Non-intrusive load monitoring method and system based on Mamba-graph convolutional network

Abstract

The invention discloses a non-intrusive load monitoring method and a non-intrusive load monitoring system based on a Mamba-graph convolutional network. The method comprises the following steps: firstly, collecting the total power utilization load of a user and the load data of each device, and preprocessing and enhancing the collected data; secondly, constructing a user-device hierarchical dynamic graph structure data set; secondly, designing a model combining a graph convolutional network and a Mama network, connecting by using a residual module, and capturing a complex space-time dependency relationship of the power load data; dividing the dynamic graph structure data set into a training set, a verification set and a test set, and introducing a self-supervision task; after the model is trained, various evaluation indexes are adopted, the model is optimized on the verification set, the performance of the model is tested on the test set, and the performance of the model in different power consumption scenes is analyzed. And finally deploying the model to an actual application scene. According to the invention, the non-intrusive load monitoring analysis capability of the user equipment is improved, and accurate energy consumption analysis and optimization suggestions are provided for users.

Description

Technical Field

[0001] The present invention belongs to the field of non-intrusive load monitoring, and in particular relates to a non-intrusive load monitoring method and system based on a Mamba-graph convolutional network. Background Art

[0002] Non-intrusive load monitoring (NILM) is a method that analyzes the user's total load profile to infer the power usage and load consumption of individual devices without directly intervening in the device. In recent years, NILM technology has been widely used in energy conservation management, user electricity behavior analysis, and smart grid optimization.

[0003] Traditional non-intrusive load monitoring methods primarily rely on signal processing, feature extraction, and pattern matching. However, with the increasing variety and number of consumer electrical devices, the complexity of power load curves has increased significantly. The potential relationships between devices also pose significant challenges to device decomposition. Traditional methods often struggle to handle highly complex load curves due to signal overlap between devices, the extraction of low-frequency features, and the diversity of load variations, resulting in low decomposition accuracy. To address this issue, machine learning and deep learning technologies have been introduced into the field of non-intrusive load monitoring in recent years, attempting to improve the accuracy of load decomposition through data-driven approaches. Graph neural networks, in particular, have become a hot topic in non-intrusive load monitoring research due to their ability to handle complex topologies and capture inter-node dependencies. However, traditional graph neural networks still have shortcomings in capturing spatiotemporal features. For example, they struggle to model the temporal dependencies of time series data, are limited in capturing dynamic changes in node states, and lack the complexity of modeling spatiotemporal interactions. Furthermore, multi-layer network training is prone to vanishing and exploding gradients.

[0004] To this end, this patent proposes a non-intrusive load monitoring solution based on the Mamba graph convolutional network. First, by building a dynamic graph module, the system can capture complex, time-varying underlying relationships between devices in real time, addressing the shortcomings of traditional static graphs in reflecting dynamic device interactions. Furthermore, statistical methods are used to calculate the similarities and relationships between devices, generating an adjacency matrix that updates over time, ensuring that the model captures more realistic spatial information. Second, the Mamba network is used to model long-term time series data. Its design, based on a selective state-space model, efficiently extracts key temporal features and overcomes the high computational complexity and low efficiency of traditional Transformer models when processing long sequences. The Mamba network captures long-range dependencies with linear complexity, effectively addressing the information dilution and dependency challenges in long-term data. Third, to fully exploit the spatial dependencies between nodes in the dynamic graph, this solution employs a graph convolutional network for feature extraction. Graph convolutions can capture the interactions between devices through an adjacency matrix, but are prone to vanishing gradients when stacked in multiple layers. To this end, the introduction of residual connections in graph convolutional layers not only improves the problem of poor gradient propagation during deep network training but also enhances feature fusion and the overall learning ability of the model. Furthermore, the introduction of self-supervision effectively reduces the reliance on large amounts of labeled data. Through tasks such as time series reconstruction and future value prediction, the inherent spatiotemporal features of the data are automatically extracted, thereby enriching the feature representation, improving the model's robustness and generalization capabilities, and providing good initial parameters for supervised training, accelerating convergence and reducing the risk of overfitting.

[0005] During the implementation of the present invention, the inventors of this application discovered that the current non-invasive load monitoring methods have the following challenges:

[0006] 1) Difficulties in data processing and dataset construction. Non-intrusive load monitoring technology requires detailed power load data, especially load data accurate to the device level. However, constructing such a dataset and labeling the device categories is very complex and costly. In addition, differences in electricity usage habits among users, device types and power systems, and potential connections between different devices further exacerbate the difficulty of building large-scale datasets. As a result, the decomposition accuracy and generalization ability of the model are limited by the size of the dataset, making it difficult to effectively apply the model in diverse scenarios.

[0007] 2) Real-time performance requirements and high computational costs. In practical applications, non-intrusive load monitoring technologies require real-time monitoring and decomposition, requiring models capable of rapidly processing large amounts of time series data. However, complex deep learning models often require extensive computing resources for training and inference, making them difficult to run efficiently on low-power devices. Existing models face significant challenges in real-time performance and resource consumption in practical applications, hindering their widespread adoption.

[0008] 3) Load changes are highly dynamic and diverse. The on-and-off behavior of consumer electrical devices is often random and unpredictable, and the load variations across different devices are highly dynamic. Some devices dynamically adjust their load consumption based on load conditions, resulting in complex and diverse load curves. These nonlinear and dynamic load variations, especially when loads on multiple devices change simultaneously, increase the difficulty of model load decomposition.

[0009] 4) Poor generalization and cross-scenario adaptability. Different users may have significantly different electrical equipment, electricity usage habits, and usage time patterns, making it difficult for a trained model to achieve the same results in different environments. In particular, when data comes from different regions or time periods, the model's decomposition accuracy may be limited. Even if a model performs well on the training set, it may perform poorly in new user environments or electricity usage scenarios, resulting in reduced prediction accuracy in actual use. Summary of the Invention

[0010] In view of this, the present invention provides a non-intrusive load monitoring method and system based on Mamba-graph convolutional network, which is used to solve or at least partially solve the difficulties of existing non-intrusive load monitoring.

