A method and system for identifying power loads based on transfer learning

Abstract

This invention discloses a power load identification method and system based on transfer learning, mainly relating to the field of power load detection and identification technology. It includes the following steps: deleting abnormal data from the power load dataset to be trained on electrical appliances and from the actual collected power load data, and filling in the missing data in the dataset using a forward data missing value imputation method; then extracting effective electrical appliance load activations; generating a power load decomposition model, evaluating the model's performance based on a test set; and fine-tuning the training using a small amount of actual collected power load data with activated electrical appliance load data. The beneficial effects of this invention are: it provides an efficient and intelligent power load analysis and management solution while achieving accurate decomposition and monitoring of power load.

Description

Technical Field

[0001] The present invention relates to the technical field of power load detection and identification, and in particular to a power load identification method and system based on transfer learning. Background Art

[0002] Machine learning algorithms can automatically learn complex patterns and features from large amounts of data and can effectively identify the load characteristics of different electrical appliances. Compared with traditional physical model-based methods, machine learning can better adapt to changing power consumption environments and improve the accuracy of power load decomposition. Power load decomposition usually requires a large amount of labeled data to train the model, while transfer learning can learn features from a large amount of existing labeled data and migrate them to a new environment by applying the pre-trained model to a new data set.

[0003] There are some existing power load identification algorithms, but they have some shortcomings in deployment and use:

[0004] First, power load has significant time dependence, and it is usually necessary to capture long-term trends and seasonal changes, while traditional methods often have difficulty capturing temporal relationships in long time series;

[0005] Second, although traditional RNN, LSTM, and GRU have certain advantages in processing time series, they are usually only used to generate a single output, such as predicting data for one time step, and cannot effectively handle the situation where the length of the input and output sequences are asymmetric;

[0006] Third, existing non-intrusive power load decomposition systems are usually trained on foreign public data sets. Therefore, effective load decomposition can be achieved on specific data sets or for certain specific types of electrical appliances. However, when faced with new environments or unknown types of electrical appliances, these methods usually lack sufficient adaptability and versatility.

[0007] Therefore, a non-invasive power load identification method and system based on transfer learning is urgently needed to solve the above problems. Summary of the invention

[0008] The purpose of the present invention is to provide a non-invasive power load identification method and system based on transfer learning, which uses a sequence-to-sequence model to decompose the power load, builds a small program backend web service on a cloud server, and interacts with users through the small program to achieve accurate decomposition and monitoring of the power load; by combining edge computing and cloud computing, and using the small program as an interactive platform, an efficient and intelligent power load analysis and management solution is provided.

[0009] In order to achieve the above object, the present invention is implemented through the following technical solutions:

[0010] On the one hand, a method for identifying power load based on transfer learning is provided, comprising the following steps:

[0011] S1: Delete the data in the electric load data set to be trained and the abnormal data in the actually collected electric load data, and fill the missing data in the data set using the forward data missing value filling method;

[0012] S2: Detect the start and stop status of the electrical appliances in step S1 according to the minimum start time, minimum shutdown time and start power threshold of the relevant electrical appliances, and then extract the effective electrical load activation;

[0013] S3: Based on the electrical load activation data obtained in step S2, the data is divided into a training set and a test set in a ratio of 4:1, and input into a sequence-to-sequence algorithm for learning and training to generate a power load decomposition model, and the performance of the model is evaluated based on the test set;

[0014] S4: fine-tune the electric load pre-training model trained in step S3 using a small amount of electrical load activation data from the actual collected electric load data.

[0015] Preferably, the specific implementation process of the forward data missing value filling method in step S1 includes:

[0016] S11: Obtain a power load data set containing missing values, where the data set records the load values ​​of the power system and their corresponding timestamps, and the data set includes several time points: , where the time points with missing load values ​​are marked as ;

[0017] S12: Traverse the power load data set one by one and detect each time point The corresponding load value Is it a missing value? If If is a missing value, , then mark the data point as a value to be supplemented;

[0018] S13: For each missing value detected , perform the following steps:

[0019] Check time point The time point with the most recent non-missing value before ,in and is the effective load value;

[0020] The time point Load value As a time point The filling value is: ;

[0021] If there are missing values ​​at multiple consecutive time points, step S13 is repeated until all missing points are filled.

