A multi-task non-intrusive load decomposition method based on UNet structure
The UNet-based method for non-intrusive load decomposition addresses the challenges of high costs and adaptability in existing methods by using low-frequency data and synthetic datasets, enabling efficient and adaptable power consumption monitoring across multiple appliances.
Patent Information
- Application Number
- CN202210595344.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-05-29
- Publication Date
- 2024-12-13
- Estimated Expiration
- 2042-05-29
AI Technical Summary
Existing non-intrusive load decomposition methods require large amounts of labeled data, are costly, and struggle to adapt to multiple types of electrical appliances, necessitating retraining when new appliances are introduced.
A multi-task non-intrusive load decomposition method using a UNet structure that utilizes low-frequency load power data and synthetic data sets, employing a shared encoder and unique decoders for each appliance, allowing simultaneous power decomposition of multiple appliances without extensive data collection.
Reduces data acquisition costs, enhances model portability, and improves efficiency by enabling simultaneous power decomposition of multiple appliances using a single network, while maintaining accuracy and adaptability to new appliances.
Smart Images
Figure CN114970723B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of power load decomposition, and in particular relates to a multi-task non-intrusive load decomposition method based on a UNet structure. Background Art
[0002] Non-intrusive Load Disaggregate (NILD) is to analyze the total meter data of a certain electricity user to obtain the electricity consumption information of each load of the electricity user. Its main goal is to decompose the power of each electrical appliance from the total power and realize the sub-item metering of electricity power. The non-intrusive approach means avoiding the installation of sensors inside the user. Compared with the invasive approach, this is a strategy to save costs, simplify deployment steps, and protect user privacy, but it also means that it requires more complex and advanced data processing algorithms and technologies. NILD can provide electricity users with their own detailed electricity consumption, guide users to improve their electricity usage habits, and enhance their energy-saving awareness. At the same time, the power supplier can use this technology to optimize the power supply strategy and achieve a reasonable allocation of power resources.
[0003] At present, most non-intrusive load decomposition methods rely on a large amount of known labeled data for training. However, obtaining labeled data often requires a lot of cost, which greatly increases the difficulty of deploying non-intrusive load decomposition. There are many types of existing electrical appliances. A decomposition method can often only adapt to a few types of electrical appliances, while it is difficult to maintain a good decomposition effect for other electrical appliances. In addition, when new electrical appliances are added, the model may need to be retrained to be able to decompose the new target electrical appliances. Summary of the invention
[0004] The technical problem to be solved by the present invention is to provide a multi-task non-invasive load decomposition method based on the UNet structure, which uses low-frequency load power data to achieve power decomposition of electrical appliances, thereby monitoring the power consumption of electrical appliances. The low-frequency data source ensures the low cost of the data acquisition device and its compatibility with existing smart meters.
[0005] To solve the above technical problems, an embodiment of the present invention provides a multi-task non-intrusive load decomposition method based on a UNet structure, comprising the following steps:
[0006] S1. Select the target electrical appliances to be decomposed, and measure the operating power sequence of each target electrical appliance through a power collection device;
[0007] S2, using data synthesis methods to construct synthetic data sets;
[0008] S3. Use the synthetic data set to train a multi-task non-intrusive load decomposition algorithm model based on the UNet structure to obtain a model that contains optimal weight parameters and can decompose the active power sequence of the target electrical appliance.
[0009] The multi-task non-intrusive load decomposition algorithm model based on the UNet structure includes an encoder, a plurality of decoders equal to the target electrical appliances, and a plurality of output layers corresponding to the plurality of decoders one by one;
[0010] The encoder parameters are shared by all target electrical appliances and trained by all target electrical appliances;
[0011] Each of the decoders is unique to the corresponding target appliance and is independently trained by the respective target appliance;
[0012] Each of the output layers outputs a power sequence corresponding to the target electrical appliance.
[0013] The encoder includes 5 downsampling modules, the first 4 downsampling modules are composed of two convolutional layers and one pooling layer, and the last downsampling module is composed of two convolutional layers.
[0014] Each of the encoders includes four upsampling modules, and each of the upsampling modules consists of an upsampling operation, a feature concatenation operation and two convolutional layers.
[0015] Wherein, each of the output layers includes a convolution layer with a convolution kernel of 1.
