Non-intrusive load detection method based on feature fusion and structured V-I mapping and related device

Abstract

The invention discloses a non-intrusive load detection method based on feature fusion and structured V-I mapping and a related device, and the method comprises the steps: collecting voltage and current data on a target building bus through high-frequency sampling equipment, and carrying out the preprocessing of the collected data; detecting a load switching event by using a cumulative sum algorithm, and carrying out harmonic current feature extraction on the collected current data through fast Fourier transform; processing the acquired voltage and current data by adopting a structured V-I mapping method in combination with an improved convolutional self-encoding neural network to obtain V-I mapping features, and combining the harmonic current features with the V-I mapping features to construct a fused load feature vector; and adopting an improved twin network to obtain the similarity between the fused load feature vector and the existing feature vector in the database, and identifying and updating the to-be-identified load according to a similarity criterion. According to the invention, the problem of low load monitoring accuracy caused by a single characteristic can be solved.

Description

Technical Field

[0001] The present invention relates to the field of non-intrusive load detection, and specifically relates to a non-intrusive load monitoring method and related device based on feature fusion and structured V-I mapping. Technical Background

[0002] Nowadays, load monitoring has important application values in fields such as smart meters, energy management, and smart grids. By collecting and analyzing the load information of the power system, load monitoring can assist power companies and energy management departments to deeply understand the power consumption situation, thereby optimizing energy supply, improving energy utilization efficiency, and reducing environmental pollution. Compared with traditional intrusive load monitoring methods, non-intrusive load monitoring methods do not require installing sensors or monitoring devices on each load in the power system. Only by obtaining the total load data at a single power measurement point and using signal processing and algorithm technologies to identify and decompose the electrical loads, the deployment cost and complexity can be reduced.

[0003] The patent document with the application publication number CN118673966A introduces a non-intrusive load monitoring method and device based on the Transformer model and transfer learning. This method preprocesses the collected power data and constructs a model including modules such as a convolutional layer, position encoding, Transformer encoder, and fully connected layer to extract features and predict load power respectively. At the same time, combined with the transfer learning strategy, the prediction accuracy and generalization ability are improved. The patent document with the application publication number CN118568573A introduces a multi-model data analysis method and system based on non-intrusive load monitoring. This method combines a decision tree model and a deep learning model to identify electrical equipment. By classifying the total power consumption of users with the trained decision tree model, the classified sample user total power consumption data is input into the deep learning model for training, and finally, the power dynamic data of a single electrical equipment is obtained from the actual total power consumption data of users through the trained deep learning model and decision tree model to achieve load detection. The above two methods only detect the load from the perspective of a single power feature, which may lead to inaccurate load identification, sensitivity to noise and interference, and inability to identify unknown loads.

[0004] Therefore, the present invention proposes a load detection method considering multi-feature fusion to solve the problem of low accuracy of load monitoring caused by a single feature, and can detect unknown loads and update the database in real time. Summary of the Invention

[0005] The object of the present invention is to address the problem of low load recognition accuracy caused by traditional non-intrusive load monitoring considering only a single feature, and to provide a non-intrusive load detection method and related device based on feature fusion and structured V-I mapping.

[0006] To achieve the above object of the invention, the present invention provides a non-intrusive load detection method based on feature fusion and structured V-I mapping, comprising the following steps:

[0007] Step 1: Use a high-frequency sampling device to collect voltage and current data on the bus of the target building, and preprocess the collected voltage and current data;

[0008] Step 2: Use the cumulative sum algorithm to detect load switching events, and extract harmonic current features from the preprocessed current data through fast Fourier transform;

[0009] Step 3: Adopt a structured V-I mapping method combined with an improved convolutional autoencoder neural network to process the preprocessed voltage and current data to obtain V-I mapping features, and combine the harmonic current features and V-I mapping features to construct a fused load feature vector;

[0010] Step 4: Use an improved siamese network to obtain the similarity between the fused load feature vector and the existing feature vectors in the database, and identify and update the load to be identified according to the similarity criterion.

