Malicious load identification method based on load event detection

Abstract

The application discloses a malignant load identification method based on load event detection. The method comprises the following steps: extracting power information features of each type of load running alone from electrical data of normal load and malignant load running alone, and establishing a load classification algorithm; detecting load events in real time based on power information features collected by an electric energy meter in real time; the load event detection comprises the following steps: in the process of measuring power information data in real time, whether a load event occurs is determined by comparing whether real-time power information data of a previous period of time and real-time power information data of a next period of time are of the same distribution; extracting power information features with incremental information from power information data of the load event detected in real time; and using the established load classification algorithm to identify and determine malignant load of a newly added load in real time. The malignant load identification method based on load event detection has the advantages of fast response speed, high reliability, accuracy, effectiveness and easy implementation.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of load identification and power safety, in particular to a malignant load identification method based on load event detection which is easy to cause power safety hazards. BACKGROUND

[0002] At present, due to illegal use of heating resistance loads such as hot air guns and electric blankets which have heating function, fire accidents frequently occur in college dormitories. At the same time, in student dormitories, rental apartments and other places, due to the limited design capacity of the line, the use of high-power heating equipment such as induction cookers, it is easy to cause line overheating and cause fires. Therefore, in order to ensure power safety, the above-mentioned types of electrical appliances which are easy to cause safety hazards are defined as malignant loads, and the identification and detection of malignant load electrical appliances are proposed, which restricts the use of malignant load electrical appliances under the premise of ensuring the normal use of other non-malignant load electrical appliances.

[0003] The existing malignant load identification method mainly sets the power factor and power threshold as the identification standard, but these methods need to set the related threshold manually, and it is easy to cause misjudgment for the unusual load. In addition, there are also some methods that analyze high-frequency current and voltage waveform data to achieve the identification of malignant loads, but this type of method requires high-frequency acquisition equipment, which has high implementation cost and is not suitable for configuration on a regular prepayment meter, and cannot exclude the influence of background electrical equipment, so the identification accuracy is not high. In summary, the current malignant load identification method mainly faces the problems of low generalization ability and difficulty in meeting the needs of malignant load identification in actual power consumption scenarios. SUMMARY

[0004] The purpose of the present application is to overcome the shortcomings of the existing malignant load identification method, and provide a malignant load identification method based on load event detection. The malignant load identification method based on load event detection has fast response speed, high reliability, accuracy and effectiveness, and is easy to implement.

[0005] The basic logic is to first detect the event of adding a new electrical appliance in the circuit (load event), and then judge whether the event is the access of a malignant load electrical appliance through an algorithm. The specific steps are as follows: the malignant load identification method based on load event detection of the present application specifically includes:

[0006] From the electrical data of normal load and malignant load running alone, the power information characteristics of each type of load running alone are extracted to establish a load classification algorithm;

[0007] Based on the real-time power information characteristics collected by the electric energy meter, the load event is detected in real time;

[0008] Load event detection involves determining whether a load event has occurred by comparing real-time power information data from a previous period with data from a subsequent period during the real-time measurement of power information data.

[0009] Extract power information features with incremental information from power information data of real-time detected load events;

[0010] The established load classification algorithm is used to identify and judge malicious loads in real time when new loads are added.

[0011] The established classification model is compressed to enable it to be embedded in smart meters.

[0012] The same distribution judgment includes comparing the threshold of the proportion of each bin and comparing the overall distribution of the difference.

[0013] It also includes model compression of the classification model and embedding it into smart meters.

[0014] The malicious load identification method based on load event detection of the present invention has the following advantages compared with the prior art:

[0015] Its basic logic is to first detect the event of a newly added electrical appliance in the circuit (load event), and then use an algorithm to determine whether the event is a malicious load appliance connection. Through a special identification process, it can complete the identification and response within three seconds. In the load event detection part, it innovatively proposes a method for judging whether the distribution is the same, which can achieve preliminary screening of events including but not limited to malicious loads. By modifying the same distribution condition, it can also achieve the identification of loads with specific power. Its processing method can effectively extract the power information features of the corresponding electrical appliances, eliminate the influence of the power consumption background, and remove events that are obviously not malicious loads. It can reduce the frequency of subsequent model identification, reduce the CPU operating pressure, and improve the practicality in actual operation scenarios. Meanwhile, to maintain the robustness of the model under hardware constraints, the proposed method to remove deployment restrictions decomposes the original matrix by transforming the matrix, then compresses it using cosine similarity, and selects a larger amount of information through rotation and scaling, tightening the parameterization boundary. This greatly reduces the memory resource allocation and computation time required during the model judgment program's calculation, reducing the pressure on the prepaid meter hardware and enabling it to be embedded and calculated more easily, further improving its practicality. Compared with existing technologies, it has lower hardware requirements.

