District power utilization behavior identification method based on multi-dimensional data fusion

Abstract

The invention discloses a transformer area power consumption behavior identification method based on multi-dimensional data fusion, and relates to the technical field of power system intellectualization, and the method comprises the steps: collecting electrical parameter data and environmental parameter data of a transformer area, and carrying out the preprocessing of the electrical parameter data and the environmental parameter data, and obtaining a standardized data set; performing feature extraction on the electrical feature data and the environment feature data to obtain an electrical feature vector and an environment feature vector, and forming a multi-dimensional feature space by the electrical feature vector and the environment feature vector; and constructing a feature fusion model in the multi-dimensional feature space, inputting the feature fusion model into the deep learning model for training, and outputting a power utilization behavior recognition result of the transformer area. According to the method, the technical problem of heterogeneous data fusion is solved by constructing the multi-dimensional feature space, and the accuracy of power consumption behavior recognition is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent power system, and in particular to a method for identifying power consumption behavior of a substation based on multi-dimensional data fusion. Background Art

[0002] As one of the core tasks of smart grid construction, the identification of power consumption behavior in substations is of great significance to maintaining the safe operation of the power system and improving the quality of power supply services. Traditional methods of identifying power consumption behavior mainly rely on single electrical parameter analysis, such as statistical analysis methods based on load curve characteristics, threshold judgment methods based on power quality indicators, etc. In recent years, deep learning technology has been widely used in the field of power consumption behavior identification. Researchers have proposed a power consumption feature extraction method based on convolutional neural networks, a time series feature analysis method based on recurrent neural networks, and a spatial feature modeling method based on graph neural networks, which significantly improved the accuracy and generalization ability of power consumption behavior identification.

[0003] However, the existing electricity consumption behavior recognition technology still has the following problems: First, most methods only focus on the feature analysis of a single dimension, ignoring the coupling relationship between electrical parameters and environmental parameters, resulting in insufficient feature expression capabilities; second, existing feature fusion methods often adopt simple feature splicing or weighted average strategies, which cannot effectively capture the dynamic correlation between different features; third, there is a lack of adaptive evaluation mechanism for feature importance, making it difficult to dynamically adjust feature weights according to actual scenarios; finally, the characteristics of multi-source heterogeneous data are not fully considered during model training, which can easily lead to feature redundancy and overfitting problems. Summary of the invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the present invention provides a method for identifying power consumption behavior in an area based on multi-dimensional data fusion, which can solve the problems mentioned in the background technology.

[0006] To solve the above technical problems, the present invention provides the following technical solutions: a method for identifying power consumption behavior in a substation based on multi-dimensional data fusion, comprising: collecting electrical parameter data and environmental parameter data of the substation, and preprocessing the electrical parameter data and the environmental parameter data to obtain a standardized data set; the standardized data set includes electrical feature data and environmental feature data; performing feature extraction on the electrical feature data and the environmental feature data respectively to obtain an electrical feature vector and an environmental feature vector, and forming a multi-dimensional feature space with the electrical feature vector and the environmental feature vector; constructing a feature fusion model in the multi-dimensional feature space, inputting the feature fusion model into a deep learning model for training, and outputting the power consumption behavior identification result of the substation.

[0007] As a preferred solution of the method for identifying power consumption behavior of substations based on multi-dimensional data fusion described in the present invention, wherein: the feature extraction is completed by constructing a time series feature extraction module and a spatial feature extraction module; wherein the time series feature extraction module segments the electrical feature data based on the power consumption behavior cycle by setting the sliding window length, and extracts the time series feature sequence by constructing a long short-term memory network, and at the same time uses wavelet transform to perform multi-scale decomposition on the time series feature sequence to obtain feature components in different frequency domains; the spatial feature extraction module constructs a graph convolution network by calculating the spatial autocorrelation matrix of the environmental feature data and extracts spatial features, and normalizes the time series feature sequence and the spatial features to obtain a standardized feature set.

