Incremental non-intrusive decomposition method and system considering dynamic load change

Abstract

The invention discloses an incremental non-intrusive decomposition method and system considering load dynamic change, and relates to the technical field of power load management analysis, and the method comprises the steps: receiving a public data set, determining training related information based on the public data set, obtaining power utilization data and total electricity meter data of to-be-trained equipment based on the training related information, and carrying out the training of the to-be-trained equipment; inputting the power consumption data of the to-be-trained equipment and the total ammeter data into a pre-established sequence arrival point model, and training to obtain an independent decomposition model of the to-be-trained equipment; performing traversal detection on the to-be-trained equipment to obtain newly-added electric equipment, obtaining data of the newly-added electric equipment, inputting the data of the newly-added electric equipment into an independent decomposition model of the to-be-trained equipment by using a transfer learning technology, and performing fine tuning to obtain a fine-tuned decomposition model; and inputting the total ammeter data of the to-be-trained equipment into the fine-tuned decomposition model to obtain the power consumption data of the newly-added power consumption equipment.

Description

Technical Field

[0001] The present invention relates to the technical field of power load management and analysis, and in particular to an incremental non-invasive decomposition method and system that considers dynamic changes in load. Background Art

[0002] Non-intrusive load decomposition technology obtains detailed power consumption information of each appliance in the monitoring area by analyzing and processing the total power consumption data in the monitoring area. This method does not require entering the monitoring area, but only requires installing a monitoring device at the power entrance. It is easy to operate, has low operation and maintenance costs, and can well protect the user's power privacy. It has attracted much attention in recent years.

[0003] During the operation of the power system, the load often changes dynamically, and the load demand at different times is different. Economic development and technological progress have led to a significant increase in the electricity demand of the entire society, and users' electricity consumption behaviors have also become more diverse. These situations will make the load changes of the power grid more complicated, bringing new challenges to the operation and maintenance of the power system.

[0004] Faced with the newly added electrical equipment and appliances in the power grid, the previously trained load decomposition model is usually unable to correctly identify and decompose. Therefore, the model needs to be retrained to effectively cope with the changes in the incremental load in the power system. However, retraining the model is not easy. This work requires sufficient high-quality power data to improve the performance and decomposition effect of the model. New electrical equipment and appliances are usually unknown, which means that the model cannot be retrained using a complete and accurate data set, and the identification and prediction of incremental loads has become a difficult problem. In addition, the addition of new electrical equipment and appliances will also affect the model's ability to identify the original load. The original model has not been trained on the data of the incremental load, so the incremental load is equivalent to a noise signal for the model, which will make the power data to be processed more complicated, and the accuracy of the model's decomposition of the original load will also be reduced.

[0005] Currently, the accuracy tests of most non-invasive load decomposition algorithms are conducted without considering the incremental load. The prior art, with publication number CN114037178B and titled “Non-invasive load decomposition method based on unsupervised pre-trained neural network”, uses a large amount of unlabeled data from other households in the same data set to perform unsupervised pre-training on the network based on the sequence-to-point convolutional neural network architecture, and then uses a small amount of labeled data from the target household to fine-tune the network parameters. This method solves the problem that the existing neural network is not capable of extracting effective information, resulting in low load decomposition accuracy, and is suitable for situations where the target household lacks labeled data, but has not studied the decomposition of newly added electrical appliances that also lack labeled data. However, in the operation of the power system, the emergence of incremental loads is inevitable. Summary of the invention

[0006] In order to address the deficiencies mentioned in the above background technology, the purpose of the present invention is to provide an incremental non-intrusive decomposition method and system that takes into account the dynamic changes of the load, which is used to decompose the new load appearing in the power grid, improve the reliability and robustness of the current algorithm technology, and ensure the efficient and stable operation of the power system.

[0007] In a first aspect, the object of the present invention can be achieved by the following technical solution: an incremental non-intrusive decomposition method considering dynamic changes of load, the method comprising the following steps:

[0008] Receive a public data set, determine training-related information based on the public data set, obtain power consumption data and total meter data of the device to be trained based on the training-related information, input the power consumption data and total meter data of the device to be trained into a pre-established sequence-to-point model, and train to obtain a separate decomposition model for the device to be trained;

[0009] The equipment to be tested is traversed and tested to obtain the newly added electrical equipment, and the data of the newly added electrical equipment is obtained. The data of the newly added electrical equipment is input into a separate decomposition model of the equipment to be trained using transfer learning technology for fine-tuning to obtain the fine-tuned decomposition model; the total electric meter data is input into the fine-tuned decomposition model to obtain the electricity consumption data of the newly added electrical equipment.

