A data fusion mining method and system based on multi-modal power cross-domain

Abstract

The present application relates to the technical field of power data fusion, and more particularly to a data fusion mining method and system based on multi-modal power cross-domain. The method comprises the following steps: obtaining a multi-source heterogeneous power dataset; performing tensor decomposition on the multi-source heterogeneous power dataset to obtain a power core feature tensor set; performing feature selection and reconstruction on the power core feature tensor set to obtain a reconstructed power feature dataset; performing domain adaptive feature mapping on the reconstructed power feature dataset to obtain a cross-domain power feature mapping dataset; performing domain difference elimination on the cross-domain power feature mapping dataset to obtain a domain-aligned power feature dataset; and performing attention weight calculation on the domain-aligned power feature dataset to obtain a cross-domain fusion power feature dataset. The present application can significantly improve the utilization efficiency of multi-source data in the power system.

Description

Technical Field

[0001] This invention relates to the field of power data fusion technology, and in particular to a method and system for cross-domain power data fusion and mining based on multimodal power. Background Technology

[0002] In modern power systems, data sources are increasingly diverse, covering all stages from power generation, transmission, and transformation to distribution and consumption. This data includes not only traditional power parameters (such as voltage, current, and power), but also equipment operating status, environmental monitoring data, user behavior data, and various signals collected by smart sensors. This multimodal data exhibits extremely high heterogeneity and diversity, encompassing data types ranging from structured (e.g., tabular data) to semi-structured (e.g., log files) and unstructured (e.g., images, voice, and text). This complex data environment significantly increases the difficulty of deep data mining and fusion. For example, consider a smart substation scenario. This substation contains not only traditional power equipment operating data, such as measurement data from current and voltage transformers, but also images of equipment appearance captured by smart cameras, temperature and humidity data recorded by environmental sensors, and user electricity consumption behavior data collected by smart meters. This data comes from different devices and sensors, has varying formats, and is distributed across different systems. Achieving efficient fusion of this data requires addressing issues such as inconsistent data formats, dispersed data sources, and low data utilization efficiency. Summary of the Invention

[0003] Therefore, it is necessary for the present invention to provide a method and system for cross-domain data fusion mining based on multimodal power to solve at least one of the above-mentioned technical problems.

[0004] To achieve the above objectives, a multimodal power cross-domain data fusion and mining method is proposed, comprising the following steps:

[0005] Step S1: Obtain a multi-source heterogeneous power dataset; perform tensor decomposition on the multi-source heterogeneous power dataset to obtain a power core feature tensor set; perform feature selection and reconstruction on the power core feature tensor set to obtain a reconstructed power feature dataset;

[0006] Step S2: Perform domain-adaptive feature mapping on the reconstructed power feature dataset to obtain a cross-domain power feature mapping dataset; perform domain difference elimination on the cross-domain power feature mapping dataset to obtain a domain-aligned power feature dataset; calculate attention weights on the domain-aligned power feature dataset to obtain a cross-domain fused power feature dataset.

[0007] Step S3: Obtain distributed node topology data; construct a graph convolutional network from the distributed node topology data to obtain power node graph network structure data; train a distributed model on the cross-domain fused power feature dataset based on the power node graph network structure data to obtain the local power model parameter set; perform federated averaging on the local power model parameter set to obtain the global optimization model parameter set.

[0008] Step S4: Construct a teacher network for the global optimization model parameter set to obtain a power multimodal teacher model; train a student network based on the power multimodal teacher model to obtain a power lightweight student model; fuse the power lightweight student model and the power multimodal teacher model to obtain a power feature extraction fusion model.

[0009] Step S5: Obtain power multimodal quality assessment data; construct a reinforcement learning strategy for the power multimodal quality assessment data to obtain a power data dynamic fusion strategy; adjust the weights of the power feature extraction and fusion model according to the power data dynamic fusion strategy to obtain the optimal power fusion decision model; use the optimal power fusion decision model to perform semantic association mining on the multi-source heterogeneous power dataset to obtain a power multimodal semantic association dataset.

[0010] This invention extracts core features from complex, multi-source, heterogeneous power datasets through tensor decomposition and feature selection reconstruction, effectively addressing the issues of inconsistent data formats and dispersed data sources. Domain-adaptive feature mapping and domain discrepancy elimination effectively resolve the differences between multimodal data across different domains. By constructing graph convolutional networks and distributive model training, combined with federated averaging, the topological information of distributed nodes is fully utilized, improving the efficiency and robustness of model training. Distributed training and federated learning avoid the computational pressure and data transmission bottlenecks associated with centralized processing, while ensuring global optimization of the model. Through the construction of a teacher-student network and a knowledge distillation strategy, the knowledge of complex multimodal teacher models can be transferred to lightweight student models. This not only preserves the semantic information and feature extraction capabilities of the teacher model but also significantly reduces the model's computational complexity and storage requirements, improving inference efficiency and making it more suitable for real-time applications and edge computing scenarios in practical power systems. By constructing a dynamic power data fusion strategy through reinforcement learning and adjusting model weights according to the dynamic fusion strategy, dynamic optimization of the multimodal data fusion process can be achieved. Furthermore, semantic association mining using the optimal power fusion decision model can further uncover deeper semantic information in multi-source heterogeneous data. In summary, this invention can significantly improve the utilization efficiency of multi-source data in power systems.

[0011] Preferably, the present invention also provides a multimodal power cross-domain data fusion and mining system for executing the multimodal power cross-domain data fusion and mining method described above. The multimodal power cross-domain data fusion and mining system includes:

[0012] The data preprocessing module is used to acquire multi-source heterogeneous power datasets; perform tensor decomposition on the multi-source heterogeneous power datasets to obtain a set of power core feature tensors; and perform feature selection and reconstruction on the set of power core feature tensors to obtain a reconstructed power feature dataset.

[0013] The feature mapping module is used to perform domain-adaptive feature mapping on the reconstructed power feature dataset to obtain a cross-domain power feature mapping dataset; to perform domain difference elimination on the cross-domain power feature mapping dataset to obtain a domain-aligned power feature dataset; and to calculate attention weights on the domain-aligned power feature dataset to obtain a cross-domain fused power feature dataset.

[0014] The model training module is used to acquire distributed node topology data; construct a graph convolutional network on the distributed node topology data to obtain power node graph network structure data; perform distributed model training on the cross-domain fused power feature dataset based on the power node graph network structure data to obtain the local power model parameter set; and perform federated averaging on the local power model parameter set to obtain the global optimization model parameter set.

[0015] The model fusion module is used to construct a teacher network from the global optimization model parameter set to obtain a power multimodal teacher model; to train a student network based on the power multimodal teacher model to obtain a power lightweight student model; and to fuse the power lightweight student model and the power multimodal teacher model to obtain a power feature extraction fusion model.

[0016] The data mining module is used to acquire multimodal power quality assessment data; construct a reinforcement learning strategy for the multimodal power quality assessment data to obtain a dynamic power data fusion strategy; adjust the weights of the power feature extraction and fusion model according to the dynamic power data fusion strategy to obtain the optimal power fusion decision model; and use the optimal power fusion decision model to perform semantic association mining on multi-source heterogeneous power datasets to obtain a multimodal power semantic association dataset.

[0017] In this invention, the entire system significantly improves the utilization efficiency of multi-source data in the power system and reduces data processing costs through efficient fusion of multimodal data, cross-domain alignment, distributed training, and dynamic optimization. Attached Figure Description

[0018] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description taken in conjunction with the accompanying drawings:

[0019] Figure 1 A flowchart illustrating the steps of a multimodal power cross-domain data fusion and mining method according to an embodiment is shown.

[0020] Figure 2 A detailed flowchart of step S4 of one embodiment is shown. Detailed Implementation

[0021] The technical method of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0022] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0023] It should be understood that although the terms "first," "second," etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are used merely to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0024] To achieve the above objectives, please refer to Figures 1 to 2 This invention provides a method for cross-domain data fusion and mining based on multimodal power, comprising the following steps:

[0025] Step S1: Obtain a multi-source heterogeneous power dataset; perform tensor decomposition on the multi-source heterogeneous power dataset to obtain a power core feature tensor set; perform feature selection and reconstruction on the power core feature tensor set to obtain a reconstructed power feature dataset;

[0026] Step S2: Perform domain-adaptive feature mapping on the reconstructed power feature dataset to obtain a cross-domain power feature mapping dataset; perform domain difference elimination on the cross-domain power feature mapping dataset to obtain a domain-aligned power feature dataset; calculate attention weights on the domain-aligned power feature dataset to obtain a cross-domain fused power feature dataset.

[0027] Step S3: Obtain distributed node topology data; construct a graph convolutional network from the distributed node topology data to obtain power node graph network structure data; train a distributed model on the cross-domain fused power feature dataset based on the power node graph network structure data to obtain the local power model parameter set; perform federated averaging on the local power model parameter set to obtain the global optimization model parameter set.

[0028] Step S4: Construct a teacher network for the global optimization model parameter set to obtain a power multimodal teacher model; train a student network based on the power multimodal teacher model to obtain a power lightweight student model; fuse the power lightweight student model and the power multimodal teacher model to obtain a power feature extraction fusion model.

[0029] Step S5: Obtain power multimodal quality assessment data; construct a reinforcement learning strategy for the power multimodal quality assessment data to obtain a power data dynamic fusion strategy; adjust the weights of the power feature extraction and fusion model according to the power data dynamic fusion strategy to obtain the optimal power fusion decision model; use the optimal power fusion decision model to perform semantic association mining on the multi-source heterogeneous power dataset to obtain a power multimodal semantic association dataset.

[0030] In this embodiment, firstly, multi-source heterogeneous data is collected from the power system, including voltage, current, power, equipment status, ambient temperature, humidity, and equipment images. This data comes from multiple substations and power plants. Tensor decomposition is performed on the data using Python's Tensorly library to extract the core power feature tensor set. Important features are then selected using the SelectKBest method from scikit-learn, and an autoencoder is used for feature reconstruction to form a reconstructed power feature dataset. Next, a domain adaptation network based on TensorFlow is designed. This network takes the reconstructed power feature dataset as input and outputs a cross-domain power feature mapping dataset. Inter-domain differences are eliminated through adversarial training and a discriminator network to obtain a domain-aligned power feature dataset. Finally, a self-attention mechanism (such as the Transformer architecture) is used to calculate the attention weight of each feature, generating a cross-domain fused power feature dataset. Then, topological relationship data of distributed power nodes are extracted from the SCADA system. A graph convolutional network (GCN) is constructed using PyTorch Geometric. Based on the power node graph network structure data, a distributed training is performed on the cross-domain fused power feature dataset to obtain the local power model parameter set. Federated averaging is then performed using PyTorch's torch.distributed module to obtain the global optimized model parameter set. Subsequently, a power multimodal teacher model is constructed using the global optimized model parameter set. A lightweight student network is designed, and key semantic features of the teacher model are transferred to the student model through knowledge distillation. Finally, the two are fused to obtain a power feature extraction fusion model. Finally, the reliability of the output of the power feature extraction fusion model is evaluated to generate power multimodal quality assessment data. Reinforcement learning is performed using the Stable Baselines3 framework to design a dynamic power data fusion strategy. The model weights are adjusted according to this strategy to obtain the optimal power fusion decision model. This model is then used to perform semantic association mining on multi-source heterogeneous power datasets to extract a power multimodal semantic association dataset.

