Non-intrusive load identification method, system and device based on pruning and medium

Abstract

The invention relates to the technical field of power distribution network planning, and particularly discloses a pruning-based non-intrusive load identification method, system and device and a medium, and the method comprises the steps: collecting data, extracting effective information in the data through a feature extraction algorithm, forming a feature vector, and constructing a typical power curve template library based on the collected data. Constructing a non-intrusive load monitoring model which comprises a backbone model and a non-intrusive load identification model, lightening the backbone model, and training the non-intrusive load monitoring model by using the feature vector as input to obtain an optimized non-intrusive load monitoring model, and inputting the real-time operation data of the main feeder line of the power grid into the optimized non-intrusive load monitoring model, and combining with the typical power curve template library to obtain the real-time electrical load data of the branch feeder line of the power grid and the corresponding load type thereof. The pruning algorithm is introduced, the complexity and parameter quantity of the model are reduced, and the calculation efficiency and the recognition rate of the model are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of distribution network planning, and in particular to a non-intrusive load identification method, system, device and medium based on pruning. Background Art

[0002] With the development of smart grids and the demand for energy management, non-intrusive load monitoring (NILM) technology has received widespread attention in power load identification. NILM technology does not require the installation of additional sensors. It can infer the power consumption of various household appliances or equipment by analyzing the total meter data. It has the advantages of low cost and easy installation. However, traditional load identification methods are insufficient in recognition accuracy and real-time performance when processing complex and diverse power signals. Pruning, as a lightweight convolutional neural network method, has the advantages of small computational complexity and fast speed. It can significantly improve the efficiency of load identification while maintaining high recognition accuracy. It is especially suitable for large-scale real-time power monitoring scenarios. Summary of the invention

[0003] The purpose of the present invention is to provide a non-intrusive load identification method, system, device and medium based on pruning to solve the problems of low identification accuracy and low calculation efficiency in existing load identification technology.

[0004] In order to solve the above technical problems, the present invention provides a non-intrusive load identification method based on pruning, comprising the following steps:

[0005] S1, collecting real-time operation data of the main feeder of the power grid, historical operation data of the main feeder of the power grid, historical power consumption data of the substation area and historical power consumption data of users, extracting effective information therein through a feature extraction algorithm to form a feature vector;

[0006] S2. Based on the collected data, a typical power curve template library is constructed;

[0007] S3. Construct a non-intrusive load monitoring model, including a backbone model and a non-intrusive load identification model, and lightweight the backbone model;

[0008] S4, using the feature vector as input to train the non-intrusive load monitoring model to obtain an optimized non-intrusive load monitoring model;

[0009] S5. Input the real-time operation data of the main feeder of the power grid into the optimized non-intrusive load monitoring model, and combine it with the typical power curve template library to obtain the real-time power load data of the branch feeder of the power grid and its corresponding load type.

[0010] Preferably, in step S1, the collected data specifically include voltage, current, power, number of electrical devices and time.

[0011] Preferably, the step S2 includes:

[0012] S2-1, randomly take a small amount of collected data and label it as the existing labeled data, initialize it according to the existing labeled data, and determine k cluster centers;

[0013] S2-2, repeat clustering until the cluster center no longer changes;

[0014] S2-3, taking the remaining collected data as objects, calculating the distance between each object and each cluster center, and assigning each object to the nearest cluster center, and the label corresponding to each object and its nearest cluster center is a pseudo label;

[0015] S2-4, recalculate the cluster center;

[0016] S2-5. Compare and calculate the pseudo-label with the recalculated cluster center, optimize the pseudo-label, and obtain a typical power curve template library.

[0017] Preferably, the step S3 includes:

[0018] S3-1. Constructing a non-intrusive load identification model:

[0019] y t =x i,t +o i,t +`o t ;

[0020] Among them, y t represents the total power at time t, x i,t is the power of the ith electrical device at time t, o i,t is the power of other electrical equipment except the i-th electrical equipment at time t, `o i,t is random noise, satisfying `o t ~N(0,σ 2 ), i represents the number of electrical equipment;

[0021] S3-2. Construct a backbone model with pruning structure:

[0022] Set the pruning target percentage for the convolutional layer and dense layer, and evaluate the importance of the filter in each convolutional layer. The importance of the filter is:

[0023]

[0024] Among them, F irepresents the i-th filter, N is the number of filters, and p is 1 or 2. During each pruning, the less important ones are removed first, and the convolutional layers and dense layers are pruned in a certain increment. After each pruning iteration, part of the collected data is randomly selected to train the backbone model, and pruning continues until the final target pruning percentage is reached.

