A cross-regional power data collaborative analysis system based on privacy-preserving computation

Abstract

This application provides a cross-regional power data collaborative analysis system based on privacy computing, relating to the fields of smart grid and power big data processing technology. Instead of directly interacting with potentially conflicting model parameters, it utilizes common anchor point data to map the heterogeneous features of local models into a standardized logical matrix, which is then uploaded to a central server after differential privacy encryption. During the aggregation phase, the system introduces a dynamic weight evaluation mechanism based on entropy-weighted confidence and positive interactive complementarity. By decoupling data credibility and knowledge scarcity, it accurately identifies and retains long-tail specific power patterns (such as industrial impact loads) that cannot be represented by mainstream distributions, while effectively suppressing low-quality noise. The final generated global logical matrix is ​​fed back as a knowledge carrier. Through knowledge distillation technology, it guides local models to absorb complementary generalized knowledge from the entire network while maintaining the fidelity of local features, thereby constructing a regional optimization model that combines robustness and personalization.

Description

Technical Field

[0001] This application relates to the fields of smart grid and power big data processing technology, and more specifically, to a cross-regional power data collaborative analysis system based on privacy computing. Background Technology

[0002] With the deepening development of smart grids and the energy internet, the power system has accumulated massive amounts of multi-dimensional time-series operational data across all stages of generation, transmission, transformation, distribution, and consumption. Cross-regional collaborative analysis of power data is crucial for improving the accuracy of load forecasting and fault diagnosis across the entire network, and optimizing distributed energy dispatch strategies. However, due to increasingly stringent data security laws and regulations, as well as barriers to commercial privacy protection among various power operators, power data from different geographical regions often forms isolated data silos, making centralized data mining impossible through physical aggregation. Therefore, privacy-preserving computing technologies, such as federated learning, are gradually becoming mainstream in the industry. By allowing participating parties to train models locally and only exchange intermediate parameters, they attempt to achieve collaborative modeling without data leaving the domain.

[0003] However, existing general collaborative frameworks based on parameter averaging face severe statistical heterogeneity failure problems when applied to cross-regional power scenarios. The fundamental flaw lies in the fact that power data inherently possesses strong non-independent and identically distributed characteristics, with significant differences in power load patterns across different regions (e.g., square wave characteristics in industrial cities versus bimodal characteristics in residential areas). This results in optimal model parameters for different regions being vastly different in high-dimensional space. Existing techniques simply perform linear averaging of these parameters in Euclidean space, often leading to the aggregated global parameters falling into an inefficient region between two local optima. This not only fails to effectively integrate feature knowledge from different regions but also disrupts the discriminative logic already learned by the local model, triggering negative transfer. More critically, when dealing with this heterogeneity, existing techniques often employ similarity-based weighting mechanisms, defaulting to considering deviations from the mainstream as anomalies. This mechanism is prone to misjudging and suppressing high-value long-tail data (such as extreme cold wave loads or industrial shock patterns in specific regions) that deviate from the overall network statistical mean but contain key specific knowledge, as noise. This results in the final global model tending to be homogeneous and mediocre, and seriously lacking robustness and generalization ability to cope with complex power grid emergencies.

[0004] Therefore, an optimized cross-regional collaborative analysis system for power data based on privacy computing is desired. Summary of the Invention

[0005] To address the aforementioned technical problems, this application is proposed. Embodiments of this application provide a cross-regional power data collaborative analysis system based on privacy computing, comprising: The local model pre-training module is used to perform the first round of pre-training on the initialized local deep network model based on local power cleaning data to obtain the pre-trained local deep network model. The local anchor inference module is used to input the acquired public anchor waveform data into the pre-trained local deep network model for forward inference to obtain the local anchor logic matrix. The privacy-enhanced encryption module is used to perform privacy-enhanced encryption on the local anchor point logic matrix to obtain encrypted logic data packets; The global knowledge aggregation module is used to control the central server to receive encrypted logical data packets from multiple different regions, and to perform global knowledge aggregation based on correlation on multiple encrypted logical data packets to obtain a global logical matrix containing the power pattern characteristics of the entire network. The global data decryption and verification module is used to receive the global logic matrix broadcast and distributed back by the central server, and to decrypt and verify the global logic matrix in combination with the local decryption private key to obtain the global plaintext soft tag used for model guidance. The knowledge distillation optimization module is used to perform knowledge distillation closed-loop optimization on the pre-trained local deep network model based on local power cleaning data, common anchor waveform data, and global plaintext soft labels to obtain a regional optimized power model.

