A non-intrusive power load decomposition method and system based on multi-modal feature learning

Abstract

This invention relates to the field of power load decomposition technology, and discloses a non-intrusive power load decomposition method and system based on multimodal feature learning. The method involves: synchronously collecting power parameter data from smart meters, environmental parameter data from environmental sensors, and user equipment usage behavior data to obtain multimodal load monitoring data; extracting cross-modal features through the non-negative matrix decomposition layer of a first equipment status recognition model to obtain a multimodal fusion feature vector; identifying load patterns through the first decomposition layer of the first equipment status recognition model to obtain a first decomposed load matrix; performing clustering optimization through the second decomposition layer of the first equipment status recognition model to obtain a second load decomposition matrix; and performing dynamic fuzzy decision-making based on the second load decomposition matrix to obtain the equipment operating status identification result. This invention overcomes the limitations of traditional methods that rely solely on a single power signal, achieving high-precision and highly robust non-intrusive power load decomposition.

Description

Technical Field

[0001] This invention relates to the field of power load decomposition technology, and in particular to a non-intrusive power load decomposition method and system based on multimodal feature learning. Background Technology

[0002] With the rapid development of smart grid construction and the increasing complexity of user electricity consumption behavior, non-intrusive load decomposition technology, as a core technology for smart grid energy efficiency management, is playing an increasingly important role in power system optimization scheduling, user electricity consumption behavior analysis, and energy conservation and emission reduction. This technology identifies and separates the operating status and power consumption of individual electrical devices by analyzing the total load signal, providing crucial data support for the refined management of the power system.

[0003] Traditional load decomposition methods rely primarily on single electrical parameter characteristics for equipment identification. This approach is prone to misjudgment when faced with devices exhibiting similar electrical characteristics, especially when multiple devices operate simultaneously or their operating states change frequently. Power signals alone are insufficient to accurately distinguish the load contribution of different devices. Furthermore, traditional methods lack consideration for environmental factors and user behavior patterns, failing to fully utilize the contextual information of device operation. This results in low identification accuracy in complex power consumption scenarios and an inability to adapt to dynamic changes in user power consumption habits. Summary of the Invention

[0004] The main objective of this invention is to provide a non-intrusive power load decomposition method and system based on multimodal feature learning. This invention overcomes the limitations of traditional methods that rely solely on a single power signal, and achieves high-precision and highly robust non-intrusive power load decomposition.

[0005] To achieve the above objectives, this invention provides a non-intrusive power load decomposition method based on multimodal feature learning, comprising the following steps:

[0006] Simultaneously collect power parameter data from smart meters, environmental parameter data from environmental sensors, and user equipment usage behavior data to obtain multimodal load monitoring data;

[0007] The multimodal load monitoring data is input into the non-negative matrix decomposition layer of the first equipment status identification model for cross-modal feature extraction to obtain a multimodal fusion feature vector;

[0008] The multimodal fusion feature vector is input into the first decomposition layer of the first device state recognition model to perform load pattern recognition, thereby obtaining the first decomposed load matrix;

[0009] The first decomposed load matrix is ​​input into the second decomposed layer of the first equipment status identification model for cluster optimization to obtain the second load decomposed matrix;

[0010] Dynamic fuzzy decision-making is performed based on the second load decomposition matrix to obtain the equipment operating status identification result.

[0011] Optionally, in a first implementation of the first aspect of the present invention, the step of synchronously collecting power parameter data from smart meters, environmental parameter data from environmental sensors, and user equipment usage behavior data to obtain multimodal load monitoring data includes:

[0012] The three-phase voltage, three-phase current, active power, reactive power and power factor values ​​are collected synchronously by smart meters to form time-series power parameter data;

[0013] Temperature, humidity, light intensity, and carbon dioxide concentration values ​​are collected once by environmental sensors distributed indoors and outdoors to form time-series environmental parameter data;

[0014] User device usage behavior data is constructed by extracting the on / off time records, usage frequency statistics, and operation preference settings of each device from the user behavior database.

[0015] The power parameter data, the environmental parameter data, and the user equipment usage behavior data are time-aligned and synchronized to obtain multimodal load monitoring data.

[0016] Optionally, in a second implementation of the first aspect of the present invention, the step of inputting the multimodal load monitoring data into the non-negative matrix factorization layer of the first equipment status identification model for cross-modal feature extraction to obtain a multimodal fusion feature vector includes:

[0017] The multimodal load monitoring data is input into the non-negative matrix decomposition layer of the first equipment state identification model to construct a multimodal feature matrix combination that includes the observed load matrix, environmental feature matrix, and behavioral feature matrix;

[0018] Based on the environmental feature matrix and the behavioral feature matrix, a multimodal regularization constraint term is constructed for the non-negative matrix factorization algorithm in the non-negative matrix factorization layer, and the multimodal regularization constraint term is combined with the observation load matrix to construct a regularization constraint objective function;

[0019] The device feature matrix and activation coefficient matrix are obtained by minimizing the objective function with regularization constraints, and load decomposition features are generated based on the device feature matrix and activation coefficient matrix.

[0020] The load decomposition features, the environmental feature matrix, and the behavioral feature matrix are input into the attention weight calculation network in the non-negative matrix decomposition layer to calculate multimodal attention weights.

[0021] The load decomposition features are weighted and fused according to the multimodal attention weights to obtain a multimodal fused feature vector.

[0022] Optionally, in a third implementation of the first aspect of the present invention, the step of inputting the multimodal fusion feature vector into the first decomposition layer of the first device state recognition model for load pattern recognition to obtain a first decomposed load matrix includes:

[0023] The multimodal fusion feature vector is input into the first decomposition layer of the first device state recognition model. The standardized load feature vector is obtained by performing dimensional transformation and nonlinear mapping through the feature mapping sublayer in the first decomposition layer.

[0024] The standardized load feature vector is input into the load pattern recognition sublayer in the first decomposition layer for self-attention weight calculation to obtain the load pattern feature encoding vector.

[0025] The load pattern feature encoding vector is input into the pattern classification sub-layer in the first decomposition layer for sequence modeling and probability classification calculation to obtain load pattern classification probability distribution data.

[0026] The load pattern classification probability distribution data is input into the first matrix reconstruction sub-layer in the first decomposition layer. In the first matrix reconstruction sub-layer, the load pattern classification probability distribution data is reconstructed into a first decomposed load matrix with device and time dimensions.

[0027] Optionally, in a fourth implementation of the first aspect of the present invention, the step of inputting the first decomposed load matrix into the second decomposition layer of the first equipment state identification model for cluster optimization to obtain the second load decomposition matrix includes:

[0028] The first decomposed load matrix is ​​input into the second decomposed layer of the first equipment status identification model, and the normalized load feature matrix is ​​obtained by the feature normalization sublayer in the second decomposed layer.

[0029] The normalized load feature matrix is ​​input into the adaptive clustering sublayer in the second decomposition layer for density peak detection to obtain the initial clustering configuration parameters.

[0030] Based on the initial clustering configuration parameters, the normalized load feature matrix is ​​input into the deep clustering sublayer in the second decomposition layer for joint feature learning and clustering assignment to obtain the equipment load clustering assignment probability matrix.

[0031] The equipment load clustering allocation probability matrix is ​​input into the second matrix reconstruction sublayer in the second decomposition layer for probability weighting reconstruction to obtain the second load decomposition matrix.

[0032] Optionally, in a fifth implementation of the first aspect of the present invention, the step of performing dynamic fuzzy decision-making based on the second load decomposition matrix to obtain the equipment operating status identification result includes:

[0033] The second load decomposition matrix is ​​fuzzified to obtain a fuzzified load feature matrix, and a weighted fuzzy inference rule base is constructed based on the fuzzified load feature matrix.

[0034] The fuzzy load feature matrix is ​​input into the weighted fuzzy inference rule base for fuzzy inference to obtain the fuzzy output value of the running status.

[0035] Based on the load change trend and historical recognition accuracy within the current time window, the confidence weights of the weighted fuzzy inference rule base are dynamically adjusted to obtain updated rule weights. The fuzzy output value of the operating status is then recalculated based on the updated rule weights to obtain the dynamically adjusted fuzzy decision result for the equipment status.

[0036] The fuzzy decision results of the dynamically adjusted equipment status are defuzzified, and the continuous values ​​are converted into four discrete status categories: on, off, standby, and adjustment, to obtain the equipment operating status identification results.

[0037] Optionally, in a sixth implementation of the first aspect of the present invention, the non-intrusive power load decomposition method based on multimodal feature learning further includes:

[0038] The predicted state of the first device state recognition model is used as the agent's action, and the actual device state is used as the environment state. The state space transition rule is defined by the state transition probability matrix, and a multi-objective reward function is constructed based on the recognition accuracy, energy consumption prediction error and response time to obtain the reinforcement learning environment.

[0039] In the reinforcement learning environment, the device operating status identification result is compared and verified with the real device status, and the reinforcement learning reward signal is calculated.

[0040] Based on the reinforcement learning reward signal, policy learning is performed on the first device state recognition model to obtain the policy gradient vector;

[0041] The gradient update amount of each layer weight and bias in the first model parameter set is calculated based on the gradient vector of the strategy to obtain the second model parameter set;

[0042] The second device status recognition model is obtained by replacing the original parameters of the corresponding levels in the first device status recognition model with the parameters of the second model.

[0043] Optionally, in a seventh implementation of the first aspect of the present invention, the step of calculating the gradient update amount of the weights and biases of each layer in the first model parameter set according to the policy gradient vector to obtain the second model parameter set includes:

[0044] Based on the policy gradient vector, calculate the hierarchical gradient coefficients of the nonnegative matrix factorization layer, the first factorization layer, and the second factorization layer, and combine the first and second moments of the Adam optimizer to calculate the gradient update amount of the weight matrix and bias vector of each layer in the first model parameter set.

