A method and system for real-time detection of power load in a power grid

Abstract

This invention relates to the field of power grid load detection, and more specifically, to a method and system for real-time power grid load detection. The method includes: acquiring pre-processed load values ​​from the power grid, constructing a load sequence, and labeling each load value; decomposing the load sequence using Virtual Mode Decomposition (VMD) to obtain several modal components; classifying the frequency values ​​of any modal component into normal and abnormal classes according to the labels; calculating the first offset mean for the normal class and the second offset mean for the abnormal class; training a preset network based on the importance of the calculation; inputting the frequency values ​​of several modal components of the real-time load value into the corresponding trained preset network, outputting sub-probabilities; integrating all sub-probabilities to obtain the state probability, and completing the detection. The technical solution of this invention can improve the accuracy of charge detection results and provide a guarantee for the safe operation of the power grid.

Description

Technical Field

[0001] This invention relates to the field of power grid load detection. More specifically, this invention relates to a method and system for real-time power grid load detection. Background Technology

[0002] With the ever-increasing demand for electricity in modern society, the safety and reliability of power systems have become increasingly important. As the core system for power transmission and distribution, the power grid bears the crucial responsibility of ensuring a stable power supply. Load monitoring of the power grid is a vital link in ensuring the normal operation of the power system, improving grid efficiency, and reducing failure rates. Real-time monitoring of power load helps grid operators predict and respond to changes in power demand, thereby optimizing power distribution and dispatch. Through accurate load monitoring, potential problems in the power system can be identified and early warnings issued, reducing the occurrence of grid failures.

[0003] Power load monitoring effectively reflects the operating status of the power system. Traditional monitoring methods rely heavily on manual inspections and periodic data collection, which is not only inefficient but also unable to respond promptly to sudden events in the power grid. Power grid load fluctuates significantly, especially during peak electricity consumption periods when load changes rapidly. Achieving real-time load monitoring, timely response, and effective dispatch has become a major challenge in power grid management.

[0004] A Chinese patent application with publication number CN117034177A discloses an intelligent monitoring method for abnormal power load data, which includes obtaining the probability that the target data is abnormal power load data based on historical data and target data; obtaining the local reachability distance of the target data within the incremental window of the target data based on the probability that the target data is abnormal power load data, and simultaneously obtaining the abnormal power load data within the incremental window of the target data; and calculating the degree of abnormality of the abnormal power load data within the target dataset within the incremental window of the target data.

[0005] However, power grid load signals are typically complex signals composed of multiple frequency components. The charge value at each moment is not only affected by the current load but may also be affected by interference from other frequency components. The aforementioned application only focuses on the overall charge value at each moment, ignoring the signal's spectral and time-varying characteristics. This simplistic treatment of the overall load value fails to effectively distinguish the influence of different frequency components, leading to inaccurate charge detection results. Summary of the Invention

[0006] To address the problem of inaccurate charge detection results, the present invention provides solutions in the following aspects.

[0007] In a first aspect, this invention discloses a method for real-time detection of power grid load, comprising: acquiring preprocessed load values ​​of the power grid, constructing a load sequence, and labeling each load value, the labels including normal and abnormal; decomposing the load sequence based on VMD to obtain several modal components, obtaining the frequency value of any modal component at each sampling time based on wavelet transform, classifying the frequency values ​​of any modal component into normal and abnormal classes according to the labels, calculating the first offset mean of the normal class and the second offset mean of the abnormal class; calculating the importance of the modal component based on the first offset mean and the second offset mean, training a preset network based on the importance, with one preset network corresponding to one modal component, inputting the frequency values ​​of several modal components of the real-time load value into the corresponding trained preset network, outputting sub-probabilities, integrating all sub-probabilities to obtain the state probability, and completing the detection.

[0008] By combining VMD and wavelet transform, multi-level analysis of power grid load signals is performed, effectively extracting the dynamic frequency characteristics of the power grid load. Furthermore, by classifying and processing the frequency values ​​of modal components under different labels, the normal and abnormal states of the power grid load can be accurately identified. By calculating the offset mean and filtering modal components according to importance weights, the model can focus on the most critical features for load status judgment, improving the accuracy and robustness of the power grid load detection system. Finally, the frequency value of each modal component is input into a preset network for real-time status prediction, ensuring that the system can detect the power grid load status in a timely and accurate manner and identify potential anomalies in advance.

[0009] Preferably, the preprocessing includes: acquiring load values ​​according to a preset sampling interval, and performing noise reduction, missing value filling, and outlier deletion processing on the acquired load values.

[0010] Preferably, the plurality of modal components includes: obtaining a set of charge frequencies by applying a Fourier transform to the charge sequence, and clustering the charge frequency set to obtain clusters, wherein the number of clusters is the same as the number of modal components.