[0025] Preferably, the step S4 includes:

[0026] S4-1, randomly initializing the parameters of the non-intrusive load monitoring model, including selecting the initialization strategy, initializing the weights, setting the learning rate and the optimizer;

[0027] S4-2, calculate the output of the non-intrusive load monitoring model and the actual loss;

[0028] S4-3. Update network parameters using gradient descent:

[0029]

[0030] Among them, α is the update rate of the objective value function neural network, W′ is the updated weight, W0 is the initial weight, b′ is the updated bias, b0 is the initial bias, and f is the actual loss;

[0031] S4-4, repeating steps S4-1 to S4-3 multiple times to obtain a trained non-intrusive load monitoring model;

[0032] S4-5. Randomly take part of the collected data as a test set to evaluate the performance of the trained non-intrusive load monitoring model, and optimize the non-intrusive load monitoring model according to the evaluation results to obtain an optimized non-intrusive load identification model.

[0033] Preferably, in step S5, the real-time operation data of the main feeder of the power grid is input into the optimized non-intrusive load monitoring model to obtain the missing real-time power load data of the branch feeder of the power grid and the initial load identification result, the initial load identification result is compared with the typical power curve template library to determine the load type, and the typical power curve template library is used to complete the missing data to obtain complete real-time power load data of the branch feeder of the power grid and its corresponding load type.

[0034] A non-intrusive load identification system based on pruning, comprising:

[0035] The data collection and analysis module is used to collect the real-time operation data of the main feeder of the power grid, the historical operation data of the main feeder of the power grid, the historical power consumption data of the substation area and the historical power consumption data of the users, and extract the effective information therein through the feature extraction algorithm to form a feature vector;

[0036] A typical power curve template library construction module is used to construct a typical power curve template library based on the collected data;

[0037] A non-intrusive load monitoring model building module is used to build a non-intrusive load monitoring model, including a backbone model and a non-intrusive load identification model, and to lighten the backbone model;

[0038] An optimization module, used for training the non-intrusive load monitoring model using the feature vector as input to obtain an optimized non-intrusive load monitoring model;

[0039] The application module is used to input the real-time operation data of the main feeder of the power grid into the optimized non-intrusive load monitoring model, and combine it with the typical power curve template library to obtain the real-time power load data of the branch feeder of the power grid and its corresponding load type.

[0040] A non-intrusive load identification device based on pruning comprises a memory and a processor, wherein the memory is used to store a computer program, and the computer program is used to execute the method described in any one of claims 1 to 6 when loaded by the processor.

[0041] A readable storage medium, wherein a computer program is stored in the storage medium, and the computer program is suitable for executing the method according to any one of claims 1 to 6 when loaded by a processor.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] The present invention introduces a pruning algorithm into the non-intrusive load monitoring model, prunes the redundant parameters in the neural network, and optimizes the structural design of the model. Pruning reduces the complexity and parameter amount of the model by deleting unnecessary filter connections, thereby improving the computational efficiency and recognition rate of the model. In this way, the computational complexity and storage requirements of the model are significantly reduced, allowing it to run more efficiently in edge devices and embedded systems with limited computing resources. Compared with traditional models, the pruned lightweight model has obvious advantages in hardware resource requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 This is a flow chart of an embodiment of the present invention;

[0045] Figure 2 This is a flow chart of a typical power curve template library according to an embodiment of the present invention;

[0046] Figure 3 Schematic diagram of the backbone model of the non-intrusive load monitoring model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0047] The following embodiments of the present invention are described in further detail in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0048] This embodiment is specifically a non-intrusive load identification method based on pruning, such as Figure 1 As shown, the following steps are included:

[0049] S1. Collect real-time operation data of the main feeder of the power grid, historical operation data of the main feeder of the power grid, historical power consumption data of the substation area and historical power consumption data of users. The collected data are specifically voltage, current, power, number of electrical equipment and time. The collected data are processed, including pre-processing operations such as denoising and normalization. The effective information is extracted through feature extraction algorithm to form a feature vector.