[0016] Based on the hardware limitations and timeliness of smart meters, this paper innovatively proposes to compress and optimize the model judgment process. While maintaining the similarity between the performance and the original model, it greatly reduces the memory resource allocation and computing time required in the model judgment program by tightening the parameterization boundary, thereby improving feasibility and practicality. It achieves accurate identification and judgment of malicious loads based on the second level, and the implementation cost is low. Attached Figure Description

[0017] Figure 1 This is a flowchart of an embodiment of the malicious load identification method based on load event detection of the present invention; Figure 2 This is a flowchart of an embodiment of the real-time load event detection method in the malicious load identification method based on load event detection of the present invention. Detailed Implementation

[0018] The present invention will now be described in a clearer and more complete manner with reference to the accompanying drawings and specific embodiments. It should be noted that, without conflict, the various embodiments or technical features described below can be arbitrarily combined to form new embodiments.

[0019] Please see Figures 1-2 In this embodiment (taking a prepaid electricity meter in a school dormitory building as an example), the malicious load identification method based on load event detection includes the following steps:

[0020] S101: Extract the power information features of each type of load when it is running alone from the electrical data of normal load and severe load, and establish a load classification model based on this.

[0021] First, define the electrical information characteristics of the load when it is operating alone: ​​under the condition of background power consumption, there exists an electrical characteristic in the circuit. H If the condition is the same as that in a circuit where only that load (appliance device) is working, then this electrical characteristic is considered... H It is defined as the power information characteristic of the load itself.

[0022] By definition, the electrical characteristics of a load H This includes, but is not limited to, changes in active power ΔP, reactive power ΔQ, apparent power ΔS, power factor ΔPF, and the k-th harmonic content Δh before and after connecting a new load. k Their mathematical definitions are as follows:

[0023] , The superscripts after and before indicate after and before the new load is connected, respectively.

[0024] , The superscripts after and before indicate after and before the new load is connected, respectively.

[0025] , The superscripts after and before indicate after and before the new load is connected, respectively.

[0026] ;

[0027] , The superscripts "after" and "before" indicate after the new load is connected and before it is connected, respectively. k express k Secondary harmonics.

[0028] The specific steps for establishing a load classification model by extracting power information features of incremental load information for each type of load from electrical data of normal and severe loads operating independently include:

[0029] S101-1: Collect electrical data of various normal and malicious load electrical appliances when they are working alone. These electrical data include: active power, reactive power, apparent power, power factor, and harmonic content of each current. Mark these electrical data as belonging to normal load or malicious load, with normal load marked as 0 and malicious load marked as 1.

[0030] S101-2: Preprocess the power information data across the aforementioned multiple dimensions using the Kernel Principal Component Analysis (KPCA) algorithm. The kernel function selected in the KPCA process is not limited to Gaussian kernels, cosine similarity kernels, or polynomial kernels.

[0031] S101-3: Divide the preprocessed power consumption data of normal load and severe load into training set and test set;

[0032] S101-4: Using the Support Vector Machine (SVM) model, a binary classification model for load type is built on the training set. This model is used to predict the test set data and compared with the true label values ​​marked in S101-1. The hyperparameters of the SVM model are optimized based on the F1 score of the model's prediction results on the test set, and finally the model with the best performance is obtained.

[0033] S102: Based on the real-time power information data collected by the electricity meter, detect load events (i.e. load access events) in real time and save the power information data before and after the load access event occurs;

[0034] The basic logic of load event detection is to determine whether a load event has occurred by comparing the power information data measured in real time with that measured in a later time period to see if they exhibit the same distribution. This real-time power information data includes active and reactive power. The determination of the same distribution mainly consists of two steps: comparing the percentage threshold of different power distribution boxes and comparing the overall distribution of the difference. This approach effectively extracts the power information characteristics of the corresponding electrical appliances, eliminates the influence of background electricity consumption, and removes events that are clearly not malicious loads. This reduces the frequency of subsequent model recognition, decreases CPU workload, and improves practicality in real-world operating scenarios.