[0031] Preferably, step S1 includes the following steps:

[0032] Step S11: Collect multimodal power data from the target power grid to obtain the original multimodal power dataset, and align the original multimodal power dataset with timestamps to obtain the time-series power dataset;

[0033] Specifically, smart meters can be used to collect user electricity consumption data (such as electricity consumption and power), current transformers and voltage transformers can be used to collect operating parameters of power equipment (such as current and voltage), environmental sensors can be used to record environmental data such as temperature and humidity within the substation, and high-definition cameras can be used to capture image data of the equipment's appearance. These data are stored in different local databases, forming the raw multimodal power dataset. A distributed time synchronization system (such as a GPS clock synchronization module) is used to ensure that the time base of all sensors and data acquisition devices is consistent. The GPS clock synchronization module is connected to each sensor device, and the internal clocks of the sensors are calibrated via a network time protocol. Data preprocessing software (such as the Pandas library in Python) is used to read the raw data files from each data source and perform alignment operations based on the timestamp field. For example, for smart meter data and current transformer data, sampling alignment is performed every 15 minutes; for image data, it is marked every hour, ultimately resulting in a time-series power dataset.

[0034] Step S12: Based on the preset data quality assessment indicators, perform anomaly detection on the time-series power dataset to obtain a power data anomaly label set, and perform data cleaning on the time-series power dataset according to the power data anomaly label set to obtain a multi-source heterogeneous power dataset.

[0035] Specifically, data quality assessment metrics can be selected, such as data completeness, consistency, and accuracy. Taking current data as an example, the completeness metric requires that the current data have no missing values ​​in the time series; the consistency metric requires that the trend of current data changes matches the trend of voltage data; and the accuracy metric determines whether the current data is within a reasonable range by comparing it with historical data. The NumPy and SciPy libraries in Python are used to check the completeness of the current data, and missing values ​​are filled in using interpolation methods. For consistency detection, the Pearson correlation coefficient is used to calculate the correlation between current and voltage data; when the correlation coefficient is below 0.8, it is marked as abnormal data. For accuracy detection, an anomaly detection model based on a Long Short-Term Memory (LSTM) network is constructed. Historical current data sequences are input, and the output is a judgment on whether the current data is abnormal. By training an LSTM model, it can learn the normal variation pattern of current data. When data is detected to be outside the normal range, it is marked as abnormal data and recorded in the power data anomaly marker set. For data points marked as abnormal, if the data is missing, it is filled in using linear interpolation or based on the average of surrounding data; if the data fluctuates abnormally, it is smoothed using median filtering or a sliding window averaging method. The final result is a multi-source heterogeneous power dataset.

[0036] Step S13: Identify the data types of the multi-source heterogeneous power dataset to obtain a power data type feature set, and standardize the multi-source heterogeneous power dataset according to the power data type feature set to obtain a standard heterogeneous power dataset.

[0037] Specifically, the multi-source heterogeneous power dataset contains structured data (such as numerical data like current and voltage), semi-structured data (such as equipment log files), and unstructured data (such as equipment appearance images). For structured data, Pandas is used to read the data files, and the data type of the data columns is checked to identify numerical and timestamp data. For semi-structured data, regular expressions are used to match key fields in the log files (such as timestamps, device IDs, and error codes) and extract them as structured data. For unstructured data (such as image data), the OpenCV library is used to load the image files, identify them as image data, and extract features such as image size and color channels to form a power data type feature set. For numerical data, the Z-score normalization method is used to convert the data into standardized data with a mean of 0 and a standard deviation of 1 by calculating the mean and standard deviation. For timestamp data, it is uniformly converted to UTC time format. For log data, the extracted key fields are numerically encoded; for example, the error code "ERROR_001" is mapped to the value 1. For image data, pixel values ​​are scaled to between 0 and 1 through normalization. Through these operations, a standard heterogeneous power dataset is finally obtained.

[0038] Step S14: Construct tensors from the standard heterogeneous power dataset to obtain a multidimensional power tensor dataset;

[0039] Specifically, the standard heterogeneous power dataset contains various data, including current, voltage, temperature, humidity, and images of equipment appearance. Tensor manipulation tools from the TensorFlow framework were used for tensor construction. The numerical data such as current, voltage, temperature, and humidity were arranged in a time series to form a four-dimensional tensor, where the dimensions are time, equipment number, data type (current, voltage, temperature, humidity), and data value, respectively. For example, assuming the dataset contains current, voltage, temperature, and humidity data for 10 devices at 100 time points, the constructed tensor shape would be [100, 10, 4, 1], where the last dimension represents the data value. For the equipment appearance image data, since it is a two-dimensional image, it needs to be converted into a tensor form compatible with numerical data. The OpenCV library was used to convert the image data into grayscale images and flatten them into one-dimensional vectors. Assuming each image is 224×224 pixels, the length of the flattened vector is 50,176. These vectors are arranged in a time series to form a three-dimensional tensor with the shape [100, 10, 50176], where 100 represents the time point, 10 represents the device number, and 50176 represents the length of the one-dimensional vector of the image data. Finally, the numerical data tensor and the image data tensor are merged into a multidimensional power tensor dataset with the shape [100, 10, 50180], where the last dimension contains the features of both the numerical and image data.

[0040] Step S15: Perform rank evaluation on the multidimensional power tensor dataset to obtain the power tensor rank evaluation dataset;

[0041] Specifically, we can assume the multidimensional power tensor dataset has the shape [100, 10, 50180], where 100 represents the time point, 10 represents the device number, and 50180 represents the feature dimension (including numerical and image data features). Using Tensorly, a rank evaluation library, we load the multidimensional power tensor dataset into Tensorly and convert it to a Tensorly-supported tensor format. We then use Tensorly's rank evaluation tools, such as the `tensorly.decomposition.parafac` function, to perform a preliminary rank estimate. This function evaluates the rank by calculating the tensor's CP decomposition (Canberra-Polyadic Decomposition). In practice, we set the initial rank estimation range to [10, 50] and gradually adjust the rank value to evaluate the reconstruction error at different rank values. By calculating the reconstruction error (e.g., mean squared error, MSE), we select a suitable rank value that minimizes the reconstruction error. For example, when the rank value is 20, the reconstruction error is 0.05, while when the rank value increases to 30, the reconstruction error only decreases slightly to 0.045. Therefore, a rank value of 20 is selected as the optimal rank value, and this rank value is recorded in the power tensor rank evaluation dataset.

[0042] Step S16: Perform tensor decomposition on the multidimensional power tensor dataset based on the power tensor rank evaluation dataset to obtain the power core feature tensor set;

[0043] Specifically, the CP decomposition method can be selected for tensor decomposition. The multidimensional power tensor dataset is input into the `tensorly.decomposition.parafac` function of Tensorly, and the rank value (20) obtained in step S15 is passed as a parameter. The function decomposes the multidimensional power tensor into a set of factor matrices and a core tensor. The shapes of the factor matrices are [100,20] (time dimension), [10,20] (equipment dimension), and [50180,20] (feature dimension), respectively, and the shape of the core tensor is [20,20,20]. Through the above operations, these core feature tensors are combined into a set of power core feature tensors.

[0044] Step S17: Perform feature selection and reconstruction on the core power feature tensor set to obtain the reconstructed power feature dataset.

[0045] For a detailed implementation process of this embodiment, please refer to the sub-step of step S17.

[0046] This invention ensures the consistency and integrity of data from different sensors and devices across the time dimension through timestamp alignment. Anomaly detection and data cleaning effectively identify and remove noise, outliers, and erroneous records. Data type identification and standardization unify different types of data (such as structured, semi-structured, and unstructured data) into a standardized framework. Tensor construction and rank evaluation transform multidimensional power datasets into a more compact tensor form and evaluate their inherent rank structure. Feature selection and reconstruction stages identify the key features most valuable for power system analysis and decision-making.

[0047] Preferably, step S17 includes the following steps:

[0048] Step S171: Evaluate the feature importance of the core power feature tensor set to obtain the power feature importance score set;

[0049] Specifically, we can assume the shape of the power core feature tensor set is [20, 20, 20], where each element represents the weight or strength of a feature. The power core feature tensor set is flattened into a two-dimensional matrix, where each row represents a sample (e.g., data from a time point or device), and each column represents a feature. Then, the importance of the features is evaluated using either `RandomForestClassifier` or `RandomForestRegressor` from scikit-learn (choosing a classification or regression model depending on the data type). In this embodiment, assuming the goal is to predict the fault state of power equipment, `RandomForestClassifier` is chosen. The model parameters are set as follows: 100 trees, automatic maximum number of features, and a maximum depth of 10. The flattened feature matrix and corresponding labels (e.g., whether the device is faulty) are input into the model for training. After training, the `feature_importances_` attribute of the model is called to obtain the importance score for each feature. These scores range from 0 to 1, with higher values ​​indicating greater feature importance. For example, suppose that in the feature scoring results, the score for the first feature is 0.25, the score for the second feature is 0.15, and the score for the 20th feature is 0.01. Store the score values ​​as a set of power feature importance scores.

[0050] Step S172: Perform dynamic feature threshold iteration based on the power feature importance score set to obtain the power feature selection threshold set;

[0051] Specifically, we can assume that the power feature importance score set contains scores for 20 features, ranging from 0.01 to 0.25. Choose an initial feature threshold, for example, 0.1. Mark all features with scores below 0.1 as "unimportant" and remove them from the dataset. Retrain a simple machine learning model (such as a logistic regression model) using the remaining features. In this example, we use the LogisticRegression model from scikit-learn with default parameters. Calculate the model's average precision and recall using cross-validation. Gradually adjust the feature threshold, for example, from 0.1 to 0.15, repeating the feature selection and model training process described above. Record the model performance metrics after each adjustment. For example, when the threshold is 0.1, the model's precision is 0.85 and recall is 0.80; when the threshold increases to 0.15, the precision improves to 0.88, and the recall slightly decreases to 0.78. Continue adjusting the threshold until a balance is found that optimizes the model performance metrics. Assuming the final threshold is determined to be 0.18, the model's accuracy is 0.90 and recall is 0.85. This threshold is recorded as the power feature selection threshold set.