[0011] Based on the above technical problems, the present invention adopts the following technical solutions:

[0012] A non-intrusive load monitoring method based on Mamba-graph convolutional network, comprising:

[0013] Step S1: Collect user power load data, including the user's total load and the load data of each device;

[0014] Step S2: pre-processing the collected power load data;

[0015] Step S3: constructing a dynamic graph structure using the pre-processed power load data;

[0016] Step S4: constructing a Mamba-graph convolution deep learning model that combines the Mamba network and the graph convolution network, and embedding a self-supervised task module to perform non-intrusive load monitoring on the data in the dynamic graph structure;

[0017] Step S5: Divide the collected user power load data into a training set, a validation set, and a test set, and train the Mamba-graph convolutional deep learning model to obtain an optimal model;

[0018] Step S6: Evaluate and test the generalization ability of the optimal model;

[0019] Step S7: deploying the optimal model to an actual non-intrusive load monitoring application system and optimizing it according to specific application feedback; deploying the optimal model to the cloud for real-time load monitoring.

[0020] Furthermore, the step S1 includes the following steps:

[0021] Step S1.1: Setting a sampling frequency, collecting user total load data from smart meters or other sensors at the sampling frequency, and recording the load consumption of users and the load consumption of each device in a time series;

[0022] Step S1.2: Collect long-term user total load and load data of each device and store them in a structured data format.

[0023] Furthermore, the pre-processing in step S2 includes:

[0024] Step S2.1: Fill missing values in the user's total load and each device's load data using a linear interpolation method;

[0025] Step S2.2: Normalize the total user load and the load data of each device by using the minimum-maximum normalization method to scale the data to the [0, 1] interval and convert the data to the same dimension;

[0026] Step S2.3: Adding Gaussian noise to the total user load and the load data of each device to simulate measurement errors and environmental fluctuations to enhance the robustness of the model;

[0027] Step S2.4: performing time series truncation on the total user load and the load data of each device, and randomly selecting a time window of fixed length to generate a new sample.

[0028] Furthermore, step S3 includes:

[0029] Step S3.1: Organize the power load data into a dynamic graph according to the time series, where each node represents a power load data; the central node represents the total user load, and several sub-nodes correspond to the load of each device;

[0030] Step S3.2: Determine the connection relationship between nodes by calculating the similarity or mutual relationship between each device, define the edges between nodes, and set the weight of the edges based on the frequency of simultaneous power-on or power-off of devices and the correlation of load changes, thereby constructing a dynamic graph structure.

[0031] Furthermore, the Mamba-graph convolution deep learning model in step S4 is specifically as follows:

[0032] Step S4.1: Use the Mamba network to process the time series features and combine each time series feature with the node feature to form a spatiotemporal joint representation;

[0033] Step S4.2: Stack multiple graph convolutional layers, based on a multi-layer graph convolutional network, combined with residual connections to ensure that the temporal features are aligned with the graph structure;

[0034] Step S4.3: Design the output layer through a fully connected layer to map the results of graph convolution to specific load prediction values, and finally obtain the predicted load value of each node.

[0035] Furthermore, the self-supervision task module in step S4 is designed as follows:

[0036] The self-supervised task module includes a time series reconstruction subtask and a future value prediction subtask;

[0037] The time series X with load data after preprocessing is T}, randomly mask some time steps to form a partial observation sequence X mask ;

[0038] The autoencoder is constructed by combining the Mamba network with the graph convolution module and the residual connection module. It is responsible for extracting the spatiotemporal implicit features z of dynamic graphs and time series data, and reconstructing the original sequence based on the spatiotemporal implicit features based on the decoder.

[0039] The future value prediction task includes data segmentation and model prediction:

[0040] Divide the time series with load data into past window and future forecast window;

[0041] Using the spatiotemporal model composed of the same Mamba network combined with the graph convolution module and the residual connection module, the past window X t-W+1:t Encode and output the predicted value of the future prediction window

[0042] Furthermore, step S5 specifically includes:

[0043] Step S5.1: The pre-processed training set load data is used as input to the Mamba-graph convolutional deep learning model for training to predict the load of each device at a future time.

[0044] Step S5.2: define the loss function;

[0045] Step S5.3: adjusting the parameters using a gradient descent method and an optimization algorithm;

[0046] Step S5.4: Calculate the loss based on the loss function and denormalize the predicted value and the true value;

[0047] Step S5.5: Use the validation set to evaluate model performance during training and perform model validation.

[0048] Step S5.6: Based on the training results, save the best model and record the optimal model parameters.

[0049] Furthermore, step S6 specifically includes:

[0050] Step S6.1: Based on the Mamba-graph convolutional deep learning model, the change in total load is distributed to the predicted value of each device to obtain the load consumption forecast of each device at a future time;

[0051] Step S6.2: Evaluate the trained Mamba-graph convolutional deep learning model on different test sets and datasets of different electricity usage scenarios; and calculate the accuracy of the model.

[0052] Step S6.3: Plot the predicted value and actual value of load consumption of each device on the same graph for visual comparison.

[0053] Furthermore, step S7 specifically includes:

[0054] Step S7.1: Deploy the optimal model to the cloud server and design a real-time data input interface for the Mamba-graph convolutional deep learning model to ensure that it can receive real-time data from power monitoring equipment;

[0055] Step S7.2: Continuously tune the optimal model.

[0056] On the other hand, the present invention also provides a non-intrusive load monitoring system based on Mamba-graph convolutional network, comprising:

[0057] Data acquisition, processing and enhancement module: This module collects user power load data, including the user's total load and the load data of each device, and pre-processes the collected power load data;

[0058] Dynamic graph structure construction module: This module uses the pre-processed power load data to construct a dynamic graph structure;

[0059] Load forecasting decomposition model construction module: This module is used to build a Mamba-graph convolutional deep learning model that combines the Mamba network and the graph convolutional network, and embeds a self-supervised task module to perform non-intrusive load monitoring on the data in the dynamic graph structure;

[0060] Data result acquisition module: This module is responsible for inputting the processed power load data into the trained Mamba-graph convolutional deep learning model and obtaining the load forecast results of each device. The system stores the forecast results for subsequent data analysis and energy consumption decision-making;

[0061] Data Result Visualization Module: This module is used to visualize the prediction results output by the Mamba-graph convolutional deep learning model. It generates personalized energy-saving suggestions and energy consumption analysis reports based on the electricity usage characteristics of different users. The reports contain multiple charts and key indicators to help users intuitively manage their electricity usage.

