Multi-branch correction non-intrusive load detection method based on conditional random field supervision
Through a multi-branch neural network based on conditional random field, the existing hidden Markov model has solved the accuracy and efficiency of load detection, and achieved more efficient electrical usage status and power curve detection, suitable for non-invasive loads Test.
Patent Information
- Application Number
- CN202410736430.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-06-07
- Publication Date
- 2024-11-08
AI Technical Summary
The existing non-invasive load detection method based on the Hidden Markov model has not yet achieved satisfactory results in accuracy and efficiency, and it is difficult to accurately detect the use status and power curve of electrical appliances.
Using a multi-branch neural network based on conditional random fields, the state branch layer, confidence branch layer and power curve branch layer are used to model the usage state and power curve curve of electrical appliances, and the result correction is performed through the correction layer, using sliding windows and volumes. The product module extracts features, combines conditional random field and cross entropy loss function for training to form a detection model.
It improves the accuracy and efficiency of electrical appliance detection, and can accurately detect the usage status and power curve of electrical appliances in a non-invasive manner. It is suitable for places where intrusive detection is difficult to perform, providing better detection results.
Smart Images

Figure CN118916623A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a multi-branch modified non-intrusive load detection method based on conditional random field supervision. Background Art
[0002] Traditional load monitoring methods often require the installation of sensor equipment on the user side, which not only increases the burden on users, but also makes monitoring costs high. Therefore, non-intrusive load monitoring (NILM) technology came into being. It installs monitoring equipment at the power entrance to monitor voltage, current and other signals to analyze the type and operation of a single load in the load cluster, i.e., the target electrical equipment. This method overcomes the limitations of traditional methods, but how to improve the accuracy and efficiency of load monitoring is still a problem to be solved.
[0003] In order to solve the accuracy and efficiency problems, the existing methods have combined the NILM method with deep neural networks. As a new research direction in the field of machine learning, deep learning has achieved remarkable results in the fields of image recognition and natural language processing. Deep learning aims to learn the inherent laws and representation levels of sample data, so that machines can have analytical learning capabilities like humans and recognize data such as text, images and sounds. This powerful machine learning method provides a new solution for the field of load monitoring.
[0004] Non-intrusive load detection is a type of Blind Source Separation (BSS) task, which is to extract multiple data sources from a single mixed observation data, so BSS is a highly underdetermined problem. Since the power signal of an electrical appliance is largely determined by two factors: 1. The usage status of the appliance. 2. The transition between usage states. Therefore, most of the existing mainstream NILM method solutions combined with deep neural networks use the establishment of a hidden Markov statistical model to model these two factors, thereby solving the highly underdetermined problem of BSS. However, the NILM method based on the hidden Markov model is mainly based on the idea of "direct conversion", and has not yet achieved good results in solving accuracy and efficiency problems. Summary of the invention
[0005] The present invention provides a multi-branch modified non-intrusive load detection method based on conditional random field supervision to solve the problem of poor detection effect of the NILM method of the prior art hidden Markov model.
