Non-intrusive load monitoring method and device, computer equipment and storage medium

Abstract

The invention relates to a non-intrusive load monitoring method and device, computer equipment and a storage medium. The method comprises the steps of extracting an active power sequence and a reactive power sequence from ammeter data, taking the active power sequence and the reactive power sequence as input, calling a trained hybrid non-intrusive load monitoring model, and obtaining a state detection result and an energy injection sequence prediction result of electric equipment, when the hybrid non-intrusive load monitoring model is called, the following steps are executed: extracting and splicing active power features in an active power sequence and reactive power features in a reactive power sequence to obtain a comprehensive feature sequence, capturing a long-term dependency relationship in the comprehensive feature sequence to obtain a multi-scale feature sequence, and calculating the load of the hybrid non-intrusive load monitoring model based on the multi-scale feature sequence. And respectively predicting the state of the electric equipment and the energy injection of the micro-inverter to obtain a state detection result and an energy injection sequence prediction result of the electric equipment. By adopting the method, the power consumption condition of the electric equipment can be accurately monitored.

Description

Technical Field

[0001] The present application relates to the technical field of load monitoring, and in particular to a non-intrusive load monitoring method, apparatus, computer equipment, computer-readable storage medium, and computer program product. Background Art

[0002] Non-Intrusive Load Monitoring (NILM) technology is a method to identify and estimate the power consumption of each electrical device by analyzing the total power meter data. The traditional NILM task is to identify the power consumption of each appliance based on the total power data. From the perspective of signal processing, it belongs to the blind source separation problem. Blind Source Separation (BSS) refers to the process of separating the original source signal from the observed mixed signal only in the case of unknown source signals and mixed processes.

[0003] With the popularization of distributed energy, the electricity consumption of modern households has become more complicated, which undoubtedly increases the difficulty of signal separation. Therefore, for modern smart home environments that include distributed energy, the current NILM task is difficult to accurately monitor the electricity consumption of electrical equipment. Summary of the invention

[0004] Based on this, it is necessary to provide a non-intrusive load monitoring method, device, computer equipment, computer-readable storage medium and computer program product that can accurately monitor the power consumption of electrical equipment in response to the above technical problems.

[0005] In a first aspect, the present application provides a non-intrusive load monitoring method, comprising:

[0006] Extract active power series and reactive power series from electric meter data;

[0007] Taking the active power sequence and the reactive power sequence as input, calling the trained hybrid non-intrusive load monitoring model to obtain the state detection result of the electrical equipment and the energy injection sequence prediction result;

[0008] The hybrid non-intrusive load monitoring model performs the following steps when invoked:

[0009] Extracting active power features in the active power sequence and reactive power features in the reactive power sequence, and splicing the active power features and the reactive power features to obtain a comprehensive feature sequence;

[0010] Capturing the long-term dependency in the comprehensive feature sequence to obtain a multi-scale feature sequence;

[0011] Based on the multi-scale feature sequence, the state of the electrical equipment and the energy injection of the micro-inverter are predicted respectively, so as to obtain the state detection result of the electrical equipment and the energy injection sequence prediction result;

[0012] The hybrid non-intrusive load monitoring model is trained based on historical active power sequences and historical reactive power sequences.

[0013] In one of the embodiments, the hybrid non-intrusive load monitoring model includes an encoder, a converter model, a sequence-to-point module and a sequence-to-sequence module.

[0014] In one embodiment, the encoder includes a first convolutional encoder and a second convolutional encoder; the extracting the active power feature in the active power sequence and the reactive power feature in the reactive power sequence includes:

[0015] Calling a first convolution encoder to perform one-dimensional convolution on the active power sequence to extract active power features in the active power sequence;

[0016] The second convolution encoder is called to perform one-dimensional convolution on the reactive power sequence to extract reactive power features from the reactive power sequence features.

[0017] In one embodiment, capturing the long-term dependency in the comprehensive feature sequence to obtain a multi-scale feature sequence includes:

[0018] The comprehensive feature sequence is input into the converter model, and the converter model captures the long-term dependencies in the comprehensive feature sequence through a multi-head attention mechanism to obtain a multi-scale feature sequence.

[0019] In one embodiment, the state of the electric device and the energy injection of the micro-inverter are predicted based on the multi-scale feature sequence to obtain the state detection result of the electric device and the energy injection sequence prediction result, including:

[0020] The multi-scale feature sequence output by the converter model is input into the sequence-to-point module, and the sequence-to-point module maps the last time step of the multi-scale feature sequence to the power device state prediction task through a fully connected layer to obtain a state detection result of the power device;

[0021] The multi-scale feature sequence output by the converter model is input into the sequence-to-sequence module, and the sequence-to-sequence module maps the multi-scale features of each time step in the multi-scale feature sequence to the energy injection prediction task through a fully connected layer to obtain an energy injection sequence prediction result.

[0022] In one embodiment, after obtaining the state detection result of the electric device and the energy injection sequence prediction result, the method further includes:

[0023] Normalizing the energy injection sequence prediction result to obtain a normalized energy injection sequence prediction result;

[0024] Adaptively filter the normalized energy injection sequence prediction result, map the normalized energy injection sequence prediction result to a predefined power output gear, determine the output power level of the microinverter, and the output gear corresponds to the output power level one by one.

[0025] In one embodiment, extracting the active power sequence and the reactive power sequence from the electric meter data includes:

[0026] Preprocess the meter data;

[0027] Active power sequences and reactive power sequences with fixed lengths are extracted from the preprocessed meter data using a sliding window method.