[0045] The gradient update amount is passed from the output layer to the input layer layer by layer through the backpropagation algorithm, and the numerical range of the gradient update amount is limited based on the gradient clipping mechanism to obtain the hierarchical gradient allocation result.

[0046] Based on the hierarchical gradient allocation results, the weights of the device feature matrix and the activation coefficient matrix in the non-negative matrix decomposition layer are updated respectively. The weight parameters and bias parameters of the feature mapping sublayer, load pattern recognition sublayer, pattern classification sublayer and first matrix reconstruction sublayer in the first decomposition layer are updated. The weight parameters and bias parameters of the feature normalization sublayer, adaptive clustering sublayer, deep clustering sublayer and second matrix reconstruction sublayer in the second decomposition layer are updated to obtain the updated parameter matrix of each level.

[0047] The updated parameter matrices at each level are reorganized and indexed according to the network hierarchy of the first device state recognition model to obtain the second model parameter set.

[0048] Optionally, in an eighth implementation of the first aspect of the present invention, the step of replacing the original parameters of the corresponding levels in the first device state recognition model with the second model parameter set to obtain the second device state recognition model includes:

[0049] The weight parameters and bias parameters in the second model parameter set are classified and labeled according to the layer name, sub-layer identifier and parameter type, and matched with the original parameters of the non-negative matrix factorization layer, the first decomposition layer and the second decomposition layer in the first device state recognition model to obtain the parameter replacement mapping relationship.

[0050] According to the parameter replacement mapping relationship, the device feature matrix weights, activation coefficient matrix weights and attention weights in the non-negative matrix decomposition layer are replaced and updated respectively. The weight matrix, bias vector and activation function parameters of each sub-layer in the first decomposition layer are replaced and updated. The weight matrix, bias vector and cluster center parameters of each sub-layer in the second decomposition layer are replaced and updated respectively, so as to obtain the model hierarchical structure with complete parameter replacement.

[0051] Perform network connectivity verification on the model hierarchy structure after the parameter replacement is completed to obtain a network configuration that passes the structure verification.

[0052] Based on the network configuration that has passed the structural verification, the forward propagation path and backward propagation path between the non-negative matrix factorization layer, the first factorization layer and the second factorization layer are reconstructed using a computational graph reconstruction algorithm to obtain the second device state recognition model.

[0053] The present invention also provides a non-intrusive power load decomposition system based on multimodal feature learning, comprising:

[0054] The synchronous acquisition unit is used to synchronously acquire power parameter data from smart meters, environmental parameter data from environmental sensors, and user equipment usage behavior data to obtain multimodal load monitoring data.

[0055] The feature extraction unit is used to input the multimodal load monitoring data into the non-negative matrix decomposition layer of the first equipment status identification model for cross-modal feature extraction to obtain a multimodal fusion feature vector;

[0056] The load pattern recognition unit is used to input the multimodal fusion feature vector into the first decomposition layer of the first equipment state recognition model to perform load pattern recognition and obtain the first decomposed load matrix.

[0057] The clustering optimization unit is used to input the first decomposed load matrix into the second decomposition layer of the first equipment state identification model for clustering optimization to obtain the second load decomposition matrix;

[0058] The dynamic fuzzy decision unit is used to perform dynamic fuzzy decision-making based on the second load decomposition matrix to obtain the equipment operating status identification result.

[0059] In summary, the technical solution provided by this invention constructs multimodal load monitoring data by simultaneously collecting power parameters, environmental parameters, and user behavior data. This overcomes the limitations of traditional methods that rely solely on a single power signal, enabling a comprehensive characterization of equipment operation features from multiple dimensions. A non-negative matrix factorization algorithm combined with multimodal regularization constraints and an attention weight calculation network achieves optimal fusion of cross-modal information and significantly improves feature representation capabilities. A hierarchical progressive architecture, constructed through load pattern recognition in the first decomposition layer and clustering optimization in the second decomposition layer, enables gradual optimization from coarse-grained to fine-grained levels. A dynamic fuzzy decision-making mechanism effectively handles uncertainties in load decomposition, enhancing the system's adaptability to complex scenarios. The introduction of a reinforcement learning environment for policy learning and adaptive parameter updates allows the model to continuously optimize and improve based on actual performance. A refined hierarchical gradient allocation and parameter replacement mechanism ensures the accuracy of model reconstruction and the integrity of the network structure, achieving high-precision, highly robust, and non-intrusive power load decomposition. Attached Figure Description

[0060] Figure 1 This is a schematic diagram of the steps of a non-intrusive power load decomposition method based on multimodal feature learning in one embodiment of the present invention;

[0061] Figure 2 This is a block diagram of a non-intrusive power load decomposition system based on multimodal feature learning in one embodiment of the present invention.

[0062] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0063] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0064] Reference Figure 1 This embodiment provides a non-intrusive power load decomposition method based on multimodal feature learning, including the following steps:

[0065] S1 synchronously collects power parameter data from smart meters, environmental parameter data from environmental sensors, and user equipment usage behavior data to obtain multimodal load monitoring data;

[0066] Using a unified time base as the synchronization thread, the system achieves parallel acquisition of power, environmental, and behavioral parameters throughout the entire process through the joint collaboration of smart meter systems, environmental sensor networks, and user behavior databases. Regarding power parameter acquisition, smart meter devices installed in user distribution boxes or meter nodes continuously acquire basic electrical parameters such as three-phase voltage, three-phase current, active power, reactive power, and power factor at a preset sampling frequency (e.g., 1Hz), forming a time-series power parameter data stream reflecting the overall load status of the equipment. Simultaneously, in the environmental data acquisition path, multiple sensor nodes are deployed in key indoor areas (such as kitchens, living rooms, and areas with concentrated appliances) and external ventilation openings or window areas to acquire key parameters reflecting changes in the physical environment, including temperature, humidity, light intensity, and carbon dioxide concentration. Data from each channel is written into the environmental sensor time series at equal periods, forming a time-series environmental parameter data with continuity and spatial variability. Meanwhile, the construction of user behavior data relies on long-term accumulated equipment operation logs and user preference settings. By accessing smart gateways, mobile terminal control records, and usage habit learning systems, historical on / off timestamp sequences, typical usage frequency statistics, and user-personalized settings (such as timing strategies, temperature settings, power limits, etc.) of various electrical appliances are extracted from the behavior database. A device usage probability matrix or behavior configuration vector with daily periodic characteristics is generated through behavior time indexing, thus constituting user equipment usage behavior data. Before data fusion, the above three types of heterogeneous data undergo time alignment processing. That is, using the power parameter sampling time as the global time base, environmental parameters and behavior data are interpolated and mapped according to timestamps. Simultaneously, a sliding window mechanism (such as a 60-second window with 50% overlap) is used for data segmentation and organization. Through the above processing, the original multi-source heterogeneous data is uniformly converted into a data block cluster with aligned time indexes. Each time window contains a set of three-dimensional structured data: power parameter vector, environmental parameter vector, and behavior feature vector, collectively constituting multimodal load monitoring data.

[0067] S2, input the multimodal load monitoring data into the non-negative matrix decomposition layer of the first equipment status identification model to extract cross-modal features and obtain a multimodal fusion feature vector;

[0068] Specifically, multimodal load monitoring data is input into the non-negative matrix decomposition layer constructed within the first equipment state identification model. Based on this, a triple data structure is established, comprising an observed load matrix of power parameters, an environmental impact factor feature matrix, and a user behavior pattern feature matrix, forming a multimodal feature matrix combination with strong representational capabilities. This combination uses the observed load matrix as the basic carrier to capture the actual power response information of the equipment; simultaneously, it uses the environmental feature matrix and behavioral feature matrix as external constraint dimensions to provide an auxiliary interpretation mechanism for physical disturbances and behavioral drives. Under this structure, the constructed non-negative matrix decomposition objective not only focuses on minimizing the traditional matrix approximation reconstruction error objective but also introduces multimodal regularization constraint terms to achieve modeling and coordination of the correlation between various modal features. The environmental feature matrix and behavioral feature matrix establish regularization constraints with the equipment feature matrix and activation coefficient matrix to be solved, respectively. By constructing the regularization objective function of the non-negative matrix decomposition, the deviation between the feature factor matrix and the environmental behavior vector is incorporated into the optimization objective, thereby achieving structural guidance for fusion consistency. This objective function retains the mathematical properties of nonnegativity and low-rank decomposition, and improves the interpretability and behavioral adaptability of the analytical structure by introducing modal coupling terms. During the algorithm's solution process, an iterative update mechanism is used to jointly minimize and iteratively optimize the device feature matrix and activation coefficient matrix, obtaining decomposition features with clear physical meaning. The device feature matrix represents the power characteristic shape of each typical load mode, while the activation coefficient matrix represents the activation degree of each device on the time axis. This decomposition result, along with the original environmental feature matrix and behavioral feature matrix, is input into an attention weight calculation network embedded in the nonnegative matrix decomposition layer. The latter constructs a learnable attention mechanism based on the influence weights of different modal features on load changes at the current moment. This mechanism assigns dynamic response weights to each modal feature through softmax normalization and feature channel weighting operations, establishing a selective and sensitive mapping of power load characteristics to external factors such as ambient temperature, illumination, and user switching habits under different scenarios. Based on this attention mechanism, a weighted summation and fusion operation is performed on the load decomposition features, and the power feature backbone and the environmental and behavioral feature branches are fused in the feature space to generate a multimodal fusion feature vector.