[0022] Preferably, in step S2, the specific implementation process of extracting effective electrical load activation includes:

[0023] S21: Set is the active power data sequence, where Indicates The active power at a time point is defined as the start power threshold. , and create an all-0 state matrix of the same size as the input data and a binary state array , by taking each time point With the preset threshold Compare and generate a binary array ,in The specific definition of is: ;

[0024] S22: Identify the appliance startup event by analyzing the changes in the state array and close events , start event collection and close event collection They are defined as:

[0025]

[0026] ;

[0027] S23: If the first opening event occurs before the first closing event, the state matrix Set all elements from the start to the first close event in to 1, and remove the open event from the open event set;

[0028] If the last shutdown event occurs after the last startup event, the state matrix also needs to be updated accordingly and the minimum shutdown time threshold is set. ;

[0029] If the time interval between the opening and closing events is less than , then ignore this closing event;

[0030] If the power-on duration is less than , the event is considered invalid. For a valid start event, the element of the corresponding time period in the state matrix is ​​set to 1;

[0031] Output a list of valid start events and the updated state matrix .

[0032] Preferably, the power load decomposition model in step S3 is pre-trained, specifically: the data is standardized so that the mean of the input data tends to 0, including: standardizing the data by a maximum and minimum value standardization method, inputting the data into a sequence-to-sequence algorithm for training, and using an Adam optimizer to update the model weights, wherein the sequence-to-sequence algorithm includes an encoder, a decoder and a fully connected layer.

[0033] Preferably, the encoder includes an input layer, which is a standardized active power segment of an electric load with a feature dimension of 1. The active power value of the active power segment of the electric load is input into the LSTM layer to extract time series features and return its hidden state and cell state. In the decoder, the output state of the encoder is used as input, and the last hidden state of the encoder is copied multiple times through the RepeatVector method to initialize the decoder input sequence; the LSTM layer generates a target sequence according to the encoder state and returns the hidden state of each time step; the fully connected layer is applied to each time step through the output layer to map the output of the LSTM to the actual output dimension.

[0034] Preferably, in the sequence-to-sequence algorithm, an encoder maps an input sequence into a context vector of fixed dimension, and a decoder generates an output sequence according to the context vector.

[0035] Preferably, the specific implementation process of step S4 includes:

[0036] S41: Extract features from the input sequence using the pre-trained encoder and pass these features to the decoder;

[0037] S42: After the LSTM layer of the decoder, add a fully connected layer to adapt to the data characteristics of the current task.

[0038] On the other hand, an identification system of a power load identification method based on transfer learning is provided, comprising: a terminal acquisition module, an edge device module and a cloud platform module;

[0039] The terminal acquisition module is used to: collect overall load data from the circuit bus and upload it to the edge device module;

[0040] The edge device module is used to: receive, decode and analyze data in real time, complete real-time decomposition of power load, and upload the inference results;

[0041] The cloud platform module is used to receive and store data, provide Web services for mini-programs, interact with mini-programs, and support the distribution of model images.

[0042] Preferably, the edge device module includes:

[0043] The data acquisition and processing unit is used to: delete the abnormal data in the power load data set to be trained and the actually collected power load data of the electrical appliances, and fill the missing data in the data set using the forward data missing value filling method; detect the start and stop state of the electrical appliances according to the minimum start time, minimum shutdown time and start power threshold of the relevant electrical appliances, and then extract the effective electrical load activation;

[0044] The model building and evaluation unit is used to: divide the acquired electrical load activation data into a training set and a test set in a ratio of 4:1, input the sequence-to-sequence algorithm for learning and training, generate a power load decomposition model, and evaluate the performance of the model based on the test set;

[0045] The model training and data output unit is used to fine-tune the pre-trained power load model using a small amount of electrical load activation data from the actually collected power load data.

[0046] Compared with the prior art, the beneficial effects of the present invention are:

[0047] 1. Improved accuracy: This method has significant advantages over traditional methods. This method is better at processing time series data, can capture the long-term and short-term dependencies of power load signals, and adapt to complex load change trends. Secondly, through its deep neural network structure, it has strong nonlinear modeling capabilities, can effectively fit the complex relationships between power equipment, and can maintain good generalization capabilities when power load data is incomplete or erroneous.

[0048] 2. Strong transferability and generalization performance: Using the transfer learning method, the actual collected power load data is applied to the model trained based on the public power load data set. By fine-tuning the model, the adaptability and generalization ability of the model on the new data set are further optimized. This can not only improve the performance of the model when dealing with new environments and different data distributions, but also effectively reduce the training cost of the model, reduce the computing resources and time required for training from scratch, and significantly improve the efficiency of model deployment.