[0011] Further, the specific steps of Step 1 are as follows:

[0012] Step 1-1: Set the sampling frequency to 4 kHz to capture the high-frequency harmonic characteristics and fine waveform changes during load operation;

[0013] Step 1-2: Preprocess the collected signal in the sampling device, remove the high-frequency noise and DC offset in the signal, and extract the frequency band of interest to ensure that the extracted data covers all operating conditions of the load;

[0014] Step 1-3: Store the collected high-frequency voltage and current data in the cloud database for subsequent load feature extraction and load feature recognition algorithms.

[0015] Further, in Step 2,

[0016] The extracted harmonic current features include the effective value, maximum value of the steady-state current, and the 1st, 3rd, 5th, and 7th harmonic current values and their contents. The effective value and maximum value are calculated from the current sampling data, and the calculation steps are as follows:

[0017] Step 2-1: Apply fast Fourier transform to the preprocessed current signal i(n) to obtain the frequency-domain signal in complex form:

[0018]

[0019] where: k = 0, 1, ..., N - 1 is the frequency index;

[0020] Step 2-2: Calculate the spectral amplitude:

[0021]

[0022] where: Re(I(k)) and Im(I(k)) are the real and imaginary parts of I(k), respectively;

[0023] Step 2-3: Extract the high-order harmonic components and calculate the amplitude of each harmonic:

[0024]

[0025] A h = |I(k h )| (4)

[0026] where: h is the harmonic order, h = 1, 3, 5, 7; f base is the power grid power frequency, which is 50 Hz; A h is the amplitude of the corresponding harmonic;

[0027] Step 2-4: Calculate the total harmonic distortion rate according to the fundamental component and each harmonic component:

[0028]

[0029] where: H is the maximum harmonic order considered;

[0030] Step 2-5: Construct the current harmonic feature vector F h = [A m , A ev , A base , A3, A5, A7, THD], A m is the maximum current value; A ev is the effective current value; A base , A3, A5, A7 are the fundamental current amplitude, 3rd, 5th, and 7th harmonic amplitudes; THD is the total harmonic distortion rate.

[0031] Furthermore, Step 3 includes:

[0032] Step 3-1: First, perform periodic decomposition on the collected voltage and current data and normalize them;

[0033] Step 3-2: Map the normalized voltage and current data from three dimensions, and construct multiplication mapping, addition mapping, and subtraction mapping feature expressions respectively;

[0034] Step 3-3: Design an improved convolutional autoencoder neural network with a convolutional autoencoder neural network structure combined with depthwise separable convolution. Combine the three different forms of feature expressions obtained into three channels of a tensor as the model input of the improved convolutional autoencoder neural network, and extract V-I mapping features from different perspectives;

[0035] Step 3-4: Normalize the harmonic current features and the V-I mapping features, and assign weights to their respective importance. Linearly combine the normalized harmonic current features and V-I mapping features according to the weights to obtain a fused load feature vector.

[0036] Further, Step 4 includes:

[0037] Step 4-1: Introduce an attention mechanism into the siamese network to endow the siamese network with stronger feature extraction ability, and obtain an improved siamese network;

[0038] Step 4-2: Use the BLUED dataset to generate training positive and negative sample pairs, and input the positive and negative sample pairs into the improved siamese network for training;

[0039] Step 4-3: Input the fused feature vector and the feature vector of each known load in the database into two branches of the improved siamese network to obtain the embedding vector between the two branches;

[0040] Step 4-4: Calculate the similarity score of the obtained embedding vector through the Euclidean distance. If the similarity score with a certain load in the database is higher than the set threshold, it is determined as a known load; if all similarity scores are lower than the threshold, it is determined as an unknown load, and the database is dynamically updated. Extract the feature vector of this load, add it to the database and assign it a unique label.