[0028] In a second aspect, the present application also provides a non-intrusive load monitoring device, comprising:

[0029] A data extraction module is used to extract active power sequence and reactive power sequence from the electric meter data;

[0030] A data processing module, used to take the active power sequence and the reactive power sequence as input, call the trained hybrid non-intrusive load monitoring model, and obtain the state detection result of the electrical equipment and the energy injection sequence prediction result;

[0031] The hybrid non-intrusive load monitoring model performs the following steps when invoked:

[0032] Extracting active power features in the active power sequence and reactive power features in the reactive power sequence, and splicing the active power features and the reactive power features to obtain a comprehensive feature sequence;

[0033] Capturing the long-term dependency in the comprehensive feature sequence to obtain a multi-scale feature sequence;

[0034] Based on the multi-scale feature sequence, the state of the electrical equipment and the energy injection of the micro-inverter are predicted respectively, so as to obtain the state detection result of the electrical equipment and the energy injection sequence prediction result;

[0035] The hybrid non-intrusive load monitoring model is trained based on historical active power sequences and historical reactive power sequences.

[0036] In a third aspect, the present application further provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps in any one of the above-mentioned non-intrusive load monitoring method embodiments when executing the computer program.

[0037] In a fourth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps in any one of the above-mentioned non-intrusive load monitoring method embodiments.

[0038] In a fifth aspect, the present application further provides a computer program product, including a computer program, which, when executed by a processor, implements the steps in any one of the above-mentioned non-intrusive load monitoring method embodiments.

[0039] The above-mentioned non-intrusive load monitoring method, device, computer equipment, computer-readable storage medium and computer program product are pre-trained with a hybrid non-intrusive load monitoring model. The hybrid non-intrusive load monitoring model has a multi-task learning framework and can simultaneously perform device status prediction and energy injection estimation. In practical applications, the data of two dimensions, active power and reactive power, are considered at the same time, and the active power sequence and reactive power sequence are extracted from the meter data, so that the extracted data can comprehensively reflect the working status of the electrical appliance and the operation status of the power system. Subsequently, the active power sequence and reactive power sequence are input into the trained hybrid non-intrusive load monitoring model, and the model converts the features in the active power sequence and The features in the reactive power sequence are spliced ​​together to obtain a comprehensive feature sequence. Then, the long-term dependencies in the comprehensive feature sequence are captured, and complex time dependencies are modeled to obtain a multi-scale feature sequence. Finally, based on the multi-scale feature sequence, the state of the electrical equipment and the energy injection of the micro-inverter are predicted simultaneously to obtain the state detection results of the electrical equipment and the energy injection sequence prediction results. The entire solution only needs to collect meter data to identify the state of the electrical equipment and estimate the energy injection, so as to obtain comprehensive and accurate monitoring results. It can be applied to modern smart home environments that include distributed energy. In addition, the above solution can be stable and robust on data from different households, different time periods, and different types of electrical appliances. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the drawings required for use in the embodiments of the present application or related technical descriptions will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0041] Figure 1is an application environment diagram of a non-intrusive load monitoring method in one embodiment;

[0042] Figure 2 is a flow chart of a non-intrusive load monitoring method in one embodiment;

[0043] Figure 3 is a schematic diagram of a data processing process involved in a model in an embodiment;

[0044] Figure 4 is a schematic diagram of a data processing process involved in a model in another embodiment;

[0045] Figure 5 is a flow chart of a non-intrusive load monitoring method in another embodiment;

[0046] Figure 6 is a structural block diagram of a non-intrusive load monitoring device in one embodiment;

[0047] Figure 7 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0048] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0049] The non-intrusive load monitoring method provided in the embodiment of the present application can be applied to Figure 1 In the application environment shown in the figure, the electric meter 100 and the terminal 102 communicate with the server 104 through the network. The data storage system can store the data that the server 104 needs to process. The data storage system can be integrated on the server 104, or placed on the cloud or other network servers. A hybrid non-intrusive load monitoring model is deployed on the server 104.

[0050] Specifically, the electric meter 100 transmits the collected meter data to the server 104. When there is a demand for load monitoring, the operator sends a load monitoring message to the server 104 through the terminal 102. The server 104 responds to the message, extracts the active power sequence and the reactive power sequence from the electric meter data, and uses the active power sequence and the reactive power sequence as input to call the trained hybrid non-intrusive load monitoring model. The hybrid non-intrusive load monitoring model extracts the active power features in the active power sequence and the reactive power features in the reactive power sequence, concatenates the active power features and the reactive power features to obtain a comprehensive feature sequence, captures the long-term dependency in the comprehensive feature sequence, and obtains a multi-scale feature sequence. Based on the multi-scale feature sequence, the state of the electrical equipment and the energy injection of the micro-inverter are predicted respectively, and the state detection result of the electrical equipment and the energy injection sequence prediction result are obtained.

[0051] Among them, the electric meter 100 can be but not limited to a smart electric meter, and the terminal 102 can be but not limited to various personal computers, laptops, smart phones, tablet computers, Internet of Things devices and portable wearable devices. The Internet of Things devices can be smart speakers, smart TVs, smart air conditioners, smart car devices, projection devices, etc. Portable wearable devices can be smart watches, smart bracelets, head-mounted devices, etc. The head-mounted devices can be virtual reality (VR) devices, augmented reality (AR) devices, smart glasses, etc. The server 104 can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides cloud computing services.

[0052] In an exemplary embodiment, Figure 2 As shown, a non-intrusive load monitoring method is provided, which is applied to Figure 1 The server 104 in the example is used for explanation, and the following steps (hereinafter referred to as S) 200 to S400 are included. Among them:

[0053] S200, extracting active power sequence and reactive power sequence from the electric meter data.