[0035] Taking a sampling frequency of 5Hz and collecting 10 seconds of data as an example, the data from the first period is the array of measured values ​​stored in the first 5 seconds, and the data from the second period is the array of measured values ​​stored in the last 5 seconds. The measured values ​​from the two periods together constitute the most recent power information data measurement value, and the array length is 25.

[0036] The smart meter records a series of recent active and reactive power measurements. First, these measurements are smoothed using a recursive median filter. Then, a sliding window bilateral algorithm is used to determine if the power information data in the same dimension over two consecutive time periods are identically distributed. If this condition is met, it indicates a different distribution and a load event has occurred; otherwise, no load event has occurred. The specific steps of the identical distribution determination method include:

[0037] S102-1: First, divide the data into equal-interval bins for both the first and second segments to obtain the corresponding distribution proportions. Then, calculate the absolute value of the difference in the proportions within the same period. Next, the absolute values ​​of the differences in the distribution proportions of each bin are added together to obtain the distribution corresponding to both sides of the sliding window. Finally, a reasonable threshold is given by batch experimental data to make a preliminary judgment on the same distribution.

[0038] S102-2: Obtain the preliminary same distribution judgment results filtered in step S102-1, then take the median of the two data segments after recursive median smoothing filtering and perform difference processing, and calculate the corresponding standard deviation. Then, further same distribution judgment is performed by comparing whether the difference is greater than three times the sum of standard deviations. Here, the sum of three standard deviations can be modified according to the specific needs of load identification.

[0039] In this embodiment, after each completion of the above load event detection and judgment, a preliminary judgment of malicious load will be performed. The latest measured power information data of the next moment will be introduced in the manner of head-in and tail-out (i.e., in the order of time before and after), and the earliest historical measured power information data will be discarded. The measured power information data will be updated and continuously cycled to finally realize real-time load event detection.

[0040] Following the load event detection described above, a preliminary judgment of malicious loads is performed. This filters out events that are clearly not malicious loads or are normal fluctuations in the circuit, reducing the frequency of subsequent malicious load identification algorithm calls and computational burden. The preliminary judgment of malicious loads can be adjusted by changing the threshold of the minimum power change of load events, thus meeting the different accuracy requirements of the aforementioned load event identification methods. Compared to existing technologies, this approach increases the requirements for triggering malicious load judgments without affecting recognition accuracy, thereby reducing hardware memory consumption.

[0041] Once the load event detection method identifies a load connection, it will save the original data of various dimensions such as active power, reactive power, and apparent power before and after the load connection, which will be used to extract the features required by the malicious load identification algorithm.

[0042] S103: Extract data with incremental information from the real-time detected load event power information data as power information feature data.

[0043] In this embodiment, the basic data used in the malicious load identification and classification model established in step S101 includes the following dimensions: change in active power ΔP, change in reactive power ΔQ, change in apparent power ΔS, change in power factor ΔPF, and changes in the content of the first five odd-order current harmonics Δh3 and Δh5 before and after the new load is connected. Therefore, the power information features with incremental information that need to be extracted in this step are the above six dimensions of data, which should be the same as the basic data used to establish the malicious load identification and classification model.

[0044] The current harmonic characteristics are directly measured and analyzed by the energy chip inside the smart meter. The calculation method for the change in power factor ΔPF before and after a new load connection event is as follows: , where ΔP is the change in active power and ΔQ is the change in reactive power.

[0045] For this invention, the load event detection method described in steps S102-S103 can also be summarized as a preliminary screening method for malicious loads. Without affecting the recognition accuracy, the requirements for triggering malicious load judgment are increased, the consumption of hardware running memory is reduced, and events that are obviously not malicious loads or normal loop fluctuations can be filtered out, thereby reducing the calling frequency and computational pressure of subsequent malicious load recognition algorithms.

[0046] S104: Use the load classification model established in step S101 to judge the newly added load as a malicious load in real time.