[0006] In order to achieve the above object, the technical solution adopted by the present invention is:
[0007] The multi-branch modified non-intrusive load detection method based on conditional random field supervision includes the following steps:
[0008] Step 1: Obtain a total power data sequence of each target electrical appliance in a historical time period, wherein the total power data sequence includes real power data of the corresponding target electrical appliance in multiple sampling periods in the historical time period, and extract multiple real usage states of the corresponding target electrical appliance in the historical time period based on the acquired data, and construct a data set corresponding to each target electrical appliance with the total power data sequence and multiple real usage states;
[0009] Step 2, generating a multi-branch neural network based on a conditional random field, wherein the multi-branch neural network includes a usage state branch layer, a confidence branch layer, a power curve branch layer, and a correction layer, and using the data set of each target electrical appliance obtained in step 1 to train the generated multi-branch neural network respectively, thereby obtaining a detection model corresponding to each target electrical appliance;
[0010] During training, the total power data sequence in each target electrical appliance data set is divided into data of multiple windows according to the same time period length through a sliding window, and the data of each window is respectively input into the usage status branching layer, and the usage status branching layer predicts the probability value of the usage status prediction result of the corresponding target electrical appliance at each time point position in each window, and the usage status branching layer outputs the probability value of the usage status prediction result of the corresponding target electrical appliance at each time point position in each window, that is, the usage status prediction result sequence of the corresponding target electrical appliance in each window;
[0011] The corresponding target electrical equipment usage status prediction result sequence of each window obtained by the usage status branch layer and the data of each window are respectively input into the confidence branch layer, and the confidence branch layer obtains and outputs the confidence of the probability value of the usage status prediction result of the corresponding target electrical equipment at each time point in each window;
[0012] Furthermore, the data of each window is also input into the power curve branch layer respectively, and the power curve branch layer predicts the power value prediction result of the corresponding target electrical equipment at each time point in each window, and the power curve prediction result of the corresponding target electrical equipment in each window is obtained according to the power value prediction result;
[0013] The usage state prediction result sequence of the corresponding target electrical equipment in each window obtained by the usage state branch layer, the confidence of the usage state prediction result probability value of the corresponding target electrical equipment at each time point in each window obtained by the confidence branch layer, and the power curve prediction result of the corresponding target electrical equipment in each window obtained by the power curve branch layer are respectively input into the correction layer, and the correction layer corrects the power curve prediction result of the corresponding target electrical equipment in each window according to the usage state prediction result sequence of the corresponding target electrical equipment in each window and the confidence of the usage state prediction result probability value of the corresponding target electrical equipment at each time point in each window, thereby obtaining the corrected power curve prediction result of the corresponding target electrical equipment in each window;
[0014] Perform multiple rounds of training, during which the parameters of the multi-branch neural network are updated by back propagation based on the loss function calculation results until the multi-branch neural network converges to an optimal state, thereby obtaining a trained multi-branch neural network as a detection model corresponding to each target electrical device;
[0015] Step 3: Input the current total power data sequence of each target electrical appliance to be detected into the corresponding detection model obtained in step 2, and use the corresponding detection model to predict the current prediction result of the usage status of the corresponding target electrical equipment and the current prediction result of the power curve of the corresponding target electrical equipment.
[0016] Furthermore, in step 1, a mean shift clustering algorithm is used based on the acquired data to extract multiple real usage states of the corresponding target electrical appliance in the historical time period.
[0017] Furthermore, in step 2, the state branch layer and the power curve branch layer first extract the features of the data in each window through the convolution modules of various branches, wherein the convolution module of the state branch layer extracts features related to the usage status, and the convolution module of the power curve branch extracts features related to the power curve of the target electrical equipment.
[0018] Furthermore, the loss function in step 2 includes a conditional field random term and a cross entropy loss term used by the state branch layer, and a mean square error term used by the power curve branch layer.
[0019] The present invention provides a multi-branch corrected non-intrusive load detection method based on conditional random field supervision, which uses a multi-branch neural network based on conditional random field to separately model the power curve and usage status and confidence of each target electrical appliance, and finally uses the usage status characteristics to further guide the decomposition result of the power curve.
[0020] Compared with the prior art, the present invention has the following advantages:
[0021] 1. The power curve of each target electrical appliance decomposed by the present invention can be used for various tasks, such as detecting whether illegal electrical appliances are used in the venue, whether a certain electrical appliance in the venue needs maintenance, whether there are safety hazards, etc.;
[0022] 2. In some places where invasive detection is difficult to carry out (for example, places involving privacy, highly complex and high-risk places, etc.), the present invention can provide a solution, which has the advantage of high accuracy of usage status detection in a non-invasive manner, and thus can achieve better detection results. BRIEF DESCRIPTION OF THE DRAWINGS
[0023] Figure 1 It is a schematic diagram of a multi-branch neural network in an embodiment of the present invention. DETAILED DESCRIPTION
[0024] The present invention is further described below in conjunction with the accompanying drawings and embodiments.