[0010] In combination with the first aspect, in certain implementations of the first aspect, the method further includes: the training-related information in the public data set includes: a set training time, a designated specific family, and selected training equipment.

[0011] In combination with the first aspect, in some implementations of the first aspect, the method further includes: the process of acquiring the power consumption data and the total electric meter data of the device to be trained includes two methods:

[0012] One is to directly extract the main power readings from the public data set within a set time, which includes the total power consumption of all electrical appliances in operation within the home during that period of time. This is real data measured directly from the power inlet.

[0013] The other is to extract the power consumption data of the selected individual training devices from the original data set in sequence and add them all up as the total meter data.

[0014] In combination with the first aspect, in some implementations of the first aspect, the method further includes: the pre-established sequence-to-point model is as follows:

[0015] f seq2point :X t:t+W-1 →y τ

[0016] Where W is the length of the sliding window; X t:t+W-1 represents the input sequence with a window length of W, which is the total meter power sequence; τ is the time midpoint of the sliding window, that is, τ = t + W / 2; y τ represents the power value of the target appliance at the center of the prediction sliding window, f seq2point Represents the mapping relationship between the input sequence and the output sequence.

[0017] In combination with the first aspect, in some implementations of the first aspect, the method further includes: the structure of the separate decomposition model of the to-be-trained device is as follows:

[0018] Input layer: convert the preprocessed data set into a sliding window with a sequence length of 99 and input it into the convolutional neural network;

[0019] First convolution layer: the number of convolution kernels is 30; the convolution kernel size is 10; the activation function is ReLU;

[0020] Second convolution layer: the number of convolution kernels is 30; the convolution kernel size is 8; the activation function is ReLU;

[0021] The third convolution layer: the number of convolution kernels is 40; the convolution kernel size is 6; the activation function is ReLU;

[0022] The fourth convolution layer: the number of convolution kernels is 50; the convolution kernel size is 5; the activation function is ReLU;

[0023] Dropout layer: set the dropout rate to 20%;

[0024] The fifth convolution layer: the number of convolution kernels is 30; the convolution kernel size is 8; the activation function is ReLU;

[0025] Dropout layer: set the dropout rate to 20%;

[0026] Flatten layer: flattens the input data and transitions from the convolutional layer to the fully connected layer;

[0027] Fully connected layer: 1024 neurons are set, and the activation function is ReLU;

[0028] Dropout layer: set the dropout rate to 20%;

[0029] Fully connected layer: Set 1 neuron and output the prediction result.

[0030] In combination with the first aspect, in certain implementations of the first aspect, the method further includes: when the pre-established sequence-to-point model is trained, for each device to be trained, the corresponding electricity consumption data is extracted according to the name of the appliance as the label data of the model, the input data is set to the established total electricity meter data, and the fit function in Keras is called to train the model according to the input data and the label data and update the state, including the weight and bias of the model, to obtain a separate decomposition model for the device to be trained.

[0031] In combination with the first aspect, in some implementations of the first aspect, the method further includes: the process of traversing and detecting the equipment to be trained: first traverse the test equipment to detect whether it has appeared in the training data, and determine whether it is a newly added electrical appliance. If it is not a newly added electrical appliance, the decomposition model trained is matched according to the name of the electrical appliance for testing, and the power consumption data of the electrical appliance is decomposed from the input total electric meter data by calling the predict function in Keras.

[0032] In combination with the first aspect, in some implementations of the first aspect, the method also includes: when it is detected that the test device does not appear in the training process, it is determined to be a new electrical appliance, and a trained model is selected as a pre-training model of the new electrical appliance, and the model is migrated. Since the convolution layer of the convolutional neural network extracts mostly basic features, they are directly applied to the model of the new electrical appliance. For the fully connected layer, the data of the new electrical appliance is used for retraining, and the model is fine-tuned to learn the deep features of the incremental load.