[0008] As a preferred solution of the method for identifying power consumption behavior of substations based on multidimensional data fusion described in the present invention, wherein: the multidimensional feature space is constructed by a feature mapping matrix including a main feature mapping layer and a sub-feature mapping layer; the main feature mapping layer constructs a main feature mapping vector by calculating the self-attention score of the time series feature sequence and extracts the main feature time dependency; the sub-feature mapping layer calculates the topological features of the graph structure by constructing a correlation graph of spatial features and extracts the local correlation pattern of spatial features to obtain the feature mapping correlation; when the feature mapping correlation exceeds a first preset threshold, the corresponding feature weight is increased to a preset upper limit; when the feature mapping correlation is lower than a second preset threshold, the corresponding feature weight is reduced to a preset lower limit; when the feature mapping correlation is between the first preset threshold and the second preset threshold, the feature weight is adjusted by a linear mapping function.

[0009] As a preferred solution of the method for identifying power consumption behavior of substations based on multi-dimensional data fusion described in the present invention, the feature fusion model divides the input features into multiple feature subspaces by constructing a multi-head attention mechanism, and merges the attention outputs of each feature subspace after calculating the attention weight in each feature subspace; at the same time, a feature fusion network including a residual connection module, a layer normalization layer and a feature dropout mechanism is designed, and feature adaptive fusion is achieved by calculating the feature importance score and performing multi-scale feature aggregation, and the fused features are standardized.

[0010] As a preferred solution of the method for identifying power consumption behavior of substations based on multidimensional data fusion described in the present invention, the training process of the deep learning model is optimized by constructing a mixed loss function including classification cross entropy loss, feature reconstruction loss term and contrastive learning loss term, and a regularization strategy including weight decay term, early stopping mechanism and dropout is designed; the model convergence is accelerated by using adaptive learning rate adjustment strategy, gradient clipping and batch normalization, and the model performance is evaluated by constructing multiple cross-validation sets, calculating comprehensive evaluation indicators and implementing model integration strategy.

[0011] As a preferred solution of the method for identifying power consumption behavior of substations based on multi-dimensional data fusion described in the present invention, wherein: the preprocessing is completed by outlier detection and removal, data standardization processing and missing value interpolation and completion; the electrical parameter data includes basic electrical parameter data, load curve data and power quality data; the environmental parameter data includes meteorological environment parameters, geographical location parameters and time period parameters; the power consumption behavior identification results include normal power consumption behavior type, power theft behavior type, faulty power consumption behavior type and load mutation behavior type.

[0012] As a preferred solution of the method for identifying power consumption behavior of substations based on multi-dimensional data fusion described in the present invention, wherein: the deep learning model dynamically adjusts the prediction results by constructing a model optimization module during the deployment process after training; the model optimization module includes a time series calibration unit and a spatial feature compensation unit; the time series calibration unit calculates the time calibration coefficient by establishing a historical prediction deviation sequence, and corrects the current prediction result based on the time calibration coefficient; the spatial feature compensation unit constructs a feature compensation vector by calculating the spatial correlation matrix of environmental features, and triggers an abnormal review mechanism when the amplitude of the feature compensation vector exceeds the compensation threshold; when the model optimization module detects continuous deviations in the prediction results, it updates the model parameters through online learning, and weightedly fuses the updated parameters with the original parameters to form a new model parameter set.

[0013] To further solve the above technical problems, the present invention provides the following technical solutions: a substation electricity consumption behavior identification system based on multi-dimensional data fusion, comprising: a data processing module, used to collect electrical parameter data and environmental parameter data of the substation, and pre-process the electrical parameter data and the environmental parameter data to obtain a standardized data set; a feature processing module, used to extract features from the electrical feature data and the environmental feature data, respectively, to obtain electrical feature vectors and environmental feature vectors, and form a multi-dimensional feature space with the electrical feature vectors and the environmental feature vectors; a behavior identification module, used to construct a feature fusion model in the multi-dimensional feature space, input the feature fusion model into a deep learning model for training, and output the substation electricity consumption behavior identification result.

[0014] A computer device includes a memory and a processor, wherein the memory stores a computer program, and is characterized in that when the processor executes the computer program, the steps of the method for identifying power consumption behavior of a substation based on multi-dimensional data fusion as described above are implemented.