[0025] This embodiment discloses a multi-branch modified non-intrusive load detection method based on conditional random field supervision, comprising the following steps:
[0026] Step 1: Obtain the total power data sequence of each target electrical appliance in the historical time period. The historical time period described in this embodiment includes multiple sampling periods, so the total power data sequence of each target electrical appliance in the historical time period obtained includes the real power data of the corresponding target electrical appliance in multiple sampling periods in the historical time period. Taking one year as an example of the historical time period, the sampling period is set to one day, so the total power data sequence of each target electrical appliance in one year includes the power data sampled by the target electrical appliance in a total of 365 sampling periods. In this embodiment, let the total power data sequence of the i-th target electrical appliance in the historical time period be X i , and the actual power data of the i-th target electrical equipment in each sampling period is Y i .
[0027] According to the total power data series of each target appliance in the historical time period, the mean shift clustering algorithm is used to extract the actual usage status of the corresponding target appliance in the historical time period. The mean shift clustering algorithm is a density-based clustering algorithm that can automatically discover the clustering structure in the data. The core idea of the algorithm is to continuously move the center position of the data point so that the data points gather in the area with the highest density. The calculation process of the mean shift clustering algorithm is shown in formula (1):
[0028]
[0029] Taking the i-th target electrical equipment as an example, in formula (1): K(·) is the kernel function, and in this embodiment, a radially symmetric kernel is used. d is the d-dimensional window bandwidth parameter. n is the total power data sequence X of the i-th target electrical equipment input i x is the total power data sequence X of the i-th target electrical equipment. i .x i is the total power data sequence X of the i-th target electrical equipment i (i.e., x in formula (1)) is the real power data of the i-th position. m(x) is calculated by formula (1), which is the M of the i-th target electrical equipment. i The actual usage status.
[0030] Therefore, for the total power data sequence of each target electrical device, this embodiment extracts multiple real usage states of the corresponding target electrical device in the historical time period through the mean shift clustering algorithm. Then, the data set corresponding to each target electrical device is constructed based on the total power data sequence of each target electrical device and multiple real usage states.
[0031] Step 2: Generate a multi-branch neural network based on conditional random fields.
[0032] like Figure 1 As shown, the multi-branch neural network of this embodiment is a neural network structure with multiple parallel branches. Each branch can process different input data or perform different tasks, and then the outputs of multiple branches are fused or integrated in the subsequent part of the network. This structure enables the multi-branch neural network to use multiple information sources or perform multiple operations at the same time, thereby improving the performance and accuracy of the model.
[0033] Conditional Random Field (CRF) is a discriminative probability model. It is a type of random field and is often used to label or analyze sequence data, such as natural language text or biological sequences. CRF mainly solves three problems: probability calculation problem of CRF, learning problem of CRF and decoding problem of CRF.
[0034] In a conditional random field, the distribution of the random variable Y is the conditional probability, and the given observation value is the random variable X. The layout of the graph model can be given arbitrarily, but the commonly used layout is a linked architecture. This architecture has efficient algorithms available for calculation in training, inference, or decoding.
[0035] The conditional random field breaks the observation independence assumption and homogeneous Markov assumption of the hidden Markov model, making the relationship between the input vector and the output vector more obvious, so it performs better on sequence data.
[0036] In this embodiment, the generated multi-branch neural network includes a usage state branch layer, a confidence branch layer, a power curve branch layer, and a correction layer. In this embodiment, the generated multi-branch neural network is trained respectively using the data set of each target appliance obtained in step 1, thereby obtaining a detection model corresponding to each target appliance. The specific description is as follows:
[0037] (A) During training, the present embodiment uses a sliding window method to divide the total power data sequence in each target electrical appliance data set into multiple windows of data according to the same time period length w, and the data in each window b are all total power data subsequences of the corresponding target electrical appliance device centered on the center of the window b, and the data contained in the corresponding windows can overlap.
[0038] For the data in each window, in this embodiment, the state branch layer and the power curve branch layer first extract the features of the data in each window through the convolution modules of various branches, among which the convolution module of the state branch layer extracts features related to the usage status, and the convolution module of the power curve branch extracts features related to the power curve of the target electrical equipment.