[0050] S2, based on the collected data, such as Figure 2 As shown, a typical power curve template library is constructed:

[0051] S2-1, randomly take a small amount of collected data and label it as the existing labeled data, initialize it according to the existing labeled data, and determine k cluster centers;

[0052] S2-2, repeat clustering until the cluster center no longer changes;

[0053] S2-3, taking the remaining collected data as objects, calculating the distance between each object and each cluster center, and assigning each object to the nearest cluster center, and the label corresponding to each object and its nearest cluster center is a pseudo label;

[0054] S2-4, recalculate the cluster center;

[0055] S2-5. Compare and calculate the pseudo-label with the recalculated cluster center, optimize the pseudo-label, and obtain a typical power curve template library.

[0056] S3. Construct a non-intrusive load monitoring model, including a backbone model and a non-intrusive load identification model, and lightweight the backbone model:

[0057] S3-1. The goal of non-intrusive load identification is to identify the power consumption of a single load from the total electrical load information. Considering the powerful representation ability of neural networks for end-to-end mapping, deep learning models usually directly learn the mapping relationship between the total power sequence and the power sequence of the i-th electrical equipment over a period of time. There is no need to consider the impact of electrical equipment that users are not interested in and other unknown electrical equipment on the decomposition effect. Therefore, for a time node in the sequence, the non-intrusive load identification model is:

[0058] y t =xi,t +o i,t +`o t ;

[0059] Among them, y t represents the total power at time t, x i,t is the power of the ith electrical device at time t, o i,t is the power of other electrical equipment except the i-th electrical equipment at time t, `o i,t is random noise, satisfying `o t ~N(0,σ 2 ), i represents the number of electrical equipment;

[0060] S3-2. Build a backbone model with pruning structure. The overall architecture is as follows Figure 3 As shown, there are an input layer, five convolutional layers, a fully connected layer, and an output layer;

[0061] Pruning is a technique used to reduce the size and complexity of deep learning models. It removes less important parameters to achieve lightweight models, reduce inference time and improve deployment efficiency. In this embodiment, filter pruning is used. Filter pruning is to remove unimportant filters in the convolutional layer based on predefined criteria. Neuron pruning can also be used. Neuron pruning is to selectively remove neurons with smaller contributions in the dense layer. By focusing on the most important features, pruning can improve model efficiency without significantly reducing model accuracy.

[0062] Set pruning target percentages for convolutional layers and dense layers. In this embodiment, the pruning targets for convolutional layers and dense layers are set to 30%, 60%, and 90%. Evaluate the importance of the filters in each convolutional layer. The importance of the filters is:

[0063]

[0064] Among them, F i represents the i-th filter, N is the number of filters, and p is 1 or 2. Each time pruning is performed, the less important ones are removed first. The convolutional layers and dense layers are pruned in a certain increment. This embodiment takes an increment of 10%. After each pruning iteration, a portion of the collected data is randomly taken to retrain the backbone model to restore its accuracy. Retraining can adjust the model to a smaller set of parameters to ensure model performance. Continue pruning until the final target pruning percentage is reached, which can reach up to 90% for convolutional layers and dense layers. At each pruning level, the performance of the model on the test data is evaluated to ensure that the expected accuracy and efficiency requirements are met.

[0065] The optimal number of clusters is determined by the elbow method and combined with the constructed typical power curve template library for load identification. The result of the load identification model is compared with the typical power curve template library to determine the identified load type. In the case of missing actual load power, the missing measurement data of the corresponding period is supplemented by the typical power curve template library.

[0066] S4, using the feature vector as input to train the non-intrusive load monitoring model to obtain an optimized non-intrusive load monitoring model;

[0067] S4-1. Randomly initialize the parameters of the non-intrusive load monitoring model, including selecting the initialization strategy, initializing the weights, setting the learning rate and the optimizer. According to the activation function of each layer and the characteristics of the network, select the appropriate initialization method and weight initialization method. In this example, the Xavier random initialization method is used, the learning rate is set to 0.001, and the optimizer uses the Adam method.