[0015] A computer-readable storage medium having a computer program stored thereon, characterized in that when the computer program is executed by a processor, the steps of the method for identifying power consumption behavior of a substation based on multi-dimensional data fusion as described above are implemented.

[0016] Beneficial effects of the present invention: The present invention significantly improves the quality and reliability of multi-source data by combining the data preprocessing strategies of outlier detection, standardization and missing value interpolation; through the sliding window design based on the electricity consumption behavior cycle in the time series feature extraction module, combined with the multi-scale decomposition of long short-term memory network and wavelet transform, the fine characterization of the time series characteristics of electricity consumption behavior is realized; the spatial feature extraction module effectively captures the spatial dependency of environmental features by using the spatial autocorrelation matrix and graph convolution network; the design of the feature mapping matrix establishes a dynamic association mechanism between features through the main and secondary feature mapping layers, solving the problem of insufficient expression of feature correlation in traditional methods; the multi-head attention mechanism realizes the parallel processing and adaptive fusion of feature subspaces, and the introduction of residual connection and layer normalization enhances the training stability of the deep model; the design of the mixed loss function and regularization strategy improves the generalization performance of the model; finally, the dynamic optimization of the prediction results is realized through time series calibration and spatial feature compensation. This multi-level feature processing and fusion scheme effectively solves the problems of insufficient feature expression, poor feature fusion effect, and weak model generalization ability in the prior art, and significantly improves the accuracy and reliability of the identification of power consumption behavior in the substation area. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0018] Figure 1 This is a schematic diagram of the overall process of a method for identifying power consumption behavior in a substation area based on multi-dimensional data fusion proposed by the present invention; Figure 2 This is a schematic diagram of the overall structure of a system for identifying power consumption behavior in a substation area based on multi-dimensional data fusion proposed by the present invention; Figure 3 This is a computer equipment diagram in a method for identifying power consumption behavior in an area based on multi-dimensional data fusion proposed by the present invention. DETAILED DESCRIPTION

[0019] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.

[0020] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0021] Example 1, reference Figure 1 , which is an embodiment of the present invention, provides a method for identifying power consumption behavior in a substation based on multi-dimensional data fusion.

[0022] Figure 1 The overall process diagram of a method for identifying power consumption behavior in a substation based on multi-dimensional data fusion is shown, which includes the following steps: S1: Collect electrical parameter data and environmental parameter data of the substation, and pre-process the electrical parameter data and environmental parameter data to obtain a standardized data set.

[0023] Specifically, electrical parameter data includes basic electrical parameter data, load curve data and power quality data. Environmental parameter data includes meteorological environment parameters, geographic location parameters and time period parameters. Standardized data sets include electrical characteristic data and environmental characteristic data.

[0024] The preprocessing is completed by outlier detection and removal, data standardization and missing value interpolation.

[0025] First, collect electrical parameter data and environmental parameter data during the operation of the substation. Specifically, electrical parameter data is obtained in real time through collection equipment such as smart meters, distribution terminal equipment (FTU / DTU), and power quality online monitoring devices deployed at distribution transformers and key nodes in the substation. Among them, smart meters collect basic electrical parameter data at intervals of 15 minutes, distribution terminal equipment monitors load curve changes in real time, and power quality monitoring devices continuously collect power quality disturbance data. Meteorological data in environmental parameter data is collected through equipment such as meteorological stations and temperature and humidity sensors around the substation, geographic location information is obtained through the substation GIS system, and time period parameters are automatically generated through the system clock. All collection devices transmit data to the data collection master station through the communication gateway, and encrypted transmission is used to ensure data security.

[0026] The collected raw data is preprocessed. The preprocessing process includes three key steps: First, outlier detection and removal. The 3σ criterion based on statistical features is combined with the box plot method to identify outliers, and the density-based local outlier factor (LOF) algorithm is introduced to detect outliers in the multidimensional feature space. For the detected outliers, the appropriate removal strategy is selected according to the temporal continuity characteristics of the data.

[0027] Second, data standardization. For feature parameters of different dimensions, the minimum-maximum normalization method is used to map the data to the [0,1] interval. For features with magnitude differences, the Z-score normalization method is used to process the data so that the data meets the distribution characteristics of a mean of 0 and a variance of 1.