[0033] In a second aspect, in order to achieve the above-mentioned object, the present invention discloses an incremental non-intrusive decomposition system considering dynamic changes of load, comprising:

[0034] A model training module is used to receive a public data set, determine training related information based on the public data set, obtain power consumption data and total meter data of the equipment to be trained based on the training related information, input the power consumption data and total meter data of the equipment to be trained into a pre-established sequence-to-point model, and train to obtain a separate decomposition model for the equipment to be trained;

[0035] The model fine-tuning module is used to traverse and detect the equipment to be tested, obtain the newly added electrical equipment, obtain the data of the newly added electrical equipment, and use the transfer learning technology to input the data of the newly added electrical equipment into the separate decomposition model of the equipment to be trained for fine-tuning to obtain the fine-tuned decomposition model; input the total electric meter data into the fine-tuned decomposition model to obtain the electricity consumption data of the newly added electrical equipment.

[0036] In another aspect of the present invention, in order to achieve the above-mentioned purpose, a terminal device is disclosed, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein the memory stores a computer program capable of running on the processor, and when the processor loads and executes the computer program, an incremental non-intrusive decomposition method that takes into account dynamic changes in the load as described above is adopted.

[0037] Beneficial effects of the present invention:

[0038] The present invention takes into account the impact of complex and changeable user behaviors on load monitoring and decomposition in the actual operation of the power system, and uses transfer learning and fine-tuning technology to construct an incremental non-invasive load decomposition model and system to achieve effective decomposition of new electrical appliances. Compared with traditional non-invasive load decomposition technology, the reliability and robustness of the present invention can still maintain a good level in scenarios where the load changes dynamically; the present invention focuses on the generalization performance of non-invasive load decomposition algorithms between different electrical equipment, so as to perform effective model migration when constructing the decomposition model of new electrical appliances. Most previous algorithm technologies focused on the generalization performance between different families and houses, and there was little research on the transferability of algorithms between different electrical appliances; when decomposing new electrical appliances, the present invention solves the problem of lack of sufficient available labeled data for new loads through model migration technology, and also reduces the dependence of model training on a large amount of labeled data, improves the efficiency and speed of training, and significantly reduces the heavy computational burden. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0040] Figure 1 It is a schematic flow chart of the method of the present invention;

[0041] Figure 2 It is a schematic diagram of the model input and output in the training phase and the testing phase of the present invention;

[0042] Figure 3is a schematic diagram of the present invention using transfer learning technology to establish a new load decomposition model and system;

[0043] Figure 4 It is a schematic diagram of the system structure of the present invention. DETAILED DESCRIPTION

[0044] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0045] Embodiment 1:

[0046] like Figure 1 As shown, an incremental non-intrusive decomposition method considering dynamic changes of loads includes the following steps:

[0047] S101: receiving a public data set, determining training-related information based on the public data set, obtaining power consumption data and total meter data of the device to be trained based on the training-related information, inputting the power consumption data and total meter data of the device to be trained into a pre-established sequence-to-point model, and training to obtain a separate decomposition model for the device to be trained;

[0048] The training-related information in the public data set includes: the set training time, the specific family specified, and the selected training equipment;

[0049] The process of obtaining the power consumption data and total electric meter data of the equipment to be trained includes two methods:

[0050] One is to directly extract the main power readings from the public data set within a set time, which includes the total power consumption of all electrical appliances in operation within the home during that period of time. This is real data measured directly from the power inlet.

[0051] The other is to extract the power consumption data of the selected individual training devices from the original data set in sequence and add them all up as the total meter data;

[0052] The pre-established sequence-to-point model is as follows:

[0053] f seq2point :X t:t+W-1 →y τ

[0054] Where W is the length of the sliding window; X t:t+W-1represents the input sequence with a window length of W, which is the total meter power sequence; τ is the time midpoint of the sliding window, that is, τ = t + W / 2; y τ represents the power value of the target appliance at the center of the prediction sliding window, f seq2point Represents the mapping relationship between the input sequence and the output sequence.

[0055] The structure of the separate decomposition model of the device to be trained is as follows:

[0056] Input layer: convert the preprocessed data set into a sliding window with a sequence length of 99 and input it into the convolutional neural network;

[0057] The first convolution layer: the number of convolution kernels is 30; the convolution kernel size is 10; the activation function is ReLU;

[0058] Second convolution layer: the number of convolution kernels is 30; the convolution kernel size is 8; the activation function is ReLU;

[0059] The third convolution layer: the number of convolution kernels is 40; the convolution kernel size is 6; the activation function is ReLU;

[0060] The fourth convolution layer: the number of convolution kernels is 50; the convolution kernel size is 5; the activation function is ReLU;

[0061] Dropout layer: set the dropout rate to 20%;

[0062] The fifth convolution layer: the number of convolution kernels is 30; the convolution kernel size is 8; the activation function is ReLU;

[0063] Dropout layer: set the dropout rate to 20%;

[0064] Flatten layer: flattens the input data and transitions from the convolutional layer to the fully connected layer;

[0065] Fully connected layer: 1024 neurons are set, and the activation function is ReLU;

[0066] Dropout layer: set the dropout rate to 20%;

[0067] Fully connected layer: Set 1 neuron and output the prediction result.