[0039] In the convolution modules of the state branch layer and the power curve branch layer, the convolution kernel slides on the total power data subsequence in each input window b to extract the features of the data in the current window b. The convolution kernel performs a convolution operation when sliding on the input image. The convolution operation can be expressed as formula (2):
[0040]
[0041] In formula (2): is the total power data subsequence in each window b of the input convolution kernel. σ(·) is the activation function. Common activation functions include Sigmoid, Tanh, ReLU, etc. W represents the learnable weight parameter, and β represents the learnable bias term. F conv (x) represents the features extracted by the convolution kernel operation.
[0042] The state branch layer and the power curve branch layer are used to perform their respective subsequent processing flows based on the corresponding features extracted from the data of each window b, and the data of each window b is also directly input into the confidence branch layer.
[0043] (B) The usage state branch layer predicts the probability value of the usage state prediction result of the corresponding target electrical equipment at each time point in each window based on the features corresponding to the data input into each window b, and the usage state branch layer outputs the probability value of the usage state prediction result of the corresponding target electrical equipment at each time point in each window, and the probability values of these usage state prediction results constitute the usage state prediction result sequence of the corresponding target electrical equipment in each window. The calculation process of the usage state branch layer is as follows:
[0044] First define the use state sequence
[0045] in: Represents the sequence of prediction results using the state The value range of is a positive real number and is a real vector of dimension w. Represents the sequence of predicted results using the state The value range of M is 0 to 1 i ×τ-dimensional closed interval, τ is the τth time point in the current window b. q is the length of the current window b.
[0046] The usage state branch layer outputs a usage state distribution by analyzing the features of the current window b data. The process can be expressed as formula (3):
[0047]
[0048] In formula (3): F state (·) represents the usage state branch layer. t,w is a w-dimensional real vector, corresponding to the subsequence power data of the original power data X centered at time point t and with a length of w.
[0049] is the prediction result sequence of the usage status of the corresponding target electrical equipment in the current window b, and has in: , indicating the time point The usage status prediction results on . ) indicates a time point The usage status prediction results on .
[0050] q is the length of the current window b, and the position t of the current window b on the corresponding total power data subsequence X is represented by the position of the window center data on X.
[0051] (C) The usage status prediction result sequence of the corresponding target electrical equipment in each window obtained by the usage status branch layer and the data of each window b (i.e., the total power data subsequence of window b) are respectively input into the confidence branch layer, and the confidence branch layer obtains and outputs the confidence of the probability value of the usage status prediction result of the corresponding target electrical equipment at each time point in each window.
[0052] The main purpose of the confidence branch layer is to further correct the probability value of the usage state prediction result predicted by the usage state branch layer. The confidence branch layer contains an attention layer, and its overall calculation process is as shown in formulas (4)-(8). The attention layer first uses the total power data subsequence in the current window b to generate the corresponding query vector Q, which can be expressed as formula (4):
[0053] Q=F linear (x) (4)
[0054] In formula (4): F linear (·) represents a fully connected layer in the attention layer.
[0055] The confidence branch layer reuses the usage status prediction result sequence of the corresponding target electrical equipment in the current window b output by the usage status branch layer Generate the corresponding key-value vector K and value vector V. This process can be expressed as formula (5) and formula (6):
[0056]
[0057] In formula (5) and (6), F linear (·) represents an independent fully connected layer in the confidence branch layer.
[0058] Then, the confidence branch layer calculates the product of Q and K, and uses the Softmax function to scale the product of Q and K to obtain the attention score Score. This process can be expressed as formula (7):
[0059] Score = F softmax (QK) (7)
[0060] In formula (7): And L is the length of the sequence QK.
[0061] Finally, the confidence branch layer multiplies the attention score Score to the value vector V, that is, the confidence of the probability value of the usage status prediction result of the corresponding target electrical equipment at each time point in each window is obtained. This confidence evaluates the credibility of the probability value of the usage status prediction result of the corresponding target electrical equipment in the usage status prediction result sequence output by the usage status branch layer. The process can be expressed as formula (8):
[0062]
[0063] In formula (8): for confidence level.
