features, which can improve the accuracy of the identification results. At the same time, the identification results include four aspects of data: prediction center point, prediction length, confidence level and prediction category, which improves the comprehensiveness of power quality problem identification.

[0110] Step 15: Construct a loss function based on all recognition results, and use the loss function to optimize the lightweight network model, feature fusion model, and prediction model to obtain the optimized lightweight network model, optimized feature fusion model, and optimized prediction model.

[0111] The loss function described above is used to describe the accuracy of all recognition results.

[0112] Specifically, the loss function is:

[0113] loss=loss_iou+loss_obj+loss_cls

[0114] Where loss represents the value of the loss function, loss_iou represents the location loss function, loss_obj represents the confidence loss function, and loss_cls represents the classification loss function.

[0115] loss_iou = 1 - IOU 2

[0116]

[0117] Where IOU represents the intersection-over-union ratio (IoU) between all predicted center points and the true center points, and the IoU represents the intersection-over-union ratio between all predicted lengths and the true lengths; y represents the set of true confidence scores for all predicted transient power quality problems. Let represent the set of confidence levels for all predicted transient power quality problems, and let c represent the set of true categories for transient power quality problems in all power quality signal sequences. This represents the set of all prediction categories for predicting transient power quality problems.

[0118] For example, the true confidence level of the above-mentioned prediction of transient power quality problems can be obtained by means of manual methods.

[0119] In some embodiments of this application, optimization algorithms such as Adam can be used to optimize the lightweight network model, feature fusion model, and prediction model using a loss function, resulting in optimized lightweight network model, optimized feature fusion model, and optimized prediction model. The optimization termination condition can be that the number of optimization iterations reaches a preset number, or the value of the loss function is less than a preset loss value.

[0120] It is worth mentioning that by constructing a loss function, the accuracy of the recognition results can be described. Optimizing the lightweight network model, feature fusion model, and prediction model using the loss function can improve their performance, thereby improving the accuracy of the recognition results.

[0121] The loss function described above will be illustrated with a specific example below.

[0122] like Figure 6 As shown, Figure 6 The horizontal axis represents the number of iterations, and the vertical axis represents the loss value. The curves represent the values ​​of the confidence loss function (obj_loss), the classification loss function (cls_loss), the location loss function (reg_loss), and the total loss function (total_loss), respectively.

[0123] Step 16: Obtain multiple current power quality signals of the distribution network to be identified, and use the optimized lightweight network model, the optimized feature fusion model and the optimized prediction model to identify all current power quality signals to obtain the final identification result of the distribution network to be identified.

[0124] The final identification results include the prediction center point, prediction length, confidence level, and prediction category corresponding to multiple predicted transient power quality problems. The distribution network to be identified refers to the distribution network for which transient power quality problems need to be identified, and the multiple current power quality signals refer to all power quality signals of the distribution network to be identified during the current time period. For example, if the current time is 10:00, then the current time period can be from 9:30 to 10:00.

[0125] Specifically, based on all current power quality signals, the optimized lightweight network model is used to obtain low-level and high-level features. Then, the optimized feature fusion model is used to fuse the low-level and high-level features to obtain the final features. Finally, based on the final features and high-level features, the prediction model is used to obtain the final recognition result.

[0126] For example, after obtaining the final identification results, the power quality of the distribution network to be identified in the current time period can be analyzed based on the prediction center point, prediction length, confidence level, and prediction category corresponding to all the final predicted transient power quality problems. According to the number, prediction category, prediction center point, prediction length, and confidence level of the final predicted transient power quality problems, timely maintenance can be carried out on the distribution network to be identified to improve its power quality. For example, if voltage sags occur more frequently than other types among all the final predicted transient power quality problems, it indicates that the distribution network to be identified is prone to voltage sags and requires maintenance to address this issue.

[0127] It is worth mentioning that by acquiring low-level and high-level features, the information of the power quality signal sequence can be represented at two levels, increasing the comprehensiveness of the information. The final feature is obtained by fusing the low-level and high-level features, so that the final feature contains information from both low-level and high-level features, thus improving the information richness of the final feature. The identification result is obtained based on the information-rich final feature and high-level feature, which can improve the accuracy of the identification result. At the same time, the identification result includes four aspects of data: predicted center point, predicted length, confidence level and predicted category, which improves the comprehensiveness of power quality problem identification.

