A non-intrusive load decomposition method based on noise inference model

Abstract

The application provides a non-intrusive load decomposition method based on a noise inference model, and relates to the fields of power technology and artificial intelligence. The non-intrusive load decomposition method based on the noise inference model obtains active power consumption data of resident electricity; obtains a power data sequence with a proper length by using a sliding window method; trains a deep learning model by using a training algorithm based on a noise diffusion process; starts from random noise subject to a Gaussian distribution, and restores a target load element power curve by using a noise inference algorithm; the application processes the non-intrusive load decomposition into a conditional generation task of a load element power curve, models the distribution of load element power data by using a noise inference model, and provides a more stable training structure compared with a load decomposition algorithm based on a generative adversarial network.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electric power and artificial intelligence, in particular to a non-intrusive load decomposition method based on a noise inference model. BACKGROUND

[0002] With the continuous improvement of electrification and digitization in various fields of society, electric energy has gradually become the main form of terminal energy consumption in modern society, and electric power data also plays a very important role in energy big data. However, the existing electric power load data acquisition hardware system can only provide total power consumption data of users, which limits the demand side management measures such as load forecasting and energy saving guidance to some extent. Therefore, researchers and engineers have proposed a non-intrusive load decomposition technology (Non-Intrusive Load Monitoring, NILM) to decompose the electric energy consumption of each electrical appliance in the house according to the information provided by the total power meter. The detailed power consumption monitoring scheme based on non-intrusive load monitoring has the advantages of simplicity, low cost, high reliability, good data integrity, etc., and can obtain high-quality detailed power consumption information while making full use of existing infrastructure, which meets the concept of sustainable development.

[0003] The non-intrusive load decomposition task can be regarded as a task of generating power curves of individual load elements, so many scholars have studied the application of deep generative models in the field of load decomposition. Researchers from the University of Bath proposed a deep latent generative model based on variational recurrent neural networks (VRNNs), which performs sequence-to-multiple sequence prediction to generate continuous rather than discrete power consumption signals. A research group from Shandong University uses conditional generative adversarial networks (cGAN) to avoid artificial design of loss functions. Researchers from the National Technical University of Athens unified the autoencoder and generative adversarial network architecture, enabling the autoencoder to achieve nonlinear power signal source separation. Training a generative adversarial network requires achieving Nash equilibrium, and the design of adversarial training leads to poor stability of the network during training. SUMMARY

[0004] (I) Technical problems solved

[0005] In view of the deficiencies of the prior art, the present application provides a non-intrusive load decomposition method based on a noise inference model, which eliminates the training of the discriminator network, overcomes the problem of high training difficulty of load decomposition methods based on generative adversarial networks from the model structure, and at the same time ensures the accuracy of the load decomposition result.

[0006] (II) Technical solutions

[0007] To achieve the above object, the present application is implemented by the following technical solutions: a non-intrusive load decomposition method based on noise inference model, comprising:

[0008] obtaining resident active power consumption data;

[0009] obtaining target active power data sequence;

[0010] training the deep learning model parameters based on the target active power data sequence using the training algorithm based on noise diffusion process;

[0011] starting from random noise subject to Gaussian distribution, and combining the deep learning model after parameter training, the noise inference algorithm restores the power curve of the target load element.

[0012] Preferably, the resident active power consumption data includes aggregated active power data and active power data of load elements.

[0013] Preferably, the length of the target active power data sequence is at least twice the length of the single operation of the load element.

[0014] Preferably, the noise inference algorithm includes a forward process and a reverse process; wherein the forward process is a fixed Markov chain, and Gaussian noise is added to a clean target load element power curve until the curve loses significant features; the reverse process is a process of restoring the initial distribution step by step from random noise subject to Gaussian distribution, and the reverse transition distribution needs to be predicted by a deep learning model.

[0015] Preferably, the training algorithm of the noise diffusion process includes:

[0016] define linear noise control parameters

[0017] select samples x0~q(x) from the training set;

[0018] randomly select time step t value in {1, 2, … T};

[0019] take random noise ε~N(0, I);

[0020] use gradient descent method to train the model to minimize the distance between the actual noise and the model predicted noise wherein α t = 1-β t ,

[0021] repeat the above steps until the model converges;

[0022] The noise diffusion process step number is T, and T hidden variables are involved, and the hidden variable dimension is the same as the original power data dimension.