[0064] (D) After the power curve branch extracts the corresponding features based on the data of each window b, the total power data in the current window b is input into the power curve branch layer, and the power curve branch layer predicts the power value prediction result of the corresponding target electrical equipment at each time point in the current window b based on the extracted features, and then obtains the power curve prediction result of the corresponding target electrical equipment in the current window b. The calculation process of the power curve branch layer is as follows:
[0065] Similar to the usage status branch, first define the power sequence of the corresponding target appliance
[0066] in: Represents the total power sequence of the corresponding target electrical appliance The value range of is a positive real number and is a real vector of dimension w. It can be further described as M i ×τ-dimensional positive real vector sequence, τ is the τth time point in window b. q is the length of window b.
[0067] The power curve branching layer analyzes the total power data subsequence in the current window b and outputs the power sequence of the current window b. This process can be expressed as formula (9):
[0068]
[0069] In formula (9): F power (·) represents the power curve branch layer.
[0070] is the power value prediction result sequence of the corresponding target electrical equipment on the current input window b, and has in:
[0071] q is the length of window b, and the position t of window b on the original total power data X is represented by the position of the window center data on X.
[0072] Finally, we get The corresponding target electrical equipment power curve prediction result is included in the current window b, which is composed of the power value prediction result of the corresponding target electrical equipment at each time point position obtained by the power curve branch layer prediction.
[0073] (E) The usage status prediction result sequence of the corresponding target electrical equipment in each window obtained by the usage status branch layer, the confidence value of the usage status prediction result of the corresponding target electrical equipment at each time point in each window obtained by the confidence branch layer, and the power curve prediction result of the corresponding target electrical equipment in each window obtained by the power curve branch layer are respectively input into the correction layer, and the correction layer corrects the power curve prediction result of the corresponding target electrical equipment in each window according to the usage status prediction result sequence of the corresponding target electrical equipment in each window and the confidence value of the usage status prediction result of the corresponding target electrical equipment at each time point in each window, thereby obtaining the corrected power curve prediction result of the corresponding target electrical equipment in each window.
[0074] In this embodiment, the correction layer receives three types of inputs, namely, the output of the state branch layer Output of the confidence branch layer And the output of the power curve branch layer
[0075] Correction layer first uses right A weighted calculation is performed, and then the weighted usage status prediction result sequence is used to correct the power curve prediction result of the corresponding target electrical equipment in the current window b of the power curve branch layer prediction. The whole process is as shown in formula (10):
[0076]
[0077] In formula (10): is the corrected power curve prediction result of the corresponding target electrical equipment in the current window b, and its length is q. +ψ is a learnable bias parameter, which is updated and learned through neural network back propagation.
[0078] (F) performing multiple rounds of training, during which the parameters of the multi-branch neural network are updated by back propagation based on the loss function calculation results until the multi-branch neural network converges to an optimal state, thereby obtaining a trained multi-branch neural network as a detection model corresponding to each target electrical device.
[0079] In this embodiment, when the multi-branch neural network performs multiple rounds of training, the multi-branch neural network's own parameters are updated based on the calculated loss function and using back propagation until the multi-branch neural network converges to the optimal state, and the multi-branch neural network in the optimal state is used as the detection model corresponding to the corresponding target electrical equipment.
[0080] The loss function includes the conditional random field term and the cross entropy loss term of the state branch layer, and the mean square error loss term of the power curve branch layer. The specific description is as follows:
[0081] (X) Loss function using state branch layer
[0082] Conditional random field term: This regular term consists of two components, one is called the estimation score, which is used to evaluate the error of the usage state; the other is the transition score, which considers the transition probability between every two adjacent usage states. The conditional random field term can be expressed as formula (11):
[0083]
[0084] In formula (11), E[x] represents the expectation of x. To estimate the score. For transfer score. Represents the probability value of the usage status prediction result of the corresponding target electrical equipment at time step t. Θ is the state transfer matrix. Represents the conditional random field item corresponding to the target electrical equipment.