[0062] Model optimization and control module: This module is used to combine actual application feedback, enrich the data set, improve the optimization algorithm, and adjust the hyperparameters to continuously iterate the training model, monitor the model performance, and further improve the accuracy of load decomposition and the generalization ability of the system.

[0063] Compared with the prior art, the present invention has the following beneficial effects:

[0064] In summary, the present invention utilizes smart meters and sensor systems to collect high-frequency, multi-level power data, and adopts advanced data preprocessing and enhancement technologies to ensure the integrity, standardization, and diversity of the data set; through the dynamic graph structure construction module, the complex and dynamic correlation relationships between devices are captured in real time, and the changes in the load status of different devices are accurately reflected; the Mamba network is used to efficiently extract the time dependency in long time series data, and the multi-layer graph convolutional network and residual connection are combined to realize the in-depth mining of spatial features. At the same time, self-supervision tasks are embedded to further reduce the dependence on large-scale labeled data, significantly improve the accuracy of load decomposition, model robustness and generalization ability, so as to better adapt to the diverse electricity usage scenarios of users and reduce the cost of data collection and labeling. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 This is a flow chart of a non-intrusive load monitoring method based on Mamba-graph convolutional network of the present invention;

[0066] Figure 2 This is a schematic diagram of the calculation results of the power load of each device obtained by model decomposition in a specific embodiment of a non-invasive load monitoring method based on a Mamba-graph convolutional network of the present invention. DETAILED DESCRIPTION

[0067] The inventors of this application have discovered through extensive research and practice that current non-invasive load monitoring methods have the following challenges:

[0068] 1) Difficulties in data processing and dataset construction. Non-intrusive load monitoring technology requires detailed power load data, especially load data accurate to the device level. However, constructing such a dataset and labeling the device categories is very complex and costly. In addition, differences in electricity usage habits among users, device types and power systems, and potential connections between different devices further exacerbate the difficulty of building large-scale datasets. As a result, the decomposition accuracy and generalization ability of the model are limited by the size of the dataset, making it difficult to effectively apply the model in diverse scenarios.

[0069] 2) Real-time performance requirements and high computational costs. In practical applications, non-intrusive load monitoring technologies require real-time monitoring and decomposition, requiring models capable of rapidly processing large amounts of time series data. However, complex deep learning models often require extensive computing resources for training and inference, making them difficult to run efficiently on low-power devices. Existing models face significant challenges in real-time performance and resource consumption in practical applications, hindering their widespread adoption.

[0070] 3) Load changes are highly dynamic and diverse. The on-and-off behavior of consumer electrical devices is often random and unpredictable, and the load variations across different devices are highly dynamic. Some devices dynamically adjust their load consumption based on load conditions, resulting in complex and diverse load curves. These nonlinear and dynamic load variations, especially when loads on multiple devices change simultaneously, increase the difficulty of model load decomposition.

[0071] 4) Poor generalization and cross-scenario adaptability. Different users may have significantly different electrical equipment, electricity usage habits, and usage time patterns, making it difficult for a trained model to achieve the same results in different environments. In particular, when data comes from different regions or time periods, the model's decomposition accuracy may be limited. Even if a model performs well on the training set, it may perform poorly in new user environments or electricity usage scenarios, resulting in reduced prediction accuracy in actual use.

[0072] Based on the above considerations, the present invention proposes a non-intrusive load monitoring method based on Mamba-graph convolutional network, which mainly addresses the current problems of lack of accurate device-level load decomposition model, difficulty in modeling complex relationships between users and devices, difficulty in processing load data with time series characteristics, insufficient model generalization ability, and poor model interpretability.

[0073] In order to achieve the above object, the main concepts of the present invention are as follows:

[0074] The present invention first automatically collects user power load data through smart meters or sensor systems, covering time series data of the user's total power load and the load of each device; then, the collected data is preprocessed, including missing value filling, feature scaling and standardization, and combined with data enhancement techniques such as noise addition and time series truncation to ensure the integrity, standardization and diversity of the data set, while providing high-quality input for self-supervision tasks. Next, the preprocessed data is converted into dynamic graph data to construct a user-device hierarchical structure, in which the central node represents the total user load and the child nodes correspond to the load of each device. The similarity or mutual relationship between the devices is calculated by statistical methods between the nodes, and the edge weights are set according to features such as simultaneous on / off of the devices to generate a symmetric adjacency matrix reflecting the actual spatiotemporal dynamic dependency. Based on this dynamic graph structure, the present invention designs a deep learning model that integrates the Mamba network and the graph convolution residual structure. The model first uses the Mamba network to efficiently capture the temporal dependency in long time series data, then extracts the spatial features in the dynamic graph through a multi-layer graph convolution module, and adopts residual connections to achieve effective fusion of features at each layer and smooth transmission of gradients. Furthermore, this model embeds a self-supervised task module within its structure. Through tasks such as time series reconstruction and future value prediction, it automatically learns the inherent spatiotemporal structure of the data during pre-training. This further strengthens feature representation, reduces reliance on large-scale labeled data, and improves model robustness. Ultimately, after supervised fine-tuning, the model is able to accurately decompose total user load and predict future load consumption by individual devices, forming an efficient, non-intrusive load monitoring solution that fully exploits time series features and captures dynamic spatial dependencies.

[0075] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0076] Example 1

[0077] like Figure 1 As shown, this embodiment provides a non-intrusive load monitoring method based on Mamba-graph convolutional network, and the establishment of the method includes:

[0078] Step S1: Power load data collection. In this example, the REFIT dataset is used for the experiment. This dataset contains power load data for multiple users, including total load and load information for each device. By collecting the total load of a specific user and detailed load data for each device, the model input is ensured to have comprehensive and authentic power load information, providing a solid data foundation for subsequent load decomposition tasks.