[0028] Third, missing value interpolation. For missing values ​​in time series data, moving average interpolation, linear interpolation and other methods are used to fill in the missing values ​​according to the time series correlation characteristics of the data. For missing values ​​in non-time series data, the K nearest neighbor interpolation algorithm based on similar samples is used for estimation.

[0029] Through the above preprocessing steps, a standardized data set is obtained, which contains processed electrical characteristic data and environmental characteristic data. The standardized data set eliminates abnormal interference, dimension differences and missing value effects in the original data, laying the foundation for subsequent feature extraction and fusion.

[0030] In addition, to ensure data quality, a data quality assessment mechanism was established during the preprocessing process. The preprocessing results were evaluated by calculating indicators such as data completeness and efficiency to ensure that the standardized data set met the requirements of subsequent analysis. At the same time, key parameters in the preprocessing process (such as outlier determination threshold, interpolation window size, etc.) were dynamically optimized to improve the adaptability and reliability of the preprocessing.

[0031] S2: Extract features from the electrical feature data and the environmental feature data respectively to obtain an electrical feature vector and an environmental feature vector, and construct a multi-dimensional feature space with the electrical feature vector and the environmental feature vector.

[0032] Specifically, feature extraction is completed by constructing a temporal feature extraction module and a spatial feature extraction module.

[0033] Among them, the time series feature extraction module segments the electrical feature data by setting the sliding window length based on the power consumption behavior cycle, and extracts the time series feature sequence by constructing a long short-term memory network. At the same time, the wavelet transform is used to perform multi-scale decomposition on the time series feature sequence to obtain characteristic components in different frequency domains.

[0034] The spatial feature extraction module constructs a graph convolutional network and extracts spatial features by calculating the spatial autocorrelation matrix of environmental feature data, and normalizes the temporal feature sequence and spatial features to obtain a standardized feature set.

[0035] Furthermore, a multi-dimensional feature space is constructed by a feature mapping matrix including a primary feature mapping layer and a secondary feature mapping layer.

[0036] The main feature mapping layer constructs the main feature mapping vector by calculating the self-attention score of the temporal feature sequence and extracts the time dependency of the main features; the secondary feature mapping layer calculates the topological features of the graph structure by constructing the correlation graph of the spatial features and extracts the local correlation patterns of the spatial features to obtain the feature mapping correlation.

[0037] When the feature mapping correlation exceeds the first preset threshold, the corresponding feature weight is increased to the preset upper limit; when the feature mapping correlation is lower than the second preset threshold, the corresponding feature weight is reduced to the preset lower limit; when the feature mapping correlation is between the first preset threshold and the second preset threshold, the feature weight is adjusted through a linear mapping function.

[0038] It should be noted that this step first constructs a dual-channel feature extraction architecture, including a temporal feature extraction module and a spatial feature extraction module, which process electrical feature data and environmental feature data respectively.

[0039] The time series feature extraction module is designed for the time series characteristics of electrical characteristic data. First, the sliding window length is set based on the typical periodic characteristics of the power consumption behavior of the substation. By analyzing the historical data statistics, it is found that the power consumption pattern on weekdays and holidays has a 24-hour periodicity, so the basic sliding window length is set to 24 hours. At the same time, considering the load change characteristics caused by seasonal changes, an adaptive window adjustment mechanism is introduced, and the window length can be dynamically adjusted within the range of 24 hours to 168 hours (one week). For the segmented electrical characteristic data, a two-layer long short-term memory network (LSTM) is constructed to extract time series features. The first layer of LSTM is responsible for capturing the characteristics of short-term power consumption behavior, and the number of hidden layer nodes is set to 128; the second layer of LSTM is used to extract long-term power consumption patterns, and the number of hidden layer nodes is set to 64. In order to further analyze the characteristics of power consumption behavior at different time scales, the discrete wavelet transform is applied to the time series feature sequence output by LSTM for multi-scale decomposition. The db4 wavelet basis function is selected to decompose the feature sequence into four scales, corresponding to hourly, intraday, daytime and periodic change characteristics.