[0085] Cross entropy loss term: Cross entropy is a commonly used loss function for classification problems. Its calculation process is as shown in formula (12):
[0086]
[0087] In formula (12): is the actual usage status of the corresponding target appliance at time step t in the current window b, which is extracted in step 1 using the mean shift clustering algorithm. is the sequence of usage state prediction results predicted by the usage state branch layer. Represents the cross entropy loss term of the corresponding target electrical equipment.
[0088] (Y) Loss function of the power curve branch layer
[0089] The loss function of the power curve branch layer is the mean square error loss, which calculates the mean square error between the predicted value and the true value and can be expressed as formula (13):
[0090]
[0091] In formula (13): Y t i is the real power sequence of the corresponding target electrical equipment, collected in step 1. The power value prediction result sequence of the corresponding target electrical appliance predicted by the power curve branch layer.
[0092] (Z) Overall loss function
[0093] The overall loss function is the weighted sum of the above items, as shown in formula (14):
[0094]
[0095] In formula (14): is the total loss function under the i-th target appliance. 1 ,Α 2 ,Α 3 is an adjustable weight, and there is 1 +Α 2 +Α 3 =1.
[0096] Step 3: Input the current total power data sequence of each target electrical appliance to be detected into the corresponding detection model obtained in step 2, and use the corresponding detection model to predict the current prediction result of the usage status of the corresponding target electrical equipment and the current prediction result of the power curve of the corresponding target electrical equipment.
[0097] This embodiment proposes a multi-branch neural network model based on conditional random field supervision. One branch of the multi-branch neural network model is used to capture multiple usage states of the device, one branch is used to model the confidence of the usage state probability, and one branch is used to predict the power consumption data under each state. Finally, the outputs of the three branches are used again to correct the power consumption data and use it as the result of the decomposition. And in the model training stage, for the branch that captures the usage state, this embodiment uses the conditional random field term to replace the cross entropy term commonly used in classification problems, so as to achieve the purpose of explicitly calculating the correlation between states. Experiments on the actual collected non-invasive load detection data set show that the method designed in this embodiment is effective and superior to other methods. The specific experimental description is as follows:
[0098] This experiment uses NVIDIA GTX 4090 GPU, and the model code is built based on Python and Pytorch toolbox. This experiment uses Adam optimizer.
[0099] (1) This experiment will evaluate the model's capabilities on the public dataset REDD
[0100] REDD consists of power readings from six homes, where aggregate power and individual appliance power are recorded every 1 second and 3 seconds, respectively. Four appliances (microwave, washing machine, dishwasher, refrigerator) in houses 1, 2, and 3 are selected for evaluation. This experiment uses three evaluation metrics, namely MAE as shown in formula (16), SAE as shown in formula (17), and SAE δ As shown in formula (18):
[0101]
[0102] In formulas (16), (17), and (18), δ represents a time period, T δRepresents splitting the original sequence into T δ time period. δ is the length of δ. Y is the true value, is the predicted value of the model. y is the true value within the corresponding δ, is the predicted value of the model within the corresponding δ.
[0103] (2) Experimental performance comparison
[0104] In this experiment, the multi-branch neural network model proposed in this embodiment is evaluated. The related studies compared include: FHMM, LSTM, S2P, BERT4NILM, SGN.
[0105] Table 1 shows the evaluation results of six prediction models including the multi-branch neural network model of this embodiment on the REDD dataset, and the best result is shown in bold.
[0106] Table 1 Comparative experimental results of six methods
[0107]
[0108] As shown in Table 1, the experimental results show that the multi-branch neural network model proposed in this embodiment achieves the best results under the four target appliances, which proves the practicality of this embodiment.
[0109] The preferred embodiments of the present invention are described in detail above in conjunction with the accompanying drawings. The embodiments described in the present invention are merely descriptions of the preferred embodiments of the present invention, and do not limit the concept and scope of the present invention. The various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction, and such combinations should also be regarded as the contents disclosed in the present disclosure as long as they do not violate the concept of the present invention. In order to avoid unnecessary repetition, the present invention will not further describe various possible combinations.
[0110] The present invention is not limited to the specific details of the above-mentioned embodiments. Within the technical concept of the present invention and without departing from the design concept of the present invention, various modifications and improvements made to the technical solution of the present invention by technical personnel in this field should fall within the protection scope of the present invention. The technical contents for which protection is sought in the present invention have been fully recorded in the claims.