[0079] Step S2: Data preprocessing and enhancement. The load data in the REFIT dataset is preprocessed, and missing values are filled using linear interpolation to ensure data integrity and continuity. The load data is feature scaled and each load value is normalized to the range of 0 to 1 to eliminate differences in loads across different devices. At the same time, data enhancement techniques such as noise addition and time series truncation are used to improve the standardization and diversity of the dataset. In addition, the preprocessed data provides high-quality input for self-supervision tasks, supporting the automatic extraction of intrinsic spatiotemporal features through subsequent tasks such as time series reconstruction and future value prediction.

[0080] Step S3: Dynamic graph structure construction and convolution layer design. Based on the preprocessed data, a dynamic graph is constructed with the total user load as the central node and the load of each device as a sub-node, forming a user-device hierarchical structure. Between nodes, based on the correlation between devices or load change trends, similarities or mutual relationships are calculated through statistical methods, and edge weights are set based on characteristics such as simultaneous device on / off to generate a symmetric adjacency matrix. This dynamic graph not only provides a foundation for subsequent graph convolution operations, but also supports the extraction of effective features in spatial relationship modeling for self-supervised tasks.

[0081] Step S4: Load forecast decomposition model construction. A spatiotemporal deep model combining the Mamba network and the graph convolution residual structure is designed to perform non-invasive load monitoring on dynamic graph data. First, the Mamba network is used to efficiently capture the temporal dependency features in long time series data; secondly, a multi-layer graph convolution module is used to extract the spatial dependencies in the dynamic graph, and the residual connection is used to achieve efficient feature fusion and gradient transfer; in addition, a self-supervised task module (including time series reconstruction tasks and future value prediction tasks) is embedded in the model structure to further strengthen the learning of long-series dynamic patterns. Finally, the model decomposes the total user load, accurately predicts the future load consumption of each device, and improves the model robustness and generalization ability through joint optimization of self-supervised tasks;

[0082] Step S5: Model training. The pre-processed power load data is divided into a training set, a validation set, and a test set, and the data is input into the constructed model for training. During the training process, the training set is first used to perform self-supervised pre-training tasks, automatically extracting spatiotemporal features to provide good initial parameters for the model; then, in a supervised environment, the denormalized predicted value is compared with the actual value through loss functions such as mean square error (MSE), and the Adam optimizer is used to update the model parameters. According to the performance of the validation set, the hyperparameters and network structure are dynamically adjusted to finally obtain the optimal model;

[0083] Step S6: Model evaluation and generalization capability testing. The model's performance in load decomposition tasks is tested using various evaluation metrics, such as mean square error and mean absolute error, to ensure model prediction accuracy and practical usability. Furthermore, the model's generalization capability in different power usage scenarios is analyzed to verify its applicability to various types of power data.

[0084] Step S7: Model Deployment and Scenario Optimization. The trained model is deployed to an actual non-intrusive load monitoring system, typically on a cloud server, to provide real-time power load monitoring and decomposition. Based on actual application feedback, the dataset is continuously expanded and the model is continuously trained. Hyperparameters are fine-tuned to further improve prediction accuracy and system stability.

[0085] In one embodiment, step S1 specifically includes:

[0086] Step S1.1: In this embodiment, the power dataset REFIT comes from smart meters and other sensors installed in the user. These devices can record the total load in the time series and the load consumption of each device. Specifically, the non-intrusive load data is obtained through the smart meter, including the total consumption of all the user's electrical devices and the load usage of each device, providing basic data for the subsequent load decomposition task. The original dataset has twelve tags: time (Time), timestamp (Unix), total user power load (Aggregate), and nine different types of equipment power load (Appliance1-9).

[0087] Step S1.2: To ensure data timeliness and accuracy, this embodiment sets a sampling frequency of once per minute. This frequency is high enough to capture the load consumption details of the device, reflecting its on / off status and dynamic load changes, ensuring that the model can identify and learn the power usage patterns of each device. In specific implementations, the sampling frequency can be flexibly adjusted according to the needs of different scenarios to meet different time resolution requirements.

[0088] Step S1.3: During the data collection process, the user's total load and the load data for each device are recorded over a long period of time. In this embodiment, the collected data is stored as a CSV file to facilitate subsequent data preprocessing and analysis. This data storage method facilitates the management, query, and analysis of power data, providing efficient data support for subsequent model training and evaluation.

[0089] In one embodiment, step S2 specifically includes:

[0090] Step S2.1: In this embodiment, the collected power data is first preprocessed. Missing values in the data are filled using linear interpolation. This method uses an interpolation algorithm to estimate and fill missing values, ensuring the continuity and integrity of the time series data and preventing incomplete data from affecting model training.

[0091] Linear interpolation assumes that the change between two known points is linear, and the missing value is calculated through the linear relationship of the known points. The formula is as follows:

[0092]

[0093] Where (x1, y1) and (x2, y2) are known data points, and x is the point where the missing value is located;

[0094] Step S2.2: To adapt to the numerical range of the model, the total load and each equipment load data are normalized. In this embodiment, the minimum-maximum normalization method is used to scale the original data to the [0,1] interval. The formula is as follows:

[0095]

[0096] Among them, x is the original data, x min and x max are the minimum and maximum values of the data set, respectively, and x′ is the normalized data. This process ensures that the load data of different equipment are in the same dimension, so that the model can more effectively learn the features of the load decomposition task;

[0097] Step S2.3: Add Gaussian noise to the normalized data to simulate measurement errors and environmental fluctuations. This can enhance the robustness of the model and make it perform better in the face of uncertainty in the real environment:

[0098] d′=d+∈,∈~N(0,σ 2 )

[0099] This can improve the model's tolerance to noise and enhance its generalization ability;

[0100] Step S2.4: Truncate the long time series data and randomly select a fixed-length time window to generate a new sample. This can be achieved by randomly selecting the starting time point:

[0101] W i =[d i ,d i+1 ,...,d o+W-1 ]

[0102] This results in multiple overlapping time windows for model training and prediction.

[0103] In one embodiment, step S3 specifically includes:

[0104] Step S3.1: In this embodiment, the collected load data is used to construct a graph structure. The nodes of this graph structure represent the power load data of each user. It includes a central node representing the user's total power load and several sub-nodes representing the load of each device. In other words, the central node represents the total load of all the user's devices, while the sub-nodes represent the load consumed by each device, such as the refrigerator, washing machine, and air conditioner.