[0069] S3, input the multimodal fusion feature vector into the first decomposition layer of the first equipment state recognition model to perform load pattern recognition, and obtain the first decomposed load matrix;

[0070] It should be noted that the multimodal fusion feature vector is input into the first decomposition layer of the first device state recognition model. This sub-layer achieves dimensional transformation and nonlinear mapping operations through a stacked structured fully connected network. This mapping process embeds and reconstructs the original feature space through a linear weight matrix and introduces nonlinear expressive power through the ReLU activation function. Simultaneously, Dropout technology is used to regularize the network structure during training to suppress overfitting risk, thereby generating a standardized load feature vector with uniform scale, good continuity, and stable expression. The standardized load feature vector is then input into the load pattern recognition sub-layer in the first decomposition layer. This sub-layer introduces a self-attention mechanism based on the Transformer architecture. A multi-head attention computation module performs global dependency modeling on the information of each dimension in the input vector sequence. In each attention head, a dot product weighting operation is performed between the query (Q), key (K), and value (V) vectors, and positional encoding is combined to enhance the temporal context representation capability of the sequence. This mechanism can dynamically capture key temporal relationships and potential correlation activation patterns between devices in the device load features. Through stacked multi-layer encoders, the feature representation is hierarchically abstracted and multi-scale modeled to obtain a load pattern feature encoding vector representing the load behavior of each device. The load pattern feature encoding vector is input into the pattern classification sublayer in the first decomposition layer. This sublayer consists of a bidirectional LSTM network and a probability output module. The bidirectional LSTM, with its forward and backward structures, simultaneously models the state transition features of the load time series signal at past and future time points, effectively enhancing the model's ability to perceive contextual information in long sequences. The LSTM output state, after dense connection and standardization, is input into the classification layer composed of a Softmax activation function. This layer estimates the multi-class probability of the possible load patterns for each device, resulting in a discriminative load pattern classification probability distribution. This distribution reflects the activation probability of different device load states within each time window and retains the ability to remember and predict historical load trajectories. The load pattern classification probability distribution data is then input into the first matrix reconstruction sublayer. This sublayer performs structural mapping through a set of learnable linear transformation matrices and bias terms, mapping the probability distribution to specific load values. The reconstructed sublayer performs weighted aggregation of the classification probability vectors for each device, transforms the classification results into corresponding device power estimation curves through matrix multiplication, and establishes a two-dimensional data matrix with a device-time structure by combining the time and device dimensions, thereby restoring the physical magnitude of the load pattern classification results. The reconstructed result is the first decomposed load matrix, reflecting the load proportion and state change trends of different devices within a specific time window.

[0071] S4, input the first decomposed load matrix into the second decomposed layer of the first equipment status identification model for cluster optimization to obtain the second load decomposed matrix;

[0072] Specifically, the first decomposed load matrix is ​​input into the second decomposed layer of the first equipment status recognition model. The feature normalization sublayer of the second decomposed layer standardizes each feature of this matrix. This normalization process employs a batch normalization strategy, performing zero-mean and unit-variance transformations on each row vector element of the matrix to bring the distribution of all samples in the feature space back to a uniform scale, thereby eliminating clustering bias caused by inconsistencies in dimensions. Simultaneously, learnable scaling coefficients and bias terms are introduced to maintain adaptability to changes in numerical distribution, outputting a normalized load feature matrix. The normalized load feature matrix is ​​then input into the adaptive clustering sublayer of the second decomposed layer. A density peak detection algorithm based on an improved K-means++ mechanism is used to perform a holistic analysis of the normalized feature vectors. This method assesses the density of sample distribution by calculating the cosine similarity between samples, then identifies regions with significant density in the feature space using a density estimation function. Furthermore, based on the product of local density and minimum relative distance as an index function, the location and number of cluster centers are automatically selected, constructing an initial clustering configuration parameter set containing an initial set of cluster center vectors and dynamic cluster quantity estimates. Based on the initial clustering configuration parameters, the normalized load feature matrix is ​​input into the deep clustering sublayer in the second decomposition layer to perform joint feature learning and cluster assignment optimization. This sublayer consists of an embedding network and a clustering network. The embedding network is composed of three fully connected neural networks, used to compress and map the input high-dimensional normalized vector into a compact representation in a low-dimensional embedding space, enhancing the aggregation and separability between features. This low-dimensional embedded feature vector is input into the clustering network. By introducing a soft assignment mechanism, each feature vector is mapped to its probability membership distribution in each cluster center. At the same time, KL divergence is introduced as an optimization objective to minimize the relative information difference between the model's cluster label assignment and feature representation space, so as to achieve the effect of synergistic convergence of joint clustering and representation learning, resulting in an equipment load cluster assignment probability matrix. This matrix contains the probability information of each equipment load belonging to different cluster centers in each time window, and expresses the structural characteristics of fuzzy assignment in a differentiable form. The equipment load cluster assignment probability matrix is ​​then input into the second matrix reconstruction sublayer in the second decomposition layer to perform a load reconstruction process based on a probability weighting mechanism. In this process, the model aggregates the load pattern features corresponding to each cluster category, that is, it constructs a representative cluster load vector by weighted average pooling, and then weights and fuses the representative loads of each category according to the cluster allocation probability of the device at each time point, thereby generating a continuous load value matrix containing device dimension and time dimension, and obtaining the second load decomposition matrix.

[0073] S5. Based on the second load decomposition matrix, perform dynamic fuzzy decision-making to obtain the equipment operating status identification result.

[0074] The second load decomposition matrix is ​​fuzzified, mapping the original continuous numerical load data to a fuzzy set space. Specifically, various membership functions (such as triangular and Gaussian functions) are constructed to divide load intensity level intervals, mapping the load value of each device at each time point to fuzzy sets such as "low load," "medium load," and "high load," and calculating its membership degree in each set, thus forming a fuzzified load feature matrix. This matrix represents the state distribution of each device's load in the fuzzy space using fuzzy vectors. Based on this fuzzified feature matrix, a weighted fuzzy inference rule base is constructed. This rule base is defined in the form of IF-THEN rules, covering the correspondence between typical device behavior states and load changes, and assigning an initial confidence weight to each rule to reflect its reliability during system initialization. This rule base includes static discrimination rules and considers load change trends, usage periods, and the continuity of previous and subsequent states, enabling it to handle dynamic load states. After the rule base is constructed, the aforementioned fuzzy load feature matrix is ​​input into the weighted fuzzy inference rule base for fuzzy inference calculation. This stage employs a typical Mamdani fuzzy inference mechanism, performing a minimum activation operation on the membership degrees of rule conditions. Then, at the rule level, the activation results are fused using a maximum value synthesis method, and the centroid method is combined to perform numerical regression on the inference results, obtaining the fuzzy output value of each device's operating status at the current time point. This output value is a continuous variable between 0 and 1, representing the degree of fuzziness between the "on" and "off" states of the device. To improve the dynamic response capability and adaptive level of the inference, the confidence weights in the existing rule base are dynamically adjusted based on the load change trend within the current time window and historical recognition accuracy. The historical effect of each rule is evaluated using the exponentially weighted moving average method or accuracy statistics within the sliding time window, and the corresponding weights are updated based on their actual recognition errors, thereby strengthening the activation priority of efficient rules and weakening the influence weight of inefficient rules. The updated weights are reloaded into the fuzzy inference rule base, and the fuzzy inference process is executed again to generate dynamically adjusted fuzzy decision results for the device status. The dynamically adjusted fuzzy decision results are defuzzified by using a centroid defuzzification method to map continuous output values ​​into operable discrete state categories. These categories are then classified using preset state thresholds. For example, values ​​greater than 0.75 are mapped to "on", values ​​less than 0.25 are mapped to "off", and values ​​in between are further classified as "standby" or "adjustment" based on the trend of the preceding and following states. This process ultimately enables the identification and classification of the device's operating status, resulting in a real-time, stable, and intelligent state output.

[0075] In one example, power parameter data from smart meters, environmental parameter data from environmental sensors, and user equipment usage behavior data are collected simultaneously to obtain multimodal load monitoring data, including:

[0076] The three-phase voltage, three-phase current, active power, reactive power and power factor values ​​are collected synchronously by smart meters to form time-series power parameter data;

[0077] Temperature, humidity, light intensity, and carbon dioxide concentration values ​​are collected once by environmental sensors distributed indoors and outdoors to form time-series environmental parameter data;

[0078] User device usage behavior data is constructed by extracting the on / off time records, usage frequency statistics, and operation preference settings of each device from the user behavior database.

[0079] By aligning and synchronizing power parameter data, environmental parameter data, and user equipment usage behavior data over time, multimodal load monitoring data is obtained.

[0080] In this example, at the power data dimension, based on smart meters or distributed load measurement devices installed at user-end distribution nodes, core power parameter data is continuously collected at a frequency of 1Hz or higher without interfering with users' electricity consumption behavior. This includes three-phase voltage, three-phase current, active power, reactive power, and power factor. This data reflects the real-time characteristics of energy transmission and consumption in the power system and also includes dynamic fluctuations caused by changes in load combinations, forming the backbone information source for load status identification. The collected power parameters are timestamped, forming a time-series power parameter data sequence. Simultaneously, to effectively identify the correlation between power load status and the external physical environment, the system synchronously deploys multiple environmental sensing nodes. These nodes are spatially distributed in key indoor areas (such as kitchens, living rooms, bathrooms, and areas with concentrated electrical appliances) and outdoor areas such as windows and ventilation entrances to continuously collect key physical quantities reflecting dynamic changes in climate and environment. The collected parameters mainly include temperature, humidity, light intensity, and carbon dioxide concentration. These environmental parameters are stored in time-series form through independent sensing channels, forming a time-series environmental parameter data set parallel to the power parameters. At the behavioral data level, in order to construct the mapping logic between human-computer interaction and device response, the system is designed to access the user behavior database. Through channels such as smart home control terminals, mobile device App logs, and voice control response information, the system extracts the on / off time records of each target device to form a historical behavior sequence. At the same time, the system statistically models the usage frequency of each device over a certain period of time, establishes a frequency density function to characterize the device's usage activity, and combines user-set operation preference parameters (such as temperature setting, lighting brightness level, power timer setting, etc.) to construct a device behavior pattern vector. After acquiring the three types of data, the system performs time alignment and data synchronization operations on all modal data using a unified time base. This operation uses the power parameter sampling timestamp as the global reference clock, performs linear or cubic spline interpolation on the environmental parameter data, aligns it to the power parameter time step, and performs interpolation compensation for data loss caused by sampling differences to ensure temporal continuity. For user behavior data, due to its strong event-driven nature and irregular time span, the system constructs an event mapping function to match the behavior trigger time point with power load changes or sudden environmental state changes, thereby converting discontinuous behavior events into discrete labels or state mask matrices for time window modeling. The entire synchronization process uses a sliding time window strategy to organize the input data, for example, with a window length of 60 seconds and a 50% window overlap rate, which improves the traceability of temporal features while ensuring the continuity of model calculation.The aligned three types of data constitute a complete data segment within each time window, including power information vector, environmental state vector, and behavior label vector. They are then unified into the input form supported by the model through tensor splicing or feature fusion operations to form a multimodal load monitoring dataset with triple structural attributes.