[0011] Preferably, the clustering method is K-means clustering or density clustering.

[0012] Preferably, the calculation of the first offset mean of the normal class and the second offset mean of the abnormal class includes: calculating the center point of all frequency values ​​in the normal class, and taking the vector between the frequency value and the center point at any sampling time as the feature vector; for the same modal component, taking the mean of the feature vectors at all sampling times as the first offset mean, and similarly obtaining the second offset mean of the abnormal class.

[0013] By calculating the mean offsets of normal and abnormal classes and using the distance between the frequency value and the center point as a feature vector, subtle differences in load signals under different states can be effectively captured. Analyzing these feature vectors allows for the quantification of the changing trends and abnormal behaviors of the power grid load signals, thereby improving the sensitivity and identification capability of the power grid status. The mean offsets of normal and abnormal classes reflect the typical performance of the signal under different states. Through this mechanism, load signals can be classified more accurately, thus improving the precision and accuracy of the detection system.

[0014] Preferably, the importance includes: for the same modal component, the normalized result of the difference between the first offset mean and the second offset mean is used as the importance.

[0015] A large difference in the mean offset indicates a significant difference in the performance of this feature under normal and abnormal conditions. This allows the feature to provide a clearer basis for judging the power grid load status, thereby improving the accuracy and reliability of the classification model. Conversely, a small difference in the mean offset means that the feature does not change significantly under different labels, making it difficult to effectively distinguish different states of the power grid load, which may lead to a higher error rate.

[0016] Preferably, the importance includes: for the same modal component, calculating the cosine similarity between the first offset mean and the second offset mean, and using the sum of the cosine similarity and 1 through negative correlation mapping as the importance.

[0017] The importance of modal components is assessed by calculating the cosine similarity of the offset mean, thereby quantifying the similarity difference between normal and abnormal classes. Through negative correlation mapping, the difference between similarity and 1 is transformed into an importance value, effectively identifying which modal components contribute more to distinguishing between normal and abnormal grid load states. This allows for a focus on modal components with significant differences between normal and abnormal states, enhancing the sensitivity and accuracy of the detection and prediction network.

[0018] Preferably, the importance-based training of the preset network includes: taking the importance of any modality component in the history as input, taking the actual value of the sub-probability of any modality component in the history as the network label, obtaining a training set, inputting the training set into the preset network, and outputting the predicted value of the sub-probability; using the cross-entropy loss function to calculate the loss value between the predicted value of the sub-probability and the actual value of the sub-probability, iteratively updating the preset network parameters, and stopping the update when the preset network reaches the set maximum number of training iterations or the loss value is less than the set loss value, thus obtaining the trained preset network.

[0019] Preferably, the step of integrating all sub-probabilities to obtain the state probability includes: for any modal component, calculating the cumulative value of the importance of all modal components, using the ratio of the importance of the modal component to the cumulative value as the weight, calculating the product of the weight and the importance, and using the cumulative value of the product of all modal components as the state probability.

[0020] By using the ratio of the importance of each modal component to its accumulated value as a weight, and combining this with the sub-probabilities of each modal component for weighted summation, a global state probability is obtained. This method can reasonably allocate the weight of each modal component in the judgment of grid load status, ensuring that key modal components receive greater attention, thereby enhancing the accuracy and stability of state prediction. Through this weighted integration approach, the importance of each modal component can be comprehensively considered, improving the overall perception capability of grid load status and facilitating the real-time detection and accurate identification of normal and abnormal states in the grid.

[0021] Secondly, the present invention discloses a real-time power load detection system for power grids, comprising: a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, a real-time power load detection method for power grids as described in any one of the present invention is implemented.

[0022] The beneficial effects of this invention are:

[0023] This invention effectively identifies and detects potential abnormal load states in the power grid through multi-level analysis and processing of power grid load data. By performing modal decomposition and frequency feature extraction on load sequences, combined with label classification of normal and abnormal loads, the importance of modal components can be accurately calculated. Based on this, a pre-set network is trained to achieve accurate prediction and anomaly detection of real-time load. This method, by comprehensively considering the frequency variation characteristics of the load and the weights of the modal components, improves the sensitivity and accuracy of detection, ensuring the safety and stability of power grid operation. Attached Figure Description

[0024] Figure 1 This is a flowchart of a real-time power load detection method for a power grid according to an embodiment of the present invention. Detailed Implementation

[0025] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0026] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0027] Reference Figure 1A method for real-time detection of power grid load includes steps S1-S3, which are described in detail below.

[0028] S1: Obtain the preprocessed load values ​​of the power grid, construct a load sequence, and add a label to each load value, including normal and abnormal labels.

[0029] In one embodiment, sensors are used to acquire grid load values ​​in real time at preset sampling intervals, and a series of preprocessing operations are performed on the collected data, including noise reduction, filling in missing values, and deleting outliers, to ensure the accuracy and integrity of the load data, thereby constructing a stable load sequence.