[0105] In the embodiment, if the user has N devices in total, the graph structure contains N+1 nodes, where a central node represents the user's total power load P total , the remaining N nodes represent each device P i (i=1, 2, ..., N).

[0106] Therefore, the node set V can be expressed as:

[0107] V={P total , P1, P2, ..., P N}

[0108] Here P total represents the total load node, P i represents the load node of the i-th device;

[0109] Step S3.2: In this embodiment, the edges between nodes are defined based on the correlation between devices or the load change trend. Specifically, the existence of an edge can be determined by calculating the similarity or mutual correlation between different devices. If two devices are often turned on or off at the same time, or their load change trends have a strong correlation, an edge is set between the corresponding nodes. The weight of the edge is set based on the similarity of power consumption between devices, the frequency of simultaneous turning on / off, and the correlation of load changes. The weight reflects the dependency between devices and the degree of their impact on the total load. If two devices have a high load correlation in multiple time periods, the edge weight between their corresponding nodes will be set to a higher value. This design can better capture the relationship between devices and improve the decomposition accuracy of the model in non-invasive load monitoring.

[0110] In this embodiment, the edge between nodes is defined by calculating the correlation or load change between different devices. The formula for defining the edge is as follows:

[0111]

[0112] Where A and B represent the load data of different equipment respectively, Cov(A,B) represents the covariance between the two, σ A and σ B Represents the standard deviation of the two. This formula is used to calculate the similarity between devices and reflects whether their load change trends are synchronized.

[0113] The time period during which devices A and B are turned on simultaneously within the time window T is t o , then the weight ω(A, B) can be expressed as:

[0114]

[0115] In one embodiment, step S4 specifically includes:

[0116] Step S4.1: In this embodiment, in order to effectively process the time dependency in the power load data, a Mamba network is used to capture the characteristics of the equipment load changing over time.

[0117] The Mamba network is based on the principle of the Selective State Space Model (SSM) and is specifically designed to efficiently capture dependencies in long time series data. Its main working principles can be summarized as follows:

[0118] Basic form of state space model: In continuous time, a linear state space model is described as:

[0119]

[0120] y(t)=Cx(t)+Du(t)

[0121] Among them, is the hidden state of x(t), u(t) is the input signal, y(t) is the output signal, A, B, C and D are the model parameter matrices. After discretization, we can get the recursive formula:

[0122]

[0123] Generation and efficient computation of convolution kernels: By using the characteristics of the state-space model, it can be shown that its output is equivalent to performing a convolution operation on the input sequence. The convolution kernel can be expressed as:

[0124]

[0125] Then the output of the discrete sequence can be written as

[0126]

[0127] This form makes the model have linear complexity (O(N)) in calculation, which is particularly suitable for processing long sequence data;

[0128] Selective Mechanism: To better extract temporal features, the Mamba network introduces a selectivity (or gating) mechanism based on the state-space model, allowing the model parameters to be dynamically adjusted based on the input. In other words, the parameter matrices A, B, and C can be transformed into input-dependent functions, such as:

[0129] A(u),B(u),C(u)

[0130] So the convolution kernel becomes

[0131] K[k]=C(u)(e A(u)k )B(u)

[0132] This design enables the network to selectively emphasize certain important temporal dependencies based on different time steps or contextual information, effectively capturing long-distance dependencies and changing trends;

[0133] Residual connections and multi-layer stacking: In order to deepen the network and ensure effective propagation of gradients, the Mamba module is usually embedded in a residual structure. Specifically, the features output by each Mamba layer are added to the input through a residual connection:

[0134] Output=MAMBA(u)+u

[0135] This design not only helps the model capture complex temporal dependencies, but also alleviates the gradient vanishing problem during deep network training.

[0136] The Mamba network transforms long-term sequence problems into convolutional ones using a state-space model, then dynamically adjusts parameters using a selective mechanism. Ultimately, it leverages residual connections to achieve stable and efficient long-term dependency modeling. This approach offers significant advantages in extracting temporal features from long-term series data while maintaining computational efficiency.

[0137] Step S4.2: Combine graph convolution with residual connections to ensure alignment of temporal features with the graph structure. Implement a multi-layer graph convolutional network in the model and combine it with residual connections to improve the model's performance on the device load decomposition task.

[0138] Graph convolutional networks are used to process the features of nodes and their neighboring nodes and aggregate node information. The formula is:

[0139]

[0140] Among them, H (l) is the node feature matrix of the lth layer, A is the adjacency matrix, D is the node degree matrix, W (l) is the weight matrix of the lth layer, σ is the activation function, and the ReLU activation function is used;

[0141] The formula for the ReLU function is:

[0142] ReLU(x)=max(0,x)

[0143] Its meaning is that when the input x is greater than 0, the output of the function is x; when the input x is less than or equal to 0, the output of the function is 0;

[0144] The introduction of residual connections can solve the gradient vanishing problem in deep network training. The residual connection directly jumps the input to the output. The formula is:

[0145] H (l+1) =H (l) +GraphConv(H (l) )

[0146] This connection structure enables the model to retain effective input information even at a deeper level, ensuring gradient stability during training.

[0147] Stacking multiple graph convolution layers can capture deeper graph structure information. Multi-layer graph convolution enables each node to receive feature information from more distant neighboring nodes. Assuming there are L layers of graph convolution, the final output is expressed as:

[0148] H (L) =GCN L (GCN (L-1) (...GCN1(H 0)...))

[0149] Each layer is connected through residual connections and uses appropriate activation functions and regularization.

[0150] Step S4.3: Output layer design. The output layer maps the results of graph convolution to specific load prediction values through a fully connected layer, ultimately obtaining the predicted load value of each node.

[0151] Step S4.4: Self-supervised task design. The self-supervised task is divided into two subtasks: time reconstruction and future value prediction.