[0041] A non-intrusive load detection device based on feature fusion and structured V-I mapping, comprising:

[0042] A data acquisition and preprocessing module, which is used to collect voltage and current data on the bus of the target building and preprocess the collected voltage and current data;

[0043] A harmonic current feature extraction module, which is used to detect load switching events by using the cumulative sum algorithm and extract harmonic current features from the preprocessed current data through fast Fourier transform;

[0044] A load feature vector fusion module, which is used to process the preprocessed voltage and current data by using a structured V-I mapping method combined with an improved convolutional autoencoder neural network to obtain V-I mapping features, and combine the harmonic current features and the V-I mapping features to construct a fused load feature vector;

[0045] A load recognition module, which is used to obtain the similarity between the fused load feature vector and the existing feature vectors in the database by using an improved siamese network, and identify and update the load to be recognized according to the similarity criterion.

[0046] A non-intrusive load detection system based on feature fusion and structured V-I mapping, comprising: a computer-readable storage medium and a processor;

[0047] The computer-readable storage medium is used to store executable instructions;

[0048] The processor is used to read the executable instructions stored in the computer-readable storage medium and execute the non-intrusive load detection method based on feature fusion and structured V-I mapping.

[0049] A non-transitory computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the non-intrusive load detection method based on feature fusion and structured V-I mapping is implemented.

[0050] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0051] 1. Compared with the traditional non-intrusive load monitoring based on a single power feature, the present invention overcomes the limitations of single-feature detection, such as the inability to distinguish similar loads, sensitivity to noise and interference, and difficulty in identifying unknown loads, by means of multi-feature fusion, and improves the accuracy of load recognition.

[0052] 2. Compared with the traditional V-I mapping method, the present invention overcomes the limitation that the traditional V-I mapping only provides a two-dimensional trajectory map, resulting in insufficient information when input into a deep learning model and limited feature extraction ability, by means of a structured V-I mapping method. The structured V-I mapping enables the deep learning model to automatically extract more details of the load by generating multi-channel feature tensors, thereby improving the classification accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] The present invention will be further described below with reference to the drawings and embodiments:

[0054] Figure 1 is a flowchart of a non-intrusive load detection method based on feature fusion and structured V-I mapping of the present invention;

[0055] Figure 2It is the structural diagram designed in the embodiments of the present invention by combining structured V-I mapping and convolutional autoencoder neural network;

[0056] Figure 3 It is the structural diagram of the Siamese network in the embodiments of the present invention.

[0057] Specific implementation steps

[0058] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some but not all of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0059] The first aspect of the present invention provides a non-intrusive load detection method based on feature fusion and structured V-I mapping, as Figure 1 shown, including the following steps:

[0060] Step 1: Use a high-frequency sampling device to collect voltage and current data on the bus of the target building, and preprocess the collected data;

[0061] Step 2: Use the cumulative sum algorithm to detect load switching events, and extract harmonic current features from the preprocessed current data through fast Fourier transform;

[0062] Step 3: Adopt the structured V-I mapping method combined with an improved convolutional autoencoder neural network to process the preprocessed voltage and current data to obtain V-I mapping features, and combine the harmonic current features and V-I mapping features to construct a fused load feature vector;

[0063] Step 4: Use an improved Siamese network to obtain the similarity between the fused load feature vector and the existing feature vectors in the database, and identify and update the load to be identified according to the similarity criterion.

[0064] In Step 1: Use a high-frequency sampling device to collect voltage and current data on the bus of the target building, and preprocess the collected data. The specific steps are as follows:

[0065] Step 1-1: Set the sampling frequency to 4 kHz to capture the high-frequency harmonic characteristics and fine waveform changes during load operation.

[0066] Step 1-2: Preprocess the collected signals in the sampling device, remove the high-frequency noise and DC offset in the signals, extract the frequency band of interest, and ensure that the extracted data covers all working conditions of the load.