[0128] The following example illustrates the method for identifying transient power quality problems in this application.

[0129] Precision, recall, F1 score, average precision (AP), beginning power quality signal error (Begin_loss), ending power quality signal error (Over_loss), and length error (W_loss) were selected as evaluation indicators. The method was used to classify and assess the accuracy of locating four types of transient power quality problems: voltage sag, voltage swell, voltage interruption, and transient impact. The evaluation indicators obtained by the method in this application after maximum suppression with a confidence threshold of 0.6 and an intersection-with-union (IoU) threshold of 0.5 are shown in Table 1.

[0130]

[0131] Table 1

[0132] Therefore, the method of this application demonstrates excellent classification performance for all four types of transient power quality problems, exhibiting both high accuracy and high recall. This indicates that the method not only correctly identifies but also comprehensively identifies transient power quality problems that may exist in the power quality signal.

[0133] Furthermore, traditional methods locate the start and end points of transient power quality problems by determining whether the effective value of the half-cycle is within a threshold range. This method has a positioning accuracy of half a cycle, or 10 ms. As shown in Table 1, the embodiment of this invention achieves a maximum error of no more than 2.7 ms in the start and end point location, and particularly achieves a positioning accuracy of 0.05 ms for transient impacts with extremely long durations; the maximum error in duration detection is also no more than 3.5 ms, far less than the required 20 ms measurement error. This confirms that the method of this application has excellent positioning performance for transient power quality problems.

[0134] The following is an exemplary description of the device for identifying transient power quality problems provided in this application.

[0135] like Figure 7 As shown, this application embodiment provides a transient power quality problem identification device 700, which includes:

[0136] The acquisition module 701 acquires multiple power quality signal sequences of the target distribution network; each power quality signal sequence includes power quality signals at multiple historical moments, and the power quality signals are used to describe the power quality of the target distribution network.

[0137] The feature acquisition module 702 acquires the low-level and high-level features of each power quality signal sequence based on all power quality signals in each power quality signal sequence using a lightweight network model. The low-level features are used to describe the location information of all power quality signals in the power quality signal sequence, and the high-level features are used to describe the semantic information of all power quality signals in the power quality signal sequence.

[0138] The fusion module 703, for each power quality signal sequence, uses a feature fusion model to fuse the low-level and high-level features of the power quality signal sequence to obtain the final features;

[0139] The identification result acquisition module 704, based on all final features and all high-level features, uses a prediction model to obtain the identification results for each power quality signal sequence. The identification results include the prediction center point, prediction length, confidence level, and prediction category corresponding to multiple predicted transient power quality problems. The prediction length is the number of all power quality signals corresponding to the duration of the predicted transient power quality problem. The prediction center point is the center power quality signal among all power quality signals corresponding to the duration of the predicted transient power quality problem. The confidence level is used to describe the probability that the predicted transient power quality problem actually occurs. The prediction category is the category of the predicted transient power quality problem.

[0140] The optimization module 705 constructs a loss function based on all recognition results and uses the loss function to optimize the lightweight network model, feature fusion model, and prediction model, resulting in optimized lightweight network model, optimized feature fusion model, and optimized prediction model; the loss function is used to describe the accuracy of all recognition results.

[0141] The identification module 706 acquires multiple current power quality signals of the distribution network to be identified, and uses an optimized lightweight network model, an optimized feature fusion model, and an optimized prediction model to identify all current power quality signals, thereby obtaining the final identification result of the distribution network to be identified. The final identification result includes the prediction center point, prediction length, confidence level, and prediction category corresponding to multiple final predicted transient power quality problems.

[0142] It should be noted that the information interaction and execution process between the above-mentioned devices / units are based on the same concept as the method embodiments of this application. For details on their specific functions and technical effects, please refer to the method embodiments section, and they will not be repeated here.

[0143] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0144] like Figure 8 As shown, an embodiment of this application provides a terminal device, wherein the terminal device D10 of this embodiment includes: at least one processor D100 ( Figure 8 The diagram shows only one processor, a memory D101, and a computer program D102 stored in the memory D101 and executable on the at least one processor D100, wherein the processor D100 executes the computer program D102 to implement the steps in any of the above method embodiments.