[0054] Meter data refers to the power consumption information recorded by smart meters (including main meters and sub-meters), usually including parameters such as voltage, current, and frequency. For non-intrusive load monitoring, the focus is on active power and reactive power. The active power series represents the actual electrical energy consumed by electrical equipment over a period of time, in kilowatt-hours (kWh) or watts (W), reflecting the part of energy converted into useful work, while the reactive power series represents the part used to establish a magnetic field or other forms of energy storage but does not do work.

[0055] In actual applications, taking the home scene as an example, the user's home is equipped with a smart meter, which can measure and record the total power consumption in the home (i.e., meter data) in real time, including active power and reactive power, and send the recorded data to the server through the network. The user sends a load monitoring message to the server through the terminal, and the server can perform data preprocessing on the meter data, including cleaning, missing value and outlier processing, and data alignment. Subsequently, active power and reactive power are extracted from the preprocessed meter data to obtain active power sequence and reactive power sequence.

[0056] S400, taking active power sequence and reactive power sequence as input, calling the trained hybrid non-intrusive load monitoring model to obtain the state detection results of the electrical equipment and the energy injection sequence prediction results, wherein the hybrid non-intrusive load monitoring model is trained based on historical active power sequence and historical reactive power sequence.

[0057] The state detection result of the electrical equipment refers to the predicted on, off or running state of each electrical equipment. The energy injection sequence prediction result refers to the predicted amount of energy injection from general energy injection or renewable energy over a period of time, such as the energy injected by solar panels through micro inverters.

[0058] The hybrid non-intrusive load monitoring model (HybridNILM) is a deep learning model trained based on a large amount of historical active power and reactive power data. It aims to identify the working status of individual electrical appliances and predict energy injection by analyzing active power and reactive power.

[0059] In practical applications, model training methods include training from scratch and fine-tuning based on pre-trained models. Specifically, the learning rate adjustment strategy can adopt a cosine annealing learning rate scheduling strategy with hot restart, which converges quickly in the early stage of training and jumps out of the local optimum through periodic restart in the later stage. Specifically, in the model training stage, a large learning rate is used in the initial stage, generally using the default learning rate of 0.001 of the Adam optimizer, and then manually adjusted according to the training loss and performance results of a few epochs. In the second half of the training or after the loss function jumps no longer decreases, a small learning rate, a large training cycle epoch, and the storage of training evaluation indicators for each step are used, and the optimal checkpoint model is selected as the model weight according to the performance indicator. In the model fine-tuning stage, the pre-trained model weights are loaded, some layers are frozen, and only specific layers (decoder and output layer) are trained. In addition, a smaller learning rate is used for fine-tuning, generally a learning rate of about 0.0001 is used, and it is adjusted according to the actual situation. Finally, we found that around epoch 30, both the test and training performance tended to be stable, indicating that the model parameters have converged. At this time, we can select the saved optimal checkpoint as the model weight, or further reduce the learning rate to observe the changes in model performance. If the model's test performance index decreases after reducing the learning rate, it means overfitting. Through the above training and fine-tuning strategies, the HybridNILM model can effectively learn complex appliance usage patterns and energy injection patterns, and predict the state of the appliance.

[0060] After the hybrid non-intrusive load monitoring model is trained, it is deployed in the server. After the server extracts the active power sequence and the reactive power sequence, it inputs the active power sequence and the reactive power sequence into the trained hybrid non-intrusive load monitoring model (hereinafter referred to as the model), and the model predicts the state of each electrical device and the energy injection situation, and obtains the state detection results of the electrical device and the energy injection sequence prediction results.

[0061] like Figure 3 Specifically, when the hybrid non-intrusive load monitoring model is called, the following steps are performed:

[0062] S420, extracting active power features in the active power sequence and reactive power features in the reactive power sequence, and concatenating the active power features and the reactive power features to obtain a comprehensive feature sequence.

[0063] Active power features refer to specific attributes extracted from active power sequences that can reflect the working mode or state of electrical appliances, such as instantaneous power changes, average power levels, power fluctuation frequency, etc. Reactive power features refer to patterns extracted from reactive power sequences that can reveal electromagnetic characteristics related to power systems, such as starting current, harmonic components, etc. Comprehensive feature sequences refer to combining active power features and reactive power features to form a new feature vector or sequence that can more comprehensively describe the power consumption at each point in time.

[0064] In specific implementation, the model can extract the active power features in the active power sequence and the reactive power features in the reactive power sequence through the encoder, and then combine the extracted active power features and reactive power features one by one in chronological order to ensure that each pair of features corresponds to the data at the same time point, and obtain a new sequence containing all features (comprehensive feature sequence).

[0065] S440, captures the long-term dependencies in the comprehensive feature sequence and obtains a multi-scale feature sequence.

[0066] Long-term dependencies refer to the correlations between different time points over a long time span, especially those that span a long time interval but still maintain significant correlation. Multi-scale feature sequences are a series of feature representations generated by capturing patterns and trends at different time scales, which are used to adapt to the identification of different types of electrical appliances and electricity usage behaviors.

[0067] In specific implementation, the model can divide the concatenated feature sequence into time windows of different lengths, call the corresponding models used to capture long-term dependencies such as long short-term memory networks or converter models, process the data in each time window, learn how to identify important patterns at different time scales, and integrate the feature representations from different time windows to form a multi-scale feature sequence.

[0068] S460, based on the multi-scale feature sequence, respectively predict the state of the electrical equipment and the energy injection of the micro-inverter to obtain the state detection result of the electrical equipment and the energy injection sequence prediction result.

[0069] After obtaining the multi-scale feature sequence, the model can use the multi-scale feature sequence to infer which electrical appliances are currently in use and their operating status, such as on / off status, and predict the amount of power that distributed generation systems (such as solar panels) will provide to the power grid or local loads in the future based on the multi-scale feature sequence.