[0067] Step 1-3: Store the collected high-frequency voltage and current data in the cloud database for subsequent load characteristic extraction and load characteristic recognition algorithms.

[0068] In Step 2, construct the current characteristics using the effective value, maximum value, 1st, 3rd, 5th, 7th harmonics, and total harmonic distortion rate of the steady-state current. The effective value and maximum value can be directly calculated from the current sampling data, and the calculation steps are as follows:

[0069] Step 2-1: Apply the fast Fourier transform to the preprocessed current signal i(n) to obtain the frequency-domain signal in complex form:

[0070]

[0071] where: k = 0, 1,..., N-1 is the frequency index.

[0072] Step 2-2: Calculate the spectral amplitude:

[0073]

[0074] where: Re(I(k)) and Im(I(k)) are the real and imaginary parts of I(k), respectively.

[0075] Step 2-3: Extract the high-order harmonic components and calculate the amplitude of each harmonic:

[0076]

[0077] A h = |I(k h )| (4)

[0078] where: h is the harmonic order, h = 1, 3, 5, 7; f base is the power grid power frequency, which is 50Hz; A h is the amplitude of the corresponding harmonic.

[0079] Step 2-4: Calculate the total harmonic distortion rate based on the fundamental component and each harmonic component:

[0080]

[0081] where: H is the maximum harmonic order considered.

[0082] Step 2-5: Construct the current harmonic feature vector F h = [A m ,A ev ,A base ,A3,A5,A7,THD]. A m is the maximum current value; A evis the effective value of current; A base A3, A5, and A7 are the fundamental wave amplitude, 3rd, 5th, and 7th harmonic amplitudes of the current; THD is the total harmonic distortion rate.

[0083] In step 3: The structured V-I mapping method is combined with an improved autoencoder neural network to process the voltage and current signals, obtain the V-I mapping features, and combine the harmonic current features and the V-I mapping features to construct a complete load feature vector. The specific steps are as follows:

[0084] Step 3-1: Define the target device to be analyzed as i, and its current and voltage data are C i ={c i (1), c i (2),..., c i (t)} and U i ={u i (1), u i (2),..., u i (t)}. Perform periodic decomposition on the current and voltage data. Define the length of one period as l. For the current and voltage at the entire dataset level, use the normalization formula to constrain their distributions:

[0085]

[0086] In the formula: and are the means of voltage and current, and σ(U) and σ(I) are their standard deviations;

[0087] Step 3-2: Define a three-dimensional tensor where d ∈ {1, 2, 3} represents the dimension index of the tensor; x and y represent the numerical indices in the corresponding dimensions; represents the data corresponding to the jth period tensor of device i.

[0088] The first dimension of the tensor is a multiplicative mapping, representing the coupling feature expression between current and voltage; the second dimension is an additive mapping, representing the isotropic feature expression between current and voltage; the third dimension is a subtractive mapping, representing the anisotropic feature expression between current and voltage.

[0089]

[0090] In the formula: x ∈ [1, l], y ∈ [1, l].

[0091] It can be seen from the above formula that in each dimension of the tensor Q i Tj point-to-point feature expressions are established between current and voltage;

[0092] Step 3-3: Design an improved convolutional autoencoder neural network with a convolutional autoencoder neural network structure combined with depthwise separable convolution (as shown in Figure 2 ), and combine the three different forms of feature expressions obtained into three channels of a tensor as the model input of the improved convolutional autoencoder neural network to extract V-I mapping features from different perspectives.

[0093] Depthwise separable convolution divides standard convolution into two steps: depthwise convolution and pointwise convolution.