[0016] Among them, constructing the synthetic data set in step S2 includes the following steps:
[0017] S2.1. Collect the active power sequence of a single electrical appliance during operation through a power collection device to obtain an active power sequence of the target electrical appliance from the moment before it is turned on to the moment it is turned off;
[0018] S2.2. All target electrical appliances are turned off for a period of time, and the total power sequence during this period is measured;
[0019] S2.3, select a time period equal to that of step S2.2, randomly select multiple non-overlapping moments within the time period, and construct multiple active power sequences of a target electrical appliance during operation according to step S2.1;
[0020] S2.4. Repeat step S2.3 to construct multiple runtime active power sequences of each target electrical appliance as labels for training, and then superimpose the runtime active power sequences of all target electrical appliances and the total power sequence in step S2.2 to obtain a synthesized total power sequence.
[0021] Furthermore, in step S2.1, for each target electrical appliance, a number of operating active power sequences may be measured to simulate different operating durations and operating modes of the target electrical appliance.
[0022] In step S2.2, the total power sequence includes irrelevant electrical appliance power and noise.
[0023] In step S2.3, the interval between each start-up time of the target electrical appliance should be greater than the shortest operating time of the target electrical appliance.
[0024] The beneficial effects of the above technical solution of the present invention are as follows:
[0025] 1. The present invention uses only low-frequency load power data to achieve power decomposition of electrical appliances, thereby monitoring the power consumption of electrical appliances. The low-frequency data source ensures the low cost of the data acquisition device and compatibility with existing smart meters.
[0026] 2. The present invention uses the UNet structure network to construct a multi-task deep learning algorithm, which extracts input data features through a convolutional neural network and fuses shallow information and deep information through cross-layer connections.
[0027] 3. The present invention constructs a multi-task model, which can use only one network and perform one inference to obtain the operating power of multiple appliances at the same time. Compared with the previous model, in which one model only decomposes the operating power of a specific appliance, this method improves the overall efficiency of the load decomposition task and reduces the parameter scale.
[0028] 4. The present invention uses the operating power of each target electrical appliance to construct a synthetic data set for training. Compared with the method of simultaneously collecting data from multiple electrical appliances to construct a data set, the present invention reduces the complexity of the data acquisition system, saves the cost of data set construction, and improves the portability of the method. BRIEF DESCRIPTION OF THE DRAWINGS
[0029] Figure 1 It is a structural block diagram of a multi-task model for simultaneously decomposing the power of multiple target electrical appliances in the present invention;
[0030] Figure 2 It is a structural diagram of the load decomposition model based on the UNet structure in the present invention;
[0031] Figure 3 It is a general implementation flow chart of the present invention. DETAILED DESCRIPTION
[0032] In order to make the technical problems, technical solutions and advantages to be solved by the present invention more clear, a detailed description will be given below with reference to the accompanying drawings and specific embodiments.
[0033] The present invention provides a multi-task non-intrusive load decomposition method based on UNet structure, comprising the following steps:
[0034] S1. Select the target electrical appliances to be decomposed, and measure the operating power sequence of each target electrical appliance through a power collection device; multiple collections can be performed so that each target electrical appliance has multiple different power sequences corresponding to different operating modes of the target electrical appliance. In this step, multiple power collection devices are not required, and at least one power collection device is required to collect the operating power sequence of each target electrical appliance respectively.
[0035] S2, using data synthesis methods to construct synthetic data sets;
[0036] S3. Use the synthetic data set to train a multi-task non-intrusive load decomposition algorithm model based on the UNet structure to obtain a model that contains optimal weight parameters and can decompose the active power sequence of the target electrical appliance.
[0037] like Figure 2 , Figure 3 As shown, the multi-task non-intrusive load decomposition algorithm model based on the UNet structure includes an encoder, a plurality of decoders equal to the target electrical appliances, and a plurality of output layers corresponding to the plurality of decoders one by one;
[0038] The encoder parameters are shared by all target electrical appliances and trained by all target electrical appliances;
[0039] Each of the decoders is unique to the corresponding target appliance and is independently trained by the respective target appliance;
[0040] Each of the output layers outputs a power sequence corresponding to the target electrical appliance.
[0041] In the above structure, the multi-task non-intrusive load decomposition algorithm model based on the UNet structure is divided into a decoder and an encoder. The encoder parameters are shared by all target appliances, that is, the encoder is trained by all target appliances. The decoder is independent of each target appliance. Therefore, the input of the model first passes through the encoder to obtain the encoded data, and then branches are performed. Each branch decodes the encoded data to obtain the power sequence of the corresponding target appliance. N target appliances will correspond to N different branches. Each branch is the same in model structure, but has its own set of trained model parameters. When a new target appliance is added, there is no need to train the encoder parameters. Only the decoder corresponding to the appliance needs to be trained. After that, the decoder branch is added to the model, and the model can decompose the power sequence of the target appliance through this branch.