[0070] The above non-intrusive load monitoring method is pre-trained with a hybrid non-intrusive load monitoring model. The hybrid non-intrusive load monitoring model has a multi-task learning framework and can simultaneously perform device state prediction and energy injection estimation. In practical applications, the data of two dimensions, active power and reactive power, are considered at the same time, and active power sequence and reactive power sequence are extracted from the meter data, so that the extracted data can comprehensively reflect the working state of the electrical appliance and the operation of the power system. Subsequently, the active power sequence and reactive power sequence are input into the trained hybrid non-intrusive load monitoring model, and the model splices the features in the active power sequence and the features in the reactive power sequence together. First, a comprehensive feature sequence is obtained. Then, the long-term dependencies in the comprehensive feature sequence are captured, and the complex time dependencies are modeled to obtain a multi-scale feature sequence. Finally, based on the multi-scale feature sequence, the state of the electrical equipment and the energy injection of the micro-inverter are predicted at the same time, and the state detection results of the electrical equipment and the energy injection sequence prediction results are obtained. The entire scheme only needs to collect meter data to identify the state of the electrical equipment and estimate the energy injection, so as to obtain comprehensive and accurate monitoring results, which can be applied to modern smart home environments that include distributed energy. Moreover, the above scheme can be stable and robust on data from different households, different time periods and different types of electrical appliances.

[0071] In an exemplary embodiment, a hybrid non-intrusive load monitoring model includes an encoder, a converter model, a sequence-to-point module, and a sequence-to-sequence module.

[0072] In this embodiment, the overall architecture of the HybridNILM model includes the following main components:

[0073] (1) Encoders, such as CNN encoders (CNNEncoders, convolutional neural network encoders, abbreviated as convolutional encoders) and RNN encoders (recurrent neural networks).

[0074] (2) Transformer model (i.e., converter model), where the Transformer model includes the Transformer encoder (TransformerEncoder) and the Transformer decoder (TransformerDecoder).

[0075] (3) Task-specific Projection Layers, including Seq2Point (sequence to point) module and Seq2Seq (sequence to sequence) module. The Seq2Point module is used to process the status detection of electrical equipment, and the Seq2Seq module is used to predict the amount of energy injection.

[0076] like Figure 4As shown, in some embodiments, S420 includes, S422, calling the first convolution encoder to perform one-dimensional convolution on the active power sequence to extract the active power features in the active power sequence, and calling the second convolution encoder to perform one-dimensional convolution on the reactive power sequence to extract the reactive power features in the reactive power sequence features.

[0077] In this embodiment, the encoder is a CNN encoder as an example for explanation. Each input feature channel of the CNN encoder has a separate one-dimensional CNN encoder for extracting local time patterns. Each CNN encoder contains multiple one-dimensional convolutional layers, and each layer is followed by a LayerNorm operation. The design of the CNN encoder allows the model to learn local patterns specific to each feature, enhancing the model's ability to capture the unique characteristics of different electrical measurement features.

[0078] The mathematical expression of CNN encoder is as follows:

[0079]

[0080] in, It is feature channels, is the corresponding encoding representation.

[0081] In specific implementation, the server calls the first convolution encoder to perform one-dimensional convolution on the active power sequence to extract the active power features in the active power sequence, and calls the second convolution encoder to perform one-dimensional convolution on the reactive power sequence to extract the reactive power features in the reactive power sequence features.

[0082] Among them, each CNN encoder consists of multiple one-dimensional convolutional layers, layer normalization layers and activation functions. The specific processing process is as follows:

[0083] One-dimensional convolution operation: For each power sequence, the CNN encoder first performs a one-dimensional convolution operation on the input data to extract the local features of the sequence. The size of the convolution kernel is usually set to 5, and the number of channels of the convolution layer is set to 64. For the input sequence ( is the sequence length), the convolution operation is calculated as:

[0084]

[0085] in, For the The convolution kernel weights of the layer, is the bias term, Represents a convolution operation.

[0086] Layer Normalization: The output after convolution is normalized to speed up training and improve the stability of the model. The mathematical expression of layer normalization is:

[0087]

[0088] in, and They are The mean and standard deviation of and are learnable parameters.

[0089] Activation function: After each convolutional layer, the ReLU activation function is used to perform a nonlinear transformation on the output. The mathematical expression of the ReLU activation function is:

[0090]

[0091] After the above processing, the feature tensors of active power and reactive power can be obtained, with the shape of (batch size, number of channels, sequence length). Then, these two feature tensors are concatenated on the feature dimension (i.e., the number of channels dimension) to form a comprehensive feature representation, i.e., the comprehensive sequence feature, with the shape of (batch size, 2 times the number of channels, sequence length).

[0092] In this embodiment, two designed one-dimensional convolutional encoders are called to process active power and reactive power sequences respectively, which can not only efficiently extract features but also significantly improve the performance of the system. Moreover, this architecture can adjust the number of convolutional layers, the size of the convolution kernel and other hyperparameters according to specific task requirements, providing great flexibility.

[0093] like Figure 4 As shown, in an exemplary embodiment, S440 includes: S442, inputting the comprehensive feature sequence into the converter model, and the converter model captures the long-term dependencies in the comprehensive feature sequence through a multi-head attention mechanism to obtain a multi-scale feature sequence.