[0152] The time series reconstruction task includes data preprocessing and masking and autoencoder model construction:

[0153] The time series X with load data after preprocessing is T}, randomly mask some time steps (for example, some x t Set as missing or noise data), forming a partial observation sequence X mask ;

[0154] The autoencoder is constructed using a Mamba network combined with a graph convolution module and a residual connection module. The encoder is responsible for extracting the temporal and spatial implicit features of dynamic graphs and time series, while the decoder reconstructs the original sequence based on z.

[0155] The reconstruction process can be expressed as:

[0156]

[0157] Where M is the set of masked time steps, f enc (·) represents the encoder (composed of Mamba network and graph convolution module), f dec (·) is the decoder;

[0158] Define the reconstruction loss (such as mean square error, MSE). By minimizing this loss, the model automatically learns the spatiotemporal features in the time series and captures the inherent laws of equipment load.

[0159] The future value prediction task includes data segmentation and model prediction:

[0160] The time series with load data is divided into past window and future forecast window. Given the sequence X of the past W time steps t-W+1:t ={x t-W+1 ,...,x t} and the target sequence Y for the next H time steps t+1:t+H ={x t+1 ,...,x t+H};

[0161] Using the spatiotemporal model composed of the same Mamba network and graph convolution residual module, the past window X t-W+1:t Encode and output the predicted value of the next H steps

[0162] The prediction formula can be expressed as:

[0163]

[0164] In the model, it is also necessary to ensure that the temporal features are aligned with the output of the graph convolution layer. Specifically, after the node features of the graph convolution layer are output, they are combined with the temporal features processed by the Mamba network to form a joint spatiotemporal representation. By combining the temporal features captured by the Mamba network with the spatial topology information of the graph convolution, the model's performance in the device load decomposition task is improved.

[0165] In one embodiment, step S5 specifically includes:

[0166] Step S5.1: In this embodiment, the pre-processed power load data is used as the input of the model for model training and prediction. The specific input data includes total load data (representing the overall power consumption of users, as the feature input of the central node) and the device feature matrix (including the load data of each device, as the sub-node feature of the graph structure). The model performs time series prediction by inputting these features and outputs the predicted load value of each device at the future moment. The input data uses the processed REFIT dataset, which contains the real electricity consumption records of multiple users. After pre-processing, the data is input into the model as a spatiotemporal joint feature;

[0167] Step S5.2: To measure the difference between the model prediction value and the actual equipment load value, this embodiment uses the mean square error (MSE) as the loss function. MSE is a loss function commonly used in regression tasks. It calculates the square error between the predicted value and the true value and takes the average to measure the prediction accuracy of the model. Its formula is:

[0168]

[0169] Among them, P pred is the model prediction value, P true is the true value, n is the number of samples;

[0170] Step S5.3: To optimize the model parameters, this embodiment uses the gradient descent method and the Adam optimizer. The gradient descent method calculates the gradient of the loss function and gradually updates the model weights, causing the loss function to gradually decrease, thereby achieving the goal of optimizing the model. The Adam optimizer is a widely used optimization algorithm that combines momentum and adaptive learning rate to adjust model parameters and has the characteristics of rapid convergence. Its update rules are as follows:

[0171]

[0172] Among them, θ t is the model parameter, m t and v t are momentum and acceleration estimates, respectively, and η is the learning rate;

[0173] Step S5.4: Since the input data has been standardized during the data preprocessing phase, the model's predicted values and true values need to be denormalized when calculating the loss function to ensure that the calculated loss reflects the actual differences in the power data. The denormalization formula is:

[0174] P true =P′×(P max -P min )+P min

[0175] Through denormalization, when the model predicts load data, the output is consistent with the load unit of the actual user load, ensuring that the model prediction is consistent with the actual situation.

[0176] Step S5.5: During model training, this embodiment uses a validation set to evaluate model performance to avoid overfitting. Before training begins, the dataset is divided into a training set and a validation set in an 8:2 ratio. The training set is used to update model parameters, while the validation set is used to monitor the model's generalization performance.

[0177] After each training epoch, the validation set loss is calculated. If the validation set loss starts to rise while the training set loss continues to fall, this indicates that the model may be overfitting.

[0178] To prevent overfitting, this example introduces an early stopping method. When the validation set loss stops decreasing within a certain training cycle, training is stopped early, thus preventing the model from overfitting on the training set.

[0179] Step S5.6: Based on the training results, save the optimal model with the best effect and record the relevant model parameters.

[0180] In one embodiment, step S6 specifically includes:

[0181] Step S6.1: To comprehensively evaluate the model's performance in the load decomposition task, this embodiment uses multiple evaluation metrics, including mean squared error (MSE) and mean absolute error (MAE), to further assess the model's performance. These two metrics measure the difference between the model's predicted value and the actual value.

[0182] The MSE formula is as follows:

[0183]

[0184] The MAE formula is as follows:

[0185]

[0186] Among them, P pred is the model prediction value, P true is the true value, n is the number of samples;

[0187] Step S6.2: To intuitively demonstrate the prediction effect of the model, this embodiment performs a visual analysis by drawing a comparison chart of the predicted value and the actual value. The specific steps are as follows:

[0188] Figure 2 A curve chart of the decomposed predicted value of each device is drawn to observe the load consumption of the model on each device.

[0189] Step S6.3: This embodiment further tests the adaptability and stability of the model under different power usage scenarios. The specific analysis includes:

[0190] Testing electricity usage data from different scenarios: The model was tested on electricity data from multiple scenarios to analyze its generalization capabilities under different data. The tests covered scenarios with different electricity usage time periods, different equipment combinations, and different user electricity usage behaviors to analyze the model's adaptability.

[0191] Model Explanation and Improvement Suggestions: Analyze the model output to explain how the model decomposes the total load. For devices or scenarios with large prediction errors, optimize the model, such as adjusting the depth of the graph convolution layer or optimizing the processing of time series features.

[0192] In one embodiment, step S7 specifically includes:

[0193] Step S7.1: After the non-intrusive load monitoring model of this embodiment is trained, it is first necessary to select a suitable method for deployment so as to process real-time data from the user's power monitoring equipment in actual applications;

[0194] Step S7.2: Based on actual application scenarios and feedback, continuously tune the model to better adapt to different application scenarios. For example, the model can be continuously trained to optimize performance based on the ever-expanding non-intrusive load monitoring dataset in real applications. Alternatively, the convolutional layer structure of the model or the model hyperparameters can be adjusted based on user feedback on power consumption characteristics to improve the model's adaptability and accuracy.