[0040] The spatial feature extraction module focuses on processing the spatial correlation of environmental feature data. First, the spatial autocorrelation matrix of environmental feature data is calculated, and the strength of spatial correlation is evaluated using the Moran's I index. The adjacency matrix reflecting the propagation relationship of environmental features between stations is constructed. Based on the adjacency matrix, a three-layer graph convolutional network (GCN) is constructed, with feature dimensions of 64, 32, and 16 for each layer, and LeakyReLU is used as the activation function. The spatial distribution pattern and regional correlation features of environmental features are extracted through graph convolution operations.

[0041] The extracted temporal feature sequence and spatial features are normalized by the Min-Max method to obtain a standardized feature set. Furthermore, a feature mapping matrix is ​​constructed to realize the construction of a multidimensional feature space. The matrix contains two key components: the main feature mapping layer and the secondary feature mapping layer.

[0042] The main feature mapping layer focuses on the temporal dependency analysis of the temporal feature sequence. Specifically, a multi-head self-attention mechanism is used to calculate the attention score of the feature sequence. The number of heads is set to 8, and the dimension of each attention head is 32. The attention score reflects the importance of features at different time points, and the main feature mapping vector is constructed based on it. The dimension of this vector is the same as the input feature dimension.

[0043] The secondary feature mapping layer focuses on extracting local correlation patterns of spatial features. First, a correlation graph of spatial features is constructed, the correlation coefficient between features is calculated using cosine similarity, and the local connection relationship is determined using the K nearest neighbor algorithm (K=5). The aggregated features of each node are calculated through a graph neural network to obtain the feature mapping correlation that reflects the topological structure of the spatial features.

[0044] The threshold management of feature mapping correlation adopts an adaptive strategy: the first preset threshold is set to 0.8, indicating a highly correlated feature mapping relationship. When the mapping correlation exceeds this threshold, the corresponding feature weight is increased to the preset upper limit of 1.5; the second preset threshold is set to 0.3, indicating a weakly correlated feature mapping relationship. When the mapping correlation is lower than this threshold, the corresponding feature weight is reduced to the preset lower limit of 0.5. The setting of these two thresholds is based on a large amount of experimental data statistics and expert experience, and can be fine-tuned according to actual application scenarios. When the feature mapping correlation is between the two thresholds, a piecewise linear mapping function is used to dynamically adjust the feature weight. The slope of the mapping function is optimized based on historical data using the least squares method to ensure the smoothness and continuity of the feature weight adjustment.

[0045] Through the above feature extraction and mapping process, a multidimensional feature space is finally constructed, which can effectively express the spatiotemporal correlation between electrical characteristics and environmental characteristics, providing high-quality input data for subsequent deep learning model training.

[0046] S3: Construct a feature fusion model in the multi-dimensional feature space, input the feature fusion model into the deep learning model for training, and output the power consumption behavior recognition results of the substation.

[0047] Specifically, the feature fusion model divides the input features into multiple feature subspaces by constructing a multi-head attention mechanism, calculates the attention weights in each feature subspace separately, and then merges the attention outputs of each feature subspace; at the same time, a feature fusion network including a residual connection module, a layer normalization layer and a feature dropout mechanism is designed, and feature adaptive fusion is achieved by calculating the feature importance scores and performing multi-scale feature aggregation, and the fused features are standardized.

[0048] Furthermore, the training process of the deep learning model is optimized by constructing a mixed loss function including classification cross entropy loss, feature reconstruction loss term and contrastive learning loss term, and a regularization strategy including weight decay term, early stopping mechanism and dropout is designed; the model convergence is accelerated by using adaptive learning rate adjustment strategy, gradient clipping and batch normalization, and the model performance is evaluated by constructing multiple cross-validation sets, calculating comprehensive evaluation indicators and implementing model integration strategy.