[0052] Step S173: Select a threshold set based on power characteristics to perform feature filtering on the core power feature tensor set, and obtain the key power feature dataset;

[0053] Specifically, we can assume the shape of the power core feature tensor set is [20, 20, 20], and the feature selection threshold is 0.18 (as determined in step S172). The power core feature tensor set is flattened into a two-dimensional matrix, where each row represents a sample (e.g., data from a time point or device), and each column represents a feature. Based on the power feature importance score set (the result of step S171) and the feature selection threshold, features with scores higher than the threshold are selected. For example, assuming the feature importance score set is an array of length 20, the indices of features with scores higher than 0.18 are [0, 1, 3, 5, 9]. These key feature columns are selected from the flattened feature matrix using Boolean indexing to form a key power feature dataset. Finally, the key power feature dataset has a shape of [N, 5], where N is the number of samples and 5 is the number of selected key features.

[0054] Step S174: Design an autoencoder network structure for the key power feature dataset to obtain the encoder structure parameter set;

[0055] Specifically, we can assume the key power feature dataset has a shape of [N, 5], where N is the number of samples and 5 is the number of key features. The autoencoder network structure is designed using the TensorFlow framework and its API Keras. In this embodiment, the encoder structure is designed as follows: the input layer contains 5 neurons (consistent with the number of key features), the hidden layer contains 3 neurons, and the activation function is ReLU; the latent layer contains 2 neurons, and the activation function is tanh. The decoder structure is designed as follows: the input layer contains 2 neurons (consistent with the latent layer), the hidden layer contains 3 neurons, and the activation function is ReLU; the output layer contains 5 neurons (consistent with the input layer), and the activation function is sigmoid. The input layer of the encoder is defined with dimensions consistent with the number of features in the key power feature dataset. A hidden layer compresses the input features into a more compact representation; the number of neurons in the hidden layer is empirically chosen to be 3, and the activation function is ReLU. The latent layer further compresses the features into a 2-dimensional space, and the activation function is tanh, because the tanh function can limit the output value to the range [-1, 1]. The optimizer for the autoencoder is Adam. The loss function chosen is mean squared error. By training the autoencoder model, the model automatically learns how to compress input features into a low-dimensional latent space and reconstruct features from the low-dimensional space that are as close as possible to the original input. Finally, the autoencoder model parameters obtained through training (including the weights and biases of the encoder and decoder) constitute the encoder structural parameter set.

[0056] Step S175: Train the encoding of the key power feature dataset based on the encoder structure parameter set to obtain the latent power feature encoding dataset;

[0057] Specifically, we can assume the key power feature dataset has a shape of [N, 5], where N is the number of samples and 5 is the number of key features. The encoder's structural parameter set has been defined in step S174. The encoder part is extracted from the autoencoder model to construct an independent encoder model. The encoder model takes the key power feature dataset as input and outputs the latent layer feature encoding. During training, the Adam optimizer is selected, the learning rate is set to 0.001, and the loss function is the mean squared error (MSE). The training process includes the following steps: dividing the key power feature dataset into a training set and a validation set in a ratio of 8:2. The encoder is trained using the training set, with a batch size of 32 and 100 training epochs. In each training epoch, the model calculates the MSE loss between the input features and the reconstructed features and updates the encoder weights through backpropagation. Simultaneously, the model's performance is evaluated using the validation set. After training, the encoder model is able to encode the input key power feature dataset into low-dimensional latent layer features. Finally, the encoded latent features are stored as a latent power feature encoding dataset with a shape of [N,2], where 2 is the feature dimension of the latent layer.

[0058] Step S176: Decode the latent power feature encoding dataset to obtain the reconstructed power feature dataset.

[0059] Specifically, we can assume the shape of the latent power feature encoding dataset is [N, 2]. The decoder's structural parameter set has been defined in step S174. Decoder reconstruction is performed using the TensorFlow framework and its API Keras. The decoder part is extracted from the autoencoder model to construct an independent decoder model. The input to the decoder model is the latent power feature encoding dataset, and the output is the reconstructed power feature data. During reconstruction, the decoder gradually decodes the low-dimensional latent features back to the original feature space based on the trained weights and bias parameters. Specifically, the latent power feature encoding dataset is input into the decoder model. The decoder's input layer receives 2-dimensional latent features, performs a non-linear transformation through the ReLU activation function of the hidden layer, and finally reconstructs 5-dimensional power feature data through the sigmoid activation function of the output layer. The power feature dataset reconstructed by the decoder has the same shape [N, 5] as the original key power feature dataset.

[0060] This invention, through feature importance assessment and dynamic feature threshold iteration, can accurately select the most valuable key features for power system analysis. Encoding and decoding processes further extract deeper semantic information from these features. An autoencoder maps features to a low-dimensional latent space through nonlinear transformation and reconstructs features during decoding, enhancing not only the robustness and representational power of the features but also capturing the implicit structures and patterns in power data. The feature selection threshold set determined by feature importance assessment and dynamic threshold iteration allows for dynamic adjustment of feature selection criteria to adapt to different data distributions and analytical needs. The encoding training and decoding reconstruction process transforms complex feature data into a more compact and representative latent feature encoding dataset. During training, the autoencoder learns the inherent structure of the data, automatically removing the influence of noise and outliers. The reconstructed power feature dataset, processed by feature selection and autoencoder, not only retains the core information of the original data but also improves feature quality through dimensionality reduction and optimization.

[0061] Preferably, step S2 includes the following steps:

[0062] Step S21: Perform domain feature distribution quantization on the reconstructed power feature dataset to obtain a multi-domain power feature statistical data set;

[0063] Specifically, the reconstructed power feature dataset contains features from multiple domains, each corresponding to different operating states or environmental conditions of power equipment. The reconstructed power feature dataset is divided into multiple domains, for example, based on equipment type, into "power generation equipment domain," "transmission equipment domain," and "distribution equipment domain." For each domain, the statistical distribution of the features is calculated, including mean, variance, skewness, and kurtosis statistics. For example, `numpy.mean()` is used to calculate the feature mean, `numpy.var()` to calculate the variance, `scipy.stats.skew()` to calculate the skewness, and `scipy.stats.kurtosis()` to calculate the kurtosis. In practice, the reconstructed power feature dataset is assumed to have a shape of [N, D], where N is the number of samples and D is the feature dimension. For each domain, a corresponding subset is extracted, and the statistics for each feature are calculated. For example, for the power generation equipment domain, the calculated mean of feature 1 is 0.5, variance is 0.1, skewness is 0.2, and kurtosis is 3.0; for the transmission equipment domain, the calculated mean of feature 1 is 0.6, variance is 0.15, skewness is 0.1, and kurtosis is 2.8. These statistics are then summarized into a multi-domain power feature statistical data set.

[0064] Step S22: Based on the multi-domain power feature statistical data set, perform domain adaptation network mapping on the reconstructed power feature dataset to obtain the power domain mapping network parameter set;

[0065] Specifically, a domain adaptation network structure can be designed based on a multi-domain power feature statistical dataset. Assuming the feature dimension of the reconstructed power feature dataset is D, the input layer dimension of the domain adaptation network is also D. The network structure includes an input layer, several hidden layers (e.g., two fully connected layers), and an output layer. The number of neurons in the hidden layers can be set to D / 2 or D / 4, and the activation function is ReLU. The output layer has the same dimension as the input layer (i.e., D), and a linear activation function is chosen to output the mapped features. Next, a discriminator network is constructed, whose goal is to distinguish which domain the input features originate from. The input of the discriminator network is the output of the domain adaptation network, and the output is the probability distribution of the domain labels. Assuming two domains (domain A and domain B) in the multi-domain power feature statistical dataset, the output layer of the discriminator network will have two neurons, corresponding to the probability distributions of the two domains respectively. During training, an adversarial training mechanism is adopted. Specifically, this involves minimizing the discriminator's loss function (e.g., cross-entropy loss) while maximizing the confusion loss of the domain adaptation network (i.e., making the discriminator unable to distinguish between domains). During training, a learning rate of 0.001 was set, the optimizer was Adam, and the training epochs were 100. In each epoch, the parameters of the discriminator network were updated first; then the parameters of the domain adaptation network were updated to make its output features more difficult for the discriminator to distinguish. For example, assuming the initial discriminator loss was 0.6, after 100 epochs, the discriminator loss decreased to 0.3, indicating that the domain adaptation network successfully mapped features from different domains to a similar distribution space. Finally, the parameter set of the domain adaptation network obtained through training, including the network weights and biases, was used.

[0066] Step S23: Perform intra-domain sample clustering on the power domain mapping network parameter set to obtain a hierarchical intra-domain clustering feature dataset;

[0067] Specifically, hierarchical clustering can be used to cluster samples within a domain, with the scikit-learn library in Python as the specific tool. The power domain mapping network parameter set is extracted from the domain adaptation network. Clustering is then performed on the mapped features for each domain. For example, suppose the dataset contains three domains: power generation equipment, transmission equipment, and distribution equipment. For the power generation equipment domain, hierarchical clustering is performed using the AgglomerativeClustering class in scikit-learn, setting the number of clusters to K=6, the distance metric to Euclidean distance, and the Ward criterion as the clustering connection method. Through hierarchical clustering, the samples in the power generation equipment domain are divided into 6 sub-clusters, each representing a set of similar features within that domain. The above process is repeated for the transmission equipment and distribution equipment domains, respectively, to obtain the clustering results within each domain. This ultimately forms a hierarchical intra-domain clustered feature dataset. Each sample in the dataset not only contains its original features but also has a corresponding cluster label attached.

[0068] Step S24: Calculate the inter-domain distance based on the hierarchical intra-domain clustering feature dataset to obtain the cross-domain metric distance dataset;

[0069] Specifically, the scipy and scikit-learn libraries in Python can be used to calculate inter-domain distances. The cluster centers of each domain in the hierarchical domain clustering feature dataset are extracted. For each domain, the KMeans algorithm in scikit-learn is used to recalculate the cluster centers, which are then used as representative features of that domain. For example, for the six subclusters of the power generation equipment domain, the cluster centers of each subcluster are calculated, resulting in six cluster center vectors. The `scipy.spatial.distance.cdist` function is used to calculate the Euclidean distance between the cluster centers of different domains. Assume the power generation equipment domain has six cluster centers, the transmission equipment domain has five cluster centers, and the distribution equipment domain has four cluster centers. Through calculation, a cross-domain distance matrix is ​​obtained, with shapes [5,4] (distance matrix between the power generation equipment domain and the transmission equipment domain), [5,3] (distance matrix between the power generation equipment domain and the distribution equipment domain), and [4,3] (distance matrix between the transmission equipment domain and the distribution equipment domain). Finally, these distance matrices are summarized into a cross-domain metric distance dataset.

[0070] Step S25: Iteratively adjust the gradient of the power domain mapping network parameter set based on the cross-domain metric distance dataset to obtain an adaptive power domain mapping parameter set;

[0071] Specifically, iterative gradient tuning can be performed using the deep learning framework TensorFlow and its optimizer tools. A loss function is defined to measure the difference in distance between domains. For example, the MMD (Maximum Mean Difference) loss function is chosen, as it effectively measures the difference between feature distributions in different domains. During the iteration process, the Adam optimizer is used with a learning rate set to 0.0005. In practice, the cross-domain distance metric dataset is taken as input, and the loss value of the inter-domain distance (such as the MMD loss) is calculated. Backpropagation is then used to update the weights and biases of the domain mapping network. In each iteration, the optimizer adjusts the network parameters based on the gradient value of the loss function to reduce inter-domain differences. For example, assuming the initial inter-domain distance loss is 0.5, after 100 iterations, the loss value decreases to 0.1, indicating that the parameters of the domain mapping network are better aligned with the feature distributions of different domains. Finally, through multiple iterations of optimization, an adaptive power domain mapping parameter set is obtained.