[0081] In one example, multimodal load monitoring data is input into the non-negative matrix factorization layer of the first equipment condition identification model for cross-modal feature extraction, resulting in a multimodal fusion feature vector, including:

[0082] Multimodal load monitoring data is input into the nonnegative matrix decomposition layer of the first equipment state identification model to construct a multimodal feature matrix combination that includes the observed load matrix, environmental feature matrix, and behavioral feature matrix;

[0083] Based on the environmental feature matrix and the behavioral feature matrix, a multimodal regularization constraint term is constructed for the non-negative matrix decomposition algorithm in the non-negative matrix decomposition layer, and the multimodal regularization constraint term is combined with the observation load matrix to construct a regularization constraint objective function;

[0084] By performing a minimization solution using a regularized constraint objective function, the equipment feature matrix and activation coefficient matrix are obtained, and load decomposition features are generated based on the equipment feature matrix and activation coefficient matrix.

[0085] The load decomposition features, environmental feature matrix, and behavioral feature matrix are input into the attention weight calculation network in the non-negative matrix decomposition layer to calculate the multimodal attention weights.

[0086] The load decomposition features are weighted and fused according to the multimodal attention weights to obtain the multimodal fused feature vector.

[0087] In this example, power parameter data (including three-phase voltage, three-phase current, active power, reactive power, and power factor) are standardized to construct an observed load matrix. This matrix has m rows and n columns, where m is the number of time windows and n is the dimension of equipment measurement or spectral characteristics, used to express the combined response of various power loads at different time points. Simultaneously, data from environmental sensors, such as temperature, humidity, light intensity, and carbon dioxide concentration, are used to construct an environmental feature matrix. User equipment usage behavior data, including switching time periods, usage frequency, and operating habits, are encoded into a behavioral feature matrix. These three types of data are structurally aligned through a time window mechanism, forming a three-channel, multi-dimensional feature matrix combination structure with a unified time base at the input. To ensure that the decomposition result not only satisfies the low-rank approximation reconstruction of the observed load but also effectively reflects the explanatory power of environmental factors and behavioral patterns on load characteristics, a multimodal regularization constraint term is introduced into the non-negative matrix factorization algorithm, forming a regularization objective function. The basic form of non-negative matrix factorization is that an observed matrix is ​​approximated by two non-negative matrices, one being the equipment feature matrix and the other being the activation coefficient matrix. The regularization term introduced on this basis is used to strengthen the structural constraints of these two matrices in the modal coupling dimension. This regularization term consists of two parts: first, a distance regularization between the device feature matrix and the environment feature matrix, which expresses the dependence of device power consumption changes on environmental parameters by constraining the minimization of their differences; second, a fitting term between the activation coefficient matrix and the behavior feature matrix, which guides the system to learn device activation rhythms that conform to user preferences by constraining their distribution to maintain consistency with the behavioral cycle. Combining these two types of regularization conditions, the system incorporates them into the overall optimization objective, ensuring that the decomposition model maintains sensitivity and consistency to environmental states and behavioral patterns while considering load reconstruction. During the optimization process, the model uses an alternating minimization strategy or a multiplicative update method to iteratively solve for the device feature matrix and the activation coefficient matrix. In each step, one matrix is ​​fixed while the other is updated, ensuring that the non-negativity constraint is always satisfied. Simultaneously, guided by the regularization term, the model converges to a local optimum that conforms to the environmental conditions and behavioral patterns. Once the optimization reaches a stable state, the resulting device feature matrix represents the characteristic configuration of each typical device load under different environmental behavioral scenarios, while the activation coefficient matrix represents the activation intensity of this feature in each time segment. Together, they generate a high-dimensional, well-structured, and physically interpretable load decomposition feature set. The load decomposition features, along with the environmental feature matrix and behavioral feature matrix, are input into the attention weight calculation network in the non-negative matrix factorization layer. A learnable feedforward network structure with a set of parameters is used to calculate the relevance score of each modality to the current task.After the intermediate feature tensors from the decomposition are encoded along with the environmental and behavioral features, they are concatenated in the network and fed into the attention weight function to calculate their importance weights. Calculation methods include softmax-based weighted activation mechanisms or multi-head attention mechanisms. The core objective is to establish weighted channels among different feature sources to control the strength of information flow, giving higher weights to channels with greater discriminative power or those more relevant to the current context. After obtaining these multimodal attention weights, the system fuses and weights the decomposed features, environmental features, and behavioral features according to their channel weights, and generates the final multimodal fused feature vector through linear combination or tensor-level feature mapping.

[0088] In one example, the multimodal fusion feature vector is input into the first decomposition layer of the first device state recognition model for load pattern recognition, resulting in the first decomposed load matrix, including:

[0089] The multimodal fusion feature vector is input into the first decomposition layer of the first equipment state recognition model. The standardized load feature vector is obtained by performing dimensional transformation and nonlinear mapping through the feature mapping sub-layer in the first decomposition layer.

[0090] The standardized load feature vector is input into the load pattern recognition sublayer in the first decomposition layer for self-attention weight calculation, thus obtaining the load pattern feature encoding vector.

[0091] The load pattern feature encoding vector is input into the pattern classification sub-layer in the first decomposition layer for sequence modeling and probability classification calculation to obtain load pattern classification probability distribution data.

[0092] The load pattern classification probability distribution data is input into the first matrix reconstruction sub-layer in the first decomposition layer. In the first matrix reconstruction sub-layer, the load pattern classification probability distribution data is reconstructed into the first decomposed load matrix in terms of equipment and time dimensions.

[0093] In this example, the multimodal fusion feature vector is input into the first decomposition layer. A feature mapping sublayer is deployed at the beginning of this layer. This sublayer performs dimensionality transformation and nonlinear activation mapping on the input feature vector, giving it stronger expressive power and structural normalization capabilities. Specifically, multiple fully connected layers are stacked consecutively to perform linear mapping on the original fusion vector, compressing or expanding it from a high-dimensional modal channel space to a unified structural dimension. Nonlinear activation functions such as ReLU or GELU are introduced to enhance the network's ability to capture complex patterns, and a Dropout mechanism is added after each layer to prevent overfitting during training. The output after processing by the feature mapping sublayer is the standardized load feature vector. This standardized load feature vector is input into the load pattern recognition sublayer in the first decomposition layer. This sublayer introduces a self-attention mechanism to identify the internal dependencies and contextual interaction structures between load features at different time intervals. The self-attention structure can establish the strength of relationships between implicit positions across the entire time range by constructing query, key, and value vectors without explicit temporal priors, and performs global weighted aggregation of features at each time step, thereby achieving state linkage modeling between different time windows. This structure employs a multi-head attention mechanism to learn load correlations from multiple perspectives in parallel. Each attention head captures variation patterns in a specific dimension, which are then linearly combined and aggregated into a unified attention weight distribution. This attention weight is applied to a standardized load feature vector to generate a load pattern feature encoding vector. This load pattern feature encoding vector is then input into a pattern classification sublayer in the first decomposition layer to perform sequence modeling and multi-class probability classification calculations. This sublayer uses a bidirectional recurrent neural network structure, employing a bidirectional long short-term memory (Bi-LSTM) network to propagate state information both forward and backward in the time dimension, thereby comprehensively capturing the historical trends and expected evolution trajectories of load changes. In this structure, the forward LSTM unit models the load evolution process before the current time point, while the backward LSTM unit models the response patterns afterward. The outputs of both are concatenated in the channel dimension to form a unified temporal state representation. The time-series encoding result is input into a fully connected classification output layer containing multiple nodes, and the Softmax function is used to transform it into a probability distribution for the corresponding device category or load pattern category, thereby generating load pattern classification probability distribution data. This data describes the probability value of the input load feature belonging to various typical load patterns within the current time window. The load pattern classification probability distribution data is input into the first matrix reconstruction sublayer in the first decomposition layer. In this sublayer, the system uses a learnable transformation structure to restore the classification result into a two-dimensional load decomposition matrix with device and time dimensions. The matrix reconstruction process is based on a weighted summation strategy, using the load pattern probability distribution of each device as a weighting coefficient, mapping it to its corresponding load pattern reference vector, and forming a load intensity value sequence through linear combination between channels.During this process, the system uses a trainable linear transformation matrix and bias terms to adjust the weighted results so that the output conforms to the physical characteristics of the load. The output of each time window is arranged according to the device dimension to form a matrix structure of device rows × time columns, which is the first decomposed load matrix.

[0094] In one example, the first decomposed load matrix is ​​input into the second decomposed layer of the first equipment state identification model for cluster optimization to obtain the second load decomposed matrix, which includes:

[0095] The first decomposed load matrix is ​​input into the second decomposed layer of the first equipment status identification model, and then standardized through the feature normalization sublayer in the second decomposed layer to obtain the normalized load feature matrix.

[0096] The normalized load feature matrix is ​​input into the adaptive clustering sublayer in the second decomposition layer for density peak detection to obtain the initial clustering configuration parameters.