[0049] During the deployment process after training, the deep learning model dynamically adjusts the prediction results by constructing a model optimization module. The model optimization module includes a time series calibration unit and a spatial feature compensation unit. The time series calibration unit calculates the time calibration coefficient by establishing a historical prediction deviation sequence, and corrects the current prediction result based on the time calibration coefficient; the spatial feature compensation unit constructs a feature compensation vector by calculating the spatial correlation matrix of environmental features, and triggers the abnormal review mechanism when the amplitude of the feature compensation vector exceeds the compensation threshold; when the model optimization module detects continuous deviations in the prediction results, it updates the model parameters through online learning, and weightedly fuses the updated parameters with the original parameters to form a new model parameter set.

[0050] Furthermore, the electricity usage behavior identification result includes normal electricity usage behavior type, electricity theft behavior type, faulty electricity usage behavior type and load mutation behavior type.

[0051] It should be noted that this step first designs a feature fusion model in the constructed multi-dimensional feature space. The model uses a multi-head attention mechanism to divide the input features into 8 feature subspaces, and the dimension of each subspace is 1 / 8 of the original feature dimension. In each feature subspace, the attention weight is obtained by calculating the dot product between the query vector (Query), the key vector (Key), and the value vector (Value), and the Softmax function is used for normalization. The attention outputs of each feature subspace are merged through the concat operation, and then the fused feature representation is obtained through linear mapping.

[0052] The core architecture of the feature fusion network includes a residual connection module, a layer normalization layer, and a feature dropout mechanism. The residual connection module uses a short-circuit connection method to ensure that the deep network can effectively learn feature residual information and alleviate the gradient vanishing problem. The layer normalization operation standardizes the feature distribution of each sample and accelerates the network training process. The feature dropout mechanism randomly inactivates 30% of neurons to improve the robustness of the model. By calculating the feature importance score based on the SHAP value, the adaptive fusion of features is achieved, and the feature weights with an importance score higher than 0.7 will be doubled. Multi-scale feature aggregation adopts a pyramid pooling structure, setting four scale pooling layers, and the pooling kernel sizes are 1×1, 2×2, 4×4, and 8×8 respectively.

[0053] The training process of the deep learning model is optimized using an innovative hybrid loss function. This loss function consists of three components: the classification cross entropy loss is used to evaluate the classification accuracy of the model; the feature reconstruction loss term uses the mean square error to measure the difference between the original features and the reconstructed features; the contrastive learning loss term uses the InfoNCE loss function to enhance the model's ability to distinguish similar samples. The weight ratio of the three loss terms is 0.5:0.3:0.2, which can be dynamically adjusted according to actual task requirements.

[0054] The regularization strategy design includes three key mechanisms: the weight decay term uses L2 regularization, and the coefficient is set to 0.0001; the early stopping mechanism is triggered when the performance of the validation set has not improved for 5 consecutive epochs; the dropout rate is set to 0.5. To accelerate model convergence, the Adam optimizer is used in combination with the cosine annealing learning rate scheduling strategy, and the initial learning rate is set to 0.001; the gradient clipping threshold is set to 5.0 to prevent gradient explosion; the momentum parameter of the batch normalization layer is set to 0.9. The model evaluation uses 5-fold cross validation, and the model performance is comprehensively evaluated by calculating indicators such as accuracy, precision, recall, and F1 score, and an ensemble learning strategy is used to fuse the prediction results of multiple models.

[0055] The optimization module in the model deployment phase contains two core units. The time series calibration unit calculates the time calibration coefficient by analyzing the forecast deviation sequence of the past 30 days using the exponentially weighted moving average method. The calibration window size is set to 24 hours and the smoothing coefficient is 0.3. When the absolute value of the calibration coefficient exceeds 0.1, the forecast result correction mechanism is triggered. The spatial feature compensation unit calculates the spatial correlation matrix of environmental features based on the Pearson correlation coefficient and constructs a feature compensation vector with the same dimension as the original feature. The compensation threshold is set to twice the standard deviation. When any component of the compensation vector exceeds the threshold, the abnormal review mechanism is activated and secondary confirmation is performed through the expert system.

[0056] When it is detected that the average absolute error of the prediction results exceeds 20% at 12 consecutive time points, the model optimization module starts the online learning update process. The update process uses the small batch stochastic gradient descent method, the learning rate is set to 1 / 10 of the initial value, and each update uses the data of the last 24 hours. The updated parameters are weighted and fused with the original parameters. The fusion weight is dynamically determined according to their performance on the validation set. Generally, the weight of the new parameter fluctuates between 0.3-0.7.