[0072] Step S26: Perform feature projection on the reconstructed power feature dataset according to the adaptive power domain mapping parameter set to obtain the cross-domain power feature mapping dataset;

[0073] Specifically, the reconstructed power feature dataset can be input into an optimized domain mapping network. The structure of the domain mapping network is consistent with the structure defined in step S22, but the parameters have been updated to an adaptive power domain mapping parameter set. Through the forward propagation of the network, the input features are mapped to a new feature space. The feature vector of each sample is input into the domain mapping network, and after nonlinear transformation of the hidden layer and linear transformation of the output layer, the mapped feature vector is obtained. For example, assuming the dimension of the input features is D = 10, after passing through the domain mapping network, the dimension of the output features is compressed to D. 1 =5. Through the above operations, the features of all samples are projected to obtain the cross-domain power feature mapping dataset. The shape of this dataset is [N,D]. 1 ], where D 1 It is the feature dimension after mapping.

[0074] Step S27: Perform domain difference elimination on the cross-domain power feature mapping dataset to obtain a domain-aligned power feature dataset, and calculate attention weights on the domain-aligned power feature dataset to obtain a cross-domain fused power feature dataset.

[0075] For a detailed implementation process of this embodiment, please refer to the sub-step of step S27.

[0076] This invention aligns power characteristic data from different domains to the same feature space through domain feature distribution quantization and domain-adaptive network mapping, significantly reducing inter-domain differences. Hierarchical clustering and inter-domain distance calculation enable in-depth mining of the data's intrinsic structure and distribution characteristics. Iterative gradient adjustment dynamically modulates the parameters of the domain mapping network, adaptively adapting to differences between domains. Feature projection and domain discrepancy elimination further eliminate inconsistencies in cross-domain data. Attention weight calculation automatically identifies and highlights the most valuable key features for power system analysis.

[0077] Preferably, step S27 includes the following steps:

[0078] Step S271: Construct a discriminator network on the cross-domain power feature mapping dataset to obtain an adversarial discriminator network;

[0079] Specifically, the discriminator network can be built using the deep learning framework TensorFlow and its API Keras. The discriminator network structure is designed as follows: First, define the input layer of the discriminator network, whose dimension is consistent with the feature dimension. Then, perform a non-linear transformation on the input features through two hidden layers, and finally output the probability that a sample belongs to a certain domain through the sigmoid activation function. For example, suppose the input feature dimension is D. 1If the input layer has 10 neurons, the discriminator network structure can be represented as follows: Input layer: 10 neurons; Hidden layer 1: 64 neurons, activation function is ReLU; Hidden layer 2: 32 neurons, activation function is ReLU; Output layer: 1 neuron, activation function is Sigmoid. Through the above steps, a network model capable of performing domain discrimination on input features is constructed.

[0080] Step S272: Perform adversarial training on the cross-domain power feature mapping dataset based on the adversarial discriminator network to obtain the domain adversarial loss power dataset;

[0081] Specifically, a cross-domain electricity feature mapping dataset can be input into an adversarial discriminator network to calculate the discriminator's output probability. Simultaneously, an adversarial loss function is defined. For example, a binary cross-entropy loss function can be chosen, which measures the difference between the discriminator's output probability and the true domain label. During adversarial training, the parameters of both the discriminator network and the domain mapping network are updated simultaneously. The specific operations are as follows: Updating the discriminator network: The parameters of the discriminator network are updated using the Adam optimizer by minimizing the discriminator's loss function (i.e., maximizing the discriminator's discriminative ability). For example, a learning rate of 0.001 and 50 training epochs can be set. Updating the domain mapping network: The parameters of the domain mapping network are updated using the same optimizer by maximizing the discriminator's loss function (i.e., minimizing the discriminator's discriminative ability). For example, during training, assuming the initial discriminator loss is 0.6, after adversarial training, the discriminator loss gradually decreases to 0.3, indicating that the domain mapping network has successfully generated features that are difficult to distinguish. Finally, the domain adversarial loss electricity dataset obtained through adversarial training records the loss value for each training session.

[0082] Step S273: Adjust the features of the cross-domain power feature mapping dataset based on the domain adversarial loss power dataset to obtain the domain-aligned power feature dataset;

[0083] Specifically, the domain adversarial loss power dataset can be analyzed to extract the loss value for each iteration. Based on the trend of the loss value, the feature dimensions that need adjustment are determined. For example, if the loss value of a certain feature dimension is high, it indicates that the feature has significant differences across different domains and needs adjustment. In practice, a feature adjustment module is defined, taking the cross-domain power feature mapping dataset and the domain adversarial loss dataset as inputs. The goal of the feature adjustment module is to minimize the domain adversarial loss while maintaining the representational power of the features. Gradient descent is used to adjust the features. For example, for feature dimensions with high loss values, the feature values ​​are gradually adjusted by calculating the gradient of the loss function to make them more consistent across different domains. Assuming the initial domain adversarial loss is 0.3, after feature adjustment, the loss is reduced to 0.1, indicating that the feature adjustment is effective. Finally, the adjusted feature dataset is called the domain-aligned power feature dataset, and its shape remains [N, D]. 1 ].

[0084] Step S274: Calculate attention weights on the domain-aligned power feature dataset to obtain a multimodal power feature weight dataset;

[0085] Specifically, attention weights can be computed using the deep learning framework TensorFlow and its API Keras. An attention mechanism module is designed, taking a domain-aligned power feature dataset as input. The goal of the attention mechanism is to assign a weight to each feature. In practice, an attention layer is defined, consisting of a fully connected layer and a softmax activation function. The input dimension of the fully connected layer is D. 1 The output dimension is also D. 1 The activation function chosen is ReLU. The Softmax activation function normalizes the output values ​​into weights, ranging from [0,1]. For example, suppose the input feature dimension is D. 1 =10, the structure of the attention layer can be represented as follows: fully connected layer: input dimension 10, output dimension 10, activation function is ReLU; softmax activation function: normalizes the output value into weights; through the above steps, a weight value is calculated for each feature. Finally, the calculated weight values ​​are combined with the domain-aligned power feature dataset to obtain the multimodal power feature weight dataset.

[0086] Step S275: Construct a weight optimization target based on the multimodal power feature weight dataset to obtain the power feature weight optimization target dataset, and perform parameter iterative updates on the multimodal power feature weight dataset based on the power feature weight optimization target dataset to obtain the optimized power feature weight dataset;

[0087] Specifically, the TensorFlow deep learning framework can be used to construct the weight optimization objective. A weight optimization objective function is defined, aiming to minimize the variance of the feature weights while maximizing the correlation between the feature weights and the target variable (such as power system fault prediction labels). The weight optimization objective function can be expressed as: Loss = α·Var(w) - β·Corr(w,y); where w is the feature weight, y is the target variable, and α and β are hyperparameters, set to 0.1 and 0.9, respectively. The Adam optimizer is used to iteratively update the feature weights. The multimodal power feature weight dataset is input into the optimization objective function, the loss value is calculated, and the weight parameters are updated through backpropagation. For example, if the initial feature weight distribution is relatively dispersed with a variance of 0.2, after 100 iterations, the variance decreases to 0.05, while the correlation between the weights and the target variable increases from 0.6 to 0.8. Finally, the power feature weight optimization objective dataset is obtained, and the updated weight values ​​are stored as the optimized power feature weight dataset.

[0088] Step S276: Based on the optimized power feature weight dataset, the domain-aligned power feature dataset is weighted and fused to obtain an initial fused power feature dataset. The feature consistency of the initial fused power feature dataset is then verified to obtain an inter-domain feature consistency evaluation dataset.

[0089] Specifically, the domain-aligned power feature dataset and the optimized power feature weight dataset can be multiplied element-wise to obtain a weighted feature matrix. For example, assuming the feature dataset has a shape of [1000, 10] and the weight dataset has a shape of

[10] , element-wise multiplication yields the initial fused power feature dataset, which also has a shape of [1000, 10]. Cosine similarity is used as the consistency evaluation metric for feature consistency verification of the initial fused power feature dataset. Specifically, the cosine similarity between samples from different domains is calculated to generate an inter-domain feature consistency evaluation dataset. For example, assuming the dataset contains samples from two domains, calculating the cosine similarity yields a consistency evaluation dataset with a shape of [500, 500], where each element represents the similarity between a sample and other samples. Finally, through weighted fusion and consistency verification, the initial fused power feature dataset and the inter-domain feature consistency evaluation dataset are obtained.

[0090] Step S277: Perform consistency calibration on the initial fused power feature dataset based on the inter-domain feature consistency assessment dataset to obtain the cross-domain fused power feature dataset.

[0091] Specifically, the inter-domain feature consistency assessment dataset can be analyzed to calculate the consistency score for each feature dimension. The method involves evaluating the similarity of each feature dimension across different domains by calculating the mean and standard deviation of the cosine similarity. If the consistency score of a feature dimension is lower than a preset threshold (e.g., 0.8), it is considered inconsistent and requires calibration. For inconsistent feature dimensions, the following calibration methods are used: Normalization: Inconsistent feature dimensions are normalized to limit their values ​​to [0,1]. Weighted average calibration: Inconsistent feature dimensions are calibrated using a weighted average based on the inter-domain consistency score. Specifically, for the value of a feature dimension across different domains, the weights are dynamically adjusted based on its consistency score. For example, if the consistency score of a feature dimension between domain A and domain B is 0.7, which is lower than the threshold of 0.8, then a weighted average is applied to that feature dimension. Through these steps, the final feature dataset after consistency calibration is called the cross-domain fused power feature dataset.

[0092] This invention effectively identifies and eliminates inter-domain differences in cross-domain feature data by constructing an adversarial discriminator network and conducting adversarial training. Through weight calculation and optimization, it can automatically identify and highlight the key features most valuable for power system analysis. By constructing the weight optimization objective and iteratively updating the parameters, the feature weights can be dynamically adjusted to ensure that the fusion process always progresses in the optimal direction. By verifying and calibrating the feature consistency of the initial fused dataset, potential inconsistencies in the fusion process can be further eliminated.