[0049] 3. Cloud-edge collaboration: This system combines the powerful data processing capabilities of cloud computing with the real-time data analysis advantages of edge computing to distribute the power load decomposition tasks to edge devices, which helps to reduce the burden on the cloud platform, reduce dependence on network resources, and improve the overall stability of the system. At the same time, the model image distribution function provided by the cloud platform makes model updates and maintenance more convenient, and can quickly deploy new models or upgrade existing models, so as to better cope with the needs of continuous iteration of electrical equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 is a flow chart of the method of the present invention;

[0051] Figure 2 It is a flow chart of pre-training of the power load classification model of the present invention;

[0052] Figure 3 It is a schematic diagram of the system framework of the present invention;

[0053] Figure 4 It is a client interaction schematic diagram of the present invention;

[0054] Figure 5 It is a flow chart of the power load decomposition cloud-edge-end collaborative architecture of the present invention;

[0055] Figure 6 It is a flow chart of edge device data processing of the present invention. DETAILED DESCRIPTION

[0056] The present invention will be further described below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms fall within the scope limited by the application equally.

[0057] Example:

[0058] like Figure 1 As shown, this embodiment provides a non-intrusive power load identification method based on transfer learning, comprising the following steps:

[0059] S1: Delete the data in the electric load data set to be trained and the abnormal data in the actually collected electric load data, and fill the missing data in the data set using the forward data missing value filling method;

[0060] S2: Detect the start and stop status of the electrical appliances in step S1 according to the minimum start time, minimum shutdown time and start power threshold of the relevant electrical appliances, and then extract the effective electrical load activation;

[0061] Since different power loads have different activation durations and activity levels, directly using the original load data for training may result in an imbalance in the ratio of data in the activation period to the inactive period in the training samples. In this way, the model may be more inclined to learn the data features in the inactive state, making it difficult to accurately identify different types of power loads, resulting in a decrease in the accuracy of power load decomposition;

[0062] S3: Based on the electrical load activation data obtained in step S2, the data is divided into a training set and a test set in a ratio of 4:1, and input into a sequence-to-sequence algorithm for learning and training to generate a power load decomposition model, and the performance of the model is evaluated based on the test set;

[0063] S4: fine-tuning the electric load pre-training model trained in step S3 with a small amount of electrical load activation data from the actual collected electric load data to enhance its generalization ability on a specific data set;

[0064] Due to the different brands and models of electrical appliances and people's daily electricity usage habits, the characteristics of power loads in different regions may vary significantly. This diversity can lead to reduced accuracy when the data provided by local power companies is input into the model trained with public datasets. Therefore, transfer learning is used to enhance the applicability and accuracy of the model in new areas.

[0065] Among them, in step S1, the specific implementation process of the forward data missing value filling method is:

[0066] Step S11: Obtain a power load data set containing missing values. The data set records the load values ​​of the power system and their corresponding timestamps. The data set may contain multiple time points. , where the load values ​​at certain time points are There may be missing items, marked as " ”;

[0067] Step S12: Traverse the power load data set one by one and detect the The corresponding load value Is it a missing value? If is a missing value, that is , then mark the data point as a value to be supplemented;

[0068] Step S13: For each missing value detected , perform the following steps:

[0069] Check time point The time point with the most recent non-missing value before ,in and is the effective load value;

[0070] The time point Load value As a time point The filling value is: ;

[0071] If there are missing values ​​for multiple consecutive time points, repeat this step until all missing points are filled.

[0072] In step S2, the power load start and stop point detection is performed to extract load activation, and the specific implementation process is as follows:

[0073] Step S21: Set is the active power data sequence, Indicates The active power at a time point is defined as the start power threshold. , and create an all-0 state matrix of the same size as the input data and a binary state array First, by With the preset threshold Compare and generate a binary array ,in The specific definition of is: ;

[0074] Step S22: Identify the start event of the appliance by analyzing the changes in the state array and close events , start event collection and close event collection They are defined as:

[0075]

[0076] ;

[0077] Step S23: If the first opening event occurs before the first closing event, the state matrix All elements from the beginning to the first closing event are set to 1, and the opening event is removed from the opening event set. If the last closing event occurs after the last opening event, the state matrix also needs to be updated accordingly. In order to avoid accidental shutdown of the appliance, a minimum shutdown time threshold is set. If the time interval between the on and off events is less than the threshold, the off event is ignored. If the duration of the on time is less than the preset minimum on time , the event is considered invalid. For valid start events, the elements of the corresponding time period in the state matrix are set to 1. Finally, a list containing valid start events and the updated state matrix is ​​output. .