[0094] Depthwise convolution designs convolutional kernels with the same number of input channels. Each convolutional kernel is only responsible for one channel, extracts features in the channel and generates corresponding feature maps. Here, convolution processing is performed on three channels. The number of input channels, the number of convolutional kernels, and the number of feature map channels generated during depthwise convolution are all 3. The formula is as follows:

[0095] Y DC (i,j,c) = ∑ m,n X(i+m,j+n,c)·W DC (m,n,c) (8)

[0096] In the formula: X is the feature input map; W DC is the depthwise convolutional kernel.

[0097] The main purpose of the pointwise convolution part operation is to perform feature weighted combination on the output of the depthwise convolution part. Its convolutional kernel size is 1*1*M, where M is the number of channels in the previous layer, which is 3 here. The number of convolutional kernels in the pointwise convolution part is the same as the number of output channels. The formula is:

[0098] Y PC (i,j,c') = ∑ c Y DC (i,j,c)·W PC (c,c') (9)

[0099] In the formula: W PC is the pointwise convolutional kernel.

[0100] The present invention trains the above model based on the pytorch deep learning framework and uses stochastic gradient descent. Combine the three different forms of feature expressions obtained into three channels of a tensor as the input of the model to obtain the final output vector F V-I = [f1,f2,...,f k , where f i represents a certain specific hidden characteristic in the V-I mapping image, and k is taken as 32;

[0101] Step 3-4: Normalize the harmonic current characteristics and the V-I mapping characteristics. Use the Z-Score normalization method. The Z-Score normalization formula is as follows:

[0102]

[0103] In the formula: x i is the original eigenvalue; μ x is the mean of the characteristics; σ x is the standard deviation of the characteristics.

[0104] Perform Z-Score normalization on each component in F h and F V-I respectively to obtain F′ h =[A′ m ,A′ ev ,A′ base ,A′3,A′5,A′7,THD′] and F′ V-I =[f′1,f′2,...,f′ k , and assign weights ω h and ω V-I to the normalized harmonic current characteristics and V-I characteristics respectively. Set ω h and ω V-I to be 0.6 and 0.4 respectively to obtain the fused load feature vector F=[ω h ·F′ h , ω V-I ·F′ V-I .

[0105] Step 4: Use the improved Siamese network to obtain the similarity between the fused load feature vector and the existing feature vectors in the database, and identify and update the load to be identified according to the similarity criterion. The specific steps are as follows:

[0106] Step 4-1: By introducing an attention mechanism into the Siamese network, endow the Siamese network with stronger feature extraction ability to obtain the improved Siamese network (as shown in Figure 3 );

[0107] Step 4-2: Use the BLUED dataset to generate positive and negative sample pairs for training - the positive sample pairs are two sets of characteristics of the same load, and the negative sample pairs are two sets of characteristics of different loads. Input the positive and negative sample pairs into the improved Siamese network and train through the output embedding vectors and the contrast loss function:

[0108]

[0109] Where: D is the Euclidean distance between two embedding vectors; X and Y are the embedding vectors output by the model; m is the threshold boundary; y is the similarity label, when y = 1, it represents a positive sample pair, and when y = 0, it represents a negative sample pair;

[0110] Step 4-3: Input the fused feature vector and the feature vector of each known load in the database into the two branches of the improved Siamese network to obtain the embedding vectors between the two branches;

[0111] Step 4-4: Calculate the similarity score S of the obtained embedding vectors through the Euclidean distance:

[0112]

[0113] If the similarity score with a certain load in the database is higher than the set threshold σ s , it is determined as a known load; if all similarity scores are lower than the threshold, it is determined as an unknown load, and the database is dynamically updated, the feature vector of this load is extracted, added to the database and a unique label is assigned to it. Embodiment

[0114] To verify the superiority of the present invention, the PLAID dataset is used as the test dataset. The PLAID dataset contains the voltage and current waveforms of 235 individual loads of 11 different types from more than 55 households in the United States, and the sampling frequency is 30 kHz. 200 steady-state cycles are extracted from the voltage and current data of each load to construct 100 sample data, making it closer to the actual application. 11 independent loads are extracted from the database, assuming that 6 of them are known loads and the other 5 are unknown loads, and the recognition ability of the unknown loads is observed.