[0057] The final electricity consumption behavior recognition results are divided into four categories: normal electricity consumption behavior type (including residential electricity consumption, industrial production electricity consumption and other conventional electricity consumption modes), electricity theft behavior type (including abnormal electricity consumption behaviors such as wiring around the meter and intentional damage to the metering device), faulty electricity consumption behavior type (including abnormalities caused by non-human factors such as equipment aging and line failures) and load mutation behavior type (including drastic load changes caused by short-term high-power equipment startup and sudden failures of power-consuming equipment). The model's recognition accuracy for these four types of behaviors reached 95%, 90%, 88% and 92% respectively.

[0058] In summary, the present invention realizes the deep fusion of electrical features and environmental features by constructing a multi-dimensional feature space, designs an adaptive feature weight adjustment mechanism to improve the robustness of the model, adopts a multi-head attention mechanism and a feature dropout strategy to optimize the feature extraction effect, and introduces a hybrid loss function and a regularization strategy to improve the generalization ability of the model. The present invention improves the accuracy and reliability of electricity consumption behavior recognition, enhances the model's adaptability to complex scenarios, and realizes efficient fusion and dynamic optimization of features.

[0059] Example 2, reference Figure 2 , which is an embodiment of the present invention, provides a substation electricity consumption behavior identification system based on multi-dimensional data fusion.

[0060] like Figure 2 As shown in FIG. 1 , it is a schematic diagram of the overall structure of the system, including: A data processing module, used for collecting electrical parameter data and environmental parameter data of the substation, and preprocessing the electrical parameter data and the environmental parameter data to obtain a standardized data set; A feature processing module, used to extract features from the electrical feature data and the environmental feature data respectively to obtain an electrical feature vector and an environmental feature vector, and to form a multi-dimensional feature space with the electrical feature vector and the environmental feature vector; The behavior recognition module is used to construct a feature fusion model in the multi-dimensional feature space, input the feature fusion model into the deep learning model for training, and output the power consumption behavior recognition result of the substation.

[0061] Example 3, reference Figure 3, is an embodiment of the present invention, which is different from the previous embodiment in that: if the function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0062] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute instructions), or in conjunction with such instruction execution systems, devices or apparatuses. For the purposes of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in conjunction with such instruction execution systems, devices or apparatuses.

[0063] More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection with one or more wires (electronic device), a portable computer disk case (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be a paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering or, if necessary, processing in another suitable manner, and then stored in a computer memory.

[0064] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0065] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for identifying power consumption behavior in a substation based on multi-dimensional data fusion, characterized in that: include: Collecting electrical parameter data and environmental parameter data of the station area, and preprocessing the electrical parameter data and the environmental parameter data to obtain a standardized data set; the standardized data set includes electrical characteristic data and environmental characteristic data; Extracting features from the electrical feature data and the environmental feature data respectively to obtain an electrical feature vector and an environmental feature vector, and forming a multi-dimensional feature space with the electrical feature vector and the environmental feature vector; A feature fusion model is constructed in the multi-dimensional feature space, the feature fusion model is input into a deep learning model for training, and a result of identifying the power consumption behavior of the substation is output.

2. The method for identifying power consumption behavior in a substation area based on multi-dimensional data fusion according to claim 1, characterized in that: The feature extraction is completed by constructing a temporal feature extraction module and a spatial feature extraction module; The time series feature extraction module segments the electrical feature data by setting the sliding window length based on the power consumption behavior cycle, extracts the time series feature sequence by constructing a long short-term memory network, and uses wavelet transform to perform multi-scale decomposition on the time series feature sequence to obtain feature components in different frequency domains; The spatial feature extraction module constructs a graph convolution network and extracts spatial features by calculating the spatial autocorrelation matrix of the environmental feature data, and normalizes the temporal feature sequence and the spatial feature to obtain a standardized feature set.