[0006] Compared with existing technologies, the cross-regional power data collaborative analysis system based on privacy computing provided in this application abandons the traditional direct averaging paradigm of model parameters and constructs a federated knowledge distillation closed-loop architecture based on common anchor point data. Each regional node pre-trains a model based on local data and uploads the prediction logic matrix of the common anchor point with privacy encryption. The central server utilizes a dynamic weighting mechanism based on entropy weight confidence and positive interaction complementarity to decouple data credibility and knowledge scarcity during the aggregation process, thereby accurately identifying and retaining long-tail power patterns with high specificity while effectively suppressing low-quality noise interference. The generated global soft tags containing network-wide consensus are then transmitted back to the local server to guide the local model in multi-objective optimization updates. This concept effectively solves the problems of global model generalization failure and negative transfer caused by the heterogeneity of power data statistics, achieving lossless capture of key heterogeneous scenario features such as heavy industrial loads, and improving the global adaptability and robustness of the collaborative model while ensuring data privacy. Attached Figure Description

[0007] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0008] Figure 1 This is a system block diagram of a cross-regional power data collaborative analysis system based on privacy computing, according to an embodiment of this application. Figure 2 This is a schematic diagram of data flow in a cross-regional power data collaborative analysis system based on privacy computing according to an embodiment of this application; Figure 3 This is a block diagram of the local anchor inference module in a cross-regional power data collaborative analysis system based on privacy computing, according to an embodiment of this application. Figure 4 This is a block diagram of the global knowledge aggregation module in a cross-regional power data collaborative analysis system based on privacy computing, according to an embodiment of this application. Figure 5 This is a block diagram of a dynamic weight estimation unit in a cross-regional power data collaborative analysis system based on privacy computing, according to an embodiment of this application. Figure 6 This is a block diagram of the knowledge distillation optimization module in a cross-regional power data collaborative analysis system based on privacy computing, according to an embodiment of this application. Detailed Implementation

[0009] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0010] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not specifically singular and may include plural forms. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0011] While this application makes various references to certain modules of the systems according to embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The modules described are merely illustrative, and different aspects of the systems and methods may use different modules.

[0012] Flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0013] Existing cross-regional power collaborative analysis mainly relies on parameter averaging federated learning algorithms. However, when faced with the inherent strong statistical heterogeneity of power data, simple parameter aggregation is prone to causing model performance degradation due to gradient conflicts. Furthermore, traditional similarity-based weighting mechanisms often misclassify high-value specific loads that deviate from the mainstream distribution as noise and suppress them, resulting in a severe lack of robustness of the global model under complex operating conditions. Therefore, this application proposes a cross-regional power data collaborative analysis system based on privacy computing. This system abandons the traditional paradigm of directly interacting with model parameters and innovatively uses common anchor data as a medium for feature alignment, mapping the heterogeneous features of local models into standardized logical matrices and performing privacy encryption. In the core aggregation stage, this solution introduces a dynamic weight evaluation mechanism based on entropy-weighted confidence and positive interaction complementarity. By decoupling data credibility and knowledge scarcity, it uses information entropy to eliminate low-quality noise nodes on the one hand, and orthogonal projection to accurately identify and retain specific power patterns that cannot be represented by the mainstream distribution on the other hand. Finally, by backfeeding the aggregated global logic matrix and performing multi-objective knowledge distillation, the local model is guided to efficiently absorb complementary generalized knowledge from the entire network while maintaining the fidelity of regional features, thereby constructing a regional optimized power model that is both robust and personalized.

[0014] Figure 1 This is a system block diagram of a cross-regional power data collaborative analysis system based on privacy computing, according to an embodiment of this application. Figure 2 This is a schematic diagram of data flow in a cross-regional power data collaborative analysis system based on privacy computing, according to an embodiment of this application. Figure 1 and Figure 2As shown, the cross-regional power data collaborative analysis system 100 based on privacy computing according to an embodiment of this application includes: a local model pre-training module 110, used to perform the first round of pre-training processing on an initialized local deep network model based on local power cleaning data to obtain a pre-trained local deep network model; a local anchor point inference module 120, used to input the acquired public anchor point waveform data into the pre-trained local deep network model for forward inference to obtain a local anchor point logic matrix; a privacy enhancement encryption module 130, used to perform privacy enhancement encryption on the local anchor point logic matrix to obtain an encrypted logic data packet; and a global knowledge aggregation module 140, used to control the central... The server receives encrypted logic data packets from multiple different regions and performs global knowledge aggregation based on correlation on the multiple encrypted logic data packets to obtain a global logic matrix containing the power pattern characteristics of the entire network; the global data decryption and verification module 150 is used to receive the global logic matrix broadcast and distributed back by the central server, and decrypt and verify the global logic matrix in combination with the local decryption private key to obtain global plaintext soft labels for model guidance; the knowledge distillation optimization module 160 is used to perform knowledge distillation closed-loop optimization on the pre-trained local deep network model based on local power cleaning data, common anchor waveform data and global plaintext soft labels to obtain a regional optimized power model.