[0023] Preferably, the deep learning model comprises an information embedding part, a core decomposition part and a feature mapping part.

[0024] Preferably, the input information of the information embedding part comprises single load element power data x t , total power data x d and time step t; for the single load element power data and the total power data, a one-dimensional convolutional network with a Swish activation function is used for feature extraction; for the time step, first, an embedding method based on a sine-cosine function is used for transformation, and then a fully connected network with a Swish activation function is used for feature extraction. The calculation formula of the sine-cosine embedding is as follows:

[0025] t embedding =[sin(10 -[0:63]×γ ×C×t), cos(10 -[0:63]×γ ×C×t)].

[0026] Preferably, the core decomposition part is composed of N residual layers with consistent structure, and in each residual layer, the three types of embedded data are first fused by summation operation; then the fused data is processed using a bidirectional dilated convolutional network structure, and in the i-th residual layer, the spacing between the dilated convolution kernel points is 2 imodn ; then a gating unit is activated; finally, the vector is split into two parts, which are respectively used as the residual output and the skip connection output of the layer; the feature mapping part takes the sum of the skip connection outputs of all residual layers as input, and includes two one-dimensional convolutional layers, the first convolutional layer does not change the dimension of the processed data, and the second convolutional layer transforms the output data into a one-dimensional sequence form.

[0027] Preferably, the deep learning model uses Lognorm as the loss function, and the neural network parameter optimization target is to minimize

[0028] Preferably, the noise inference algorithm comprises:

[0029] Starting from random noise x T ~N(0, I);

[0030] Take t=T;

[0031] Take random noise z~N(0, I);

[0032] Neural network predicts noise ε θ ;

[0033] Computing variance of reverse transfer distribution

[0034] Computing

[0035] t = t-1, and repeat the above steps until t = 1.

[0036] (III) Beneficial Effects

[0037] The non-intrusive load decomposition method based on a noise inference model, the load decomposition algorithm of the present application processes the non-intrusive load decomposition into the condition generation task of the load element power curve, and uses the noise inference model to model the distribution of the load element active power data. Since the noise inference model does not need to train the discriminator, compared with the load decomposition algorithm based on the generative adversarial network, the present application provides a more stable training structure; in the UK-DALE data set, the microwave oven, washing machine, kettle and refrigerator are selected as the research objects, and the houses 1, 3, 4, 5 and 6 are selected as the training set, and the house 2 is selected as the test set, and the model is trained and tested in the Python programming platform using the Pytorch deep learning framework. The performance of the model of the present application in the accuracy, F1 score, mean absolute error and mean relative error reaches the advanced level in the field. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 The method flowchart of the present application;

[0039] Figure 2 The forward process and reverse process schematic diagram of the noise inference model of the present application;

[0040] Figure 3 The neural network input and output information schematic diagram provided by the present application;

[0041] Figure 4 The deep learning model schematic diagram in the embodiment of the present application. DETAILED DESCRIPTION

[0042] The technical solutions in the embodiments of the present application will be described clearly and completely below in combination with the drawings of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0043] EMBODIMENT

[0044] The application provides a non-intrusive load decomposition method based on a noise inference model, which decomposes a non-intrusive load into a generation task of a load element power curve, and uses a noise inference method to model the distribution of load element power data, thereby improving the decomposition accuracy and reducing the difficulty of model training.

[0045] Please refer to Figure 1 In some embodiments, a non-intrusive load decomposition method based on a noise inference model is provided, comprising:

[0046] obtaining resident active power consumption data;

[0047] obtaining a target active power data sequence by using a sliding window method;

[0048] training a deep learning model parameter based on the target active power data sequence by using a training algorithm based on a noise diffusion process;

[0049] starting from random noise subject to a Gaussian distribution, and combining the deep learning model after parameter training, a noise inference algorithm is used to restore the power curve of the target load element.

[0050] Further, the active power consumption data is collected by a second-level power sampling device, and includes aggregated active power data and active power data of main load elements.

[0051] Further, the active power sequence length is at least twice the length of a single operation of the load element.