[0068] S4-2, calculate the output of the non-intrusive load monitoring model and the actual loss;

[0069] S4-3. Update network parameters using gradient descent:

[0070]

[0071] Wherein, α is the update rate of the objective value function neural network. In this embodiment, it is set to 0.0001, W′ is the updated weight, W0 is the initial weight, b′ is the updated bias, b0 is the initial bias, and f is the actual loss;

[0072] S4-4, repeating steps S4-1 to S4-3 multiple times to obtain a trained non-intrusive load monitoring model. In this embodiment, the number of repetitions is set to 10,000;

[0073] S4-5. Randomly take part of the collected data as a test set to evaluate the performance of the trained non-intrusive load monitoring model, and optimize the non-intrusive load monitoring model according to the evaluation results to obtain an optimized non-intrusive load identification model. In this embodiment, MSE, MAE, and MPAE are used as methods to evaluate model performance.

[0074] S5. Input the real-time operation data of the main feeder of the power grid into the optimized non-intrusive load monitoring model, and combine it with the typical power curve template library to obtain the real-time power load data of the branch feeder of the power grid and its corresponding load type. Input the real-time operation data of the main feeder of the power grid into the optimized non-intrusive load monitoring model to obtain the missing real-time power load data of the branch feeder of the power grid and the initial load identification results, compare the initial load identification results with the typical power curve template library to determine the load type, and use the typical power curve template library to complete the missing data to obtain the complete real-time power load data of the branch feeder of the power grid and its corresponding load type.

[0075] A non-intrusive load identification system based on pruning, comprising:

[0076] The data collection and analysis module is used to collect the real-time operation data of the main feeder of the power grid, the historical operation data of the main feeder of the power grid, the historical power consumption data of the substation area and the historical power consumption data of the users, and extract the effective information therein through the feature extraction algorithm to form a feature vector;

[0077] A typical power curve template library construction module is used to construct a typical power curve template library based on the collected data;

[0078] A non-intrusive load monitoring model building module is used to build a non-intrusive load monitoring model, including a backbone model and a non-intrusive load identification model, and to lighten the backbone model;

[0079] An optimization module, used for training the non-intrusive load monitoring model using the feature vector as input to obtain an optimized non-intrusive load monitoring model;

[0080] The application module is used to input the real-time operation data of the main feeder of the power grid into the optimized non-intrusive load monitoring model, and combine it with the typical power curve template library to obtain the real-time power load data of the branch feeder of the power grid and its corresponding load type.

[0081] A non-intrusive load identification device based on pruning comprises a memory and a processor, wherein the memory is used to store a computer program, and the computer program is used to execute the above method when loaded by the processor.

[0082] A readable storage medium stores a computer program, wherein the computer program is suitable for executing the above method when loaded by a processor.

[0083] The embodiments of the present invention are given for the purpose of illustration and description, and are not intended to be exhaustive or to limit the invention to the disclosed forms. Many modifications and variations will be apparent to those of ordinary skill in the art. The embodiments are selected and described in order to better illustrate the principles and practical applications of the present invention and to enable those of ordinary skill in the art to understand the present invention and thereby design various embodiments with various modifications suitable for specific uses.

Claims

1. A non-intrusive load identification method based on pruning, characterized in that: The following steps are involved: S1, collecting real-time operation data of the main feeder of the power grid, historical operation data of the main feeder of the power grid, historical power consumption data of the substation area and historical power consumption data of users, extracting effective information therein through a feature extraction algorithm to form a feature vector; S2. Based on the collected data, a typical power curve template library is constructed; S3. Construct a non-intrusive load monitoring model, including a backbone model and a non-intrusive load identification model, and lightweight the backbone model; S4, using the feature vector as input to train the non-intrusive load monitoring model to obtain an optimized non-intrusive load monitoring model; S5. Input the real-time operation data of the main feeder of the power grid into the optimized non-intrusive load monitoring model, and combine it with the typical power curve template library to obtain the real-time power load data of the branch feeder of the power grid and its corresponding load type.

2. The non-intrusive load identification method based on pruning according to claim 1 is characterized in that: In step S1, the collected data specifically include voltage, current, power, number of electrical devices and time.