[0015] In the aforementioned cross-regional power data collaborative analysis system 100 based on privacy computing, the local model pre-training module 110 is used to perform the first round of pre-training processing on the initialized local deep network model based on local power cleaning data to obtain the pre-trained local deep network model. It should be noted that, given that the weight parameters of the deep neural network are usually randomly generated during the initialization phase, the model does not yet possess any feature extraction and discrimination capabilities for power load waveforms. If inference is directly performed on the common anchor point data based on the untrained random model, the output logic matrix will only present meaningless random noise. This not only fails to reflect the unique load patterns of the region, such as industrial shocks or residential bi-peak loads, but also causes the subsequent similarity assessment and weighted aggregation mechanism based on anchor point responses to completely fail due to a lack of effective signals. Therefore, the technical solution of this application first utilizes the multi-dimensional time-series features such as voltage, current, and active power in the local power cleaning data to perform the first round of pre-training processing on the initialized local deep network model. Predicted values ​​are calculated through forward propagation and updated through backpropagation combined with the real labels until the loss function converges. This approach endows the local model with an initial understanding of local regional power data, enabling its weight parameters to initially fit the local load distribution patterns. This ensures that when the model subsequently encounters common anchor point data, it can output a statistically significant logistic distribution based on its learned local experience. Through this processing, the cold start problem in federated collaboration can be effectively solved, ensuring that the generated local anchor point logistic matrix truly reflects the model's understanding and preference for standard waveforms within the region. This lays a solid data foundation for the central server to subsequently use entropy values ​​to assess model confidence and accurately identify high-value heterogeneous nodes.

[0016] More specifically, in a specific example of this application, the local model pre-training module 110 includes: a multidimensional data cleaning and standardization unit, used to perform multidimensional power data cleaning and standardization on the local original power dataset to obtain local power cleaned data; and a local model pre-training unit, used to divide the local power cleaned data into training and validation sets and input them into an initialized local deep network model for forward propagation calculation, and to perform backpropagation update based on the deviation between the predicted value and the true label until the loss function converges, so as to obtain the pre-trained local deep network model.

[0017] Accordingly, the multidimensional data cleaning and standardization unit is used to perform multidimensional power data cleaning and standardization on the local raw power dataset to obtain local cleaned power data. It should be noted that due to the influence of electromagnetic interference, transmission packet loss, or equipment failure on the sensor equipment deployed at the power system site, the collected local raw power dataset often contains missing values, abnormal noise, and physical quantities with inconsistent dimensions, such as voltage at the kilovolt level and current at the ampere level. Such data quality defects and magnitude differences can cause deep network models to struggle to capture effective features during training and even exhibit gradient oscillations that prevent convergence. Based on this, the technical solution of this application first performs multidimensional power data cleaning and standardization on the local raw power dataset to obtain local cleaned power data, thereby eliminating noise interference from the data source and unifying the numerical distribution range of different physical dimensions, laying a standardized data foundation for constructing a high-precision power load model. Through the above processing, the quality and stability of the input data can be effectively improved, accelerating the gradient descent convergence speed during the backpropagation of the subsequent deep network model, and preventing bias in dominant features caused by differences in dimensions.

[0018] More specifically, in a concrete example of this application, the process first performs data cleaning on the multi-dimensional time-series features of voltage, current, and active power contained in the local raw power dataset. A linear interpolation algorithm is used to fill in timestamp gaps or lost values ​​caused by network latency to maintain the continuity of the time-series data. Simultaneously, statistical threshold judgment rules are used to identify and remove abnormal noise points that deviate from the normal operating range to purify the sample space. Subsequently, a standardization transformation is performed on the cleaned multi-dimensional data. The mean and standard deviation of each dimension of voltage, current, and active power are calculated, and the Z-Score standardization formula is used to map the original physical values ​​to a dimensionless distribution with a mean of 0 and a variance of 1. This eliminates the differences between different power physical quantities and ensures that each dimension of features has a balanced contribution in the subsequent neural network weight update.