[0052] Further, please refer to Figure 2 In some embodiments, the noise inference model includes a forward process and a reverse process. The forward process is a process of adding Gaussian noise to a clean load element power curve x0, which is a fixed Markov chain, and the closed-form transition distribution expression of the noise addition process is:

[0053]

[0054] wherein x t is the noise-added data at the tth step, is the noise at the corresponding step. The reverse process is a process of restoring the initial distribution step by step from random noise x T subject to a Gaussian distribution, and the deep learning model predicts the noise ε θ at the corresponding position, and then calculates the previous term x t-1 :

[0055]

[0056] Further, the training algorithm based on the noise diffusion process comprises:

[0057] Define linear noise control parameters

[0058] Select a sample x0~q(x) from the training set;

[0059] Randomly select a time step t value in {1, 2, … T};

[0060] Take random noise ε~N(0, I);

[0061] Train the model using gradient descent to minimize the distance between the actual noise and the model predicted noise Where α t = 1-β t ,

[0062] Repeat the above steps until the model converges.

[0063] Further, please refer to Figure 3 , the deep learning model takes the load element power data x t , the total power data x d and the time step t as input, and takes the noise ε at the corresponding position as the target output.

[0064] Further, please refer to Figure 4 , the deep learning model includes an information embedding part, a core decomposition part and a feature mapping part.

[0065] Further, the input information of the information embedding part includes noisy single load element power data x t , total power data x d and time step t; for single load element power data and total power data, one-dimensional convolutional network with Swish activation function is used for feature extraction; for time step, first use the embedding method based on sine function for transformation, and then use two fully connected networks with Swish activation function for feature extraction. Wherein, the Swish activation function expression is as follows:

[0066] Swish(x) = x * Sigmoid(x)

[0067] The calculation formula of the sine embedding is as follows:

[0068] t embedding = [sin(10 -[0:63] ×γ×C×t), cos(10 -[0:63] ×γ×C×t)]

[0069] Further, the deep learning model, the core decomposition part is composed of N structurally consistent residual layers, in each residual layer, first, the three types of embedded data are fused by summation operation; then the fused data is processed using a bidirectional hollow convolutional network structure, in the i-th residual layer, the distance between the kernel points of the hollow convolution is 2 imodn ; then using the gate unit activation; finally, the vector is split into two parts, which are respectively the residual output and the skip connection output of the layer.

[0070] Further, the feature mapping part takes the sum of the skip connection outputs of all residual layers as input, which includes two one-dimensional convolution layers, the first convolution layer does not change the dimension of the processed data, and the second convolution layer transforms the output data into a one-dimensional sequence form.

[0071] Further, the deep learning model takes Lognorm as the loss function, and the neural network parameter optimization goal is to minimize

[0072] Further, the deep learning model uses Adam optimizer as the gradient descent algorithm.

[0073] Further, the noise inference algorithm includes:

[0074] Starting from random noise x T ~ N(0, I);

[0075] Take t=T, T-1, …, 1;

[0076] Take random noise z ~ N(0, I);

[0077] Neural network prediction noise ε θ ;

[0078] Calculate the variance of the reverse transition distribution

[0079] Calculate

[0080] Repeat the above steps until t=1.

[0081] Further, after the target load element power curve is restored, four types of indicators are used to evaluate the effect of the non-intrusive load decomposition algorithm model based on the noise inference model, the evaluation includes accuracy, F1 score, mean absolute error and mean relative error, and the evaluation index calculation formula is as follows:

[0082]

[0083]

[0084]

[0085]

[0086] Wherein, TP, FP, FN represent the number of three kinds of judgment results that the load element is actually opened and the decomposition result is also opened, the load element is actually closed and the decomposition result is opened, and the load element is closed and the decomposition result is also closed, respectively, and P and N represent the number of the load element actually opened and closed, respectively. and y t Respectively represent the model predicted load element power and the actual load element power; T is the total number of sequence points.

[0087] Embodiments of the present application can be provided as a method or a computer program product. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, etc.) containing computer usable program code. The solutions in the embodiments of the present application can be implemented in various computer languages, such as object-oriented programming languages Java and interpreted scripting languages JavaScr ipt, etc.