[0078] In step S3, the power load decomposition model is pre-trained, and the specific implementation process is as follows:

[0079] In order to improve the learning efficiency of deep learning algorithms and avoid the interference of the dimension and value range of indicators on experimental results, data is usually standardized. Then the mean of the input data tends to 0. First, the data is standardized by the maximum and minimum value normalization method, and the data is input into the sequence-to-sequence algorithm for training. The sequence-to-sequence algorithm includes an encoder, a decoder, and a fully connected layer. The loss function is the mean square error (MSE), and the Adam optimizer is used to update the model weights.

[0080] In the sequence-to-sequence algorithm, the input layer in the encoder is a segment of power load with a feature dimension of 1, which is then input into the LSTM layer to extract time series features and return its hidden state and cell state. In the decoder, the output state of the encoder is used as input. First, the last hidden state of the encoder (a vector) is copied multiple times through the RepeatVector method to initialize the decoder input sequence. The subsequent LSTM layer generates the target sequence based on the encoder state and returns the hidden state of each time step. Finally, the output layer applies a fully connected layer to each time step to map the output of the LSTM to the actual output dimension.

[0081] In the sequence-to-sequence algorithm, a decoder and an encoder are introduced. The encoder maps the input sequence into a context vector of a fixed dimension, and the decoder generates an output sequence based on this context vector.

[0082] In the encoder, for the input sequence , the encoder uses LSTM to process the input at each time step , and update the hidden state and cell state;

[0083] LSTM contains input gate and forget gate, and its calculation formula is as follows:

[0084] Encoder:

[0085] Input Gate:

[0086]

[0087] in, The input gate controls whether the current input is allowed Stored in LSTM unit, , is the weight matrix of input and hidden state, is the bias of the input gate;

[0088] Forget Gate:

[0089]

[0090] in, The forget gate controls how much information from the previous time step is discarded , , is the weight matrix of input and hidden state, is the bias of the forget gate;

[0091] Cell status update:

[0092]

[0093]

[0094] in, is the current candidate memory cell state, is the updated cell state, which combines the previous state and the currently entered information;

[0095] Hide status updates:

[0096]

[0097] in, The updated hidden state (i.e. LSTM output) is passed to the next time step and the encoder;

[0098] Decoder:

[0099] The decoder at each time step Generate Output , which is calculated in the same way as the LSTM of the encoder, except that at each time step, the decoder accepts the output of the previous time step As input, rather than being given directly from the dataset;

[0100] The initial hidden state and cell state are passed by the encoder, i.e. and At each decoding time step , the decoder’s LSTM is updated as follows:

[0101]

[0102] in, are the hidden state and cell state at the previous time step, is the output of the previous time step.

[0103] In step S4, fine-tuning transfer learning is performed based on the method in step S3, and the specific implementation method is as follows:

[0104] First, the training data is the actual collected power load data. First, in step 3, the pre-trained sequence-to-sequence algorithm model is loaded into the transfer learning algorithm. The specific implementation method is as follows:

[0105] Use the pre-trained encoder to extract features from the input sequence and pass these features to the decoder. In order to retain the feature representations learned in the pre-trained model, freeze all layers in the encoder, that is, keep the weights of these layers unchanged to ensure that the features they extract are not disturbed or overwritten by new data;

[0106] Fine-tune the decoder. After the LSTM layer of the decoder, add a fully connected layer, which will be adjusted according to the requirements of the new task to better adapt to the data characteristics of the current task, such as Figure 2 As shown in the figure, it can help the model to more accurately capture the temporal dependencies and change trends in power load data.

[0107] like Figure 3 As shown, this embodiment also proposes an identification system of a non-invasive power load identification method based on the above transfer learning, including:

[0108] The overall architecture consists of a terminal acquisition module, an edge device module, and a cloud platform module. The terminal acquisition module is responsible for collecting overall load data from the circuit bus and uploading it to the edge device. The edge device module is responsible for receiving, decoding, and analyzing data in real time, completing real-time decomposition of the power load, and uploading the inference results. The cloud platform module is responsible for receiving and storing data, providing Web services for the mini-program, and interacting with the mini-program, such as Figure 4 As shown, it also supports functions such as sending model images.