[0093] Preferably, step S3 includes the following steps:

[0094] Step S31: Collect topological relationships of distributed power nodes to obtain a node connection relationship dataset;

[0095] Specifically, the target power network comprises multiple distributed nodes, such as substations, power plants, and user-end equipment, interconnected by transmission lines. Real-time operational data of the power network, including node voltage, current, and power information, is acquired through an automated monitoring system (such as a SCADA system). Network topology analysis tools (such as MATLAB's Power System Analysis Toolbox or Python's NetworkX library) are used to model the connections between nodes. Node identifiers (such as node IDs) and connection information (such as line start and end points) are extracted from the SCADA system. For example, assuming a transmission line exists between node A and node B, this is recorded as a connection (A, B), along with the line's attributes, such as a length of 10.5 km and an impedance of 0.25 ohms. Through these steps, a node connection dataset is constructed. The dataset is stored in matrix or tabular form, with each row representing a connection, including the start node ID, end node ID, and the attributes of the connecting line.

[0096] Step S32: Evaluate the link status of distributed power nodes based on the node connection relationship dataset to obtain the node link quality dataset;

[0097] Specifically, link-state assessment can be performed using power system analysis software (such as MATLAB's SimPowerSystems) and Python's data processing tools. Attributes for each line are extracted from the node connectivity dataset. Combined with power system operating data (such as current, voltage, and power), the losses, voltage drop, and transmission efficiency of each line are calculated. For each line, its transmission loss and voltage drop are calculated. For example, assuming the impedance of line (A, B) is 0.25 ohms and the current through the line is 100 amperes, then the line loss is I. 2 ×R=(100) 2 ×0.25 = 2500 watts. Simultaneously, the voltage drop across the line is calculated as I×R = 100×0.25 = 25 volts. Based on the calculation results, the link quality of each line is evaluated. For example, the link quality metric is defined as transmission efficiency divided by the sum of loss and voltage drop. Assuming the input power of line (A,B) is 100 kW and the output power is 95 kW, then the transmission efficiency is 95%. Link quality can be calculated using the above metrics. Finally, the link quality metrics for each line are stored as a node link quality dataset. The dataset contains the starting node ID, ending node ID, loss, voltage drop, transmission efficiency, and link quality information for each line.

[0098] Step S33: Calculate the connection weights of the distributed power nodes based on the node link quality dataset to obtain the node connection weight dataset, and construct a dynamic adjacency matrix based on the node connection weight dataset to obtain the distributed node topology data.

[0099] Specifically, the link quality index for each line can be extracted from the node link quality dataset. Higher link quality indicates higher transmission efficiency and lower loss, thus assigning a higher weight. For example, the connection weight can be defined as a normalized value of the link quality, ranging from [0,1]. Assuming the link quality of line (A,B) is 0.85, the link quality of line (B,C) is 0.88, and the link quality of line (C,D) is 0.82, the normalized weights are 0.85, 0.88, and 0.82, respectively. A dynamic adjacency matrix is ​​constructed based on the node connection weight dataset. The adjacency matrix is ​​a two-dimensional matrix where rows and columns represent nodes, and elements represent the connection weights between nodes. The construction process of the dynamic adjacency matrix is ​​as follows: For node A, which is connected to node B with a weight of 0.85, the position from A to B in the adjacency matrix is ​​recorded as 0.85; similarly, the position from B to A is also recorded as 0.85. Node B is connected to node C with a weight of 0.88, therefore the position record from B to C is 0.88, and the position record from C to B is 0.88. Node C is connected to node D with a weight of 0.82, therefore the position record from C to D is 0.82, and the position record from D to C is 0.82. For nodes that are not directly connected, the corresponding position record in the adjacency matrix is ​​0. Finally, through connection weight calculation and the construction of a dynamic adjacency matrix, the distributed node topology data is obtained.

[0100] Step S34: Based on the distributed node topology data, plan the feature extraction layer for the distributed power nodes to obtain the parameter set of the power node graph convolutional layer;

[0101] Specifically, the feature extraction layer can be planned using the deep learning framework TensorFlow and its API Keras, combined with the Graph Neural Network (GNN) toolkit. The structural parameters of the graph convolutional layer are defined, including the input feature dimension, hidden layer dimension, and output feature dimension. For example, suppose the input feature dimension is D. in =10 (such as information like node voltage and current), the hidden layer dimension is D hiddden =16, output feature dimension is D out =8. Constructing Graph Convolutional Layers. The core operation of graph convolutional layers is to aggregate the feature information of neighboring nodes and perform a linear transformation using the weight matrix. In TensorFlow, `tf.keras.layers.Dense` is used to implement the transformation of the weight matrix and combine it with the adjacency matrix for feature aggregation. For example, the graph convolution operation for the hidden layer can be represented as: H (L+1) =σ(AH (L) W (L) ); where H (L) W is the node feature matrix of the Lth layer. (L)Here, A is the weight matrix, A is the adjacency matrix, and σ is the activation function (such as ReLU). Through the above steps, the parameter set of the power node graph convolutional layer is defined, including the input feature dimension, hidden layer dimension, output feature dimension, and activation function.

[0102] Step S35: Construct a heterogeneous message passing mechanism based on the parameter set of the power node graph convolutional layer to obtain the power node message passing rule set;

[0103] Specifically, a heterogeneous message passing mechanism can be constructed using the deep learning framework PyTorch and its graph neural network library PyTorch Geometric. The input and output feature dimensions of the message passing are defined based on the parameter set of the power node graph convolutional layers. For example, the input feature dimension is D. in =10, hidden layer dimension is D hiddden =16, output feature dimension is D out =8. In PyTorch Geometric, the MessagePassing module is used to implement a custom message passing mechanism. For example, define a message passing rule where each node aggregates the features of its neighboring nodes and performs a linear transformation through a learnable weight matrix. The specific rule is as follows: Message function: For each edge (i,j), compute the message m from node i to node j. ij :m ij =W msg ·h i ; where h i W is a feature of node i. msg This is the message weight matrix. Aggregation function: Aggregates the messages of all neighboring nodes of node j: m j =AGGREGATE({m ij |i∈N(j)}); where N(j) is the set of adjacent nodes of node j, and the aggregation method can be summation, averaging, or maximum value. Update function: Update the characteristics of node j based on the aggregated message: h j ` =W update (h j +m j ); where W update This involves updating the weight matrix. By defining these message passing rules, a message passing rule set for power nodes is constructed.

[0104] Step S36: Construct a graph neural architecture based on the power node message passing rule set to obtain power node graph network structure data;

[0105] Specifically, the structure of the graph convolutional layer can be defined based on the power node message passing rule set. In PyTorchGeometric, the GCNConv module can be used to implement graph convolution operations, or a custom graph convolutional layer based on MessagePassing can be defined. For example, a two-layer graph convolutional network can be defined, where the input feature dimension of the first layer is D. in =10, hidden layer dimension is D hiddden =16, the output feature dimension of the second layer is D out =8. The specific operation is as follows: First graph convolutional layer: input feature dimension is 10, output feature dimension is 16. The ReLU activation function is used to perform a non-linear transformation on the output features. Second graph convolutional layer: input feature dimension is 16, output feature dimension is 8. The output features of this layer will be used as the final node feature representation of the graph network. By stacking these two graph convolutional layers, the power node graph network structure is constructed.

[0106] Step S37: Based on the power node graph network structure data, perform distributed model training on the cross-domain fused power feature dataset to obtain the local power model parameter set, and perform federated averaging on the local power model parameter set to obtain the global optimized model parameter set.

[0107] For a detailed implementation process of this embodiment, please refer to the sub-step of step S37.

[0108] This invention characterizes the node connectivity and link quality of power networks through topology relationship acquisition and link state assessment. By calculating connection weights and dynamically constructing an adjacency matrix, the connection strength between nodes can be adaptively adjusted according to the actual link quality. Through feature extraction layer planning and the construction of a heterogeneous message passing mechanism, node features can be efficiently extracted and information can be transmitted between nodes. By employing distributed model training combined with federated averaging, model training can be performed in parallel on distributed power nodes, and global optimization can be achieved through a federated learning mechanism. Through the construction of a graph neural network architecture and distributed training, the model can better adapt to the complex topology and dynamic changes of power networks.

[0109] Preferably, step S37 includes the following steps:

[0110] Step S371: Divide the cross-domain fusion power feature dataset into distributed training batches based on the power node graph network structure data to obtain the distributed training batch parameter set;

[0111] Specifically, based on the power node graph network structure data, the data volume and computational resource allocation for each node are determined. For example, assuming the entire dataset contains 100,000 samples distributed across 10 distributed nodes, each node has approximately 10,000 samples. The data for each node is further divided into multiple batches. Assuming each batch size is 1,000 samples, each node will be divided into 10 batches. A distributed training batch parameter set is defined, including the size of each batch, the number of batches, and the node allocation. For example, for node 1, its training batch parameter set can be represented as: node ID is 1, batch size is 1,000, and the total number of batches is 10. Similarly, corresponding batch parameter sets are generated for other nodes.

[0112] Step S372: Asynchronously and parallelly schedule the cross-domain fused power feature dataset according to the distributed training batch parameter set to obtain the parallel training task allocation dataset;

[0113] Specifically, training tasks can be assigned to each node based on the distributed training batch parameter set. For example, assuming each node has 10 batches, and each batch contains 1,000 samples, training tasks will be assigned to each node in batch order. Asynchronous parallel scheduling is then initiated. In asynchronous mode, each node executes its assigned training task independently without waiting for other nodes to complete. For example, node 1 starts processing the first batch, node 2 starts processing the first batch simultaneously, and so on. After completing the training of a batch, each node immediately sends the results to the master node for aggregation before continuing to process the next batch. Finally, a parallel training task allocation dataset is generated, recording the task allocation and completion status of each node.

[0114] Step S373: Configure the distributed optimizer according to the dataset allocated for the parallel training task to obtain the distributed optimizer configuration parameter set;

[0115] Specifically, an appropriate optimizer can be selected based on the training workload and computational resources of each node. For example, for an electricity feature training task, the Adam optimizer can be chosen. Optimizer parameters, including learning rate and weight decay, are configured for each node. Assume an initial learning rate of 0.001 and a weight decay of 0.0001. In a distributed environment, a synchronization mechanism for the optimizer also needs to be configured. For example, using PyTorch's DistributedDataParallel module, setting gradient_as_bucket_view = True optimizes gradient synchronization efficiency, and using the all_reduce operation of the torch.distributed module ensures that the gradients of all nodes are updated synchronously after each iteration. Finally, the generated distributed optimizer configuration parameter set will include the optimizer type, learning rate, weight decay, and synchronization mechanism parameters for each node.

[0116] Step S374: Design a learning rate scheduling strategy for the distributed optimizer configuration parameter set to obtain the power characteristic scheduling learning rate parameter set;

[0117] Specifically, the learning rate scheduler can be selected based on the characteristics of the training task and the features of the distributed environment. For example, the torch.optim.lr_scheduler.CosineAnnealingLR scheduler can be chosen. Assuming the total number of iterations for the training task is 1000, the learning rate scheduler period is set to 500, and the minimum learning rate is set to 1... e-6 Configure the parameters for the learning rate scheduler. For example, the initial learning rate is 0.001, the number of cycles is 500, and the minimum learning rate is 1. e-6 During training, the learning rate will be adjusted according to the cosine annealing strategy. When the number of iterations reaches 500, the learning rate will gradually decrease to a minimum of 1. e -6 Then, the next cycle begins again. Through the above steps, the final generated power characteristic scheduling learning rate parameter set will include the type of learning rate scheduler, the initial learning rate, the cycle length, and the minimum learning rate parameter.