Claims
1. A multi-branch modified non-intrusive load detection method based on conditional random field supervision, characterized in that: The following steps are involved: Step 1: Obtain a total power data sequence of each target electrical appliance in a historical time period, wherein the total power data sequence includes real power data of the corresponding target electrical appliance in multiple sampling periods in the historical time period, and extract multiple real usage states of the corresponding target electrical appliance in the historical time period based on the acquired data, and construct a data set corresponding to each target electrical appliance with the total power data sequence and multiple real usage states; Step 2, generating a multi-branch neural network based on a conditional random field, wherein the multi-branch neural network includes a usage state branch layer, a confidence branch layer, a power curve branch layer, and a correction layer, and using the data set of each target electrical appliance obtained in step 1 to train the generated multi-branch neural network respectively, thereby obtaining a detection model corresponding to each target electrical appliance; During training, the total power data sequence in each target electrical appliance data set is divided into data of multiple windows according to the same time period length through a sliding window, and the data of each window is respectively input into the usage status branching layer, and the usage status branching layer predicts the probability value of the usage status prediction result of the corresponding target electrical appliance at each time point position in each window, and the usage status branching layer outputs the probability value of the usage status prediction result of the corresponding target electrical appliance at each time point position in each window, that is, the usage status prediction result sequence of the corresponding target electrical appliance in each window; The corresponding target electrical equipment usage status prediction result sequence of each window obtained by the usage status branch layer and the data of each window are respectively input into the confidence branch layer, and the confidence branch layer obtains and outputs the confidence of the probability value of the usage status prediction result of the corresponding target electrical equipment at each time point in each window; Furthermore, the data of each window is also input into the power curve branch layer respectively, and the power curve branch layer predicts the power value prediction result of the corresponding target electrical equipment at each time point in each window, and the power curve prediction result of the corresponding target electrical equipment in each window is obtained according to the power value prediction result; The usage state prediction result sequence of the corresponding target electrical equipment in each window obtained by the usage state branch layer, the confidence of the usage state prediction result probability value of the corresponding target electrical equipment at each time point in each window obtained by the confidence branch layer, and the power curve prediction result of the corresponding target electrical equipment in each window obtained by the power curve branch layer are respectively input into the correction layer, and the correction layer corrects the power curve prediction result of the corresponding target electrical equipment in each window according to the usage state prediction result sequence of the corresponding target electrical equipment in each window and the confidence of the usage state prediction result probability value of the corresponding target electrical equipment at each time point in each window, thereby obtaining the corrected power curve prediction result of the corresponding target electrical equipment in each window; Perform multiple rounds of training, during which the parameters of the multi-branch neural network are updated by back propagation based on the loss function calculation results until the multi-branch neural network converges to an optimal state, thereby obtaining a trained multi-branch neural network as a detection model corresponding to each target electrical device; Step 3: Input the current total power data sequence of each target electrical appliance to be detected into the corresponding detection model obtained in step 2, and use the corresponding detection model to predict the current prediction result of the usage status of the corresponding target electrical equipment and the current prediction result of the power curve of the corresponding target electrical equipment.
2. The multi-branch modified non-intrusive load detection method based on conditional random field supervision according to claim 1 is characterized in that: In step 1, a mean shift clustering algorithm is used based on the acquired data to extract multiple real usage states of the corresponding target appliance in a historical time period.
3. The multi-branch modified non-intrusive load detection method based on conditional random field supervision according to claim 1 is characterized in that: In step 2, the state branch layer and the power curve branch layer first extract the features of the data in each window through the convolution modules of various branches, among which the convolution module of the state branch layer extracts features related to the usage status, and the convolution module of the power curve branch extracts features related to the power curve of the target electrical equipment.
4. The multi-branch modified non-intrusive load detection method based on conditional random field supervision according to claim 1 is characterized in that: The loss function described in step 2 includes the conditional field random term used by the state branch layer, the cross entropy loss term, and the mean square error term used by the power curve branch layer.