[0115] A load feature library is constructed using 6 known loads to obtain the 7-dimensional harmonic current characteristics of the load at steady state; and the V-I trajectory characteristics of each load are obtained using the improved convolutional autoencoder neural network trained by the present invention. The encoding part of the network model is used for feature extraction to obtain a 32-dimensional V-I feature vector, and the final load feature vector is obtained through feature fusion; a total of 100 samples are randomly selected from the steady-state voltage and current data of 11 independent loads as test data. According to the above method, a 7-dimensional harmonic current feature vector and a 32-bit V-I trajectory feature vector are obtained, and then the improved Siamese network is used for load recognition.

[0116] To compare the recognition results with and without feature fusion, comparison is made with only the harmonic current feature and only the V-I trajectory feature. Table 1 shows the recognition effect when only the harmonic current feature is used, Table 2 shows the recognition effect when only the V-I trajectory feature is used, and Table 3 shows the recognition effect when the two features are fused.

[0117] Load 1 - Fluorescent lamp, Load 2 - Hair dryer, Load 3 - Microwave oven, Load 4 - Air conditioner, Load 5 - Laptop, Load 6 - Vacuum cleaner, Load 7 - Incandescent lamp, Load 8 - Washing machine, Load 9 - Fan, Load 10 - Heater, Load 11 - Refrigerator.

[0118] Table 1

[0119]

[0120]

[0121] Table 2

[0122]

[0123] Table 3

[0124]

[0125]

[0126] As shown in Table 1, when only using the harmonic current characteristics, the unknown load 7 is identified as load 4 because the harmonic current characteristics of the unknown load 7 are very close to those of load 4 in the database; as shown in Table 2, when only using the V-I trajectory characteristics, since the unknown load 10 (heater) and load 2 (hair dryer) are both resistive loads and their V-I trajectories are very close, the unknown load 10 is identified as load 2 in the database. As shown in Table 3, when using the harmonic current characteristics and the V-I trajectory characteristics simultaneously, the complementarity between the harmonic current characteristics and the V-I trajectory characteristics can be utilized to overcome the deficiencies of a single characteristic, thereby improving the recognition accuracy.

[0127] To verify the superiority of the structured V-I mapping compared with the traditional V-I mapping, as well as the superiority of the improved convolutional autoencoder neural network designed in the present invention, a comparison is made here with the support vector machine model, and Table 4 shows the comparison results of the recognition effects.

[0128] Table 4

[0129]

[0130] As shown in Table 4, compared with the traditional V-I mapping, in terms of recognition accuracy, the structured V-I mapping has a better accuracy rate, and although the designed improved convolutional autoencoder neural network takes slightly longer training time than the support vector machine, its accuracy rate is much superior to that of the support vector machine. In summary, the non-intrusive load monitoring model proposed in the present invention can better extract the load characteristics and at the same time ensure a high recognition accuracy during load recognition.

[0131] On the other hand, the present invention provides a non-intrusive load detection device based on feature fusion and structured V-I mapping, including:

[0132] A data acquisition and preprocessing module, configured to acquire voltage and current data on the bus of the target building, and preprocess the acquired voltage and current data;

[0133] A harmonic current feature extraction module, configured to detect load switching events by using the cumulative sum algorithm, and extract harmonic current features from the preprocessed current data through fast Fourier transform;

[0134] A load feature vector fusion module, configured to process the preprocessed voltage and current data by using the structured V-I mapping method combined with an improved convolutional autoencoder neural network to obtain V-I mapping features, and combine the harmonic current features and the V-I mapping features to construct a fused load feature vector;

[0135] A load identification module, configured to obtain the similarity between the fused load feature vector and the existing feature vectors in the database by using an improved siamese network, and identify and update the load to be identified according to the similarity criterion.