[0068] When the pre-established sequence-to-point model is trained, for each device to be trained, the corresponding electricity consumption data is extracted according to the name of the appliance as the label data of the model, the input data is set to the established total electricity meter data, and the fit function in Keras is called to train the model according to the input data and label data and update the state, including the weight and bias of the model, to obtain a separate decomposition model for the device to be trained.

[0069] S102: Perform traversal detection on the equipment to be tested to obtain newly added electrical equipment, obtain the data of the newly added electrical equipment, and use transfer learning technology to input the data of the newly added electrical equipment into a separate decomposition model of the equipment to be trained for fine-tuning to obtain a fine-tuned decomposition model; input the total electric meter data into the fine-tuned decomposition model to obtain the electricity consumption data of the newly added electrical equipment.

[0070] The process of traversing and detecting the equipment to be trained is as follows: first, traverse the test equipment to detect whether it has appeared in the training data, and determine whether it is a newly added electrical appliance. If it is not a newly added electrical appliance, match the trained decomposition model according to the name of the electrical appliance for testing, and decompose the electrical appliance's electricity consumption data from the input total electric meter data by calling the predict function in Keras.

[0071] When it is detected that the test equipment does not appear in the training process, it is determined to be a new appliance, and a trained model is selected as the pre-training model of the new appliance to migrate the model. Since the convolutional layer of the convolutional neural network extracts mostly basic features, it can be directly applied to the model of the new appliance. For the fully connected layer, a small amount of data from the new appliance is used for retraining. Through fine-tuning, the model learns the deep features of the incremental load and achieves better decomposition effect.

[0072] In order to more intuitively illustrate the decomposition effect of incremental load, we first conduct a test without incremental load. The selected electrical appliances are all regarded as basic appliances, and sufficient data is used for conventional model training and decomposition testing. The evaluation index obtained is the decomposition of each device by the model and system in the scenario without new appliances, which is called the basic index. Four appliances are selected from the No. 1 household in the REDD dataset for basic testing, and the evaluation index of model decomposition is shown in Table 1.

[0073] Table 1 Model evaluation indicators of basic electrical appliances on the REDD dataset

[0074]

[0075] Electric lights, refrigerators, and dishwashers have achieved good decomposition effects on the model and system proposed in the present invention. Therefore, the power data of these three appliances are used to pre-train the incremental load model to achieve knowledge transfer from the source domain to the target domain and complete the decomposition of the incremental load. First, dishwashers and refrigerators are selected as basic appliances, electric lights as incremental loads, and the decomposition model of dishwashers is used as the pre-training model of incremental loads to conduct decomposition experiments on newly added appliances. The decomposition results are shown in Table 2. Dishwashers and refrigerators are used as basic loads, so their decomposition indicators are almost unchanged compared with the benchmark indicators in Table 1. Focusing on the decomposition indicators of the newly added appliance "electric lights", comparative calculations show that when the total meter data is constructed using the direct extraction method, the average absolute error of the electric lights increases by 4.65%, the root mean square error increases by 1.98%, and the F1 index decreases by 0.26%; when the total meter data is artificially synthesized, the root mean square error increases by 28.33%, and the F1 index decreases by 5.01%. Compared with the benchmark indicators, the decomposition result of the electric light decreases very little, which shows that the present invention successfully completes the decomposition of the incremental load.

[0076] Table 2 Model evaluation indicators of electric lights on the REDD dataset (pre-trained model with dishwasher data)

[0077]

[0078] Still selecting the light as the incremental load, replacing the pre-trained model with the decomposition model of the refrigerator, the results are shown in Table 3. Analysis of the evaluation indicators of the incremental load shows that when the total meter data is constructed by direct extraction, the mean absolute error increases by 3.58%, the root mean square error increases by 7.55%, and the F1 index decreases by 0.52%; when the total meter data is synthesized, the root mean square error increases by 13.26%, and the F1 index decreases by 6.12%. The decomposition effect is almost the same as the benchmark index, so it can also be considered that the model and system designed by the present invention achieve the decomposition of the incremental load.