[0145] Specifically, when the processor D100 executes the computer program D102, it acquires multiple power quality signal sequences of the target distribution network. Then, based on all power quality signals in each power quality signal sequence, it uses a lightweight network model to acquire the low-level and high-level features of each power quality signal sequence. Next, for each power quality signal sequence, it uses a feature fusion model to fuse the low-level and high-level features to obtain the final features. Then, based on all the final features and all the high-level features, it uses a prediction model to obtain the identification result of each power quality signal sequence. Then, it constructs a loss function based on all the identification results and uses the loss function to optimize the lightweight network model, the feature fusion model, and the prediction model, obtaining optimized lightweight network model, optimized feature fusion model, and optimized prediction model. Finally, it acquires multiple current power quality signals of the distribution network to be identified, and uses the optimized lightweight network model, optimized feature fusion model, and optimized prediction model to identify all current power quality signals, obtaining the final identification result of the distribution network to be identified. By acquiring low-level and high-level features, the information of the power quality signal sequence can be represented at two levels, increasing the comprehensiveness of the information. The low-level and high-level features are fused to obtain the final features, which contain information from both low-level and high-level features, thus improving the information richness of the final features. Based on the information-rich final and high-level features, the identification results can be obtained, which can improve the accuracy of the identification results. At the same time, the identification results include four aspects of data: predicted center point, predicted length, confidence level, and predicted category, which improves the comprehensiveness of power quality problem identification.

[0146] The processor D100 can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0147] In some embodiments, the memory D101 may be an internal storage unit of the terminal device D10, such as a hard disk or memory of the terminal device D10. In other embodiments, the memory D101 may be an external storage device of the terminal device D10, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the terminal device D10. Furthermore, the memory D101 may include both internal and external storage units of the terminal device D10. The memory D101 is used to store the operating system, applications, bootloader, data, and other programs, such as the program code of the computer program. The memory D101 can also be used to temporarily store data that has been output or will be output.

[0148] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps described in the various method embodiments above.

[0149] This application provides a computer program product that, when run on a terminal device, enables the terminal device to implement the steps described in the various method embodiments above.

[0150] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of this application can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include at least: any entity or device capable of carrying the computer program code to a transient power quality problem identification method / terminal device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electrical carrier signals or telecommunication signals.

[0151] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0152] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0153] The above description is the preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principles described in this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method of identifying a transient power quality problem, characterized by, The method comprises the following steps: obtaining a plurality of power quality signal sequences of a target power distribution network; each power quality signal sequence comprises power quality signals at a plurality of historical time points, and the power quality signals are used to describe the power quality of the target power distribution network; based on all power quality signals in each power quality signal sequence, obtaining low-level features and high-level features of each power quality signal sequence by using a lightweight network model; the low-level features are used to describe the position information of all power quality signals in the power quality signal sequence, and the high-level features are used to describe the semantic information of all power quality signals in the power quality signal sequence; for each power quality signal sequence, respectively, fusing the low-level features and the high-level features of the power quality signal sequence by using a feature fusion model to obtain final features; based on all final features and all high-level features, obtaining an identification result of each power quality signal sequence by using a prediction model; the identification result comprises a predicted center point, a predicted length, a confidence and a predicted category corresponding to a plurality of predicted transient power quality problems, the predicted length is the number of all power quality signals corresponding to the duration of the predicted transient power quality problem, the predicted center point is the center power quality signal of all power quality signals corresponding to the duration of the predicted transient power quality problem, the confidence is used to describe the probability of the real occurrence of the predicted transient power quality problem, and the predicted category is the category of the predicted transient power quality problem; constructing a loss function according to all identification results, and optimizing the lightweight network model, the feature fusion model and the prediction model by using the loss function to obtain an optimized lightweight network model, an optimized feature fusion model and an optimized prediction model; the loss function is used to describe the accuracy of all identification results; obtaining a plurality of current power quality signals of a power distribution network to be identified, and identifying all current power quality signals by using the optimized lightweight network model, the optimized feature fusion model and the optimized prediction model to obtain a final identification result of the power distribution network to be identified; the final identification result comprises a predicted center point, a predicted length, a confidence and a predicted category corresponding to a plurality of final predicted transient power quality problems.