[0030] Based on this, domain experts label the load value at each sampling moment according to the actual situation and historical data, usually dividing the load value into two categories: "normal" and "abnormal". The normal label indicates that the load value at that moment is in line with expectations or fluctuates normally, while the abnormal label indicates abnormal fluctuations caused by sudden events, equipment failures or external factors.

[0031] S2: Based on VMD, the load sequence is decomposed to obtain several modal components. Based on wavelet transform, the frequency value of any modal component at each sampling time is obtained. The frequency values ​​in any modal component are classified into normal and abnormal classes according to the label. The first offset mean of the normal class and the second offset mean of the abnormal class are calculated.

[0032] It should be noted that when analyzing power grid load data, the variational mode decomposition (VMD) algorithm can decompose the power grid load signal into multiple mode components with different center frequencies and bandwidths based on the inherent characteristics of the signal. This decomposition method helps to more accurately identify different frequency components in the power grid load signal, thereby gaining a deeper understanding of the dynamic characteristics of the load. In particular, it is especially effective for processing non-stationary signals.

[0033] However, the VMD algorithm has high computational complexity, especially when processing large-scale power grid data, which may lead to excessive consumption of computing resources. To effectively reduce computational complexity and improve processing efficiency, an appropriate number of modes can be set according to the actual characteristics of the power grid load signal. By selecting a suitable number of modes, redundant calculations can be reduced while maintaining decomposition accuracy, thereby significantly reducing the computational cost of the VMD algorithm in practical applications, while still effectively extracting dynamic information of the power grid load.

[0034] In one embodiment, a number of modal components are obtained by decomposing the load sequence based on VMD. The process of determining the number of modal components includes: using Fourier transform on the charge sequence to obtain a set of charge frequencies, and clustering the charge frequency set to obtain clusters. The number of clusters is the same as the number of modal components. The clustering method is K-means clustering or density clustering.

[0035] Cluster analysis can automatically identify the main frequency components in a power grid load sequence, avoiding the subjectivity issues caused by manually setting the number of modes. Power grid load signals are typically complex, non-stationary signals with superimposed multiple frequency components, which fluctuate with time and different operating conditions. Traditional methods often rely on manually setting the number of modes, which may not accurately reflect the dynamic characteristics of the load signal. Cluster analysis, however, can adaptively identify the frequency components in the power grid load signal and automatically adjust the number of modes based on signal changes, ensuring the rationality and accuracy of the mode count. In this way, cluster analysis can effectively avoid frequency overlap between modes, improve the accuracy of VMD decomposition, and thus more accurately extract the frequency domain characteristics of the power grid load signal, enhancing the reliability of load analysis and forecasting.

[0036] The frequency value of any modal component at each sampling time is obtained based on wavelet transform. The frequency values ​​of any modal component are classified into normal and abnormal classes according to the label. The center point of all frequency values ​​in the normal class is calculated, and the vector between the frequency value at any sampling time and the center point is used as the feature vector. For the same modal component, the mean of the feature vectors at all sampling times is used as the first offset mean. Similarly, the second offset mean of the abnormal class is obtained.

[0037] The offset mean satisfies the following relationship: , Indicates that the label is Time modal components The offset mean, Indicates that the label is Time modal components Mid-sampling time eigenvectors, Indicates that the label is Time modal components The center point, This represents the total number of sampling times.

[0038] in, Labels, including normal and abnormal, when When the condition is normal, the first offset mean is calculated; when the condition is abnormal, the second offset mean is calculated.

[0039] S3: Calculate the importance of the modal components based on the first offset mean and the second offset mean, train a preset network based on the importance, with one preset network corresponding to one modal component, input the frequency values ​​of several modal components of the real-time load value into the corresponding trained preset network, output sub-probabilities, integrate all sub-probabilities to obtain the state probability, and complete the detection.

[0040] In one embodiment, for the same modal component, the normalized result of the difference between the first offset mean and the second offset mean is taken as the importance.

[0041] Importance satisfies the following relation: , Represents modal components The importance of Represents modal components The first offset mean, Represents modal components The second offset mean, This represents the normalization function.

[0042] The greater the difference between the mean offsets under different labels, the higher the discriminative power of the feature under different labels. In other words, the more significant the role of the feature in distinguishing different states of power grid load (i.e., normal and abnormal). Therefore, the greater the difference between features, the higher the accuracy of the model's judgment, thereby improving the accuracy of power grid load state detection and prediction.

