A non-intrusive load decomposition method, device and storage medium
By employing a non-intrusive load decomposition method based on instance and batch normalization networks, and utilizing encoder-decoder architecture and attention mechanism, the problem of insufficient generalization performance under low-frequency data is solved, achieving higher load decomposition accuracy and generalization capability.
Patent Information
- Application Number
- CN202310066496.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-01-12
- Publication Date
- 2026-01-13
- Estimated Expiration
- 2043-01-12
AI Technical Summary
Existing non-intrusive load decomposition methods have insufficient generalization performance under low-frequency data, making it difficult to maintain high load decomposition accuracy in unknown scenarios.
A non-intrusive load decomposition method based on instance and batch normalized networks is adopted. It utilizes an encoder-decoder structure with fused attention mechanism, combined with sliding window and sequence-to-point mapping, and constructs a model through IBN-Net to decompose low-frequency power data.
It improves the model's generalization performance and load decomposition accuracy, reduces interference from window edge information, enhances multi-scale information fusion capabilities, and strengthens load decomposition accuracy in unknown scenarios.
Smart Images
Figure CN115983347B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of non-intrusive load monitoring in smart grids, and in particular to a non-intrusive load decomposition method, apparatus, and storage medium based on instance and batch normalized networks. Background Technology
[0002] Electricity is not only a clean and efficient secondary energy source but also an important form of consumption in national production. Precise regulation of electricity can effectively improve energy efficiency. Non-intrusive load monitoring (NILM) technology decomposes the total load information at the power inlet into the load information of each electrical device, providing technical support for efficient demand-side management of electricity. NILM technology is a low-cost energy consumption monitoring and management method that can effectively reduce energy waste in buildings and has significant practical implications for achieving carbon peaking and carbon neutrality. Therefore, numerous academic and commercial institutions worldwide have conducted research in the field of NILM, resulting in a series of algorithmic achievements with varying principles, some of which have already been applied to products. Existing research largely focuses on improving the accuracy of appliance status identification and power / energy decomposition.
[0003] Load decomposition methods can be categorized into two types based on data acquisition frequency: those based on high-frequency data and those based on low-frequency data. While high-frequency data acquisition provides richer electrical parameter information, it requires sophisticated measurement equipment, is complex, and costly, making it unsuitable for residential electricity loads. Therefore, current research focuses on improving the accuracy of load decomposition based on low-frequency data acquired by smart meters. Improving the accuracy of non-intrusive load decomposition based on low-frequency data relies on efficient load feature extraction and more precise identification methods. Although the performance of non-intrusive load decomposition methods for low-frequency sampling has significantly improved, generalization performance remains a challenging problem. Summary of the Invention
[0004] The purpose of this invention is to provide a non-intrusive load decomposition method, apparatus, and storage medium based on instance and batch normalized networks, thereby improving generalization performance and load decomposition accuracy.
[0005] The objective of this invention can be achieved through the following technical solutions:
[0006] A non-intrusive load decomposition method based on instance and batch normalized networks includes the following steps:
[0007] Step 1) Obtain the total household power and low-frequency power data of each electrical device based on the smart meter, and divide them into training set and test set;
[0008] Step 2) Perform data preprocessing on the low-frequency power data;
[0009] Step 3) Segment the preprocessed low-frequency power data based on a sliding window;
[0010] Step 4) Construct a non-intrusive load decomposition model based on instance and normalized network for each electrical device. Use an encoder-decoder structure with fused attention mechanism to map the low-frequency active power sequence of each electrical device to the point. The input of the non-intrusive load decomposition model is the total household power time series window, and the output is the point value of the power time series of the electrical device in the window. The encoder-decoder structure is based on a multi-layer IBN-Net structure that integrates instance normalization and batch normalization.
[0011] Step 5) Train a non-intrusive load decomposition model based on the segmented training set data;
[0012] Step 6) Input the segmented test set data into the trained non-intrusive load decomposition model to obtain the active power decomposition value of each electrical device, and compare it with the true value to evaluate the model.
[0013] Step 2) includes the following steps:
[0014] Step 2-1) Remove outliers from the low-frequency power data;
[0015] Step 2-2) Normalize the low-frequency power data:
[0016]
[0017] Where z is the normalized value, z t z represents the reading of the total power supply or electrical equipment at time t. p σ represents the mean of the total power supply or electrical equipment, and σ represents the standard deviation of the total power supply or electrical equipment.