[0097] Based on the initial clustering configuration parameters, the normalized load feature matrix is ​​input into the deep clustering sub-layer in the second decomposition layer for joint feature learning and clustering assignment, resulting in the equipment load clustering assignment probability matrix.

[0098] The equipment load clustering allocation probability matrix is ​​input into the second matrix reconstruction sublayer in the second decomposition layer for probability weighting reconstruction to obtain the second load decomposition matrix.

[0099] In this example, the first decomposed load matrix is ​​input into the second decomposed layer of the first device state recognition model. The feature normalization sublayer in the second decomposed layer performs standardization processing, normalizing each row of the matrix (i.e., the load distribution vector of each device at different time periods) according to the rule of zero mean and unit variance. This unifies the performance of each device on a statistical scale, preventing the high-power devices from dominating the numerical distribution and masking the detailed changes of low-power devices. The normalization operation includes mean and variance adjustment and introduces learnable scaling coefficients and bias terms, allowing the model to fine-tune the standardization process based on training data. While maintaining a uniform scale, it still retains some power characteristic patterns of the devices, resulting in a normalized load feature matrix. This normalized load feature matrix is ​​then input into the adaptive clustering sublayer in the second decomposed layer. This sublayer uses a density peak detection algorithm to identify the number and initial positions of cluster centers in the feature space. This detection algorithm constructs a local density estimation function by calculating similarity metrics (such as cosine similarity or Euclidean distance) between samples, identifies sample points in densely distributed regions of the feature space, and further calculates the minimum distance between these points and other high-density points as a relative margin metric. By jointly analyzing density intensity and relative separation, the system automatically selects a set of high-density points with large intervals as initial cluster centers, and simultaneously determines a reasonable number of clusters, forming an initial clustering configuration parameter set, which includes the initial center index, the number of clusters, and preliminary classification labels. Based on the above configuration parameters, the system inputs the normalized load feature matrix into the deep clustering sublayer in the second decomposition layer to perform joint feature learning and cluster assignment. This sublayer includes two sub-modules: an embedding network and a clustering network. The embedding network compresses and maps the input vector through a multi-layer fully connected neural network, mapping it from the original high-dimensional load space to a low-dimensional embedding space. This embedding space has stronger category aggregation capabilities and structural decoupling, making the originally ambiguous and overlapping equipment loads more distinguishable in the low-dimensional space. The clustering network performs soft clustering assignment calculations based on embedding features, transforming the embedding representation of each sample into a membership probability distribution across all categories, thus obtaining the equipment load clustering assignment probability matrix. This matrix records the probability information of each device belonging to different cluster centers at each time slice, reflecting the multiple membership degrees of equipment load under fuzzy boundary conditions. The clustering assignment probability matrix is ​​input into the second matrix reconstruction sublayer in the second decomposition layer. In this sublayer, the system performs a load matrix reconstruction process based on a probability weighting strategy, transforming the fuzzy clustering results into actual load distribution estimates. Based on the soft clustering assignment results output from the previous sublayer, the system calculates the representative load pattern vector for each cluster center, i.e., synthesizing the standard load profile of member devices in each cluster through a weighted averaging method; then, the membership probability of each device in a certain time period is weighted and combined with the load patterns of various clusters, and the final load value is reconstructed through matrix-level weighted summation.This process preserves the fuzzy information under soft classification and enhances the expressive power of the reconstruction matrix in terms of equipment status changes, load continuity and physical rationality, thereby outputting the second load decomposition matrix.

[0100] In this embodiment, the step of inputting the normalized load feature matrix into the deep clustering sublayer in the second decomposition layer for joint feature learning and cluster allocation based on the initial clustering configuration parameters to obtain the equipment load cluster allocation probability matrix includes: inputting the normalized load feature matrix into the feature encoder in the deep clustering sublayer for nonlinear transformation and dimensionality compression; mapping high-dimensional load features to low-dimensional clustering feature representations through a multi-layer fully connected network and batch normalization operations; and introducing random perturbations based on the reparameterization technique of variational autoencoders to obtain robust clustering feature encoding vectors; based on the initial clustering configuration parameters and the clustering feature encoding vectors, adaptive initialization of cluster centers is performed using the K-means++ algorithm; calculating the Euclidean distance and cosine similarity between each clustering feature encoding vector and the candidate cluster centers; and gradually determining the initial cluster center positions for each equipment category using a distance-weighted probability sampling strategy to obtain an optimized cluster center coordinate matrix; inputting the clustering feature encoding vectors and the cluster center coordinate matrix into the soft allocation calculation module in the deep clustering sublayer; and using Student's... The t-distribution kernel function calculates the membership probability of each cluster feature encoding vector to each cluster center, and controls the sharpness of soft assignment based on temperature parameters to obtain an initial equipment load soft assignment probability matrix. Based on this initial equipment load soft assignment probability matrix, a joint optimization objective function is constructed. This objective function includes a feature reconstruction loss term, a cluster assignment consistency constraint term, and a KL divergence regularization term. An alternating iterative optimization strategy is used to update the feature encoder parameters and cluster center positions, respectively, so that the feature learning process and the cluster assignment process mutually promote each other, resulting in jointly optimized feature encoder parameters and cluster center coordinates. Based on the jointly optimized feature encoder... The parameters are recalculated to obtain the cluster feature encoding vectors, and the soft assignment probabilities of each cluster feature encoding vector are recalculated based on the jointly optimized cluster center coordinates. The soft assignment probabilities are converted into differentiable hard assignment probabilities using the Gumbel-Softmax reparameterization technique to obtain accurate equipment load cluster assignment probabilities. The equipment load cluster assignment probabilities are post-processed and optimized by filtering out low-confidence assignment results through confidence threshold, eliminating outlier assignment anomalies using neighborhood consistency test, and smoothing the assignment probabilities based on temporal continuity constraints. Finally, a complete equipment load cluster assignment probability matrix is ​​constructed.

[0101] In one example, dynamic fuzzy decision-making is performed based on the second load decomposition matrix to obtain the equipment operating status identification results, including:

[0102] The second load decomposition matrix is ​​fuzzified to obtain the fuzzified load feature matrix, and a weighted fuzzy inference rule base is constructed based on the fuzzified load feature matrix.

[0103] The fuzzy load feature matrix is ​​input into the weighted fuzzy inference rule base for fuzzy inference to obtain the fuzzy output value of the running status.

[0104] Based on the load change trend and historical recognition accuracy within the current time window, the confidence weights of the weighted fuzzy inference rule base are dynamically adjusted to obtain the updated rule weights. The fuzzy output value of the operating status is then recalculated based on the updated rule weights to obtain the dynamically adjusted fuzzy decision result of the equipment status.

[0105] The fuzzy decision results of the dynamically adjusted equipment status are defuzzified, and the continuous values ​​are converted into four discrete status categories: on, off, standby, and adjustment, to obtain the equipment operation status identification results.

[0106] In this example, the load intensity of each device in the second load decomposition matrix under each time window is fuzzified and converted into a fuzzy set representation with membership degree expression. This fuzzification process is based on three preset load levels: low load, medium load, and high load, each using different membership functions to cover the load value range. Triangular or Gaussian membership functions are employed; triangular functions offer advantages such as simple structure and clear transitions, while Gaussian functions provide smoother boundary fuzzification. By calculating the membership degree for each load value under the three fuzzy levels, the system generates a three-dimensional fuzzy vector for each device and each time window. The fuzzy vectors of all devices across the entire time window constitute the fuzzified load feature matrix. A weighted fuzzy inference rule base is constructed based on the fuzzified load feature matrix, thereby establishing a formal logical mapping between fuzzy load states and device operating behavior. The basic structure of the rule base is a set of IF-THEN rules, each consisting of several premise clauses and a conclusion. The fuzzy conditions involved in the premise clauses are matched with the input values ​​through membership functions, and their activation strength is calculated during the inference process. The rule base encompasses typical device operation logic, including rules for device activation, deactivation, adjustment, and standby. Constructed through empirical modeling or data-driven approaches, each rule is assigned an initial confidence weight to reflect its reliability during the initial model training phase or its average recognition performance on historical data. The fuzzy load feature matrix is ​​input into this weighted rule base, and the system executes fuzzy inference calculations. During this phase, the system evaluates the matching degree of each rule's preconditions and uses the minimum activation principle (e.g., min operation) to obtain the rule's activation strength. The maximum synthesis method aggregates the activation values ​​of all rules in a specific output category, forming the fuzzy output value of the device in each operating state category. This output value is a continuous variable between [0,1], representing the fuzzy membership degree of the device's current state in categories such as activation, deactivation, adjustment, and standby. Furthermore, considering that the initial weights of the rules may gradually become ineffective due to environmental changes or device behavior migration, the system introduces a dynamic weight adjustment mechanism. After inference is completed, the weights of all rules are optimized and adjusted based on feedback. This dynamic adjustment mechanism is based on two key data sources: first, the trend of equipment load changes within the current time window, where the system evaluates the sensitivity of rule responses by calculating the load change gradient, rate of change, and trend direction; and second, historical recognition accuracy, which is statistically analyzed by using a sliding time window to assess the matching degree between rule application results and actual recognition results, thereby determining the reliability of each rule's performance in actual use. The system employs an exponentially weighted average algorithm or a dynamic update strategy based on a forgetting factor to continuously update the weight of each rule, improving the priority activation capability of efficient rules and reducing the impact of misjudged rules.The adjusted rule weights will be reloaded into the fuzzy inference module and take effect in the next round of inference, thereby completing the secondary calculation of the fuzzy inference output value and generating the dynamically adjusted fuzzy decision result for the device state. The dynamically adjusted fuzzy decision result for the device state is then defuzzified using the centroid method. This involves constructing a weighted average of the fuzzy output value for each device under each state category, mapping its continuous distribution centroid to a scalar output, and setting several decision threshold intervals to classify this scalar into discrete categories, such as 0.0 to 0.25 for "off", 0.25 to 0.5 for "standby", 0.5 to 0.75 for "adjustment", and 0.75 to 1.0 for "on".