2. The identification method according to claim 1, characterized in that, The lightweight network model comprises a first extraction sub-model, a second extraction sub-model, a third extraction sub-model, a fourth extraction sub-model, a fifth extraction sub-model, a sixth extraction sub-model, a first one-dimensional convolution sub-model and a second one-dimensional convolution sub-model connected in sequence; Each of the first extraction sub-model, the second extraction sub-model, the third extraction sub-model, the fourth extraction sub-model, the fifth extraction sub-model and the sixth extraction sub-model comprises an extraction sub-model; The extraction sub-model comprises a one-dimensional convolution unit and a pooling unit connected in sequence, an input end of the one-dimensional convolution unit is an input end of the extraction sub-model, and an output end of the pooling unit is an output end of the extraction sub-model. The input end of the first extraction sub-model is the input end of the lightweight network model, the output end of the one-dimensional convolution unit in the fifth extraction sub-model is the first output end of the lightweight network model, and the output end of the second one-dimensional convolution sub-model is the second output end of the lightweight network model.

3. The identification method according to claim 2, characterized in that, The first one-dimensional convolution sub-model and the second one-dimensional convolution sub-model each include one one-dimensional convolution unit.

4. The identification method according to claim 3, characterized in that, The one-dimensional convolution unit includes a one-dimensional convolution layer, a batch normalization layer, an activation layer, and an attention mechanism layer connected in sequence. The input end of the one-dimensional convolution layer is the input end of the one-dimensional convolution unit, and the output end of the attention mechanism layer is the output end of the one-dimensional convolution unit.

5. The identification method according to claim 4, characterized in that, The feature fusion model includes a one-dimensional convolution sub-model, an upsampling sub-model, and a feature splicing sub-model connected in sequence. The input end of the one-dimensional convolution sub-model and the input end of the feature splicing sub-model are both the input end of the feature fusion model, the input end of the feature splicing sub-model is connected with the first output end of the lightweight network model, the input end of the one-dimensional convolution sub-model is connected with the second output end of the lightweight network model, and the output end of the feature splicing sub-model is the output end of the feature fusion model.

6. The identification method of claim 1, wherein, The prediction model includes a first prediction sub-model, a second prediction sub-model, and a splicing sub-model. The input end of the splicing sub-model is connected with the output end of the first prediction sub-model and the output end of the second prediction sub-model, respectively.

7. The identification method according to claim 6, characterized in that, The first prediction sub-model and the second prediction sub-model each include one prediction sub-model. The prediction sub-model includes a one-dimensional convolution unit, a semantic convolution unit, a semantic prediction unit, a detection convolution unit, a detection prediction unit, a confidence prediction unit, and a splicing unit. The input end of the one-dimensional convolution unit is the input end of the prediction sub-model, and the output end of the splicing unit is the output end of the prediction sub-model. The output end of the one-dimensional convolution unit is connected with the input end of the semantic convolution unit and the input end of the detection convolution unit, respectively, the output end of the semantic convolution unit is connected with the input end of the semantic prediction unit, the output end of the detection convolution unit is connected with the input end of the detection prediction unit and the input end of the confidence prediction unit, respectively, and the input end of the splicing unit is connected with the output end of the semantic prediction unit, the output end of the detection prediction unit, and the output end of the confidence prediction unit.

8. The identification method of claim 1, wherein, The loss function is: ； wherein, denotes a value of the loss function, denotes a position loss function, denotes a confidence loss function, denotes a classification loss function: ； ； ； wherein, denotes the intersection over union between all predicted center points and the true center point and the intersection over union between all predicted lengths and the true length, denotes a set of true confidences of all predicted transient power quality problems, denotes a set of confidences of all predicted transient power quality problems, denotes a set of true classes of transient power quality problems of all power quality signal sequences, denotes a set of predicted classes of all predicted transient power quality problems.

9. A terminal device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, The processor executes the computer program to implement the transient power quality problem identification method according to any one of claims 1 to 8.

10. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 9. The computer program is executed by the processor to implement the transient power quality problem identification method according to any one of claims 1 to 8.

Images