[0042] UNet is a U-shaped neural network model based on convolution. The structural diagram of the load decomposition model based on the UNet structure is as follows: Figure 1As shown in the figure, the network model consists of three parts, namely the encoder, the decoder, and the final output layer. In the non-intrusive load decomposition task based on the UNet structure, the input of the model is the total active power time series with a length of 572 to be decomposed, that is, one-dimensional data with a shape of 572*1. The input data first passes through the encoder and is encoded into encoded data. In this part, the original input data will be automatically extracted by the convolution operation and finally encoded into encoded data. The encoded data is then decoded by the decoder to obtain the decoded data, and finally the decoded data enters the output layer for output to obtain the decomposed target electrical appliance active power sequence.
[0043] The encoder consists of 5 downsampling modules. Except for the last downsampling module, each downsampling module consists of two consecutive convolutional layers and a pooling layer. The size of the convolution kernel is 3*1, the step size is set to 1, the padding mode is vaild, and the ReLU activation function is used for activation. The pooling layer uses the maximum pooling of 2*1 size, that is, the maximum value of two adjacent local data points is taken as the output of the pooling layer. Therefore, each time it passes through the pooling layer, the size of the entire input data will be reduced to half of the original size. The last module consists of only two convolutional layers without pooling. In each module, the two convolutional layers use the same number of filters. The number of filters in the first module is 32, the second module is 64, the third is 128, the fourth is 256, and the fifth is 512. The final output encoded data shape is 28*512. The first four downsampling modules are composed of two convolutional layers plus a pooling layer. The size of the convolution kernel is 3*1, and the number of convolution kernels is 32.
[0044] After that, the encoded data output by the encoder enters the decoder, which is composed of 4 upsampling modules. Each upsampling module consists of an upsampling operation, a feature splicing operation, and two convolutional layers. The upsampling operation is to expand the input data size to twice the original data size, corresponding to the reduction of the data scale by each pooling, and is used to restore the data size. Feature splicing is to splice the output information of different layers. In order to ensure the matching of feature sizes, it is necessary to crop the two sides of the low-level feature map first. For example, for the first feature splicing, the splicing sources are the output of the convolution layer Conv8, which is 64*256 in shape, and the output of the upsampling layer Upsampling1, which is 56*256 in shape. Crop 4 units on both sides of Conv8 and then splice it with Upsampling1. Similarly, the two convolutional layers in the same upsampling module use the same number of filters, and the number of convolution kernels of the convolutional layers in the four modules are 256, 128, 64, and 32 respectively. The size of the convolution kernel is 3*1, and the ReLU activation function is used for activation. The input of the decoder is the encoded data, which has a shape of 28*512. After the first upsampling module, the data shape becomes 52*256. After the second upsampling module, the data output shape is 100*128. And so on. After the subsequent upsampling modules, the output shapes are 196*64 and 388*32 respectively. Finally, a convolution layer with a convolution kernel size of 1 is used for output, and the shape of the output data is 388*1.
[0045] The model based on the UNet structure is used for non-invasive load decomposition. The feature information of the original input active power sequence is extracted through multi-layer convolution, and the deep and shallow information are fused through cross-layer splicing. For the shallower convolution in the network, its receptive field is small and can perceive the detailed features of the input data in a short time interval, while for the deep convolution, its receptive field is large and can obtain the features of the input data in a longer time period. Feature splicing combines the two, obtaining deep feature information without ignoring shallow information.
[0046] Considering that establishing a labeled dataset requires high equipment and time costs, a data synthesis method is proposed in step S2 of the present invention to construct a synthetic dataset. Using a synthetic dataset to train a model can greatly reduce the cost of making a dataset and achieve rapid model training and deployment. Constructing a synthetic dataset includes the following steps:
[0047] S2.1. Use a power collection device to collect the active power sequence of a single electrical appliance during operation. The sampling is low-frequency sampling (1Hz), and an active power sequence of a target electrical appliance from the moment before it is turned on to the moment it is turned off is obtained. Repeat the operation to measure the active power sequence of each target electrical appliance during operation. For each target electrical appliance, several operating power sequences can be measured to simulate different operating durations and operating modes of the electrical appliance.