[0107] In one example, a non-intrusive power load decomposition method based on multimodal feature learning also includes:

[0108] The predicted state of the first device state recognition model is used as the agent's action, and the actual device state is used as the environment state. The state space transition rule is defined by the state transition probability matrix, and a multi-objective reward function is constructed based on the recognition accuracy, energy consumption prediction error and response time to obtain the reinforcement learning environment.

[0109] In a reinforcement learning environment, the device operation status recognition results are compared and verified with the real device status, and the reinforcement learning reward signal is calculated.

[0110] Based on the reinforcement learning reward signal, policy learning is performed on the first device state recognition model to obtain the policy gradient vector;

[0111] The gradient update amount of each layer weight and bias in the first model parameter set is calculated based on the policy gradient vector to obtain the second model parameter set.

[0112] The second device status recognition model is obtained by replacing the original parameters of the corresponding levels in the first device status recognition model with the parameters of the second model.

[0113] In this example, the output of the state recognition model, i.e., the predicted device state (such as on, off, adjusting, standby, etc.), is set as the expression of the action taken by the agent at each time step. Subsequently, the system uses the actual detected or historically labeled real operating state of the device as the environmental state to judge the correctness of the agent's actions and their impact on the overall system performance. Simultaneously, by establishing a state transition probability matrix, the transition rules from the current environmental state to the next environmental state are defined. This transition matrix, obtained based on statistical learning, characterizes the dynamic behavior patterns of the device under different operating conditions and is used in policy training to assist in state evaluation and future state prediction, thereby enhancing the learning model's ability to perceive trends in device state changes. During the construction of the reinforcement learning environment, to achieve multi-objective collaborative optimization of device recognition accuracy, energy consumption prediction effect, and response latency performance, a composite reward function containing three core indicators is designed. Recognition accuracy, as the most basic performance evaluation indicator, directly uses the degree of matching between the model output and the actual state as the main signal for positive rewards. Secondly, energy consumption prediction error, an indicator reflecting the model's ability to judge equipment energy load, is introduced. This is achieved by analyzing the difference between the predicted load and the actual measured load within the corresponding time period and normalizing it into a penalty term, used to constrain the accumulation of errors in load magnitude during the prediction inference process. The response time indicator measures the adjustment period required by the model after a sudden state change. A third evaluation item is constructed by statistically analyzing the model's response delay time window and comparing it with the expected response time. These three indicators, combined using a weighted fusion strategy, constitute the total reward function and are used to evaluate the performance of the current strategy during each state transition, thereby providing feedback on the value function of the action-state pair combination and constructing a reinforcement learning environment oriented towards actual business goals. After the reinforcement learning environment is officially running, the system compares and verifies the equipment operating state recognition results with the actual equipment state, evaluates the deviation between prediction and reality in real time, and quantifies the aforementioned multi-objective reward function accordingly to generate reinforcement learning reward signals. This signal is used not only for single-step action optimization but also for backtracking evaluation of long-term policy optimization paths through a cumulative discount strategy. At each time step, the system collects a four-tuple consisting of the current state, current action, immediate reward, and next state, forming an experience pool. After a certain number of steps, a policy update phase is initiated. In this phase, the system iteratively trains the current policy using a policy optimization algorithm (such as proximal policy optimization, policy gradient method, or behavior cloning). Specifically, it calculates the policy gradient vector, which is the derivative of the expected cumulative reward generated by taking actions in each state with respect to the policy parameters, serving as the direction for policy improvement.The system updates the parameters of each layer of the first device state recognition model based on the policy gradient vector. Specifically, it calculates the update amounts of the weights and biases of each layer in the model parameters using the direction of the partial derivatives in the policy gradient vector and the learning rate adjustment magnitude. These updates are then propagated layer by layer to the input using a backpropagation mechanism, thus completing the gradient correction of the overall parameter structure. During this process, the system uses an adaptive optimizer (such as Adam or RMSProp) to dynamically adjust the learning step size of different parameters to enhance convergence stability and policy generalization ability. After all parameter update operations are completed, a new set of model parameters, the second model parameter set, is obtained. This set includes optimized core parameters such as the connection weights, bias terms, and normalization factors of each neural network layer, representing the internal structural state of the model under the optimal policy in the current training phase. Using the second model parameter set as a carrier, the system performs a structural replacement operation on the first device state recognition model. The weights and bias values ​​from this parameter set are loaded one-to-one into each network layer of the original model, replacing the corresponding parameters in the original model structure, thereby generating a new recognition network structure, the second device state recognition model.

[0114] In one example, the gradient update amounts of the weights and biases of each layer in the first model parameter set are calculated based on the policy gradient vector to obtain the second model parameter set, which includes:

[0115] The hierarchical gradient coefficients of the nonnegative matrix factorization layer, the first factorization layer, and the second factorization layer are calculated based on the policy gradient vector, and the gradient update amount of the weight matrix and bias vector of each layer in the first model parameter set is calculated by combining the first-order moment and the second-order moment of the Adam optimizer.

[0116] The gradient update amount is passed from the output layer to the input layer layer by layer through the backpropagation algorithm, and the numerical range of the gradient update amount is limited based on the gradient clipping mechanism to obtain the hierarchical gradient allocation result.

[0117] Based on the hierarchical gradient allocation results, the weights of the device feature matrix and activation coefficient matrix in the non-negative matrix decomposition layer are updated respectively. The weight parameters and bias parameters of the feature mapping sublayer, load pattern recognition sublayer, pattern classification sublayer and first matrix reconstruction sublayer in the first decomposition layer are updated. The weight parameters and bias parameters of the feature normalization sublayer, adaptive clustering sublayer, deep clustering sublayer and second matrix reconstruction sublayer in the second decomposition layer are updated to obtain the updated parameter matrix of each level.

[0118] The updated parameter matrices at each level are reorganized and indexed according to the network hierarchy of the first device state recognition model to obtain the parameter set of the second model.

[0119] In this example, the policy gradient vector is projected onto each structural module inside the model to construct a gradient distribution map corresponding to the policy. Based on the different functional modules in the model structure, the gradient vector is refined into three main branches, which correspond to the hierarchical gradient coefficient calculation paths of the non-negative matrix factorization layer, the first factorization layer, and the second factorization layer, respectively. In the nonnegative matrix factorization layer, the system calculates the gradient update amounts of the weights of the device feature matrix and the activation coefficient matrix, respectively. Therefore, guided by the policy gradient, the system performs first-order partial derivative operations on the two core substructures of this layer and records the gradient direction and magnitude corresponding to each parameter. Similarly, in the first decomposition layer, the feature mapping sublayer, load pattern recognition sublayer, pattern classification sublayer, and first matrix reconstruction sublayer also calculate the corresponding weight and bias term gradients based on the input data received in the forward propagation and the loss propagation path. The second decomposition layer includes a feature normalization sublayer, an adaptive clustering sublayer, a deep clustering sublayer, and a second matrix reconstruction sublayer. These sublayers are relatively complex and nest multiple nonlinear mappings and fuzzy clustering mechanisms. Therefore, their gradient estimation needs to be jointly differentiated by combining the residual flow path of the hierarchical structure and the sensitivity of the clustering probability distribution. After gradient estimation is completed in all sublayers, the system obtains a gradient allocation structure containing multiple subspaces, which is used to drive the parameter update behavior of each submodule in the next step. The system incorporates the Adam optimizer for gradient update calculation. The Adam optimizer, an adaptive optimization algorithm based on first-order moments (i.e., the historical gradient mean) and second-order moments (i.e., the historical squared gradient mean), can independently maintain the gradient update trajectory for each parameter unit and perform non-linear scaling on gradients of different magnitudes by dynamically adjusting the learning rate. In this stage, the system initializes the first-order moment estimates (representing the cumulative gradient expectation in the current parameter direction) and second-order moment estimates (representing the variance trend in the current parameter direction) for each weight matrix and bias vector to be updated, respectively, and calculates the final update step size and numerical correction value for each parameter node in conjunction with the policy gradient direction. Based on this, preparatory operations are performed before parameter updates, namely, backpropagation of the gradients of each sub-layer from the model output to the input layer by layer, ensuring that the error derivatives of the weights and biases in each layer are structurally consistent with the true gradient direction, thereby realizing the vector flow of distributed gradients. To ensure numerical stability during the update process and prevent gradient explosion, a gradient pruning mechanism is introduced during backpropagation. This mechanism performs magnitude limiting on all backpropagated gradient vectors, truncating gradient values ​​exceeding a threshold to a preset upper limit and proportionally adjusting the overall direction of the gradient vectors to suppress anomalous numerical expansion while preserving gradient trends. This mechanism yields a normalized, hierarchical gradient distribution result.Based on the hierarchical gradient allocation results, the weight matrices and bias vectors in the three main structural modules are updated according to the functional attributes of each sub-layer: In the non-negative matrix decomposition layer, the weights of the equipment feature matrix represent the primitive features of each typical load form, and the weights of the activation coefficient matrix represent the load activation intensity in each time window. Therefore, the system updates them respectively according to the gradient results guided by the strategy to adapt to the changing trends of equipment load structure and power fluctuations; In the first decomposition layer, the feature mapping sub-layer maps multimodal fusion features to the standard structural space, the load pattern recognition sub-layer uses a self-attention mechanism to capture temporal correlations, and the pattern classification sub-layer outputs equipment status. The system maps classification probabilities to load values ​​in the first matrix reconstruction sublayer, thus the updates of these four sublayers directly affect the model's ability to abstract features and model states. The system synchronously corrects all weight and bias parameters. In the second decomposition layer, the feature normalization sublayer adjusts the input data standardization scale to match the changed load distribution, the adaptive clustering sublayer dynamically adjusts the cluster center initialization strategy, and the deep clustering sublayer jointly optimizes the embedding space structure and soft classification probability. The second matrix reconstruction sublayer reconstructs the clustering results into the actual load matrix. Therefore, the system uniformly updates all structural parameters based on the hierarchical gradient results. After completing the above parameter update operations, the system structurally reorganizes and indexes all updated sublayer parameter matrices, including the device feature matrix, activation coefficient matrix, various weight matrices, and bias vectors, according to the existing network hierarchy in the first device state recognition model. This ensures that each parameter node is correctly positioned in the final model while maintaining consistency in network connections, dependencies, and tensor dimensions, ultimately forming the second model parameter set.