[0019] Accordingly, the local model pre-training unit is used to divide the local power cleaning data into training and validation sets and input them into the initialized local deep network model for forward propagation calculation. It then performs backpropagation updates based on the deviation between the predicted values ​​and the true labels until the loss function converges, thus obtaining the pre-trained local deep network model. It should be noted that, due to the random distribution of internal weight parameters in the initial construction stage of the deep neural network, the model has not yet established a mapping relationship between multi-dimensional features such as voltage, current, and active power and power fault categories, lacking the ability to discriminate load patterns in specific local areas. Based on this, the technical solution of this application further divides the local power cleaning data into training and validation sets and inputs them into the initialized local deep network model for forward propagation calculation. It then performs backpropagation updates based on the deviation between the predicted values ​​and the true labels until the loss function converges. This utilizes local historical data for supervised learning of the model, forcing the model weight parameters to evolve towards fitting the local load feature space. Through the above processing, the model can be effectively endowed with the initial representation capability of local power data, ensuring that it can extract regionally specific waveform features such as industrial impact loads or new energy reverse power transmission, laying the parameter foundation for the subsequent generation of high-quality anchor point logic matrices.

[0020] More specifically, in a specific example of this application, the implementation process first performs data splitting and forward mapping operations, dividing the standardized local power cleaning data into mutually exclusive training and validation subsets according to a preset ratio. The time-series samples from the training subset are then batch-inputted into the input layer of a local deep network model, undergoing feature extraction and nonlinear transformation through a multi-layer neural network. In this embodiment, considering the temporal characteristics of power data, the local deep network model can specifically employ a one-dimensional convolutional neural network or a long short-term memory network. For example, a 1D-CNN containing three convolutional layers and two pooling layers is used to extract local mutation features of the waveform, followed by two fully connected layers for fault classification mapping. Normalized prediction probabilities belonging to various operating conditions / event categories (including load mode categories or fault categories) are generated at the output layer. Subsequently, error measurement and backpropagation operations based on cross-entropy are performed. The cross-entropy loss function is used to calculate the degree of difference between the predicted distribution and the true label distribution. This function constructs the optimization objective by measuring the log-likelihood error between the predicted output and the true label when the model parameters are input to local samples, and the gradient vector of the loss function with respect to each weight parameter is calculated using the chain rule. Finally, iterative optimization and convergence determination operations are performed. The model weights are dynamically adjusted according to the negative gradient direction using the stochastic gradient descent algorithm. After each iteration, the generalization performance of the model is evaluated using the validation subset. Once the loss function value tends to stabilize or the validation accuracy no longer improves, the training process is terminated and a pre-trained model with solidified local feature knowledge is output.

[0021] In the aforementioned cross-regional power data collaborative analysis system 100 based on privacy computing, the local anchor inference module 120 is used to input the acquired common anchor waveform data into a pre-trained local deep network model for forward inference to obtain a local anchor logic matrix. It should be noted that, given the inherent significant non-independent and identically distributed characteristics of cross-regional power data, and the vast differences in power load patterns across different geographical regions, directly aggregating model parameters can lead to severe gradient conflicts and negative migration phenomena. Furthermore, the participating parties cannot directly interact with the original data for feature alignment. Based on this, the technical solution of this application further inputs the acquired common anchor waveform data into a pre-trained local deep network model for forward inference to obtain a local anchor logic matrix. This utilizes the common anchor data containing standard waveform features as a medium for knowledge interaction, stimulating the pre-trained local model to output its original confidence level and probability distribution regarding fault category judgment, thereby explicitly transforming the local-specific load characteristic knowledge into a standardized logic vector. Through the above processing, a unified feature mapping benchmark can be effectively established between physically isolated data silos. Without leaking the privacy of local original data, the parameters of heterogeneous models can be transformed into a general knowledge space without loss. This provides quantifiable data support for the central server to accurately evaluate the confidence and complementarity of models in each region.