[0118] Step S375: Calculate the distributed gradient of the model for the distributed power nodes based on the power feature scheduling learning rate parameter set to obtain the power node gradient update dataset;

[0119] Specifically, distributed gradient computation of the model can be performed using the deep learning framework PyTorch and its distributed training module torch.distributed. The learning rate parameter set is scheduled according to the power feature data, configuring the learning rate for the current iteration for each node. For example, in the 100th iteration of training, the learning rate scheduler adjusts the learning rate to 0.0005 according to a preset cosine annealing strategy. Each node independently performs forward and backward propagation computations on its assigned training batch. During forward propagation, the model generates a predicted output based on the input power feature data; during backward propagation, the loss function (such as mean squared error or cross-entropy loss) between the predicted output and the true label is calculated, and the gradient of the model parameters is calculated according to the chain rule. For example, assuming a node's current batch contains 1000 samples and the model's loss on that batch is 0.35, the gradient calculated through backpropagation will be used to update the model parameters. After each node completes the gradient computation, the calculated gradient is stored as a power node gradient update dataset, which contains the gradient information of each node on the current batch.

[0120] Step S376: Asynchronously aggregate the gradients of the power node gradient update dataset to obtain a global gradient aggregate dataset, and update the model parameters based on the global gradient aggregate dataset to obtain a local model update parameter set;

[0121] Specifically, a global gradient aggregation strategy can be defined using PyTorch's `torch.distributed` module. In asynchronous mode, each node immediately sends its gradient to the master node (or parameter server) after completing its local gradient calculation. The master node is responsible for collecting gradients from all nodes and performing aggregation operations. For example, using an average aggregation strategy, the gradients of all nodes are averaged element-wise to obtain the global gradient. This global gradient aggregation dataset will be used to update the model parameters. The model parameters are updated based on the global gradient aggregation dataset. Specifically, the global gradient is applied to the model's current parameters, and the optimizer (such as Adam) completes the parameter update. For example, assuming the current model parameters are `params`, the optimizer updates the model parameters based on the global gradient `global_grad` and the current learning rate (e.g., 0.0005), resulting in new model parameters `new_params`. Finally, the updated model parameters `new_params` are broadcast back to each node as the local model update parameter set.

[0122] Step S377: Perform a weighted average of the local power model parameter set to obtain a federated average weight dataset, and perform federated aggregation of the local power model parameter set based on the federated average weight dataset to obtain a global optimization model parameter set.

[0123] Specifically, the weights of each node can be determined. Weights can be allocated based on the amount of data a node has or its model performance. For example, suppose there are four nodes in a distributed system, each with 1000, 1200, 1100, and 900 samples respectively. The weight of each node can be set as the proportion of its data volume to the total data volume. With a total data volume of 4200 samples, the weight of node 1 would be 1000 / 4200, the weight of node 2 would be 1200 / 4200, and so on. A weighted average of the model parameters for each node is then calculated. This involves multiplying the model parameters of each node by its corresponding weight, and then summing the weighted parameters of all nodes. Through these steps, a federated average weight dataset is obtained, reflecting each node's contribution to the global model. Finally, the weighted average parameters are used as the new parameters for the global model and broadcast back to each node. Each node updates its local model using the globally optimized model parameter set, preparing for the next round of training.

[0124] This invention, through distributed training batch partitioning and asynchronous parallel scheduling, fully utilizes the computing resources of distributed power nodes to achieve efficient parallel training of cross-domain fused power feature data. Distributed optimizer configuration assigns the most suitable optimization strategy to each distributed node. Distributed gradient computation ensures that the model on each node is updated with the optimal learning rate during training. Asynchronous gradient aggregation and federated averaging mechanisms effectively integrate local model parameters across distributed nodes.

[0125] Preferably, step S4 includes the following steps:

[0126] Step S41: Construct a neural network on the parameter set of the global optimization model to obtain a multimodal teacher network, and enhance the multimodal power features of the multimodal teacher network to obtain a multimodal power teacher model;

[0127] Specifically, a multimodal teacher network can be constructed using the deep learning framework TensorFlow and its API Keras. First, the input layer of the network is defined to accommodate different modalities of power data, such as voltage, current, power, and image data. For structured data (e.g., voltage, current), the input layer can be a fully connected layer; for image data, it can be a convolutional layer. Next, a multimodal fusion layer is constructed to fuse features from different modalities. For example, a fully connected layer with 128 neurons can be used, with ReLU activation function selected, mapping features from different modalities to a shared feature space. An attention mechanism module is introduced to achieve cross-modal feature enhancement. For example, a self-attention layer is used to assign weights based on the similarity between features, highlighting important feature information. Through the above steps, after neural network construction and cross-modal feature enhancement, a multimodal teacher model for power is obtained.

[0128] Step S42: Perform feature hierarchical decomposition on the power multimodal teacher model to obtain the teacher semantic knowledge feature set;

[0129] Specifically, the structure of the multimodal power teacher model can be analyzed to determine the feature extraction layers at different levels. For example, assuming the teacher model contains multiple convolutional and fully connected layers, the convolutional layers extract low-level features (such as edges and textures), while the fully connected layers extract high-level semantic features (such as equipment status and fault modes). A specific feature extraction layer is selected as the benchmark for decomposition. For example, the output of the second-to-last fully connected layer can be chosen as the representative of the high-level semantic features. By inserting hook functions into the model, the feature output of that layer is extracted. Assuming the output dimension of this layer is 64, the extracted features can be represented as a 64-dimensional vector, with each dimension corresponding to a semantic feature. Through the above steps, the teacher semantic knowledge feature set is obtained.

[0130] Step S43: Perform feature selection on the teacher semantic knowledge feature set to obtain the key power representation knowledge feature set, and perform neural architecture search on the key power representation knowledge feature set to obtain the power lightweight student model.

[0131] Specifically, information gain-based methods can be used to evaluate the importance of each feature in the teacher's semantic knowledge feature set. For example, the mutual information between each feature and the target variable (e.g., the state of power equipment) is calculated, and the top 20% of features with the highest information gain are selected as the key power representation knowledge feature set. Assuming the teacher's semantic knowledge feature set has a dimension of 100, the 20 features with the highest information gain are selected as key features after information gain evaluation. Neural architecture search (NAS) methods based on reinforcement learning, such as ENAS (Efficient Neural Architecture Search), are used. During the search process, the search space is defined, including different types of neural network layers (e.g., convolutional layers, fully connected layers) and the connection methods between layers. For example, the search space includes the number of filters in the convolutional layers (e.g., 32, 64, 128), the kernel size (e.g., 3×3, 5×5), and whether batch normalization is used, among other hyperparameters. The optimal network structure is automatically searched using the ENAS algorithm. For example, after multiple rounds of searching, a network structure was found consisting of two convolutional layers (64 filters per layer, with a kernel size of 3×3) and two fully connected layers (128 neurons per layer). Finally, a power-lightweight student model was constructed through feature selection and neural architecture search.

[0132] Step S44: Align and extract features between the lightweight power student model and the multimodal power teacher model to obtain the cross-modal feature alignment loss dataset;

[0133] Specifically, a feature alignment target can be defined. Assume the teacher model's feature extraction layer outputs a 64-dimensional feature, while the corresponding layer in the student model outputs a 32-dimensional feature. A fully connected layer is used to map the student model's features to the same dimension as the teacher model. For example, adding a fully connected layer maps the 32-dimensional features to 64 dimensions. The feature alignment loss is then calculated. Mean squared error (MSE) is chosen as the loss function to measure the difference between the student model's features and the teacher model's features. For example, assuming the teacher model's features are `teacher_features` and the student model's aligned features are `student_features`, the feature alignment loss can be expressed as: `alignment_loss = MSE(teacher_features, student_features)`. By calculating the feature alignment loss for each sample, a cross-modal feature alignment loss dataset is generated, which records the errors made by the student model in learning the teacher model's features.

[0134] Step S45: Design a knowledge distillation strategy based on the cross-modal feature alignment loss dataset to obtain a multi-level power knowledge distillation strategy parameter set, and perform distillation optimization on the power lightweight student model based on the multi-level power knowledge distillation strategy parameter set to obtain the model distillation optimization parameter set;

[0135] Specifically, knowledge distillation strategies can be designed using the deep learning framework TensorFlow and its API Keras. A distillation loss function is defined, combining the feature alignment loss and the soft target (softmax output) of the teacher model. For example, the distillation loss can be expressed as a weighted sum of the feature alignment loss and the soft target loss, with weights of 0.5 and 0.5 respectively. Assuming the soft target output of the teacher model is `teacher_softmax` and the output of the student model is `student_output`, the soft target loss can be measured using KL divergence.

[0136] distillation_loss=θ·MSE(teacher_features,student_features)+(1-θ)·KL(teacher_softmax,student_output);

[0137] Here, θ represents the weight of the feature alignment loss, set to 0.5. Based on the cross-modal feature alignment loss dataset, the parameters in the distillation strategy are adjusted. For example, if the feature alignment loss is high for certain modalities, the feature alignment weight for those modalities can be increased. Through multiple rounds of experiments, the optimal set of distillation strategy parameters, including feature alignment weights, soft target weights, and the learning rate, is determined. Finally, the lightweight student model for power is distilled and optimized based on the multi-level power knowledge distillation strategy parameter set, resulting in the model distillation optimization parameter set.

[0138] Step S46: Compress the model distillation optimization parameter set to obtain a lightweight compressed model parameter set, and construct an inference model for the power multimodal teacher model based on the lightweight compressed model parameter set to obtain a power feature extraction and fusion model.

[0139] Specifically, model compression can be performed using the deep learning framework TensorFlow and its model compression tools, such as the TensorFlow Model Optimization Toolkit. The optimized student model is then pruned, removing redundant neurons and connections. For example, a pruning ratio of 50% can be set, removing 50% of the weights while maintaining the model's structure. Pruning significantly reduces the number of parameters and computational complexity. The pruned model is then quantized, converting the weights from floating-point numbers to low-bit-width integers (e.g., 8-bit integers). For example, TensorFlow's dynamic quantization tool can be used to quantize the model's weights and activation functions into 8-bit integers. Finally, an electric feature extraction fusion model is constructed based on the lightweight compressed model parameter set.