[0120] In one example, the original parameters of the corresponding levels in the first device state recognition model are replaced with the parameters of the second model based on the second model parameter set to obtain the second device state recognition model, including:

[0121] The weight parameters and bias parameters in the second model parameter set are classified and labeled according to the layer name, sub-layer identifier and parameter type, and matched with the original parameters of the non-negative matrix factorization layer, the first decomposition layer and the second decomposition layer in the first device status recognition model to obtain the parameter replacement mapping relationship.

[0122] Based on the parameter replacement mapping relationship, the network parameters are replaced and updated by calculating the weights of the device feature matrix, activation coefficient matrix, and attention weight in the non-negative matrix factorization layer. The weight matrix, bias vector, and activation function parameters of each sub-layer in the first decomposition layer are replaced and updated. The weight matrix, bias vector, and cluster center parameters of each sub-layer in the second decomposition layer are replaced and updated, resulting in the model hierarchical structure after parameter replacement.

[0123] Perform network connectivity verification on the model hierarchy structure after parameter replacement to obtain the network configuration that passes the structure verification;

[0124] Based on the network configuration that has passed the structural verification, the forward propagation path and backward propagation path between the non-negative matrix factorization layer, the first factorization layer and the second factorization layer are reconstructed using a computational graph reconstruction algorithm to obtain the second device state recognition model.

[0125] In this example, the weight and bias parameters in the second model parameter set are ternarily labeled according to the naming convention of "layer name-sublayer identifier-parameter type". The layer name distinguishes whether the parameter belongs to a non-negative matrix factorization layer, the first factorization layer, or the second factorization layer. The sublayer identifier indicates its substructure module, such as a feature mapping sublayer, a clustering sublayer, or a matrix reconstruction sublayer. The parameter type distinguishes whether it is a weight matrix, bias vector, activation function parameter, cluster center coordinates, or attention weight, thus forming a parameter labeling system. Based on this ternary labeling system, the system performs a matching operation between all parameters in the second model parameter set and the original parameters in the first device status recognition model. Specifically, the layer name is used as the first-level matching criterion to place the updated parameters into their target layer; the sublayer identifier is used as the second-level index condition to locate the target module; and finally, the parameter type is used as the refined mapping condition to achieve a precise correspondence between the updated parameters and the original parameters, establishing a parameter replacement mapping relationship. The non-negative matrix factorization layer undergoes a replacement operation, including updating the network parameters such as device feature matrix weights, activation coefficient matrix weights, and attention weights used for multimodal feature fusion according to the mapping relationship, ensuring that the core parameters representing the load structure and modality fusion capability in the factorization layer are consistent with the reinforcement learning training results. Next, the system sequentially replaces the parameters of multiple sub-layer structures in the first factorization layer, including the weight matrix and bias vector of the fully connected layer in the feature mapping sub-layer, the coefficient parameters of activation functions (such as ReLU and GELU), the self-attention structure parameters in the load pattern recognition sub-layer, and the LSTM hidden state weights and soft... The system first replaces the weights of the classification layer and the linear mapping weights and bias terms in the first matrix reconstruction sublayer, ensuring that the parameters in the entire sequence modeling process are fully covered both numerically and topologically. Then, it sequentially replaces each sublayer in the second decomposition layer, including the normalization scaling factor and offset factor in the feature normalization sublayer, the initial cluster vector parameters and density threshold in the adaptive clustering sublayer, the embedding network weights and clustering assignment layer parameters in the deep clustering sublayer, and the aggregation weight matrix used for cluster weighted summation in the second matrix reconstruction sublayer. Through this step, all original parameters are systematically replaced and updated based on the mapping relationship, forming a model hierarchy structure with updated parameters. The system then performs structural connectivity verification on this model hierarchy structure to ensure that all sublayers maintain valid network graph connections after the parameter updates. The verification process mainly includes input-output dimension matching verification, confirmation of connection relationships between layers, tensor dimension consistency check, activation function connectivity confirmation, and gradient path integrity verification. The specific process is completed automatically by building a structure validator. When the system finds that the shape of the input tensor of any layer does not match the output tensor of the previous layer, or that there are problems such as undefined discontinuous connections, duplicate indices, or null weight pointers, an anomaly report is triggered.Only after all connection paths, parameter structures, and dimensional interfaces have passed verification is the system marked as a structurally validated network configuration. Based on the structurally validated model configuration, computational graph reconstruction is performed. According to the topological order of each layer in the network, the system uses a computational graph reconstruction algorithm to fully construct the forward and backward propagation paths between the non-negative matrix factorization layer, the first factorization layer, and the second factorization layer. The goal of forward propagation path reconstruction is to apply the updated parameters in each layer to the forward feature calculation path of the current sample, ensuring that the feature input can smoothly pass through each sub-layer for convolution, mapping, classification, reconstruction, and other operations, ultimately outputting the load recognition result. Backpropagation path reconstruction, based on the gradient flow mechanism of subsequent training iterations, ensures that each sub-layer, guided by the loss function, can accurately calculate its gradient influence on the final prediction result and propagate it back to the previous layer, thus providing structural guarantees for model training. This process involves not only node reconnection operations in the computational graph but also address updates of parameter mapping relationships, gradient buffer initialization, and gradient path registry updates, ensuring that the model versions before and after are structurally identical and functionally seamless. The system completes the entire process from structural labeling, parameter mapping, replacement and update, structural verification to graph structure reconstruction of the second model parameter set, generating the second equipment status recognition model.

[0126] Reference Figure 2 This embodiment provides a non-intrusive power load decomposition system based on multimodal feature learning, including:

[0127] Synchronous acquisition unit 1 is used to synchronously acquire power parameter data from smart meters, environmental parameter data from environmental sensors, and user equipment usage behavior data to obtain multimodal load monitoring data.

[0128] Feature extraction unit 2 is used to input multimodal load monitoring data into the non-negative matrix decomposition layer of the first equipment status identification model for cross-modal feature extraction, and obtain multimodal fusion feature vector;

[0129] The load pattern recognition unit 3 is used to input the multimodal fusion feature vector into the first decomposition layer of the first equipment state recognition model to perform load pattern recognition and obtain the first decomposed load matrix.

[0130] Clustering optimization unit 4 is used to input the first decomposed load matrix into the second decomposed layer of the first equipment status identification model for clustering optimization to obtain the second load decomposed matrix.

[0131] The dynamic fuzzy decision unit 5 is used to perform dynamic fuzzy decision-making based on the second load decomposition matrix to obtain the equipment operating status identification result.

[0132] In this embodiment, the specific implementation of each unit in the above system embodiment is described in the above method embodiment, and will not be repeated here.

[0133] This invention constructs multimodal load monitoring data by simultaneously collecting power parameter data, environmental parameter data, and user behavior data. This data comprehensively characterizes equipment operation features from multiple dimensions, overcoming the limitations of traditional methods that rely solely on power signals. It provides richer and more complementary information sources for load decomposition, significantly enhancing the system's ability to perceive complex power consumption scenarios. By employing a non-negative matrix factorization algorithm combined with multimodal regularization constraints and an attention weight calculation network, it effectively extracts and fuses correlation features between different modalities. Through automatic learning of the importance weights of each modality's data, it achieves optimal fusion of cross-modal information, improving the discriminative power and robustness of feature representations. Through load pattern recognition in the first decomposition layer and cluster optimization in the second decomposition layer, a hierarchical processing mechanism from coarse-grained to fine-grained is constructed. The first decomposition layer is responsible for identifying basic load patterns, while the second decomposition layer further optimizes cluster allocation. This progressive architecture gradually refines and optimizes the load decomposition results, improving the accuracy and stability of equipment identification. A weighted fuzzy inference rule base is constructed based on a fuzzy load feature matrix, and the confidence weights are dynamically adjusted according to load change trends and historical recognition accuracy. This effectively handles the uncertainty and fuzziness in the load decomposition process, enhancing the system's adaptability to scenarios with ambiguous boundaries and state transitions. By constructing a reinforcement learning environment, the device state recognition results are compared and verified with the actual state. Policy learning and parameter updates are performed based on a multi-objective reward function, enabling the model to automatically adjust and optimize according to actual application effects. A hierarchical gradient allocation is performed using the Adam optimizer combined with a gradient pruning mechanism, and the accuracy of model reconstruction is ensured through parameter replacement mapping relationships and network connection verification. Precise parameter updates are achieved for the non-negative matrix decomposition layer, the first decomposition layer, and the second decomposition layer, ensuring the effectiveness of model optimization and the integrity of the network structure.

[0134] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, system, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, system, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, system, article, or method that includes that element.

[0135] The above description is only a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A non-intrusive power load disaggregation method based on multi-modal feature learning, characterized in that, include: Simultaneously collect power parameter data from smart meters, environmental parameter data from environmental sensors, and user equipment usage behavior data to obtain multimodal load monitoring data; The multimodal load monitoring data is input into the non-negative matrix decomposition layer of the first equipment status identification model for cross-modal feature extraction to obtain a multimodal fusion feature vector; The multimodal fusion feature vector is input into the first decomposition layer of the first device state recognition model to perform load pattern recognition, thereby obtaining the first decomposed load matrix; The first decomposed load matrix is ​​input into the second decomposed layer of the first equipment status identification model for cluster optimization to obtain the second load decomposed matrix; Dynamic fuzzy decision-making is performed based on the second load decomposition matrix to obtain the equipment operating status identification result.