[0136] On the other hand, the present invention provides a non-intrusive load detection system based on feature fusion and structured V-I mapping, including: a computer-readable storage medium and a processor;

[0137] The computer-readable storage medium is used to store executable instructions;

[0138] The processor is configured to read the executable instructions stored in the computer-readable storage medium and execute the non-intrusive load detection method based on feature fusion and structured V-I mapping described in the first aspect.

[0139] On the other hand, the present invention provides a non-transitory computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the non-intrusive load detection method based on feature fusion and structured V-I mapping described in the first aspect.

[0140] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of a complete hardware embodiment, a complete 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-ROM, optical storage, etc.) containing computer-usable program code.

[0141] This application is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to 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 the processors of general-purpose computers, special-purpose computers, embedded processors, or other programmable data processing devices to generate a machine, such that the instructions executed by the processors of the computer or other programmable data processing devices generate means for implementing the functions specified in one or more of the flows Figure 1 one or more flows and / or blocks Figure 1 or means for implementing the functions specified in one or more of the blocks.

[0142] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, such that the instructions stored in the computer-readable memory generate a manufactured article including instruction means that implement the functions specified in one or more of the flows Figure 1 one or more flows and / or blocks Figure 1 or means for implementing the functions specified in one or more of the blocks.

[0143] These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process, so that the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one or more of the flows Figure 1 one or more flows and / or blocks Figure 1 or means for implementing the functions specified in one or more of the blocks.

[0144] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the above embodiments, those of ordinary skill in the art should understand that: it is still possible to modify the specific implementation manners of the present invention or make equivalent replacements, and any modification or equivalent replacement that does not depart from the spirit and scope of the present invention shall be covered by the protection scope of the claims of the present invention.

Claims

1. A non-intrusive load detection method based on feature fusion and structured V-I mapping, characterized in that It includes the following steps: Step 1: Use a high-frequency sampling device to collect voltage and current data on the bus of the target building, and preprocess the collected voltage and current data; Step 2: Use the cumulative sum algorithm to detect load switching events, and extract harmonic current characteristics from the preprocessed current data through fast Fourier transform; Step 3: Adopt a structured V-I mapping method combined with an improved convolutional autoencoder neural network to process the preprocessed voltage and current data, obtain V-I mapping characteristics, and combine the harmonic current characteristics and V-I mapping characteristics to construct a fused load feature vector; Step 4: Use an improved siamese network to obtain the similarity between the fused load feature vector and the existing feature vectors in the database, and identify and update the load to be recognized according to the similarity criterion.

2. The non-intrusive load detection method based on feature fusion and structured V-I mapping according to claim 1, wherein: The specific steps of Step 1 are as follows: Step 1-1: Set the sampling frequency to 4 kHz to capture the high-frequency harmonic characteristics and fine waveform changes during load operation; Step 1-2: Preprocess the collected signal in the sampling device, remove the high-frequency noise and DC offset in the signal, extract the frequency band of interest, and ensure that the extracted data covers all operating conditions of the load; Step 1-3: Store the collected high-frequency voltage and current data in the cloud database for subsequent load feature extraction and load feature recognition algorithms.