[0048] S2.2, measure the total power sequence for a period of time. During this period of time, each target appliance is turned off, so the total power sequence includes the power of irrelevant appliances and noise;
[0049] S2.3. Select a time period equal to that of step S2.2, randomly select multiple non-overlapping moments within the time period, simulate the start-up process of the appliance, and construct multiple active power sequences of a target appliance during operation according to step S2.1; the constraint conditions for starting the target appliance are: each operation does not overlap, and the interval between the start-up moments should be greater than the shortest operating time of the appliance.
[0050] S2.4. Repeat step S2.3 to construct multiple runtime active power sequences of each target electrical appliance as labels for training, and then superimpose the runtime active power sequences of all target electrical appliances and the total power sequence in step S2.2 to obtain a synthesized total power sequence.
[0051] The advantages of the above data synthesis method are: (1) there is no need for multiple power acquisition devices, and at least only one device is needed to collect the power sequence of each target appliance and the total power sequence respectively; (2) when a new target appliance needs to be added, it is only necessary to measure the operating power sequence of the new target appliance, and use the method of step S2.3 to construct the power sequence of the target appliance, and finally superimpose it into the previous total power sequence, without the need to rebuild the data set.
[0052] The present invention provides a multi-task non-intrusive load decomposition method based on UNet structure, which is used to monitor power consumption, measure and record the operating power of each appliance, and only uses the low-frequency power of 1Hz at the power inlet, with low collection cost. The present invention can decompose multiple target appliances at the same time, without having to perform the decomposition process separately for each target appliance (see Figure 3 ).
[0053] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.
Claims
1. A multi-task non-intrusive load decomposition method based on UNet structure, characterized in that: The steps include: S1. Select the target electrical appliances to be decomposed, and measure the operating power sequence of each target electrical appliance through a power collection device; S2, using data synthesis methods to construct synthetic data sets; S3. Use the synthetic data set to train a multi-task non-intrusive load decomposition algorithm model based on the UNet structure to obtain a model that contains optimal weight parameters and can decompose the active power sequence of the target electrical appliance; Constructing the synthetic data set in step S2 includes the following steps: S2.
1. Collect the active power sequence of a single electrical appliance during operation through a power collection device to obtain an active power sequence of the target electrical appliance from the moment before it is turned on to the moment it is turned off; S2.
2. All target electrical appliances are turned off for a period of time, and the total power sequence during this period is measured; S2.3, select a time period equal to that of step S2.2, randomly select multiple non-overlapping moments within the time period, and construct multiple active power sequences of a target electrical appliance during operation according to step S2.1; S2.
4. Repeat step S2.3 to construct multiple runtime active power sequences of each target electrical appliance as labels for training, and then superimpose the runtime active power sequences of all target electrical appliances and the total power sequence in step S2.2 to obtain a synthesized total power sequence.
2. According to the UNet structure-based multi-task non-intrusive load decomposition method of claim 1, it is characterized in that: The multi-task non-intrusive load decomposition algorithm model based on the UNet structure includes an encoder, a plurality of decoders equal to the target electrical appliances, and a plurality of output layers corresponding to the plurality of decoders one by one; The encoder parameters are shared by all target electrical appliances and trained by all target electrical appliances; Each of the decoders is unique to the corresponding target appliance and is independently trained by the respective target appliance; Each of the output layers outputs a power sequence corresponding to the target electrical appliance.
3. The multi-task non-intrusive load decomposition method based on UNet structure according to claim 2 is characterized in that: The encoder includes 5 downsampling modules, the first 4 downsampling modules are composed of two convolutional layers and one pooling layer, and the last downsampling module is composed of two convolutional layers.
4. The multi-task non-intrusive load decomposition method based on UNet structure according to claim 2 is characterized in that: Each of the encoders includes four upsampling modules, and each of the upsampling modules consists of an upsampling operation, a feature concatenation operation and two convolutional layers.
5. The multi-task non-intrusive load decomposition method based on UNet structure according to claim 2 is characterized in that: Each of the output layers includes a convolution layer with a convolution kernel of 1.
6. The multi-task non-intrusive load decomposition method based on UNet structure according to claim 1 is characterized in that: In step S2.1, for each target electrical appliance, a number of operating active power sequences may be measured to simulate different operating durations and operating modes of the target electrical appliance.
7. The multi-task non-intrusive load decomposition method based on UNet structure according to claim 1 is characterized in that: In step S2.2, the total power sequence includes irrelevant electrical appliance power and noise.
8. The multi-task non-intrusive load decomposition method based on UNet structure according to claim 1 is characterized in that: In step S2.3, the interval between each start-up time of the target electrical appliance should be greater than the shortest operating time of the target electrical appliance.