[0109] The specific implementation process of the power load decomposition cloud-edge-end collaborative architecture is as follows: Figure 5 As shown, specifically:

[0110] Create UDP server: The code first starts the UDP server by creating a socket object, which is bound to the local host and port 2049, ready to receive data from the acquisition device;

[0111] Establish database connection: The code establishes a connection with the database through SQLite3 to store and manage processed data;

[0112] Receive and process data: The server keeps listening in an infinite loop to ensure that it can continuously receive data. When receiving data from the client, it first parses it into JSON format. It decodes the received Base64-encoded data and converts it into a readable string.

[0113] Extracting key information: From the decoded data, the code extracts key fields such as device ID, timestamp, and active power. This process involves parsing specific byte sequences and converting them into corresponding data types (e.g., converting timestamp and active power from byte data to integers).

[0114] Data storage: The extracted key information is stored in the corresponding tables in the database through SQL insert commands, and these commands are executed to ensure that the data is successfully written;

[0115] Database operations and connection closing: After each data processing is completed, the code will commit the changes to the database and close the currently used database cursor. After all data processing is completed, the server closes the database connection, such as Figure 6 As shown;

[0116] The specific implementation of the power load real-time reasoning program is as follows:

[0117] Get the original data from the corresponding appliance data table in the database for filling. For example, if the window size is set to 160 data points, then in the first decomposition, the first 159 cells will be filled with historical data, and the latest data point will be added to the 160th cell. This inference program realizes real-time inference of data from multiple power load collection devices. Through Python's threading library, multiple threads are created and managed, and each thread is responsible for processing the data flow of a device.

[0118] The multi-threaded architecture enables the system to process data streams from multiple devices in parallel, greatly improving the efficiency and response speed of data processing. In order to ensure the stability and performance of the system, each thread is configured to read the latest data of the corresponding device from the database every 1 second. This not only ensures the real-time nature of data processing, but also avoids the resource consumption and system performance degradation that may be caused by frequent access to the database.

[0119] The edge device module includes:

[0120] The data acquisition and processing unit is used to: delete the abnormal data in the power load data set to be trained and the actually collected power load data of the electrical appliances, and fill the missing data in the data set using the forward data missing value filling method; detect the start and stop state of the electrical appliances according to the minimum start time, minimum shutdown time and start power threshold of the relevant electrical appliances, and then extract the effective electrical load activation;

[0121] The model building and evaluation unit is used to: divide the acquired electrical load activation data into a training set and a test set in a ratio of 4:1, input the sequence-to-sequence algorithm for learning and training, generate a power load decomposition model, and evaluate the performance of the model based on the test set;

[0122] The model training and data output unit is used to fine-tune the pre-trained power load model using a small amount of electrical load activation data from the actually collected power load data.

[0123] The above is a specific description of the preferred implementation of the present invention, but the present invention is not limited to the described embodiments. Those skilled in the art may make various equivalent modifications or substitutions without violating the spirit of the present invention. These equivalent modifications or substitutions are all included in the scope defined by the claims of this application.

Claims

1. A method for identifying power loads based on transfer learning, characterized in that: The following steps are involved: S1: Delete the data in the electric load data set to be trained and the abnormal data in the electric load data actually collected, and fill the missing data in the data set using a forward data missing value filling method, wherein the forward data missing value filling method comprises: S11: Obtain a power load data set containing missing values, where the data set records the load values ​​of the power system and their corresponding timestamps, and the data set includes several time points: , where the time points with missing load values ​​are marked as ; S12: Traverse the power load data set one by one and detect each time point The corresponding load value Is it a missing value? If If is a missing value, , then mark the data point as a value to be supplemented; S13: For each missing value detected , perform the following steps: Check time point The time point with the most recent non-missing value before ,in and is the effective load value; The time point Load value As a time point The filling value is: ; If there are missing values ​​at multiple consecutive time points, repeat step S13 until all missing points are filled; S2: Detect the start and stop status of the electrical appliances in step S1 according to the minimum start time, minimum shutdown time and start power threshold of the relevant electrical appliances, and then extract the effective electrical load activation; S3: Based on the electrical load activation data obtained in step S2, the data is divided into a training set and a test set in a ratio of 4:1, and input into a sequence-to-sequence algorithm for learning and training to generate a power load decomposition model, and the performance of the model is evaluated based on the test set; S4: fine-tuning the electric load pre-training model trained in step S3 using a small amount of electric load activation data from the actually collected electric load data, wherein the training data for the fine-tuning training is the actually collected electric load data, and the fine-tuning training includes: In step S3, the pre-trained sequence-to-sequence algorithm model is loaded into the transfer learning algorithm. The specific implementation method is as follows: Use the pre-trained encoder to extract features from the input sequence and pass these features to the decoder; in order to retain the feature representations learned in the pre-trained model, freeze all layers in the encoder, that is, keep the weights of these layers unchanged to ensure that the features they extract are not interfered with or covered by new data; Fine-tune the decoder part: After the LSTM layer of the decoder, add a fully connected layer, which will be adjusted according to the requirements of the new task to better adapt to the data characteristics of the current task.