3. The method for identifying power consumption behavior in a substation area based on multi-dimensional data fusion according to claim 2, characterized in that: The multidimensional feature space is constructed by a feature mapping matrix including a primary feature mapping layer and a secondary feature mapping layer; The main feature mapping layer constructs a main feature mapping vector by calculating the self-attention score of the temporal feature sequence and extracts the main feature time dependency; the secondary feature mapping layer calculates the topological features of the graph structure by constructing a correlation graph of spatial features and extracts the local correlation pattern of spatial features to obtain the feature mapping correlation; When the feature mapping correlation exceeds the first preset threshold, the corresponding feature weight is increased to the preset upper limit; when the feature mapping correlation is lower than the second preset threshold, the corresponding feature weight is reduced to the preset lower limit; when the feature mapping correlation is between the first preset threshold and the second preset threshold, the feature weight is adjusted through a linear mapping function.

4. The method for identifying power consumption behavior in a substation area based on multi-dimensional data fusion according to claim 3, characterized in that: The feature fusion model divides the input features into multiple feature subspaces by constructing a multi-head attention mechanism, and merges the attention outputs of each feature subspace after calculating the attention weight in each feature subspace. At the same time, a feature fusion network including a residual connection module, a layer normalization layer and a feature dropout mechanism is designed. Feature adaptive fusion is achieved by calculating feature importance scores and multi-scale feature aggregation is performed, and the fused features are standardized.

5. The method for identifying power consumption behavior in a substation area based on multi-dimensional data fusion according to claim 4, characterized in that: The training process of the deep learning model is optimized by constructing a mixed loss function including classification cross entropy loss, feature reconstruction loss term and contrastive learning loss term, and a regularization strategy including weight decay term, early stopping mechanism and dropout is designed; the model convergence is accelerated by using adaptive learning rate adjustment strategy, gradient clipping and batch normalization, and the model performance is evaluated by constructing multiple cross-validation sets, calculating comprehensive evaluation indicators and implementing model integration strategy.

6. The method for identifying power consumption behavior in a substation area based on multi-dimensional data fusion according to claim 5, characterized in that: The preprocessing is completed by outlier detection and removal, data standardization and missing value interpolation and completion; The electrical parameter data includes basic electrical parameter data, load curve data and power quality data; The environmental parameter data includes meteorological environmental parameters, geographical location parameters and time period parameters; The power usage behavior identification result includes normal power usage behavior type, power theft behavior type, faulty power usage behavior type and load mutation behavior type.

7. The method for identifying power consumption behavior in a substation area based on multi-dimensional data fusion according to claim 6, characterized in that: The deep learning model dynamically adjusts the prediction results by constructing a model optimization module during the deployment process after training; The model optimization module includes a time series calibration unit and a spatial feature compensation unit; The time series calibration unit calculates a time calibration coefficient by establishing a historical prediction deviation sequence, and corrects the current prediction result based on the time calibration coefficient; The spatial feature compensation unit constructs a feature compensation vector by calculating the spatial correlation matrix of environmental features. When the amplitude of the feature compensation vector exceeds the compensation threshold, the abnormal review mechanism is triggered. When the model optimization module detects continuous deviations in the prediction results, it updates the model parameters through online learning, and at the same time, weightedly fuses the updated parameters with the original parameters to form a new model parameter set.

8. A system for identifying power consumption behavior in a substation based on multi-dimensional data fusion, based on the method for identifying power consumption behavior in a substation based on multi-dimensional data fusion according to any one of claims 1 to 7, characterized in that: include, A data processing module, used for collecting electrical parameter data and environmental parameter data of the substation, and preprocessing the electrical parameter data and the environmental parameter data to obtain a standardized data set; A feature processing module, used to extract features from the electrical feature data and the environmental feature data respectively to obtain an electrical feature vector and an environmental feature vector, and to form a multi-dimensional feature space with the electrical feature vector and the environmental feature vector; The behavior recognition module is used to construct a feature fusion model in the multi-dimensional feature space, input the feature fusion model into the deep learning model for training, and output the power consumption behavior recognition result of the substation.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method for identifying power consumption behavior of a substation based on multi-dimensional data fusion according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for identifying power consumption behavior of a substation based on multi-dimensional data fusion according to any one of claims 1 to 7 are implemented.

Images