3. The non-intrusive load detection method based on feature fusion and structured V-I mapping according to claim 1, wherein: In Step 2, The extracted harmonic current characteristics include the effective value, maximum value of the steady-state current, and the values and contents of the 1st, 3rd, 5th, and 7th harmonic currents. The effective value and maximum value are calculated from the current sampling data, and the calculation steps are as follows: Step 2-1: Apply fast Fourier transform to the preprocessed current signal i(n) to obtain the frequency-domain signal in complex form: where: k = 0, 1,..., N-1 is the frequency index; Step 2-2: Calculate the spectral amplitude: where: Re(I(k)) and Im(I(k)) are the real and imaginary parts of I(k) respectively; Step 2-3: Extract the high-order harmonic components and calculate the amplitude of each harmonic: A h = |I(k h )| (4) where: h is the harmonic order, h = 1, 3, 5, 7; f base is the power grid power frequency, which is 50 Hz; A h is the amplitude of the harmonic of the corresponding order; Step 2-4: Calculate the total harmonic distortion rate according to the fundamental wave component and each harmonic component: where: H is the maximum harmonic order considered; Step 2-5: Construct the current harmonic feature vector F h = [A m , A ev , A base , A3, A5, A7, THD], where A m is the maximum current value; A ev is the effective current value; A base , A3, A5, and A7 are the fundamental wave amplitude, 3rd, 5th, and 7th harmonic amplitudes of the current; THD is the total harmonic distortion rate.

4. The non-intrusive load detection method based on feature fusion and structured V-I mapping according to claim 1, wherein: Step 3 includes: Step 3-1: First, decompose the collected voltage and current data periodically and normalize it; Step 3-2: Map the normalized voltage and current data from three dimensions, and construct multiplication mapping, addition mapping, and subtraction mapping feature expressions respectively; Step 3-3: Design an improved convolutional autoencoder neural network with the structure of a convolutional autoencoder neural network combined with depthwise separable convolution, combine the three different forms of feature expressions obtained into three channels of a tensor as the model input of the improved convolutional autoencoder neural network, and extract V-I mapping characteristics from different angles; Step 3-4: Normalize the harmonic current characteristics and the V-I mapping characteristics, assign weights to their respective importance, and linearly combine the normalized harmonic current characteristics and V-I mapping characteristics according to the weights to obtain a fused load feature vector.

5. The non-intrusive load detection method based on feature fusion and structured V-I mapping according to claim 1, wherein: Step 4 includes: Step 4-1: Obtain an improved Siamese network by introducing an attention mechanism into the Siamese network to endow the Siamese network with stronger feature extraction capabilities; Step 4-2: Use the BLUED dataset to generate positive and negative sample pairs for training, and input the positive and negative sample pairs into the improved Siamese network for training; Step 4-3: Input the fused feature vectors and the feature vectors of each known load in the database into the two branches of the improved Siamese network to obtain the embedding vectors between the two branches; Step 4-4: Calculate the similarity scores of the obtained embedding vectors through the Euclidean distance. If the similarity score with a certain load in the database is higher than the set threshold, it is determined as a known load; if all similarity scores are lower than the threshold, it is determined as an unknown load, and the database is dynamically updated. Extract the feature vectors of this load, add them to the database, and assign a unique label to it.

6. A non-intrusive load detection device based on feature fusion and structured V-I mapping, characterized in that, Including: A data acquisition and preprocessing module, which is used to acquire the voltage and current data on the bus of the target building and preprocess the acquired voltage and current data; A harmonic current feature extraction module, which is used to detect load switching events by using the cumulative sum algorithm and extract harmonic current features from the preprocessed current data through fast Fourier transform; A load feature vector fusion module, which is used to process the preprocessed voltage and current data by using a structured V-I mapping method combined with an improved convolutional autoencoder neural network to obtain V-I mapping features, and combine the harmonic current features and the V-I mapping features to construct a fused load feature vector; A load identification module, which is used to obtain the similarity between the fused load feature vector and the existing feature vectors in the database by using an improved Siamese network, and identify and update the load to be identified according to the similarity criterion.

7. A non-intrusive load detection system based on feature fusion and structured V-I mapping, comprising: A computer-readable storage medium and a processor; The computer-readable storage medium is used to store executable instructions; The processor is used to read the executable instructions stored in the computer-readable storage medium and execute the non-intrusive load detection method based on feature fusion and structured V-I mapping according to any one of claims 1-5.

8. A non-transitory computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the non-intrusive load detection method based on feature fusion and structured V-I mapping according to any one of claims 1-5.

Images