[0043] Training a pre-defined network based on importance involves: taking the importance of any modality component in the history as input, and the actual value of the sub-probability of any modality component in the history as the network label to obtain a training set; inputting the training set into the pre-defined network to output the predicted value of the sub-probability; calculating the loss value between the predicted value and the actual value of the sub-probability using the cross-entropy loss function; iteratively updating the parameters of the pre-defined network; stopping the update when the pre-defined network reaches the set maximum number of training iterations or the loss value is less than the set loss value, thus obtaining the trained pre-defined network.

[0044] The frequency values ​​of several modal components of the real-time load value are input into the corresponding pre-trained preset network, and the sub-probabilities are output. The sub-probabilities of each modal component are obtained by traversing the network. For any modal component, the cumulative value of the importance of all modal components is calculated. The ratio of the importance of the modal component to the cumulative value is used as the weight. The product of the weight and the importance is calculated. The cumulative value of the product of all modal components is used as the state probability.

[0045] The state probabilities satisfy the following relationship: , Represents the state probability. Represents modal components The importance of Represents modal components The probability of the sub-probability.

[0046] In one embodiment, importance includes: for the same modal component, calculating the cosine similarity between the first offset mean and the second offset mean, and using the sum of the cosine similarity and 1 through a negative correlation mapping as the importance. Summing with 1 ensures that the importance of each modal component is between 0 and 1.

[0047] The system includes a processor and a memory, the memory storing computer program instructions, which, when executed by the processor, implement a steam leak detection method for a waste heat recovery system according to the first aspect of the present invention.

[0048] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0049] It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the scope of protection of this application. Therefore, the scope of protection of this patent application shall be determined by the appended claims.

Claims

1. A method for real-time detection of power load in an electrical grid, characterized in that, The method comprises the following steps: obtaining the pretreated load value of the power grid, constructing a load sequence, and labeling each load value, including normal and abnormal; based on VMD, the load sequence is decomposed to obtain a plurality of modal components, and the frequency value of any modal component at each sampling time is obtained based on wavelet transform; The frequency values ​​of any modal component are classified into normal and abnormal classes according to the label. The first offset mean of the normal class and the second offset mean of the abnormal class are calculated, satisfying the following relationship: , Indicates that the label is Time modal components The offset mean, Indicates that the label is Time modal components Mid-sampling time The frequency value, Indicates that the label is Time modal components The center point, Indicates the total number of sampling times; the importance of the modal component is calculated according to the first offset mean and the second offset mean, the preset network is trained based on the importance, one modal component corresponds to one preset network, the frequency value of the real-time load value of the plurality of modal components is input into the corresponding trained preset network, and the sub-probability is output; all sub-probabilities are integrated to obtain a state probability, including: for any modal component, the cumulative value of the importance of all modal components is calculated, the ratio of the importance of the modal component to the cumulative value is taken as the weight, the product of the weight and the sub-probability is calculated, and the cumulative value of the product of all modal components is taken as the state probability; and the detection is completed; the plurality of modal components comprise: using Fourier transform on the load sequence to obtain a load frequency set, clustering the load frequency set to obtain a clustering cluster, and the number of clustering clusters is the same as the number of modal components.

2. The method of claim 1, wherein, The preprocessing comprises: acquiring the load value according to a preset sampling interval, and performing denoising, missing value filling and abnormal value deletion on the acquired load value.

3. The method of claim 1, wherein, The clustering method is K-means clustering or density clustering.

4. The method of claim 1, wherein, The calculation of the first offset mean of the normal class and the second offset mean of the abnormal class comprises: calculating the center point of all frequency values in the normal class, and taking the vector between the frequency value at any sampling time and the center point as a feature vector; for the same modal component, the mean of the feature vectors at all sampling times is taken as the first offset mean, and the second offset mean of the abnormal class is obtained in the same way.

5. The method of claim 1, wherein, The importance comprises: for the same modal component, the difference between the first offset mean and the second offset mean is normalized to obtain the importance.

6. The method of claim 1, wherein, The importance comprises: for the same modal component, the cosine similarity of the first offset mean and the second offset mean is calculated, and the result of the cosine similarity and 1 through negative correlation mapping is taken as the importance.

7. The method of claim 1, wherein, The training of the preset network based on the importance comprises: taking the importance of any modal component in the history as input, taking the actual value of the sub-probability of any modal component in the history as the network label, obtaining a set of training sets, inputting the training sets into the preset network, and outputting the predicted value of the sub-probability; using cross-entropy loss function to calculate the loss value of the predicted value of the sub-probability and the actual value of the sub-probability, iteratively updating the preset network parameters, stopping updating when the preset network reaches the set maximum training times or the loss value is less than the set loss value, and obtaining the trained preset network.

8. A real-time detection system for power load of a power grid, characterized in that, The method comprises: a processor and a memory, the memory stores computer program instructions, and when the computer program instructions are executed by the processor, a power grid power load real-time detection method according to any one of claims 1-7 is realized.

Images