[0088] The present application is described with reference to flowcharts and / or block diagrams according to the methods, devices (systems), and computer program products of the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing apparatus to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing apparatus produce a device that implements the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in the flowcharts and / or block diagrams.

[0089] These computer program instructions can also be stored in a computer-readable memory capable of guiding a computer or other programmable data processing apparatus to work in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including instruction devices that implement the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in the flowcharts and / or block diagrams.

[0090] These computer program instructions can also be loaded into a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 Figure 1

[0091] It should be noted that, in the description, relative terms such as first and second, and the like, can be used solely to distinguish one from another without necessarily implying any actual relationship or order between or among these entities or operations. Also, the terms "comprises", "comprising", or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus. An element proceeded by "comprises... a" does not, without more constraints, exclude the existence of additional identical elements in the process, method, article, or apparatus that comprises the element.​​

Claims

1. A non-intrusive load decomposition method based on a noise inference model, characterized in that, include: Obtain active power consumption data for residential electricity consumption; Obtain the target active power data sequence; The parameters of the deep learning model are trained using a training algorithm based on the noise diffusion process, based on the target active power data sequence. Starting from random noise that follows a Gaussian distribution, and combining it with a deep learning model trained with parameters, a noise inference algorithm is used to reconstruct the power curve of the target load element. The noise inference algorithm includes a forward process and a backward process. The forward process is a fixed Markov chain that adds Gaussian noise to a clean target load element power curve until the curve loses its significant features. The backward process is the process of gradually reconstructing the initial distribution from random noise that follows a Gaussian distribution. The backward transition distribution needs to be predicted by a deep learning model. The training algorithm for the noise dispersion process includes: Define linear noise control parameters ; Samples were selected from the training set. ; Randomly select time step t values ​​from {1,2,…T}; Take random noise ; Using gradient descent to train a model that minimizes the distance between actual noise and model-predicted noise ,in , ; Repeat the above steps until the model converges; The noise dispersion process has T steps and involves T latent variables, the dimension of which is the same as the dimension of the original power data.

2. The non-intrusive load decomposition method based on a noise inference model according to claim 1, characterized in that: The active power consumption data for residential electricity includes aggregated active power data and active power data of load elements.

3. The non-intrusive load decomposition method based on a noise inference model according to claim 1, characterized in that: The length of the target active power data sequence is at least twice the duration of a single operation of the load element.

4. The non-intrusive load decomposition method based on a noise inference model according to claim 1, characterized in that: The deep learning model includes an information embedding part, a core decomposition part, and a feature mapping part.

5. The non-intrusive load decomposition method based on a noise inference model according to claim 4, characterized in that: The input information for the information embedding section includes power data of a single load element, including noise. x t Total power data x d and time step t For individual load element power data and total power data, a one-dimensional convolutional network with Swish activation function is used for feature extraction. For each time step, a transformation based on sine and cosine functions is first performed, followed by feature extraction using a fully connected network with Swish activation function. The calculation formula for sine and cosine embedding is as follows: 。 6. The non-intrusive load decomposition method based on a noise inference model according to claim 4, characterized in that: The core decomposition part consists of N The system consists of three structurally consistent residual layers. In each residual layer, the three types of embedded data are first fused through a summation operation. Then, a bidirectional dilated convolutional network structure is used to process the fused data. i In each residual layer, the spacing between the void convolution kernel points is Then, a gating unit is used for activation; finally, the vector is split into two parts, which are used as the residual output and the jump connection output of the layer, respectively; the feature mapping part takes the sum of the jump connection outputs of all residual layers as input, and includes two one-dimensional convolutional layers. The first convolutional layer does not change the dimension of the processed data, and the second convolutional layer transforms the output data into a one-dimensional sequence.

7. The non-intrusive load decomposition method based on a noise inference model according to claim 1, characterized in that: The deep learning model uses Lognorm as the loss function, and the optimization objective of the neural network parameters is to minimize... 。 8. The non-intrusive load decomposition method based on a noise inference model according to claim 1, characterized in that: The noise inference algorithm includes: From random noise Set off; Pick t = T ; Take random noise ; Neural network prediction noise ; Calculate the variance of the inverse transition distribution ; calculate ; t = t -1, and repeat the above steps until t =1.

Images