2. The non-intrusive power load decomposition method based on multimodal feature learning according to claim 1, characterized in that, The simultaneous acquisition of power parameter data from smart meters, environmental parameter data from environmental sensors, and user equipment usage behavior data yields multimodal load monitoring data, including: The three-phase voltage, three-phase current, active power, reactive power and power factor values ​​are collected synchronously by smart meters to form time-series power parameter data; Temperature, humidity, light intensity, and carbon dioxide concentration values ​​are collected once by environmental sensors distributed indoors and outdoors to form time-series environmental parameter data; User device usage behavior data is constructed by extracting the on / off time records, usage frequency statistics, and operation preference settings of each device from the user behavior database. The power parameter data, the environmental parameter data, and the user equipment usage behavior data are time-aligned and synchronized to obtain multimodal load monitoring data.

3. The non-intrusive power load decomposition method based on multimodal feature learning according to claim 2, characterized in that, The step of inputting the multimodal load monitoring data into the non-negative matrix factorization layer of the first equipment condition identification model for cross-modal feature extraction to obtain a multimodal fusion feature vector includes: The multimodal load monitoring data is input into the non-negative matrix decomposition layer of the first equipment state identification model to construct a multimodal feature matrix combination that includes the observed load matrix, environmental feature matrix, and behavioral feature matrix; Based on the environmental feature matrix and the behavioral feature matrix, a multimodal regularization constraint term is constructed for the non-negative matrix factorization algorithm in the non-negative matrix factorization layer, and the multimodal regularization constraint term is combined with the observation load matrix to construct a regularization constraint objective function; The device feature matrix and activation coefficient matrix are obtained by minimizing the objective function under the regularization constraint, and load decomposition features are generated based on the device feature matrix and activation coefficient matrix. The load decomposition features, the environmental feature matrix, and the behavioral feature matrix are input into the attention weight calculation network in the non-negative matrix decomposition layer to calculate multimodal attention weights. The load decomposition features are weighted and fused according to the multimodal attention weights to obtain a multimodal fused feature vector.

4. The non-intrusive power load decomposition method based on multimodal feature learning according to claim 1, characterized in that, The step of inputting the multimodal fused feature vector into the first decomposition layer of the first device state recognition model for load pattern recognition to obtain the first decomposed load matrix includes: The multimodal fusion feature vector is input into the first decomposition layer of the first device state recognition model. The standardized load feature vector is obtained by performing dimensional transformation and nonlinear mapping through the feature mapping sublayer in the first decomposition layer. The standardized load feature vector is input into the load pattern recognition sublayer in the first decomposition layer for self-attention weight calculation to obtain the load pattern feature encoding vector. The load pattern feature encoding vector is input into the pattern classification sub-layer in the first decomposition layer for sequence modeling and probability classification calculation to obtain load pattern classification probability distribution data. The load pattern classification probability distribution data is input into the first matrix reconstruction sub-layer in the first decomposition layer. In the first matrix reconstruction sub-layer, the load pattern classification probability distribution data is reconstructed into a first decomposed load matrix with device and time dimensions.

5. The non-intrusive power load decomposition method based on multimodal feature learning according to claim 1, characterized in that, The step of inputting the first decomposed load matrix into the second decomposition layer of the first equipment state identification model for cluster optimization to obtain the second load decomposition matrix includes: The first decomposed load matrix is ​​input into the second decomposed layer of the first equipment status identification model, and the normalized load feature matrix is ​​obtained by the feature normalization sublayer in the second decomposed layer. The normalized load feature matrix is ​​input into the adaptive clustering sublayer in the second decomposition layer for density peak detection to obtain the initial clustering configuration parameters. Based on the initial clustering configuration parameters, the normalized load feature matrix is ​​input into the deep clustering sublayer in the second decomposition layer for joint feature learning and clustering assignment to obtain the equipment load clustering assignment probability matrix. The equipment load clustering allocation probability matrix is ​​input into the second matrix reconstruction sublayer in the second decomposition layer for probability weighting reconstruction to obtain the second load decomposition matrix.

6. The non-intrusive power load decomposition method based on multimodal feature learning according to claim 1, characterized in that, The step of performing dynamic fuzzy decision-making based on the second load decomposition matrix to obtain the equipment operating status identification result includes: The second load decomposition matrix is ​​fuzzified to obtain a fuzzified load feature matrix, and a weighted fuzzy inference rule base is constructed based on the fuzzified load feature matrix. The fuzzy load feature matrix is ​​input into the weighted fuzzy inference rule base for fuzzy inference to obtain the fuzzy output value of the running status. Based on the load change trend and historical recognition accuracy within the current time window, the confidence weights of the weighted fuzzy inference rule base are dynamically adjusted to obtain updated rule weights. The fuzzy output value of the operating status is then recalculated based on the updated rule weights to obtain the dynamically adjusted fuzzy decision result for the equipment status. The fuzzy decision results of the dynamically adjusted equipment status are defuzzified, and the continuous values ​​are converted into four discrete status categories: on, off, standby, and adjustment, to obtain the equipment operating status identification results.

7. The non-intrusive power load decomposition method based on multimodal feature learning according to claim 1, characterized in that, The non-intrusive power load decomposition method based on multimodal feature learning also includes: The predicted state of the first device state recognition model is used as the agent's action, and the actual device state is used as the environment state. The state space transition rule is defined by the state transition probability matrix, and a multi-objective reward function is constructed based on the recognition accuracy, energy consumption prediction error and response time to obtain the reinforcement learning environment. In the reinforcement learning environment, the device operating status identification result is compared and verified with the real device status, and the reinforcement learning reward signal is calculated. Based on the reinforcement learning reward signal, policy learning is performed on the first device state recognition model to obtain the policy gradient vector; The gradient update amount of each layer weight and bias in the first model parameter set is calculated based on the gradient vector of the strategy to obtain the second model parameter set; The second device status recognition model is obtained by replacing the original parameters of the corresponding levels in the first device status recognition model with the parameters of the second model.

8. The non-intrusive power load decomposition method based on multimodal feature learning according to claim 7, characterized in that, The step of calculating the gradient update amounts of the weights and biases of each layer in the first model parameter set based on the policy gradient vector to obtain the second model parameter set includes: Based on the policy gradient vector, calculate the hierarchical gradient coefficients of the nonnegative matrix factorization layer, the first factorization layer, and the second factorization layer, and combine the first and second moments of the Adam optimizer to calculate the gradient update amount of the weight matrix and bias vector of each layer in the first model parameter set. The gradient update amount is passed from the output layer to the input layer layer by layer through the backpropagation algorithm, and the numerical range of the gradient update amount is limited based on the gradient clipping mechanism to obtain the hierarchical gradient allocation result. Based on the hierarchical gradient allocation results, the weights of the device feature matrix and the activation coefficient matrix in the non-negative matrix decomposition layer are updated respectively. The weight parameters and bias parameters of the feature mapping sublayer, load pattern recognition sublayer, pattern classification sublayer and first matrix reconstruction sublayer in the first decomposition layer are updated. The weight parameters and bias parameters of the feature normalization sublayer, adaptive clustering sublayer, deep clustering sublayer and second matrix reconstruction sublayer in the second decomposition layer are updated to obtain the updated parameter matrix of each level. The updated parameter matrices at each level are reorganized and indexed according to the network hierarchy of the first device state recognition model to obtain the second model parameter set.

9. The non-intrusive power load decomposition method based on multimodal feature learning according to claim 7, characterized in that, The step of replacing the original parameters of the corresponding levels in the first device state recognition model with the second model parameter set to obtain the second device state recognition model includes: The weight parameters and bias parameters in the second model parameter set are classified and labeled according to the layer name, sub-layer identifier and parameter type, and matched with the original parameters of the non-negative matrix factorization layer, the first decomposition layer and the second decomposition layer in the first device state recognition model to obtain the parameter replacement mapping relationship. According to the parameter replacement mapping relationship, the device feature matrix weights, activation coefficient matrix weights, and attention weights in the non-negative matrix decomposition layer are replaced and updated respectively. The weight matrix, bias vector, and activation function parameters of each sub-layer in the first decomposition layer are replaced and updated. The weight matrix, bias vector, and cluster center parameters of each sub-layer in the second decomposition layer are replaced and updated respectively, resulting in the model hierarchical structure with complete parameter replacement. Perform network connectivity verification on the model hierarchy structure after the parameter replacement is completed to obtain a network configuration that passes the structure verification. Based on the network configuration that has passed the structural verification, the forward propagation path and backward propagation path between the non-negative matrix factorization layer, the first factorization layer and the second factorization layer are reconstructed using a computational graph reconstruction algorithm to obtain the second device state recognition model.

10. A non-intrusive power load decomposition system based on multimodal feature learning, characterized in that, The steps for implementing the non-intrusive power load decomposition method based on multimodal feature learning as described in any one of claims 1 to 9 include: The synchronous acquisition unit is used to synchronously acquire power parameter data from smart meters, environmental parameter data from environmental sensors, and user equipment usage behavior data to obtain multimodal load monitoring data. The feature extraction unit is used to input the multimodal load monitoring data into the non-negative matrix decomposition layer of the first equipment status identification model for cross-modal feature extraction to obtain a multimodal fusion feature vector; The load pattern recognition unit is used to input the multimodal fusion feature vector into the first decomposition layer of the first equipment state recognition model to perform load pattern recognition and obtain the first decomposed load matrix. The clustering optimization unit is used to input the first decomposed load matrix into the second decomposition layer of the first equipment state identification model for clustering optimization to obtain the second load decomposition matrix; The dynamic fuzzy decision unit is used to perform dynamic fuzzy decision-making based on the second load decomposition matrix to obtain the equipment operating status identification result.

Images