[0079] Table 3 Model evaluation indicators of electric lights on REDD dataset (pre-trained model with refrigerator data)

[0080]

[0081] Table 4 Model evaluation indicators of dishwashers on REDD dataset (pre-trained model with refrigerator data)

[0082]

[0083] Table 5 Model evaluation indicators of dishwasher on REDD dataset (pre-trained model with electric light data)

[0084]

[0085] Table 6 Evaluation indicators of refrigerator models on REDD dataset (pre-trained model with lamp data)

[0086]

[0087] Table 7 Evaluation indicators of refrigerator models on REDD dataset (pre-trained model with dishwasher data)

[0088]

[0089]

[0090] The dishwasher was selected as the incremental load, and the refrigerator and the light were used as the pre-trained models, respectively, to obtain the experimental results in Tables 4 and 5. The refrigerator was selected as the newly added appliance, and the data of the dishwasher and the light were used to pre-train the decomposition model, respectively. The decomposition results are shown in Tables 6 and 7. Compared with the benchmark indicators, the decomposition results of the newly added appliances did not decrease significantly, which shows that the present invention achieves effective decomposition of the incremental load.

[0091] Embodiment 2: In the second aspect, as Figure 4 As shown, in order to achieve the above-mentioned purpose, the present invention discloses an incremental non-intrusive decomposition system considering dynamic changes of load, comprising:

[0092] The model training module 11 is used to receive a public data set, determine training related information based on the public data set, obtain power consumption data and total meter data of the equipment to be trained based on the training related information, input the power consumption data and total meter data of the equipment to be trained into a pre-established sequence-to-point model, and train to obtain a separate decomposition model of the equipment to be trained;

[0093] The model fine-tuning module 12 is used to traverse and detect the equipment to be tested, obtain the newly added electrical equipment, obtain the data of the newly added electrical equipment, and use the transfer learning technology to input the data of the newly added electrical equipment into a separate decomposition model of the equipment to be trained for fine-tuning to obtain the fine-tuned decomposition model; input the total electric meter data into the fine-tuned decomposition model to obtain the power consumption data of the newly added electrical equipment.

[0094] Based on the same inventive concept, the present invention also provides a computer device, which includes: one or more processors, and a memory for storing one or more computer programs; the program includes program instructions, and the processor is used to execute the program instructions stored in the memory. The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, which is used to implement one or more instructions, specifically used to load and execute one or more instructions in a computer storage medium to implement the above method.

[0095] It needs to be further explained that, based on the same inventive concept, the present invention also provides a computer storage medium, on which a computer program is stored, and the computer program is executed by a processor to execute the above method. The storage medium can adopt any combination of one or more computer-readable media. The computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. The computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electrical, magnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples (non-exhaustive list) of computer-readable storage media include: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present invention, a computer-readable storage medium can be any tangible medium containing or storing a program, which can be used by an instruction execution system, device or device or used in combination with it.

[0096] In the description of this specification, the description with reference to the terms "one embodiment", "example", "specific example", etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present disclosure. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0097] The above shows and describes the basic principles, main features and advantages of the present disclosure. Those skilled in the art should understand that the present disclosure is not limited by the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present disclosure. Without departing from the spirit and scope of the present disclosure, the present disclosure may have various changes and improvements, and these changes and improvements fall within the scope of the present disclosure to be protected.

Claims

1. An incremental non-intrusive decomposition method considering dynamic changes in load, characterized in that: The method comprises the following steps: Receive a public data set, determine training-related information based on the public data set, obtain power consumption data and total meter data of the device to be trained based on the training-related information, input the power consumption data and total meter data of the device to be trained into a pre-established sequence-to-point model, and train to obtain a separate decomposition model for the device to be trained; The equipment to be tested is traversed and tested to obtain the newly added electrical equipment, and the data of the newly added electrical equipment is obtained. The data of the newly added electrical equipment is input into a separate decomposition model of the equipment to be trained using transfer learning technology for fine-tuning to obtain the fine-tuned decomposition model; the total electric meter data is input into the fine-tuned decomposition model to obtain the electricity consumption data of the newly added electrical equipment.

2. The incremental non-intrusive decomposition method considering dynamic load changes according to claim 1, characterized in that: The training-related information in the public data set includes: the set training time, the specified specific family, and the selected training equipment.