[0018] The non-invasive load decomposition model is expressed as follows:
[0019] x τ =F b (Y t:t+L-1 )+∈
[0020] The input to the model is the total power sliding window sequence Y. t:t+L-1 The window width is L; the output is the midpoint element x of the corresponding target electrical equipment window sequence. τ , where τ=t+[L / 2]; F b It is to input Y t:t+L-1 Mapping to output x τ The neural network is ∈, where L is L-dimensional Gaussian noise.
[0021] The encoder in the encoder-decoder structure includes multiple first IBN-Net modules connected in sequence and a fully connected layer. Each first IBN-Net module consists of an IBN-Net and a max-pooling layer connected in sequence. The fully connected layer converts the output of the IBN-Net stack into a context vector C. The decoder has a similar structure to the encoder, consisting of the same number of second IBN-Net modules as the first IBN-Net modules and a convolutional layer. Each second IBN-Net module includes an IBN-Net, a skip connection function, and a deconvolutional layer connected in sequence. The shallow output of the encoder is input to the corresponding layer of the decoder based on the skip connection function.
[0022] The encoder receives the total power window sequence Y. t:t+L-1 Extract the power information of the target electrical equipment and map it to the context vector C, C = {h1, h2, ..., h...} L}, h t Let t be the hidden state of the input sequence generated during the encoding process. The encoding process is represented as follows:
[0023] C = f e (WY t:t+L-1 +b)
[0024] In the formula: f e is the encoding function; W and b are the function weights and biases of the encoding layer, respectively.
[0025] The attention unit between the encoder and decoder is used to capture salient activations of the target electrical equipment in the encoder output features and extract more valuable feature information for decomposition. The calculation process of the attention unit is as follows:
[0026] e t =V a T *tanh(W a h t +b a ), t=1,2,…,L
[0027]
[0028]
[0029] In the formula, V a W a and b a These are the parameters to be learned in the model;
[0030] The attention unit is a single-layer feedforward neural network, which calculates the attention weight α. t, returns the weighted average of the encoder's output over time, i.e., the context vector C'.
[0031] The decoder is used to reconstruct the sequence of the target electrical equipment. The decoding process is as follows:
[0032] x τ =f d (W'C'+b')
[0033] In the formula: f d Here, W' and b' represent the decoding function, respectively, and W' and b' represent the function weights and biases of the decoding layer. C' is a dynamically variable context vector that is generated by the attention unit between the encoder and decoder during the entire decoding process.
[0034] The non-intrusive load decomposition model uses mean squared error as the loss function and is trained by adjusting the model parameters in conjunction with the Adam optimizer.
[0035] A non-intrusive load decomposition device based on instance and batch normalization networks includes a memory, a processor, and a program stored in the memory, wherein the processor executes the program to implement the method described above.
[0036] A storage medium having a program stored thereon, which, when executed, implements the method described above.
[0037] Compared with the prior art, the present invention has the following beneficial effects:
[0038] (1) The present invention uses a sequence-to-point mapping method, which reduces the interference of window edge information and improves the generalization of the model compared with the sequence-to-sequence method.
[0039] (2) This invention applies instance and batch normalized networks with strong generalization performance to non-intrusive load decomposition, which greatly improves the load decomposition accuracy of the trained model when applied to unknown scenarios.
[0040] (3) The present invention uses an encoder-decoder structure with a fusion attention mechanism and inputs the shallow output of the encoder into the corresponding layer of the decoder using a skip connection. The use of skip connections improves the multi-scale information fusion capability of the network and helps the encoder construct a more accurate power sequence of the target electrical equipment. Attached Figure Description
[0041] Figure 1 This is a flowchart of the method of the present invention;
[0042] Figure 2 This is a schematic diagram of the non-intrusive load decomposition model in this invention;
[0043] Figure 3This is a schematic diagram of the IBN-Net structure in this invention;
[0044] Figure 4 This is a schematic diagram of the attention mechanism in this invention. Detailed Implementation
[0045] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.
[0046] This embodiment provides a non-intrusive load decomposition method based on instance and batch normalized networks, such as Figure 1 As shown, it includes the following steps:
[0047] Step 1) Based on the smart meter, acquire the low-frequency power data of the total power supply of the household and each independent electrical device over a period of time. Add time tags to the collected data, organize it into a standard active power time series, and divide it into training set and test set.
[0048] Step 2) Perform data preprocessing on the low-frequency power data.
[0049] Step 2-1) Remove outliers from the low-frequency power data;
[0050] Step 2-2) Normalize the low-frequency power data:
[0051]
[0052] Where z is the normalized value, z t z represents the reading of the total power supply or electrical equipment at time t. p σ represents the mean of the total power supply or electrical equipment, and σ represents the standard deviation of the total power supply or electrical equipment.