[0094] Following the above embodiment, the Transformer model includes a Transformer encoder and a Transformer decoder. The structure of the Transformer decoder is similar to that of the encoder, but contains an additional cross-attention layer. After the comprehensive sequence features are obtained by splicing, the comprehensive sequence features can be input into the Transformer encoder. The Transformer encoder uses a multi-head self-attention mechanism to model the time dependency in the sequence. The specific process is as follows:

[0095] Multi-head self-attention mechanism: The core of the Transformer encoder is the multi-head self-attention mechanism, which allows the model to pay attention to the information of other time steps in the sequence when calculating the representation of the current time step. Through multiple attention heads, the model can learn the dependencies between sequence elements from different subspaces. The calculation process of self-attention is:

[0096]

[0097] in, are query, key, and value matrices, respectively. is the dimension of the key.

[0098] Position encoding: Since the Transformer model itself does not contain position information, position encoding needs to be added to retain the temporal information of the sequence. The calculation method of position encoding is:

[0099]

[0100] in, For location, is the dimension index, is the model dimension.

[0101] Encoding process: The Transformer encoder processes the input comprehensive feature sequence to generate a new sequence representation, namely a multi-scale feature sequence.

[0102] In this embodiment, the multi-head attention mechanism can capture global temporal dependencies, allowing the model to understand long-term patterns in the sequence. Through the processing of the Transformer encoder, the model can better capture the hidden features in the active power and reactive power sequences, providing more effective feature representation for subsequent prediction tasks.

[0103] like Figure 4 As shown, in an exemplary embodiment, S460 includes:

[0104] S462, the multi-scale feature sequence output by the converter model is input into the sequence-to-point module, and the sequence-to-point module maps the last time step of the multi-scale feature sequence to the power equipment state prediction task through the fully connected layer to obtain the state detection result of the power equipment.

[0105] S464, the multi-scale feature sequence output by the converter model is input into the sequence-to-sequence module, and the sequence-to-sequence module maps the multi-scale features of each time step in the multi-scale feature sequence to the energy injection prediction task through a fully connected layer to obtain the energy injection sequence prediction result.

[0106] In this embodiment, the model has a multi-task learning framework. In order to enable the model to handle both the power equipment state prediction task and the energy injection estimation task at the same time, a hybrid method combining the sequence-to-point (Seq2Point) and sequence-to-sequence (Seq2Seq) strategies is used. Among them, the Seq2Point method is mainly used to predict the state of the power equipment at a certain point in time from the multi-scale feature sequence, and the Seq2Seq method is used to predict energy injection and generate an output sequence that matches the length of the input sequence, that is, the energy injection sequence prediction result. The energy injection sequence prediction result may include the estimated energy injection amount within a preset time period (such as minute by minute), as well as possible confidence intervals.

[0107] In specific implementation, the multi-scale feature sequence output by the Transformer decoder is input into the Seq2Point module and the Seq2Seq module. Since the goal of Seq2Point is to predict a single output point based on the entire input sequence, only the information of the last time step output by the Transformer decoder is used. Specifically, the features of the last time step of the multi-scale feature sequence are mapped to the electrical equipment state prediction task through the fully connected layer, that is, mapped to a low-dimensional space, and the state of each electrical appliance at the current moment is predicted, and the state prediction data of the electrical equipment is directly obtained.

[0108] Specifically, in the Seq2Point module, the feature vector of the last time step output by the Transformer decoder is It is input into the fully connected layer for projection to obtain the state prediction of each device. The mathematical expression is:

[0109] The mathematical expression is:

[0110]

[0111] in, yes The last time step of and is a learnable parameter, is the sigmoid activation function. Since the state prediction of the electrical equipment is a probability value, which represents the on / off state of the equipment (0 or 1), the Sigmoid function can be used to map the output to between (0, 1), which is suitable for probability prediction of binary classification problems.

[0112] At the same time, the Seq2Seq module processes the entire multi-scale feature sequence. Through the fully connected layer and the Sigmoid activation function, the multi-scale features of each time step in the multi-scale feature sequence are mapped to the energy injection prediction task, predicting the future energy injection sequence and obtaining the normalized energy injection sequence prediction result. The mathematical expression is:

[0113]

[0114] in, is the predicted energy injection sequence, and is a learnable parameter.

[0115] In the Seq2Seq module, after the output of the Transformer decoder passes through the fully connected layer, the Sigmoid activation function is used to compress the predicted value to between 0 and 1. The mathematical expression of this step is:

[0116]

[0117] in, is the Sigmoid activation function, and are the weights and biases of the fully connected layer.

[0118] Through the Sigmoid function, the output of the model can be mapped to (0, 1), which facilitates the subsequent mapping of continuous prediction values ​​to predefined energy levels.

[0119] In this embodiment, by combining the Seq2Point and Seq2Seq strategies, the model can simultaneously handle two tasks: power equipment status detection and energy injection prediction, thereby improving the overall performance of the model.

[0120] In the NILM project, for the fusion of multiple tasks, it is necessary to evaluate the performance of the power equipment state detection task and the energy injection estimation and optimize them by designing a loss function. Therefore, in some embodiments, a series of evaluation indicators applicable to the power equipment state detection task (classification task) and energy injection decomposition (regression task) can also be used, and corresponding multi-task loss functions are designed. Specifically, for the power equipment state detection task indicators, the following classification evaluation indicators are used:

[0121] 1. Accuracy:

[0122]

[0123] 2. Precision:

[0124]

[0125] 3. Recall:

[0126]

[0127] 4.F1 score:

[0128]

[0129] 5. False Alarm Rate (FAR):

[0130]

[0131] Among them, TP is a true positive example, TN is a true negative example, FP is a false positive example, and FN is a false negative example. From a numerical point of view, 1 represents the on state and 0 represents the off state.

[0132] These indicators provide a comprehensive performance evaluation of the model in identifying the state of electrical appliances. Among them, the newly added false alarm rate indicator is used to evaluate the reliability of the model in practical applications, thus reducing the possibility that the wrong detection of the on state of electrical appliances may lead to unnecessary intervention or wrong energy management strategies.