[0022] Figure 3 This is a block diagram of the local anchor inference module in a cross-regional power data collaborative analysis system based on privacy computing, according to an embodiment of this application. Figure 3 As shown, the local anchor inference module 120 includes: a mode switching and freezing unit 121, used to perform inference mode switching and gradient freezing processing on the pre-trained local deep network model to obtain the inference state local model; an anchor forward mapping unit 122, used to input the common anchor waveform data into the inference state local model in batches for forward mapping to obtain the original output vector of the model containing the original confidence score; and a temperature scaling and softening unit 123, used to perform temperature scaling and probability distribution softening on the original output vector of the model based on a preset distillation temperature coefficient to obtain the local anchor logic matrix.

[0023] In the aforementioned cross-regional power data collaborative analysis system 100 based on privacy computing, the mode switching and freezing unit 121 is used to perform inference mode switching and gradient freezing processing on the pre-trained local deep network model to obtain the inference state local model. It should be noted that the random deactivation mechanism and dynamic statistical characteristics of the batch normalization layer, used during the training phase of the deep network model to enhance generalization ability, introduce computational randomness, leading to inconsistent output results from multiple forward inferences of the same common anchor waveform data. Furthermore, retaining the gradient computation graph during the inference phase unnecessarily consumes computational resources and increases the risk of accidental modification of model parameters. Therefore, the technical solution of this application further performs inference mode switching and gradient freezing processing on the pre-trained local deep network model to obtain the inference state local model. This forces the fixing of the internal connection structure and statistical parameters of the deep network, eliminates the interference of random factors on the feature extraction process, and blocks the gradient update path of backpropagation from the bottom layer. Through the above processing, it can be effectively ensured that the logic matrix excited by the common anchor data strictly and definitively reflects the local model's discrimination logic for standard power waveforms, guaranteeing the stability and reliability of the cross-regional feature alignment benchmark.

[0024] More specifically, in a concrete example of this application, the implementation process first performs a mode switching operation on the computation graph state. By calling the evaluation mode interface of the deep learning computation framework, the local deep network model is switched from the training state to the inference state. During this process, the program automatically instructs the randomly deactivated layers in the network to stop working, ensuring that all neurons remain fully connected and activated during signal transmission. At the same time, the moving average and moving variance of the batch normalization layer are locked, causing it to stop real-time statistics on the input batch data and instead use global statistics fixed during the training phase for data normalization. Subsequently, a parameter gradient freezing operation is performed, traversing all weight parameter tensors in the model computation graph and marking their gradient requirement attributes as unavailable. This notifies the automatic differentiation engine to stop building the backpropagation graph for subsequent forward computation, ultimately generating a local inference state model with constant parameters that is only used for performing forward feature mapping.

[0025] In the aforementioned cross-regional power data collaborative analysis system 100 based on privacy computing, the anchor point forward mapping unit 122 is used to input the common anchor point waveform data into the local model in the inference state in batches for forward mapping to obtain the original output vector of the model containing the original confidence score. It should be noted that due to the high-dimensional non-convexity of deep neural networks, the local models trained based on heterogeneous power data from different regions exhibit huge differences in parameter space. Directly aligning the parameters is difficult to reflect the true discrimination logic of the model, and the probability distribution after normalization by the Softmax activation function often compresses the subtle feature information of non-target categories, i.e., the so-called dark knowledge, thereby losing the model's fine cognition of power waveform similarity and confusion. Based on this, the technical solution of this application further inputs the acquired common anchor point waveform data into the local model in the inference state in batches for forward mapping to obtain the original output vector of the model containing the original confidence score. This is to use the standardized public dataset to stimulate the original response of the local model in the unfiltered state and capture the model's potential association judgment and confidence tendency for various power fault modes. Through the above processing, the abstract and incommensurable model weights can be effectively transformed into feature response vectors based on a unified benchmark. While preserving the model's ability to specifically distinguish certain load patterns (such as high-impedance grounding faults), it provides a measurable and comparable standardized data carrier for subsequent global knowledge aggregation.

[0026] More specifically, in a concrete example of this application, the process first performs batch loading and tensor transformation of data. It reads a pre-set public anchor dataset containing IEEE standard node test waveforms from memory, divides it into several fixed-size micro-batch data packets according to the memory capacity limitations of the computing node, and converts these time-series waveform data into a high-dimensional tensor format adapted to the model input interface. Subsequently, it performs batch-wise forward feature extraction and mapping, sequentially inputting each micro-batch tensor into the input layer of the local model in the inference state. This drives the data flow through convolutional layers for time-frequency feature extraction and fully connected layers for high-dimensional spatial mapping, calculating the unnormalized value output by the model at the last fully connected layer. Finally, it performs vectorization extraction and concatenation of the original output, directly extracting the original values ​​that have not undergone Softmax or Sigmoid activation function processing as confidence scores. These values ​​accurately reflect the model's tendency to classify the current anchor waveform as various fault categories. These values ​​are then reassembled and concatenated according to the sample sequence to construct the original output vector of the model that can fully represent the local model's discrimination logic.