[0053] Step 3) Segment the preprocessed low-frequency power data based on a sliding window.
[0054] Step 4) Build a non-intrusive load decomposition model for each electrical device based on instances and normalized networks.
[0055] The non-intrusive load decomposition model employs an encoder-decoder structure with a fusion attention mechanism to map the low-frequency active power sequence of each electrical device to a point. The model input is a window of the total household power time series, and the output is the point value within the window of the power time series of the electrical devices, expressed as:
[0056] x τ =F b (Y t:t+L-1 )+∈
[0057] Among them, Y t:t+L-1 The total power sliding window sequence is the input to the model, with a window width of L; x τ It is the midpoint element of the corresponding target electrical equipment window sequence output by the model, where τ = t + [L / 2]; F b It is to input Y t:t+L-1 Mapping to output x τ The neural network is ∈, where L is L-dimensional Gaussian noise.
[0058] like Figure 2 As shown, the encoder and decoder adopt a symmetrical structure, both consisting of instance-batch normalization networks (IBN-Net) that integrate multiple layers of instance normalization (IN) and batch normalization (BN).
[0059] The encoder includes multiple first IBN-Net modules connected in sequence and a fully connected layer. Each first IBN-Net module consists of an IBN-Net layer and a max-pooling layer connected in sequence. The max-pooling layer is used to reduce the temporal resolution, facilitating the network to learn high-level features of the target device. The fully connected layer (Dense) converts the output of the IBN-Net stack into a context vector C. In this embodiment, the number of first IBN-Net modules is 5.
[0060] The decoder's structure is similar to the encoder, consisting of the same number of second IBN-Net modules as the first IBN-Net modules and a convolutional layer. Each second IBN-Net module includes sequentially connected IBN-Nets, skip connection functions, and deconvolutional layers. The shallow output of the encoder is input to the corresponding layer of the decoder based on the skip connection functions, which helps the decoder better fuse the features extracted by the shallow network of the encoder to construct a more accurate target electrical power sequence. The deconvolutional layer is used to progressively improve the temporal resolution and reconstruct the signal of the target device. In this embodiment, the number of second IBN-Net modules is also 5.
[0061] A schematic diagram of the IBN-Net structure is shown below. Figure 3 As shown, it includes multiple BN networks connected in sequence and an IN network, and the input of the IBN-Net and the output of the last BN network are connected by a residual connection.
[0062] During the implementation of the method, the encoder first obtains the total power window sequence Y. t:t+L-1 Extract the power information of the target electrical equipment and map it to the context vector C, C = {h1, h2, ..., h...} L}, h tLet t be the hidden state of the input sequence generated during the encoding process. The encoding process is represented as follows:
[0063] C = f e (WY t:t+L-1 +b)
[0064] In the formula: f e is the encoding function; W and b are the function weights and biases of the encoding layer, respectively.
[0065] Then, the attention unit between the encoder and decoder captures salient activations of the target electrical device in the encoder output features. This approach allows the network to implicitly detect certain events (e.g., turning the device on or off) and specific signal components (e.g., high power consumption) and assign them higher weights. A schematic diagram of the attention unit structure is shown below. Figure 4 As shown, the calculation process is as follows:
[0066] e t =V a T *tanh(W a h t +b a ), t=1,2,…,L
[0067]
[0068]
[0069] In the formula, V a W a and b a These are the model's learning parameters, which are continuously adjusted during training.
[0070] In this embodiment, the attention unit is a single-layer feedforward neural network, which calculates the attention weight α. t , returns the weighted average of the encoder's output over time, i.e., the context vector C'.
[0071] Finally, the decoder reconstructs the sequence of the target electrical equipment. The decoding process is as follows:
[0072] x τ =f d (W'C'+b')
[0073] In the formula: f d Here, W' and b' represent the decoding function, respectively, and W' and b' represent the function weights and biases of the decoding layer. C' is a dynamically variable context vector that is generated by the attention unit between the encoder and decoder during the entire decoding process.
[0074] Step 5) Train a non-intrusive load decomposition model based on the segmented training set data.
[0075] This embodiment uses the public datasets REDD and UK-DALE to train the model, selecting mean squared error (MSE) as the loss function and using the Adam optimizer to adjust the model parameters. The ModelCheckpoint callback function in Keras is used to monitor the validation set loss during training, saving the best model that emerges during training.