2. The method for identifying power load based on transfer learning according to claim 1, characterized in that: In step S2, the specific implementation process of extracting effective electrical load activation includes: S21: Set is the active power data sequence, where Indicates The active power at a time point is defined as the start power threshold. , and create an all-0 state matrix of the same size as the input data and a binary state array , by taking each time point With the preset threshold Compare and generate a binary array ,in The specific definition of is: ; S22: Identify the appliance startup event by analyzing the changes in the state array and close events , start event collection and close event collection They are defined as: ； S23: If the first opening event occurs before the first closing event, the state matrix Set all elements from the start to the first close event in to 1, and remove the open event from the open event set; If the last shutdown event occurs after the last startup event, the state matrix also needs to be updated accordingly and the minimum shutdown time threshold is set. ; If the time interval between the opening and closing events is less than , then ignore this closing event; If the power-on duration is less than , the event is considered invalid. For a valid start event, the element of the corresponding time period in the state matrix is ​​set to 1; Output a list of valid start events and the updated state matrix .

3. The method for identifying power loads based on transfer learning according to claim 2, characterized in that: The power load decomposition model in step S3 is pre-trained, specifically: the data is standardized so that the mean of the input data tends to 0, including: the data is standardized by the maximum and minimum value normalization method, the data is input into the sequence-to-sequence algorithm for training, and the Adam optimizer is used to update the model weights. The sequence-to-sequence algorithm includes an encoder, a decoder and a fully connected layer.

4. The method for identifying power loads based on transfer learning according to claim 3 is characterized in that: The encoder includes an input layer, which is a normalized active power segment of an electric load with a feature dimension of 1. The active power value of the active power segment of the electric load is input into the LSTM layer to extract time series features and return its hidden state and cell state. In the decoder, the output state of the encoder is used as input, and the last hidden state of the encoder is copied multiple times by the RepeatVector method to initialize the decoder input sequence; the LSTM layer generates a target sequence according to the encoder state and returns the hidden state of each time step; A fully connected layer is applied to each time step through the output layer to map the output of the LSTM to the actual output dimension.

5. The method for identifying power load based on transfer learning according to claim 4, characterized in that: In the sequence-to-sequence algorithm, the encoder maps the input sequence into a context vector of fixed dimension, and the decoder generates the output sequence according to the context vector.

6. An identification system based on the power load identification method based on transfer learning as claimed in claim 1, characterized in that: include: Terminal acquisition module, edge device module and cloud platform module; The terminal acquisition module is used to: collect overall load data from the circuit bus and upload it to the edge device module; The edge device module is used to: receive, decode and analyze data in real time, complete real-time decomposition of power load, and upload the inference results; The cloud platform module is used to receive and store data, provide Web services for mini-programs, interact with mini-programs, and support the distribution of model images.

7. The identification system of the power load identification method based on transfer learning according to claim 6 is characterized in that: The edge device module includes: The data acquisition and processing unit is used to: delete the abnormal data in the power load data set to be trained and the actually collected power load data of the electrical appliances, and fill the missing data in the data set using the forward data missing value filling method; detect the start and stop state of the electrical appliances according to the minimum start time, minimum shutdown time and start power threshold of the relevant electrical appliances, and then extract the effective electrical load activation; The model building and evaluation unit is used to: divide the acquired electrical load activation data into a training set and a test set in a ratio of 4:1, input the sequence-to-sequence algorithm for learning and training, generate a power load decomposition model, and evaluate the performance of the model based on the test set; The model training and data output unit is used to fine-tune the pre-trained power load model using a small amount of electrical load activation data from the actually collected power load data.

Images