[0133] For the energy injection estimation index, the following regression evaluation index is used:

[0134] 1. Root Mean Square Error (RMSE):

[0135]

[0136] 2. Mean Absolute Error (MAE):

[0137]

[0138] 3. Normalized Root Mean Square Error (NRMSE):

[0139]

[0140] in, is the time step The true value of is the corresponding predicted value, is the total number of time steps, is the average of the true values.

[0141] In order to optimize the power equipment state detection task and the energy injection estimation task simultaneously, a multi-task loss function is also designed in this embodiment:

[0142]

[0143] in, and is the weight coefficient used to balance the importance of the two tasks.

[0144] For the task of detecting the state of electrical equipment, that is, a multi-classification task, the most commonly used indicator is the cross entropy loss function. For a single sample, the cross entropy loss is defined as follows:

[0145]

[0146] in, is the number of categories, is the one-hot encoding of the true label, is the model for the category The predicted probability of .

[0147] For batch data, the average cross entropy loss is usually used:

[0148]

[0149] in, is the batch size.

[0150] Cross entropy loss works well in multi-classification problems because it directly optimizes the difference between the predicted probability distribution and the true distribution.

[0151] In addition, in some other embodiments, you can also choose to use the Dice loss function, which performs well when dealing with imbalanced categories:

[0152]

[0153] in, The model is The predicted probability of a sample, is the corresponding true label (0 or 1), is the sample size, is a small constant used to prevent the denominator from being zero.

[0154] For the energy injection estimation task, a combination of mean squared error (MSE) loss and mean absolute error (MAE) loss can be used:

[0155]

[0156]

[0157]

[0158] in, is a weight coefficient between 0 and 1 that balances the contribution of MSE and MAE. MSE is more sensitive to large errors and helps converge quickly, while MAE is more robust to outliers.

[0159] In the above embodiment, through the design of this multi-task loss function, the model can simultaneously optimize the tasks of electrical appliance state recognition and energy injection estimation, thereby improving the generalization ability and robustness of the model while maintaining high accuracy.

[0160] like Figure 5 As shown, in an exemplary embodiment, after S400, the method further includes:

[0161] S600, normalizing the energy injection sequence prediction result to obtain a normalized energy injection sequence prediction result, adaptively filtering the normalized energy injection sequence prediction result, mapping the normalized energy injection sequence prediction result to a predefined power output gear, and determining the output power level of the microinverter.

[0162] Adaptive filtering methods are a class of algorithms that can dynamically adjust their parameters (such as filter coefficients) according to the statistical characteristics of the input signal.

[0163] In this embodiment, in order to solve the problem of numerical instability in energy injection prediction and make the prediction result more consistent with the output mode of the actual microinverter, an adaptive filtering algorithm is introduced. The core idea of ​​this algorithm is to map the continuous energy injection prediction value to a predefined discrete output gear and eliminate short-term fluctuations through adaptive smoothing.

[0164] In specific implementation, after obtaining the energy injection sequence prediction result, the following processing is performed:

[0165] Normalization: First, divide the output prediction value of the microinverter (the predicted value of energy injection, hereinafter referred to as the predicted value) by the preset maximum power (determined by the nameplate parameters of the microinverter) to make it fall within the interval [0, 1]. This not only solves the problem of numerical instability, but also allows the predicted value to directly correspond to different power output gears.

[0166] Quantization: Map the continuous predicted value to a predefined discrete gear. This reflects the characteristic that actual microinverters usually have several fixed output power levels. The power output gear corresponds to the output power level one by one.

[0167] Adaptive smoothing: Identify outliers in the sequence (points that differ greatly from neighboring points) and adjust them to the average of the neighboring points. This process is repeated until no more outliers are identified or the maximum number of iterations is reached.

[0168] Requantization: After each smoothing operation, the prediction result is quantized to the nearest gear again to ensure that the output always meets the predefined output gear. The power output gear can be [0, 0.2, 0.3, 0.4, 0.7, 0.8, 1], corresponding to 7 different output power levels of the microinverter, so as to obtain discrete energy injection estimates. For example, the gear "0.2" can correspond to an energy injection estimate of 100W, and the gear "0.3" can correspond to an energy injection estimate of 500W.

[0169] In this embodiment, adaptive filtering is performed on the prediction results of the energy injection sequence, which can effectively eliminate short-term fluctuations and outliers, improve the reliability of the prediction results, and enable the energy injection prediction values ​​to directly correspond to different power output gears of the microinverter, providing important support for the application of the NILM system in smart homes that include renewable energy.

[0170] Data preprocessing is a key step to ensure data quality and consistency. Figure 5 As shown, in an exemplary embodiment, S200 includes:

[0171] S220, preprocessing the meter data, and extracting an active power sequence and a reactive power sequence with a fixed length from the preprocessed meter data by a sliding window method.

[0172] In practical applications, after obtaining the meter data, the meter data will be preprocessed. The preprocessing includes the following aspects:

[0173] 1. Data cleaning: handle missing values ​​and outliers, solve the problem of discontinuous timestamps, and handle negative value problems (check whether there is micro-reverse injection or reverse connection of the line).

[0174] 2. Time alignment: Ensure that the timestamps of the main meter and each sub-meter are aligned, and use interpolation methods to handle inconsistent sampling rates.

[0175] 3. Loadtype encoding: Use the encode_new_load_type_C function to convert the Loadtype string into the binary state of each appliance.

[0176] 4. Data resampling: Resample the data to 2-second intervals to ensure data consistency.

[0177] Subsequently, through feature engineering, active power and reactive power are extracted from the preprocessed meter data to obtain active power series and reactive power series.