[0195] This example collects a dataset called REFIT, which is specifically designed for non-intrusive load research. The dataset contains electricity load data for multiple user homes, recording the total electricity load of the users and the load usage of each device. The data is recorded at a high frequency (usually 1 Hz), which can capture subtle changes in electricity consumption and help accurately identify the on / off status of the device. The dataset covers a variety of devices, including refrigerators, washing machines, and televisions, providing a variety of electricity usage patterns. The data is collected in a real environment, reflecting the actual electricity consumption in daily life, enhancing the realism and practicality of the research. The dataset covers a long time period, including electricity consumption in different seasons and time periods, which is convenient for analysis and modeling. The specific content and features are shown in the following figure.

[0196]

[0197]

[0198] In this embodiment, the decomposition results obtained using the non-intrusive load monitoring based on the Mamba-graph convolution residual structure and the self-supervised task are as follows: Figure 2 shown.

[0199] Figure 2 The figure shows the power load of each device decomposed by the model during a specific time step. This figure demonstrates that the model can effectively identify and separate the power usage patterns of each device and extract the complex relationships between devices. This demonstrates the advantages of dynamic graph input and allows users to understand the specific power usage of devices, providing valuable data reference for energy management and energy-saving strategies.

[0200] Overall, Figure 2 Analysis shows that the proposed model, by constructing a user-device hierarchical graph data structure, achieves high prediction accuracy for non-intrusive load monitoring and enhances the ability to decompose device loads. This not only improves the practicality and effectiveness of non-intrusive load monitoring but also provides strong support for optimizing user energy management.

[0201] Example 2

[0202] This embodiment provides a non-intrusive load monitoring system based on a Mamba-graph convolutional network, including:

[0203] Data acquisition, processing and enhancement module: This module is responsible for collecting and processing user power load data to ensure that the data can provide standardized and high-quality input for subsequent models. In specific implementation, the system collects time series data of the total user load and the load of each device in real time through smart meters or sensors, and saves the data in a structured table (such as a CSV file) to ensure the efficiency and operability of data processing. In the data processing stage, the system performs missing value filling, feature scaling and standardization on the collected data to ensure the integrity and consistency of the input data; at the same time, it enriches the diversity of the data set through data enhancement techniques such as noise addition and time series truncation, and provides input support for self-supervised tasks (such as time series reconstruction and future value prediction), helping the model to automatically learn intrinsic spatiotemporal characteristics.

[0204] Graph Structure Construction Module: This module converts preprocessed power load data into a dynamic graph structure, where each node represents a power load data point. The central node represents the total user load, and several subnodes correspond to the load of each device. Between nodes, statistical methods are used to calculate inter-device correlations or load trends (for example, based on simultaneous device on / off conditions). This generates a symmetric adjacency matrix reflecting dynamic associations and initializes features for each node. This dynamic graph structure not only provides a foundation for subsequent graph convolution operations but also supports the effective extraction of key features in self-supervised spatial relationship modeling.

[0205] Load forecasting decomposition model construction module: This module is dedicated to building and training a spatiotemporal deep learning model based on the Mamba-graph convolution residual structure for non-intrusive load monitoring. The model first uses the Mamba network to efficiently capture the temporal dependency features in long time series data, solving the computational complexity and insufficient dependency capture problems in traditional long series modeling; then, it extracts the spatial dependencies in the dynamic graph through a multi-layer graph convolution module, and adopts residual connections to ensure the effective fusion of deep features and gradient transfer. In addition, self-supervised task modules (such as time series reconstruction and future value prediction tasks) are embedded in the model to automatically extract and enhance spatiotemporal features in the pre-training stage, thereby reducing dependence on large-scale labeled data and improving the robustness and generalization ability of the model; finally, the model decomposes the total user load to achieve accurate prediction of the future load consumption of each device.

[0206] Data Acquisition Module: This module is responsible for inputting pre-processed power load data into the trained model and obtaining load forecast results for each device. The model outputs the predicted data after processing and stores it for subsequent data analysis and energy consumption decision-making.

[0207] Data Results Visualization Module: This module provides a visual display of the load breakdown results generated by the model. The system compares the predicted load of each device with the actual load, generates a chart, and displays it on the user interface. This visualization includes device load breakdown diagrams, power consumption monitoring reports, energy-saving recommendations, and energy consumption analysis reports, helping users more intuitively manage and optimize power usage.

[0208] Model Optimization and Control Module: This module focuses on continuous optimization of model performance. By continuously enriching the data set, improving optimization algorithms, and dynamically adjusting hyperparameters, combined with actual application feedback, the model is continuously iteratively trained and tuned, further improving the accuracy of load decomposition and the system's generalization capabilities.

[0209] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0210] Obviously, those skilled in the art may make various changes and modifications to the embodiments of the present invention without departing from the spirit and scope of the embodiments of the present invention. Thus, if such changes and modifications of the embodiments of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is intended to include such changes and modifications.

[0211] Other parts not described in detail are prior art.

Claims

1. A non-intrusive load monitoring method based on Mamba-graph convolutional network, characterized in that: include: Step S1: Collect user power load data, including user load and load data of each device; Step S2: pre-processing the collected power data; Step S3: constructing a dynamic graph structure using the pre-processed power load data; Step S4: constructing a Mamba-graph convolution deep learning model that combines the Mamba network and the graph convolution network, and embedding a self-supervised task module to perform non-intrusive load monitoring on the data in the dynamic graph structure; Step S5: Divide the collected user power load data into a training set, a validation set, and a test set, and train the Mamba-graph convolutional deep learning model to obtain an optimal model; Step S6: Evaluate and test the generalization ability of the optimal model; Step S7: deploying the optimal model to an actual non-intrusive load monitoring application system and optimizing it according to specific application feedback; deploying the optimal model to the cloud for real-time load monitoring.