[0140] This invention effectively integrates power data features from different modalities by constructing a multimodal teacher network and performing cross-modal feature enhancement, thereby improving the model's comprehensive representation ability of multi-source heterogeneous data. By utilizing feature hierarchical decomposition and neural architecture search to construct a lightweight power student model, key semantic knowledge from the teacher model can be efficiently transferred to the student model. Feature alignment ensures that the student model accurately inherits the feature representation capabilities of the teacher model during the learning process. A multi-level power knowledge distillation strategy designed based on a cross-modal feature alignment loss dataset dynamically adjusts parameters during knowledge transfer, ensuring that the student model efficiently learns the teacher model's knowledge at different levels. Model compression further reduces the model's storage and computational requirements while retaining key feature extraction capabilities.

[0141] Preferably, step S5 includes the following steps:

[0142] Step S51: Perform a reliability assessment on the multi-source heterogeneous power dataset to obtain power multimodal quality assessment data;

[0143] Specifically, the IsolationForest algorithm in scikit-learn can be used to detect outliers in multi-source heterogeneous power datasets. For example, setting the contamination parameter of IsolationForest to 0.05 indicates that the proportion of outliers in the dataset is approximately 5%. This algorithm scores the reliability of each sample, with the reliability score ranging from 0 to 1; a value closer to 1 indicates more reliable data. Finally, based on the reliability score, high-quality data is selected as power multimodal quality assessment data. For example, data with a reliability score greater than 0.8 are selected as high-quality data.

[0144] Step S52: Perform state-space modeling on the power multimodal quality assessment data to obtain a power characteristic state-space dataset;

[0145] Specifically, state-space modeling can be performed using the statsmodels and PyEMD libraries in Python. The SARIMAX model in statsmodels is used to model time-series data (such as voltage, current, and power). For example, assuming voltage data exhibits seasonal variations, the parameters of the SARIMAX model are chosen as (p=1, D=1, q=1)×(P=1, D=1, Q=1, s=24), indicating that the model considers diurnal periodic variations. For nonlinear features (such as equipment status, ambient temperature, and humidity), the Empirical Mode Decomposition (EMD) method in the PyEMD library is used to decompose them into multiple intrinsic mode functions (IMFs). For example, EMD decomposition is performed on equipment status data to extract IMF components reflecting changes in equipment operating status. Each IMF component can be considered a state variable used to describe the dynamic behavior of the system. Finally, the output of the time-series model and the IMF components obtained from EMD decomposition are integrated into a power feature state-space dataset.

[0146] Step S53: Perform reinforcement learning on the power feature state space dataset to obtain a dynamic fusion strategy for power data;

[0147] Specifically, a reinforcement learning environment can be defined, where the state space is the power feature state space dataset, and the action space is the parameter adjustment of the data fusion strategy. For example, actions may include selecting different feature weights or adjusting the hyperparameters of the fusion model. A reinforcement learning algorithm, such as the Soft Actor-Critic (SAC) algorithm, is selected. During training, the agent learns the optimal strategy through interaction with the environment. For example, the reward function can be set as a weighted sum of the system state stability index (such as the negative value of voltage fluctuation) and the prediction accuracy. The agent's goal is to maximize the cumulative reward, i.e., to improve the accuracy of data fusion while ensuring system stability. Through multiple rounds of training, the agent gradually learns the optimal dynamic power data fusion strategy. For example, after 1000 iterations of training, the agent can dynamically adjust feature weights according to the current state to adapt to different operating conditions. Ultimately, the obtained dynamic power data fusion strategy can automatically adjust the data fusion process according to the real-time state, thereby maintaining the optimal fusion effect in different scenarios.

[0148] Step S54: Perform sensitivity evaluation on the power feature extraction and fusion model to obtain the power mode feature sensitivity set;

[0149] Specifically, the input to the power feature extraction fusion model can be defined as multimodal features, and the output as predicted values ​​of system states (such as equipment failure probability or voltage stability). The Integrated Gradients (IG) method is used to calculate the contribution of each modal feature to the model output. For example, for the voltage feature, its integral gradient with respect to the model output is calculated. Assuming the sensitivity of the voltage feature is 0.6, indicating a significant impact on the model output, while the sensitivity of the ambient temperature feature is 0.2, indicating a relatively small impact, the above steps are used to evaluate the sensitivity of all modal features, and the results are summarized into a power modal feature sensitivity set.

[0150] Step S55: Based on the power mode feature sensitivity set, the power data dynamic fusion strategy is weighted to obtain the power feature weight allocation dataset;

[0151] Specifically, each modal feature can be normalized according to the power modal feature sensitivity set, so that the sum of its weights is 1. For example, the normalized weight of the voltage feature is 0.36, the current feature is 0.30, the power feature is 0.24, the temperature feature is 0.12, and the humidity feature is 0.08. These weight values ​​are then combined with a dynamic power data fusion strategy. Specifically, the weight of each modal feature is assigned to the corresponding feature processing module in the dynamic fusion strategy. For example, in the fusion model, the voltage feature processing module will be given a higher weight (0.36), while the humidity feature processing module will have a lower weight (0.08). Through the above steps, the generated power feature weight allocation dataset will contain each modal feature and its corresponding weight value.

[0152] Step S56: Adjust the parameters of the power feature extraction and fusion model according to the power feature weight allocation dataset to obtain the optimal power fusion decision model.

[0153] Specifically, the parameters of each feature processing module in the model can be adjusted based on the weighted dataset. For example, for the voltage feature processing module, its weight in the model can be increased (e.g., by increasing the number of neurons in the relevant layers or adjusting the learning rate). Specifically, the output weight of the voltage feature processing module could be adjusted to 0.36, the output weight of the current feature module to 0.30, and so on. Next, the adjusted model is retrained. During training, the weighted feature outputs are used to optimize the model's loss function. For example, a weighted loss function can be defined where the loss contribution of each modal feature is proportional to its weight. Through the above steps, after parameter adjustment and retraining, the resulting model is the optimal power fusion decision model.

[0154] Step S57: Use the optimal power fusion decision model to perform semantic association mining on the multi-source heterogeneous power dataset to obtain the power multimodal semantic association dataset.

[0155] Specifically, multi-source heterogeneous power datasets can be input into the optimal power fusion decision model. The model can identify the relationship between voltage fluctuations and equipment status, or the impact of ambient temperature changes on power. Define the objective of semantic association mining. For example, suppose the objective is to mine the association between equipment faults and multimodal data. Extract feature representations related to equipment faults through the model's output layer. These feature representations can be activation values ​​from intermediate layers of the model, or weighted features obtained through an attention mechanism. For example, use the activation values ​​from the penultimate layer of the model as feature representations, with a dimension of 128. Use cosine similarity or Pearson correlation coefficient to measure the association strength between different modal features and equipment faults. For example, calculating the cosine similarity between voltage features and equipment fault features yields an association strength of 0.75; calculating the cosine similarity between ambient temperature features and equipment fault features yields an association strength of 0.45. Through these steps, it can be determined which modal features have a strong semantic association with equipment faults. Finally, the mined semantic association information is summarized into a power multimodal semantic association dataset. This dataset not only contains the raw data but also includes the correlation strength and semantic relationships between different modal features. For example, the dataset records a correlation strength of 0.75 between voltage features and equipment failure, indicating a strong semantic association between voltage fluctuations and equipment failure.

[0156] This invention enables the selection of high-quality, highly reliable data through reliability assessment. By utilizing state-space modeling and reinforcement learning, the data fusion strategy can be dynamically adjusted based on the real-time state of the power system. Sensitivity assessment identifies the features that have the greatest impact on model performance. Weight allocation based on these sensitive features further optimizes the data fusion strategy, enabling the model to capture key information more accurately and enhancing its adaptability to complex data environments. Parameter tuning yields the optimal power fusion decision model. Semantic association mining efficiently extracts deep semantic information from the data.

[0157] Preferably, the present invention also provides a multimodal power cross-domain data fusion and mining system for executing the multimodal power cross-domain data fusion and mining method described above. The multimodal power cross-domain data fusion and mining system includes:

[0158] The data preprocessing module is used to acquire multi-source heterogeneous power datasets; perform tensor decomposition on the multi-source heterogeneous power datasets to obtain a set of power core feature tensors; and perform feature selection and reconstruction on the set of power core feature tensors to obtain a reconstructed power feature dataset.

[0159] The feature mapping module is used to perform domain-adaptive feature mapping on the reconstructed power feature dataset to obtain a cross-domain power feature mapping dataset; to perform domain difference elimination on the cross-domain power feature mapping dataset to obtain a domain-aligned power feature dataset; and to calculate attention weights on the domain-aligned power feature dataset to obtain a cross-domain fused power feature dataset.

[0160] The model training module is used to acquire distributed node topology data; construct a graph convolutional network on the distributed node topology data to obtain power node graph network structure data; perform distributed model training on the cross-domain fused power feature dataset based on the power node graph network structure data to obtain the local power model parameter set; and perform federated averaging on the local power model parameter set to obtain the global optimization model parameter set.

[0161] The model fusion module is used to construct a teacher network from the global optimization model parameter set to obtain a power multimodal teacher model; to train a student network based on the power multimodal teacher model to obtain a power lightweight student model; and to fuse the power lightweight student model and the power multimodal teacher model to obtain a power feature extraction fusion model.

[0162] The data mining module is used to acquire multimodal power quality assessment data; construct a reinforcement learning strategy for the multimodal power quality assessment data to obtain a dynamic power data fusion strategy; adjust the weights of the power feature extraction and fusion model according to the dynamic power data fusion strategy to obtain the optimal power fusion decision model; and use the optimal power fusion decision model to perform semantic association mining on multi-source heterogeneous power datasets to obtain a multimodal power semantic association dataset.