[0178] In order to capture the dynamic characteristics of the time series, a sliding window method is used to create a fixed-length input sequence. For example, the sequence length of the active power series and reactive power series is set to 300 time steps (corresponding to 10 minutes of data), and the step size is usually set to 1, which means that only one data point is moved forward each time. Subsequently, a sliding window is applied to create an active power series and reactive power series with a fixed length of 300 time steps.

[0179] In this embodiment, the active power sequence and the active power sequence with a fixed length are obtained by using a sliding window, which can help better capture the power consumption pattern and its changes.

[0180] In some embodiments, the method further includes performing abnormal event detection according to abnormal rules to identify abnormal events.

[0181] Specifically, the following exception rules may be included:

[0182] (1) Energy consumption peak is too high: observe whether the real-time energy consumption of the appliance exceeds 1.1 times the historical peak;

[0183] (2) Accumulated energy consumption is too high: whether the energy consumption on the day exceeds 1.02 times the historical value;

[0184] (3) Excessive fluctuation in energy consumption: Check whether the fluctuation range of energy consumption on the day exceeds 1.05 times the historical value;

[0185] (4) Single-time use for too long: Observe whether the appliance's real-time single-time use time exceeds 1.5 times the longest time in history;

[0186] (5) Cumulative usage time: Count whether the usage time on the day exceeds 1.2 times the historical value;

[0187] (6) Abnormal start-up time: Observe whether the appliance is turned on earlier than the earliest time in history, or turned off later than the latest time in history;

[0188] (7) On-off frequency is too high: Statistics are collected to determine whether the On-off frequency on the day exceeds twice the historical value.

[0189] Specifically, based on the above abnormal rules, the power consumption time of each power-consuming device can be detected to detect whether there is an abnormal event and obtain the abnormal event detection result. If an abnormal event is detected, a corresponding prompt message can be pushed to notify the operation and maintenance personnel to handle it in time.

[0190] In order to make a clearer description of the non-intrusive load monitoring method provided by the present application, a specific embodiment is described below, and the specific embodiment includes the following contents:

[0191] S1, obtain total electricity meter data.

[0192] The total electricity meter data includes the electricity consumption data of air purifiers, electric heaters, multiple light bulb groups, air compressors, and air conditioners. These data reflect the different usage frequencies of various electrical appliances in the dataset.

[0193] S2, preprocessing the total meter data, and extracting the active power sequence and reactive power sequence of fixed length from the preprocessed meter data through the sliding window method.

[0194] S3, taking the active power sequence and reactive power sequence as input, calls the trained HybridNILM model to obtain the state detection results of the electrical equipment and the energy injection sequence prediction results.

[0195] It should be understood that, although the various steps in the flowcharts involved in the above-mentioned embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-mentioned embodiments can include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

[0196] Based on the same inventive concept, the embodiment of the present application also provides a non-intrusive load monitoring device for implementing the non-intrusive load monitoring method involved above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method, so the specific limitations in one or more non-intrusive load monitoring device embodiments provided below can refer to the limitations of the non-intrusive load monitoring method above, and will not be repeated here.

[0197] In an exemplary embodiment, Figure 6 As shown, a non-intrusive load monitoring device 600 is provided, comprising: a data extraction module 610 and a data processing module 620, wherein:

[0198] The data extraction module 610 is used to extract the active power sequence and the reactive power sequence from the electric meter data.

[0199] The data processing module 620 is used to take the active power sequence and the reactive power sequence as input, call the trained hybrid non-intrusive load monitoring model, and obtain the state detection result of the electrical equipment and the energy injection sequence prediction result.

[0200] When the hybrid non-intrusive load monitoring model is called, the data processing module 620 performs the following steps: extracting active power features in the active power sequence and reactive power features in the reactive power sequence, splicing the active power features and the reactive power features to obtain a comprehensive feature sequence, capturing the long-term dependencies in the comprehensive feature sequence, obtaining a multi-scale feature sequence, and predicting the state of the electrical equipment and the energy injection of the micro-inverter based on the multi-scale feature sequence, respectively, to obtain the state detection results of the electrical equipment and the energy injection sequence prediction results; wherein, the hybrid non-intrusive load monitoring model is trained based on the historical active power sequence and the historical reactive power sequence.

[0201] In an exemplary embodiment, the encoder includes a first convolutional encoder and a second convolutional encoder; the data processing module 620 is also used to call the first convolutional encoder to perform one-dimensional convolution on the active power sequence to extract the active power characteristics in the active power sequence, and call the second convolutional encoder to perform one-dimensional convolution on the reactive power sequence to extract the reactive power characteristics in the reactive power sequence characteristics.

[0202] In an exemplary embodiment, the data processing module 620 is also used to input the comprehensive feature sequence into the converter model, and the converter model captures the long-term dependencies in the comprehensive feature sequence through a multi-head attention mechanism to obtain a multi-scale feature sequence.

[0203] In an exemplary embodiment, the data processing module 620 is also used to input the multi-scale feature sequence output by the converter model into the sequence-to-point module, and the sequence-to-point module maps the last time step of the multi-scale feature sequence to the power equipment state prediction task through a fully connected layer to obtain the state detection result of the power equipment, and input the multi-scale feature sequence output by the converter model into the sequence-to-sequence module, and the sequence-to-sequence module maps the multi-scale features of each time step in the multi-scale feature sequence to the energy injection prediction task through a fully connected layer to obtain the energy injection sequence prediction result.

[0204] In an exemplary embodiment, the data processing module 620 is also used to normalize the energy injection sequence prediction results to obtain normalized energy injection sequence prediction results, adaptively filter the normalized energy injection sequence prediction results, map the normalized energy injection sequence prediction results to a predefined power output gear, determine the output power level of the microinverter, and the power output gear corresponds to the output power level one by one.