2. The non-intrusive load monitoring method based on Mamba-graph convolutional network according to claim 1, characterized in that: The step S1 specifically includes the following steps: Step S1.1: Setting a sampling frequency, collecting user total load data from smart meters or other sensors at the sampling frequency, and recording the load consumption of users and the load consumption of each device in a time series; Step S1.2: Collect long-term user total load and load data of each device and store them in a structured data format.

3. The non-intrusive load monitoring method based on Mamba-graph convolutional network according to claim 1, characterized in that: The pre-processing in step S2 includes: Step S2.1: Fill missing values in the user's total load and each device's load data using a linear interpolation method; Step S2.2: Normalize the total user load and the load data of each device by using the minimum-maximum normalization method to scale the data to the [0, 1] interval and convert the data to the same dimension; Step S2.3: Adding Gaussian noise to the total user load and the load data of each device to simulate measurement errors and environmental fluctuations to enhance the robustness of the model; Step S2.4: performing time series truncation on the total user load and the load data of each device, and randomly selecting a time window of fixed length to generate a new sample.

4. The non-intrusive load monitoring method based on Mamba-graph convolutional network according to claim 1, characterized in that: Step S3 specifically includes: Step S3.1: Organize the power load data into a dynamic graph according to the time series, where each node represents a power load data; the central node represents the total user load, and several sub-nodes correspond to the load of each device; Step S3.2: Determine the connection relationship between nodes by calculating the similarity or mutual relationship between each device, define the edges between nodes, and set the weight of the edges based on the frequency of simultaneous power-on or power-off of devices and the correlation of load changes, thereby constructing a dynamic graph structure.

5. The non-intrusive load monitoring method based on Mamba-graph convolutional network according to claim 1, characterized in that: The Mamba-graph convolution deep learning model in step S4 is specifically as follows: Step S4.1: Use the Mamba network to process the time series features and combine each time series feature with the node feature to form a spatiotemporal joint representation; Step S4.2: Stack multiple graph convolutional layers, based on a multi-layer graph convolutional network, combined with residual connections to ensure that the temporal features are aligned with the graph structure; Step S4.3: Design the output layer through a fully connected layer to map the results of graph convolution to specific load prediction values, and finally obtain the predicted load value of each node.

6. The non-intrusive load monitoring method based on Mamba-graph convolutional network according to claim 5, characterized in that: The self-supervisory task module in step S4 is specifically designed as follows: The self-supervised task module includes a time series reconstruction subtask and a future value prediction subtask; Among them, the time series reconstruction task includes data preprocessing and masking and autoencoder model construction: The time series X with load data after preprocessing is T }, randomly mask some time steps to form a partial observation sequence X mask ; The autoencoder is constructed by combining the Mamba network with the graph convolution module and the residual connection module. It is responsible for extracting the spatiotemporal implicit features z of dynamic graphs and time series data, and reconstructing the original sequence based on the spatiotemporal implicit features based on the decoder. The future value prediction task includes data segmentation and model prediction: Divide the time series with load data into past window and future forecast window; Using the spatiotemporal model composed of the same Mamba network combined with the graph convolution module and the residual connection module, the past window X t-w+1:t Encode and output the predicted value of the future prediction window 7. The non-intrusive load monitoring method based on Mamba-graph convolutional network according to claim 1, characterized in that: Step S5 specifically includes: Step S5.1: The pre-processed training set load data is used as input to the Mamba-graph convolutional deep learning model for training to predict the load of each device at a future time. Step S5.2: define the loss function; Step S5.3: adjusting the parameters using a gradient descent method and an optimization algorithm; Step S5.4: Calculate the loss based on the loss function and denormalize the predicted value and the true value; Step S5.5: Use the validation set to evaluate model performance during training and perform model validation. Step S5.6: Based on the training results, save the best model and record the optimal model parameters.

8. The non-intrusive load monitoring method based on Mamba-graph convolutional network according to claim 1, characterized in that: Step S6 specifically includes: Step S6.1: Based on the Mamba-graph convolutional deep learning model, the change in total load is distributed to the predicted value of each device to obtain the load consumption forecast of each device at a future time; Step S6.2: Evaluate the trained Mamba-graph convolutional deep learning model on different test sets and datasets of different electricity usage scenarios; and calculate the accuracy of the model. Step S6.3: Plot the predicted value and actual value of load consumption of each device on the same graph for visual comparison.

9. The non-intrusive load monitoring method based on Mamba-graph convolutional network according to claim 1, characterized in that: Step S7 specifically includes: Step S7.1: Deploy the optimal model to the cloud server and design a real-time data input interface for the Mamba-graph convolutional deep learning model to ensure that it can receive real-time data from power monitoring equipment; Step S7.2: Continuously tune the optimal model.

10. A non-intrusive load monitoring system based on Mamba-graph convolutional network, characterized in that: include: Data acquisition, processing and enhancement module: This module collects user power load data, including the user's total load and the load data of each device, and pre-processes the collected power load data; Dynamic graph structure construction module: This module uses the pre-processed power load data to construct a dynamic graph structure; Load forecasting decomposition model construction module: This module is used to build a Mamba-graph convolutional deep learning model that combines the Mamba network and the graph convolutional network, and embeds a self-supervised task module to perform non-intrusive load monitoring on the data in the dynamic graph structure; Data result acquisition module: This module is responsible for inputting the processed power load data into the trained Mamba-graph convolutional deep learning model and obtaining the load forecast results of each device. The system stores the forecast results for subsequent data analysis and energy consumption decision-making; Data Result Visualization Module: This module is used to visualize the prediction results output by the Mamba-graph convolutional deep learning model. It generates personalized energy-saving suggestions and energy consumption analysis reports based on the electricity usage characteristics of different users. The reports contain multiple charts and key indicators to help users intuitively manage their electricity usage. Model Optimization and Control Module: This module is used to combine actual application feedback, enrich the data set, improve the optimization algorithm, and adjust hyperparameters to continuously iterate the model training and monitor model performance, thereby further improving the accuracy of load decomposition and the generalization ability of the system. The Mamba-graph convolution residual structure and the non-intrusive load monitoring system for self-supervised tasks are used to perform the steps in the Mamba-graph convolution residual structure and the non-intrusive load monitoring method for self-supervised tasks according to any one of claims 1 to 9.

Images