Claims

1. A data fusion and mining method based on multimodal power cross-domain data, characterized in that, Includes the following steps: Step S1: Obtain a multi-source heterogeneous power dataset; perform tensor decomposition on the multi-source heterogeneous power dataset to obtain a set of power core feature tensors; Feature selection and reconstruction are performed on the core feature tensor set of the power industry to obtain a reconstructed power industry feature dataset; wherein, step S1 includes the following steps: Step S11: Collect multimodal power data from the target power grid to obtain the original multimodal power dataset, and align the original multimodal power dataset with timestamps to obtain the time-series power dataset; Step S12: Based on the preset data quality assessment indicators, perform anomaly detection on the time-series power dataset to obtain a power data anomaly label set, and perform data cleaning on the time-series power dataset according to the power data anomaly label set to obtain a multi-source heterogeneous power dataset. Step S13: Identify the data types of the multi-source heterogeneous power dataset to obtain a power data type feature set, and standardize the multi-source heterogeneous power dataset according to the power data type feature set to obtain a standard heterogeneous power dataset. Step S14: Construct tensors from the standard heterogeneous power dataset to obtain a multidimensional power tensor dataset; Step S15: Perform rank evaluation on the multidimensional power tensor dataset to obtain the power tensor rank evaluation dataset; Step S16: Perform tensor decomposition on the multidimensional power tensor dataset based on the power tensor rank evaluation dataset to obtain the power core feature tensor set; Step S17: Perform feature selection and reconstruction on the core power feature tensor set to obtain the reconstructed power feature dataset; wherein, step S17 includes the following steps: Step S171: Evaluate the feature importance of the core power feature tensor set to obtain the power feature importance score set; Step S172: Perform dynamic feature threshold iteration based on the power feature importance score set to obtain the power feature selection threshold set; Step S173: Select a threshold set based on power characteristics to perform feature filtering on the core power feature tensor set, and obtain the key power feature dataset; Step S174: Design an autoencoder network structure for the key power feature dataset to obtain the encoder structure parameter set; Step S175: Train the encoding of the key power feature dataset based on the encoder structure parameter set to obtain the latent power feature encoding dataset; Step S176: Reconstruct the latent power feature encoding dataset using a decoder to obtain the reconstructed power feature dataset; Step S2: Perform domain-adaptive feature mapping on the reconstructed power feature dataset to obtain a cross-domain power feature mapping dataset; perform domain difference elimination on the cross-domain power feature mapping dataset to obtain a domain-aligned power feature dataset; calculate attention weights on the domain-aligned power feature dataset to obtain a cross-domain fused power feature dataset. Step S3: Obtain distributed node topology data; construct a graph convolutional network from the distributed node topology data to obtain power node graph network structure data; train a distributed model on the cross-domain fused power feature dataset based on the power node graph network structure data to obtain the local power model parameter set; perform federated averaging on the local power model parameter set to obtain the global optimization model parameter set. Step S4: Construct a teacher network for the global optimization model parameter set to obtain a power multimodal teacher model; train a student network based on the power multimodal teacher model to obtain a power lightweight student model; fuse the power lightweight student model and the power multimodal teacher model to obtain a power feature extraction fusion model. Step S5: Obtain power multimodal quality assessment data; construct a reinforcement learning strategy for the power multimodal quality assessment data to obtain a power data dynamic fusion strategy; adjust the weights of the power feature extraction and fusion model according to the power data dynamic fusion strategy to obtain the optimal power fusion decision model; use the optimal power fusion decision model to perform semantic association mining on the multi-source heterogeneous power dataset to obtain a power multimodal semantic association dataset.

2. The data fusion and mining method based on multimodal power cross-domain as described in claim 1, characterized in that, Step S2 includes the following steps: Step S21: Perform domain feature distribution quantization on the reconstructed power feature dataset to obtain a multi-domain power feature statistical data set; Step S22: Based on the multi-domain power feature statistical data set, perform domain adaptation network mapping on the reconstructed power feature dataset to obtain the power domain mapping network parameter set; Step S23: Perform intra-domain sample clustering on the power domain mapping network parameter set to obtain a hierarchical intra-domain clustering feature dataset; Step S24: Calculate the inter-domain distance based on the hierarchical intra-domain clustering feature dataset to obtain the cross-domain metric distance dataset; Step S25: Iteratively adjust the gradient of the power domain mapping network parameter set based on the cross-domain metric distance dataset to obtain an adaptive power domain mapping parameter set; Step S26: Perform feature projection on the reconstructed power feature dataset according to the adaptive power domain mapping parameter set to obtain the cross-domain power feature mapping dataset; Step S27: Perform domain difference elimination on the cross-domain power feature mapping dataset to obtain a domain-aligned power feature dataset, and calculate attention weights on the domain-aligned power feature dataset to obtain a cross-domain fused power feature dataset.

3. The data fusion and mining method based on multimodal cross-domain power data according to claim 2, characterized in that, Step S27 includes the following steps: Step S271: Construct a discriminator network on the cross-domain power feature mapping dataset to obtain an adversarial discriminator network; Step S272: Perform adversarial training on the cross-domain power feature mapping dataset based on the adversarial discriminator network to obtain the domain adversarial loss power dataset; Step S273: Adjust the features of the cross-domain power feature mapping dataset based on the domain adversarial loss power dataset to obtain the domain-aligned power feature dataset; Step S274: Calculate attention weights on the domain-aligned power feature dataset to obtain a multimodal power feature weight dataset; Step S275: Construct a weight optimization target based on the multimodal power feature weight dataset to obtain the power feature weight optimization target dataset, and perform parameter iterative updates on the multimodal power feature weight dataset based on the power feature weight optimization target dataset to obtain the optimized power feature weight dataset; Step S276: Based on the optimized power feature weight dataset, the domain-aligned power feature dataset is weighted and fused to obtain an initial fused power feature dataset. The feature consistency of the initial fused power feature dataset is then verified to obtain an inter-domain feature consistency evaluation dataset. Step S277: Perform consistency calibration on the initial fused power feature dataset based on the inter-domain feature consistency assessment dataset to obtain the cross-domain fused power feature dataset.

4. The data fusion and mining method based on multimodal power cross-domain as described in claim 1, characterized in that, Step S3 includes the following steps: Step S31: Collect topological relationships of distributed power nodes to obtain a node connection relationship dataset; Step S32: Evaluate the link status of distributed power nodes based on the node connection relationship dataset to obtain the node link quality dataset; Step S33: Calculate the connection weights of the distributed power nodes based on the node link quality dataset to obtain the node connection weight dataset, and construct a dynamic adjacency matrix based on the node connection weight dataset to obtain the distributed node topology data. Step S34: Based on the distributed node topology data, plan the feature extraction layer for the distributed power nodes to obtain the parameter set of the power node graph convolutional layer; Step S35: Construct a heterogeneous message passing mechanism based on the parameter set of the power node graph convolutional layer to obtain the power node message passing rule set; Step S36: Construct a graph neural architecture based on the power node message passing rule set to obtain power node graph network structure data; Step S37: Based on the power node graph network structure data, perform distributed model training on the cross-domain fused power feature dataset to obtain the local power model parameter set, and perform federated averaging on the local power model parameter set to obtain the global optimized model parameter set.

5. The data fusion and mining method based on multimodal power cross-domain data according to claim 4, characterized in that, Step S37 includes the following steps: Step S371: Divide the cross-domain fusion power feature dataset into distributed training batches based on the power node graph network structure data to obtain the distributed training batch parameter set; Step S372: Asynchronously and parallelly schedule the cross-domain fused power feature dataset according to the distributed training batch parameter set to obtain the parallel training task allocation dataset; Step S373: Configure the distributed optimizer according to the dataset allocated for the parallel training task to obtain the distributed optimizer configuration parameter set; Step S374: Design a learning rate scheduling strategy for the distributed optimizer configuration parameter set to obtain the power characteristic scheduling learning rate parameter set; Step S375: Calculate the distributed gradient of the model for the distributed power nodes based on the power feature scheduling learning rate parameter set to obtain the power node gradient update dataset; Step S376: Asynchronously aggregate the gradients of the power node gradient update dataset to obtain a global gradient aggregate dataset, and update the model parameters based on the global gradient aggregate dataset to obtain a local model update parameter set; Step S377: Perform a weighted average of the local power model parameter set to obtain a federated average weight dataset, and perform federated aggregation of the local power model parameter set based on the federated average weight dataset to obtain a global optimization model parameter set.

6. The data fusion and mining method based on multimodal power cross-domain as described in claim 1, characterized in that, Step S4 includes the following steps: Step S41: Construct a neural network on the parameter set of the global optimization model to obtain a multimodal teacher network, and enhance the multimodal power features of the multimodal teacher network to obtain a multimodal power teacher model; Step S42: Perform feature hierarchical decomposition on the power multimodal teacher model to obtain the teacher semantic knowledge feature set; Step S43: Perform feature selection on the teacher semantic knowledge feature set to obtain the key power representation knowledge feature set, and perform neural architecture search on the key power representation knowledge feature set to obtain the power lightweight student model. Step S44: Align and extract features between the lightweight power student model and the multimodal power teacher model to obtain the cross-modal feature alignment loss dataset; Step S45: Design a knowledge distillation strategy based on the cross-modal feature alignment loss dataset to obtain a multi-level power knowledge distillation strategy parameter set, and perform distillation optimization on the power lightweight student model based on the multi-level power knowledge distillation strategy parameter set to obtain the model distillation optimization parameter set; Step S46: Compress the model distillation optimization parameter set to obtain a lightweight compressed model parameter set, and construct an inference model for the power multimodal teacher model based on the lightweight compressed model parameter set to obtain a power feature extraction and fusion model.

7. The data fusion and mining method based on multimodal cross-domain power data according to claim 1, characterized in that, Step S5 includes the following steps: Step S51: Perform a reliability assessment on the multi-source heterogeneous power dataset to obtain power multimodal quality assessment data; Step S52: Perform state-space modeling on the power multimodal quality assessment data to obtain a power characteristic state-space dataset; Step S53: Perform reinforcement learning on the power feature state space dataset to obtain a dynamic fusion strategy for power data; Step S54: Perform sensitivity evaluation on the power feature extraction and fusion model to obtain the power mode feature sensitivity set; Step S55: Based on the power mode feature sensitivity set, the power data dynamic fusion strategy is weighted to obtain the power feature weight allocation dataset; Step S56: Adjust the parameters of the power feature extraction and fusion model according to the power feature weight allocation dataset to obtain the optimal power fusion decision model; Step S57: Use the optimal power fusion decision model to perform semantic association mining on the multi-source heterogeneous power dataset to obtain the power multimodal semantic association dataset.

8. A data fusion and mining system based on multimodal power cross-domain data, characterized in that, For executing the multimodal cross-domain power data fusion and mining method as described in claim 1, the multimodal cross-domain power data fusion and mining system comprises: The data preprocessing module is used to acquire multi-source heterogeneous power datasets; perform tensor decomposition on the multi-source heterogeneous power datasets to obtain a set of power core feature tensors; and perform feature selection and reconstruction on the set of power core feature tensors to obtain a reconstructed power feature dataset. The feature mapping module is used to perform domain-adaptive feature mapping on the reconstructed power feature dataset to obtain a cross-domain power feature mapping dataset; to perform domain difference elimination on the cross-domain power feature mapping dataset to obtain a domain-aligned power feature dataset; and to calculate attention weights on the domain-aligned power feature dataset to obtain a cross-domain fused power feature dataset. The model training module is used to acquire distributed node topology data; construct a graph convolutional network on the distributed node topology data to obtain power node graph network structure data; perform distributed model training on the cross-domain fused power feature dataset based on the power node graph network structure data to obtain the local power model parameter set; and perform federated averaging on the local power model parameter set to obtain the global optimization model parameter set. The model fusion module is used to construct a teacher network from the global optimization model parameter set to obtain a power multimodal teacher model; to train a student network based on the power multimodal teacher model to obtain a power lightweight student model; and to fuse the power lightweight student model and the power multimodal teacher model to obtain a power feature extraction fusion model. The data mining module is used to acquire multimodal power quality assessment data; construct a reinforcement learning strategy for the multimodal power quality assessment data to obtain a dynamic power data fusion strategy; adjust the weights of the power feature extraction and fusion model according to the dynamic power data fusion strategy to obtain the optimal power fusion decision model; and use the optimal power fusion decision model to perform semantic association mining on multi-source heterogeneous power datasets to obtain a multimodal power semantic association dataset.

Images