3. The incremental non-intrusive decomposition method considering dynamic load changes according to claim 1, characterized in that: The process of obtaining the power consumption data and total electric meter data of the equipment to be trained includes two methods: One is to directly extract the main power readings from the public data set within a set time, which includes the total power consumption of all electrical appliances in operation within the home during that period of time. This is real data measured directly from the power inlet. The other is to extract the power consumption data of the selected individual training devices from the original data set in sequence and add them all up as the total meter data.

4. The incremental non-intrusive decomposition method considering dynamic load changes according to claim 1, characterized in that: The pre-established sequence-to-point model is as follows: f seq2point :X t:t+W-1 →y τ Where W is the length of the sliding window; X t:t+W-1 represents the input sequence with a window length of W, which is the total meter power sequence; τ is the time midpoint of the sliding window, that is, τ = t + W / 2; y τ represents the power value of the target appliance at the center of the prediction sliding window, f seq2point Represents the mapping relationship between the input sequence and the output sequence.

5. The incremental non-intrusive decomposition method considering dynamic load changes according to claim 1, characterized in that: The structure of the separate decomposition model of the device to be trained is as follows: Input layer: convert the preprocessed data set into a sliding window with a sequence length of 99 and input it into the convolutional neural network; First convolution layer: the number of convolution kernels is 30; the convolution kernel size is 10; the activation function is ReLU; Second convolution layer: the number of convolution kernels is 30; the convolution kernel size is 8; the activation function is ReLU; The third convolution layer: the number of convolution kernels is 40; the convolution kernel size is 6; the activation function is ReLU; The fourth convolution layer: the number of convolution kernels is 50; the convolution kernel size is 5; the activation function is ReLU; Dropout layer: set the dropout rate to 20%; The fifth convolution layer: the number of convolution kernels is 30; the convolution kernel size is 8; the activation function is ReLU; Dropout layer: set the dropout rate to 20%; Flatten layer: flattens the input data and transitions from the convolutional layer to the fully connected layer; Fully connected layer: 1024 neurons are set, and the activation function is ReLU; Dropout layer: set the dropout rate to 20%; Fully connected layer: Set 1 neuron and output the prediction result.

6. The incremental non-intrusive decomposition method considering dynamic load changes according to claim 4, characterized in that: When the pre-established sequence-to-point model is trained, for each device to be trained, the corresponding electricity consumption data is extracted according to the name of the appliance as the label data of the model, the input data is set to the established total electricity meter data, and the fit function in Keras is called to train the model according to the input data and label data and update the state, including the weight and bias of the model, to obtain a separate decomposition model for the device to be trained.

7. The incremental non-intrusive decomposition method considering dynamic load changes according to claim 1, characterized in that: The process of traversing and detecting the equipment to be trained is as follows: first, traverse the test equipment to detect whether it has appeared in the training data, and determine whether it is a newly added electrical appliance. If it is not a newly added electrical appliance, match the trained decomposition model according to the name of the electrical appliance for testing, and decompose the electrical appliance's electricity consumption data from the input total electric meter data by calling the predict function in Keras.

8. The incremental non-intrusive decomposition method considering dynamic load changes according to claim 7, characterized in that: When it is detected that the test equipment does not appear in the training process, it is determined to be a new appliance, and a trained model is selected as the pre-training model of the new appliance for model migration. Since the convolutional layer of the convolutional neural network extracts mostly basic features, they are directly applied to the model of the new appliance. For the fully connected layer, the data of the new appliance is used for retraining, and the model is fine-tuned to learn the deep features of the incremental load.

9. An incremental non-intrusive decomposition system considering dynamic changes in load, characterized in that: include: A model training module is used to receive a public data set, determine training related information based on the public data set, obtain power consumption data and total meter data of the equipment to be trained based on the training related information, input the power consumption data and total meter data of the equipment to be trained into a pre-established sequence-to-point model, and train to obtain a separate decomposition model for the equipment to be trained; The model fine-tuning module is used to traverse and detect the equipment to be tested, obtain the newly added electrical equipment, obtain the data of the newly added electrical equipment, and use the transfer learning technology to input the data of the newly added electrical equipment into the separate decomposition model of the equipment to be trained for fine-tuning to obtain the fine-tuned decomposition model; input the total electric meter data into the fine-tuned decomposition model to obtain the electricity consumption data of the newly added electrical equipment.

10. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that: The memory stores a computer program that can be run on the processor. When the processor loads and executes the computer program, an incremental non-intrusive decomposition method that takes into account dynamic changes in loads as described in any one of claims 1 to 8 is adopted.

Images