[0205] In an exemplary embodiment, the data extraction module 610 is further used to preprocess the meter data, and extract the active power sequence and reactive power sequence with fixed length from the preprocessed meter data by using a sliding window method.

[0206] Each module in the above non-intrusive load monitoring device can be implemented in whole or in part by software, hardware, or a combination thereof. Each module can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in a computer device in the form of software, so that the processor can call and execute operations corresponding to each module.

[0207] In an exemplary embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as shown in FIG. Figure 7 As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, referred to as I / O) and a communication interface. Among them, the processor, the memory and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store meter data, power equipment status detection data, and energy injection prediction sequence result data. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a non-invasive load monitoring method is implemented.

[0208] Those skilled in the art will understand that Figure 7 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0209] In an exemplary embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the steps in any one of the above-mentioned non-intrusive load monitoring method embodiments are implemented.

[0210] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in any one of the above-mentioned non-intrusive load monitoring method embodiments are implemented.

[0211] In one embodiment, a computer program product is provided, including a computer program, which implements the steps of any one of the above non-intrusive load monitoring method embodiments when executed by a processor.

[0212] It should be noted that the user information (including but not limited to user device information such as meter data, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0213] A person of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment method can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, an artificial intelligence (AI) processor, etc., but are not limited to this.

[0214] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0215] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be construed as limiting the scope of the present application. It should be noted that, for a person of ordinary skill in the art, several modifications and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A non-intrusive load monitoring method, characterized in that: The method comprises: Extract active power series and reactive power series from electric meter data; Taking the active power sequence and the reactive power sequence as input, calling the trained hybrid non-intrusive load monitoring model to obtain the state detection result of the electrical equipment and the energy injection sequence prediction result; The hybrid non-intrusive load monitoring model performs the following steps when invoked: Extracting active power features in the active power sequence and reactive power features in the reactive power sequence, and splicing the active power features and the reactive power features to obtain a comprehensive feature sequence; Capturing the long-term dependency in the comprehensive feature sequence to obtain a multi-scale feature sequence; Based on the multi-scale feature sequence, the state of the electrical equipment and the energy injection of the micro-inverter are predicted respectively, so as to obtain the state detection result of the electrical equipment and the energy injection sequence prediction result; The hybrid non-intrusive load monitoring model is trained based on historical active power sequences and historical reactive power sequences.

2. The method according to claim 1, characterized in that The hybrid non-intrusive load monitoring model includes an encoder, a converter model, a sequence-to-point module and a sequence-to-sequence module.

3. The method according to claim 2, characterized in that The encoder includes a first convolution encoder and a second convolution encoder; the extracting the active power feature in the active power sequence and the reactive power feature in the reactive power sequence includes: Calling a first convolution encoder to perform one-dimensional convolution on the active power sequence to extract active power features in the active power sequence; The second convolution encoder is called to perform one-dimensional convolution on the reactive power sequence to extract reactive power features from the reactive power sequence features.

4. The method according to claim 2, characterized in that: The capturing of the long-term dependencies in the comprehensive feature sequence to obtain a multi-scale feature sequence includes: The comprehensive feature sequence is input into the converter model, and the converter model captures the long-term dependencies in the comprehensive feature sequence through a multi-head attention mechanism to obtain a multi-scale feature sequence.

5. The method according to claim 2, characterized in that: The method of predicting the state of the electric device and the energy injection of the micro-inverter based on the multi-scale feature sequence to obtain the state detection result of the electric device and the energy injection sequence prediction result includes: The multi-scale feature sequence output by the converter model is input into the sequence-to-point module, and the sequence-to-point module maps the last time step of the multi-scale feature sequence to the power device state prediction task through a fully connected layer to obtain a state detection result of the power device; The multi-scale feature sequence output by the converter model is input into the sequence-to-sequence module, and the sequence-to-sequence module maps the multi-scale features of each time step in the multi-scale feature sequence to the energy injection prediction task through a fully connected layer to obtain an energy injection sequence prediction result.

6. The method according to any one of claims 1 to 5, characterized in that: After obtaining the state detection result of the electric device and the energy injection sequence prediction result, the method further includes: Normalizing the energy injection sequence prediction result to obtain a normalized energy injection sequence prediction result; Adaptively filter the normalized energy injection sequence prediction result, map the normalized energy injection sequence prediction result to a predefined power output gear, determine the output power level of the microinverter, and the output gear corresponds to the output power level one by one.

7. The method according to any one of claims 1 to 5, characterized in that: The step of extracting active power sequence and reactive power sequence from the electric meter data comprises: Preprocess the meter data; Active power sequences and reactive power sequences with fixed lengths are extracted from the preprocessed meter data using a sliding window method.

8. A non-intrusive load monitoring device, characterized in that: The device comprises: A data extraction module is used to extract active power sequence and reactive power sequence from the electric meter data; A data processing module, used to take the active power sequence and the reactive power sequence as input, call the trained hybrid non-intrusive load monitoring model, and obtain the state detection result of the electrical equipment and the energy injection sequence prediction result; The hybrid non-intrusive load monitoring model performs the following steps when invoked: Extracting active power features in the active power sequence and reactive power features in the reactive power sequence, and splicing the active power features and the reactive power features to obtain a comprehensive feature sequence; Capturing the long-term dependency in the comprehensive feature sequence to obtain a multi-scale feature sequence; Based on the multi-scale feature sequence, the state of the electrical equipment and the energy injection of the micro-inverter are predicted respectively, so as to obtain the state detection result of the electrical equipment and the energy injection sequence prediction result; The hybrid non-intrusive load monitoring model is trained based on historical active power sequences and historical reactive power sequences.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images