3. The non-intrusive load identification method based on pruning according to claim 2 is characterized in that: The step S2 includes: S2-1, randomly take a small amount of collected data and label it as the existing labeled data, initialize it according to the existing labeled data, and determine k cluster centers; S2-2, repeat clustering until the cluster center no longer changes; S2-3, taking the remaining collected data as objects, calculating the distance between each object and each cluster center, and assigning each object to the nearest cluster center, and the label corresponding to each object and its nearest cluster center is a pseudo label; S2-4, recalculate the cluster center; S2-5. Compare and calculate the pseudo-label with the recalculated cluster center, optimize the pseudo-label, and obtain a typical power curve template library.

4. The non-intrusive load identification method based on pruning according to claim 3 is characterized in that: The step S3 includes: S3-1. Constructing a non-intrusive load identification model: and t =x i,t +o i,t +`o t ; Among them, y t represents the total power at time t, x i,t is the power of the ith electrical device at time t, o i,t is the power of other electrical equipment except the i-th electrical equipment at time t, `o i,t is random noise, satisfying `o t ~N(0,σ 2 ), i represents the number of electrical equipment; S3-2. Construct a backbone model with pruning structure: Set the pruning target percentage for the convolutional layer and dense layer, and evaluate the importance of the filter in each convolutional layer. The importance of the filter is: Among them, F i represents the i-th filter, N is the number of filters, and p is 1 or 2. During each pruning, the less important ones are removed first, and the convolutional layers and dense layers are pruned in a certain increment. After each pruning iteration, part of the collected data is randomly selected to train the backbone model, and pruning continues until the final target pruning percentage is reached.

5. The non-intrusive load identification method based on pruning according to claim 4 is characterized in that: The step S4 includes: S4-1, randomly initializing the parameters of the non-intrusive load monitoring model, including selecting the initialization strategy, initializing the weights, setting the learning rate and the optimizer; S4-2, calculate the output of the non-intrusive load monitoring model and the actual loss; S4-3. Update network parameters using gradient descent: Among them, α is the update rate of the objective value function neural network, W′ is the updated weight, W0 is the initial weight, b′ is the updated bias, b0 is the initial bias, and f is the actual loss; S4-4, repeating steps S4-1 to S4-3 multiple times to obtain a trained non-intrusive load monitoring model; S4-5. Randomly take part of the collected data as a test set to evaluate the performance of the trained non-intrusive load monitoring model, and optimize the non-intrusive load monitoring model according to the evaluation results to obtain an optimized non-intrusive load identification model.

6. The non-intrusive load identification method based on pruning according to claim 5 is characterized in that: In step S5, the real-time operation data of the main feeder of the power grid is input into the optimized non-intrusive load monitoring model to obtain the missing real-time power load data of the branch feeder of the power grid and the initial load identification result, the initial load identification result is compared with the typical power curve template library to determine the load type, and the typical power curve template library is used to complete the missing data to obtain the complete real-time power load data of the branch feeder of the power grid and its corresponding load type.

7. A non-intrusive load identification system based on pruning, characterized in that: include: The data collection and analysis module is used to collect the real-time operation data of the main feeder of the power grid, the historical operation data of the main feeder of the power grid, the historical power consumption data of the substation area and the historical power consumption data of the users, and extract the effective information therein through the feature extraction algorithm to form a feature vector; A typical power curve template library construction module is used to construct a typical power curve template library based on the collected data; A non-intrusive load monitoring model building module is used to build a non-intrusive load monitoring model, including a backbone model and a non-intrusive load identification model, and to lighten the backbone model; An optimization module, used for training the non-intrusive load monitoring model using the feature vector as input to obtain an optimized non-intrusive load monitoring model; The application module is used to input the real-time operation data of the main feeder of the power grid into the optimized non-intrusive load monitoring model, and combine it with the typical power curve template library to obtain the real-time power load data of the branch feeder of the power grid and its corresponding load type.

8. A non-intrusive load identification device based on pruning, characterized in that: The method comprises a memory and a processor, wherein the memory is used to store a computer program, and the computer program is used to execute the method according to any one of claims 1 to 6 when loaded by the processor.

9. A readable storage medium, characterized in that: The storage medium stores a computer program, and the computer program is suitable for executing the method according to any one of claims 1 to 6 when loaded by a processor.

Images