[0076] Step 6) Input the segmented test set data into the trained non-intrusive load decomposition model to obtain the active power decomposition value of each electrical device, compare it with the actual active power value of each electrical device, calculate the load decomposition error, and evaluate the model.
[0077] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0078] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.
Claims
1. A non-intrusive load disaggregation method based on instance and batch normalization networks, characterized in that, The method comprises the following steps: Step 1) obtaining low-frequency power data of a total power supply and each electrical equipment of a family based on a smart meter, and dividing the low-frequency power data into a training set and a test set; Step 2) performing data preprocessing on the low-frequency power data; Step 3) segmenting the preprocessed low-frequency power data based on a sliding window; Step 4) constructing an instance-based and normalized network non-intrusive load decomposition model for each electrical equipment, and realizing mapping of a low-frequency active power sequence of each electrical equipment to a point by using an encoder-decoder structure with a fusion attention mechanism, wherein an input of the non-intrusive load decomposition model is a total power time sequence window of the family, and an output is a point value in a window of a power time sequence of the electrical equipment, and the encoder-decoder structure is composed of an IBN-Net structure with instance normalization and batch normalization integrated in multiple layers; The encoder extracts power information of the target electrical device from the total power window sequence Y t:t+L−1 and maps to a context vector C , , For the input sequence generated in the encoding process t The hidden state at time t, the encoding process is represented as: In the formula: is an encoding function; W , b are the function weights and bias of the encoding layer, respectively. An attention unit between the encoder and the decoder is used to capture significant activation of a target electrical equipment in an output feature of the encoder, and extract more valuable feature information for decomposition, and a calculation process of the attention unit is as follows: In the formula, , and are the parameters to be learned by the model; The attention unit is a single-layer feedforward neural network that computes attention weights , returning a weighted average over time of the output of the encoder, i.e. the context vector ; The decoder is used to reconstruct a sequence of the target electrical equipment, and a decoding process is as follows: wherein: is a decoding function, , are the function weights and bias of the decoding layer, respectively; is a dynamically variable context vector during the decoding process, which is generated by the attention unit between the encoder and the decoder. Step 5) training the non-intrusive load decomposition model based on the segmented training set data; Step 6) inputting the segmented test set data into the trained non-intrusive load decomposition model to obtain active power decomposition values of each electrical equipment, and comparing the active power decomposition values with true values to evaluate the model.
2. The non-intrusive load disaggregation method based on instance and batch normalization network according to claim 1, characterized in that, The step 2) comprises the following steps: Step 2-1) removing outliers in the low-frequency power data; Step 2-2) performing normalization processing on the low-frequency power data: wherein, z is the normalized value, z t denotes t the reading of the total power supply or of the power consuming device at the moment, z p denotes the mean value of the total power supply or of the power consuming device, The non-intrusive load decomposition model is represented as: denotes the standard deviation of the total power supply or of the power consuming device.
3. The non-intrusive load disaggregation method based on instance and batch normalization network according to claim 1, characterized in that, The encoder of the encoder-decoder structure comprises a plurality of first IBN-Net modules and a full connection layer connected in sequence, each first IBN-Net module is composed of an IBN-Net and a maximum pooling layer connected in sequence, and the full connection layer converts an output of the IBN-Net stack into a context vector C; the structure of the decoder is similar to that of the encoder, and the decoder is composed of a plurality of second IBN-Net modules and a convolution layer, each second IBN-Net module comprises an IBN-Net, a skip connection function and a deconvolution layer connected in sequence, and an output of a shallow layer of the encoder is input into a corresponding layer of the decoder based on the skip connection function. wherein the input of the model is a total power sliding window sequence Y t:t+L−1 , the window width is L ; the output is the midpoint element of the corresponding target electrical equipment window sequence x τ , wherein The non-intrusive load decomposition model adopts a mean square error as a loss function, and is trained by adjusting model parameters with an Adam optimizer. ; F b is a neural network that maps the input Y t:t+L−1 to the output x τ , and The processor implements the method of any one of claims 1-5 when executing the program. is L a d-dimensional Gaussian noise.
4. The non-intrusive load disaggregation method based on instance and batch normalization network according to claim 1, characterized in that, The program is executed to implement the method of any one of claims 1-5.
5. The non-intrusive load disaggregation method based on instance and batch normalization network according to claim 1, characterized in that, 6. An apparatus for non-intrusive load disaggregation based on instance and batch normalized networks, comprising a memory, a processor, and a program stored in the memory, wherein, 7. A storage medium having stored thereon a program, characterized by