[0047] The classification model for determining whether a newly added load is malicious involves the following steps:

[0048] S104-1: Map the original power information feature data to a new feature space through kernel principal component analysis;

[0049] S104-2: Based on the established support vector machine model, calculate the final model's judgment result according to the support vectors selected from the samples by the model.

[0050] The data representing the newly added load events to be identified is as follows: .

[0051] The prediction process of the SVM model is as follows: ,in N The size of the data sample used when building the model; For support vector labeling, if If it is a support vector, set the value to 1; otherwise, set the value to 0. For the sample Load type flag; Kernel (...) represents the kernel function; represents the bias term, obtained during the SVM model training process. The number of support vectors in an SVM model is often much smaller than the number of samples used for training, while the prediction and judgment process of an SVM model only needs to use support vectors. This characteristic of the SVM model makes it more suitable for lightweight deployment of malicious load models.

[0052] S105: Compress the determined classification model so that it can be embedded in the smart meter.

[0053] S106: Based on the identification results of newly added loads using the malicious load identification model, provide feedback such as power outages and alarms.

[0054] Based on the malicious load identification results obtained from the malicious load identification method based on load event detection, malicious load access events with a predicted probability greater than a certain threshold are directly triggered by power outage, while malicious load access events with a predicted probability less than this threshold are triggered by alarms and fed back to the administrator. The predicted probability threshold can also be modified according to the usage requirements of the electricity meter in different scenarios.

[0055] The ability to complete power outage and alarm feedback within three seconds is faster than existing technologies, making it more suitable for use in scenarios with high electrical safety requirements, such as university dormitories.

[0056] For this invention, the model established in step S104 must be robust, requiring a sufficiently large number of samples. This results in a very large transformation matrix and high memory consumption, which is a major factor limiting the deployment of the algorithm on smart meter terminals. To address this issue, step S105 of the method of this invention employs SVD decomposition without affecting accuracy, transforming the transformation matrix... K Decomposition is called the dot product of three smaller matrices. And in this process, the sample matrix is ​​analyzed using cosine similarity. The selection of the matrix involves using rotation and scaling to select a larger amount of information, further compressing the size of the three smaller matrices used to calculate the matrix. This tightens the parameterization limits, significantly reducing the memory resource allocation and computation time required for the model judgment program, reducing the pressure on the prepaid meter hardware, making it easier to embed and perform calculations normally, further improving practicality, and lowering hardware requirements.

Claims

1. A method for malignant load identification based on load event detection, characterized in that The application relates to a load classification method and system. The power information features of various types of loads are extracted from the electrical data of normal loads and malignant loads running alone, so as to establish a load classification algorithm; The dimension basic data used by the load classification algorithm includes active power variation ΔP, reactive power variation ΔQ, apparent power variation ΔS, power factor variation ΔPF and kth current harmonic content variation Δh before and after accessing a new load k ; Based on the real-time power information features collected by the electric energy meter, load events are detected in real time; The load event detection includes comparing whether the real-time power information data of a previous period and that of a later period are of the same distribution in the process of measuring the real-time power information data, so as to determine whether a load event occurs; The same distribution judgment includes: First, the two data segments are respectively binned, and the corresponding distribution proportion is obtained , then the absolute value of the proportion difference of the same period is obtained , then the absolute value of the proportion difference of each bin is added to obtain the distribution of the sliding window double side, and finally a reasonable threshold is given through batch experimental data to make a preliminary same distribution judgment; The preliminary same distribution judgment result is obtained, then the difference between the medians of the two periods of data after recursive median smoothing filtering is calculated, and the corresponding standard deviation is calculated, then the further same distribution judgment is performed by comparing whether the difference is greater than the sum of three times the standard deviation; The power information features with incremental information are extracted from the power information data of the real-time detected load events; The established load classification algorithm is used for real-time malignant load identification and judgment of newly added loads; The determined classification model is compressed, so that the classification model can be embedded in the smart electric meter.

2. The method of claim 1, wherein the method further comprises: The same distribution judgment includes bin proportion threshold comparison and difference overall distribution comparison.

3. The method of claim 2, wherein the method further comprises: The classification model is compressed and embedded in the smart electric meter.

Images