[0027] In the aforementioned cross-regional power data collaborative analysis system 100 based on privacy computing, the temperature scaling and softening unit 123, it should be noted, is crucial because conventional deep neural network classifiers, when using a standard normalized exponential function in the output layer, tend to generate extremely steep probability distributions. This results in the probability values ​​of non-target categories other than the highest confidence category being compressed to near zero. Consequently, the potential correlation information, i.e., dark knowledge, between power fault modes with similar waveform characteristics, such as high-impedance grounding and intermittent arc grounding, is severely suppressed or even lost by the local model. Based on this, the technical solution of this application further performs temperature scaling and probability distribution softening on the original output vector of the model based on a preset distillation temperature coefficient to obtain a local anchor point logic matrix. This smooths the prediction probability distribution curve and forcibly increases the probability response values ​​of non-dominant categories, thereby making the structured relationships between categories hidden in small probability values ​​explicit. Through the above processing, the information entropy density uploaded by the local model can be effectively enriched, ensuring that the deep discrimination logic of the local model for complex power conditions is completely transmitted to the central server without leaking the original data, preventing knowledge homogenization caused by excessively sharp probability distributions.

[0028] More specifically, in a concrete example of this application, the implementation process first performs temperature parameter configuration and linear scaling operations. A scalar with a value greater than 1 is selected as the distillation temperature coefficient. The value of this coefficient directly determines the smoothness of the output probability distribution and the proportion of negative label information revealed. This coefficient is used to perform a division operation on each element in the original output vector of the model, thereby reducing the numerical differences between the original confidence scores of each category as a whole. Subsequently, a softening mapping and normalization operation of the probability distribution is performed. The vector scaled by the temperature coefficient is input into a generalized normalization exponential function for nonlinear transformation. The exponent of each category score is calculated and its sum is used as the denominator for normalization processing, generating a softened probability vector with a sum of 1 and a relatively flat distribution shape. At this time, the originally weak suboptimal category probability is significantly amplified. Finally, the stacking and assembly operation of the logic matrix is ​​performed. The softening probability vector calculated for each time-series waveform sample in the common anchor point dataset is vertically spliced ​​according to the time index or batch order of the sample to construct a high-dimensional local anchor point logic matrix. This matrix encodes the complete understanding of the local model's perception of the standard power waveform feature space under a specific temperature field in a structured form.

[0029] In the aforementioned cross-regional power data collaborative analysis system 100 based on privacy computing, the privacy-enhancing encryption module 130 is used to perform privacy-enhancing encryption on the local anchor point logic matrix to obtain encrypted logic data packets. It should be noted that, although the local anchor point logic matrix desensitizes the original waveform, the high-precision probability distribution details it contains still fully record the local model's discrimination boundary information for specific power load patterns. If transmitted directly in plaintext, malicious attackers can easily use model inversion techniques to reverse-engineer sensitive local industrial or residential electricity consumption behavior characteristics. Furthermore, the direct interaction of massive amounts of floating-point data would cause severe congestion of cross-regional communication links. Based on this, the technical solution of this application further performs pruning and quantization compression preprocessing on the local anchor point logic matrix to obtain a quantized logic matrix. Noise intensity parameters are calculated based on a preset privacy budget, and a random noise matrix is ​​generated and superimposed on the quantized logic matrix to obtain a perturbation logic matrix. Finally, metadata header information containing the region ID and quantization factor is appended to the perturbation logic matrix to obtain an encrypted logic data packet. This utilizes a differential privacy mechanism to introduce statistically indistinguishable random perturbations into the original logic values, mathematically masking the precise contribution of individual data points. Simultaneously, fixed-point quantization technology significantly compresses the transmission load. Through the above processing, a defense barrier against model reverse engineering attacks can be effectively constructed, ensuring that each region participates in collaboration while strictly protecting the privacy of its local core load pattern, and reducing bandwidth consumption and communication latency in cross-regional wide area network transmission while ensuring data security.

[0030] More specifically, in a specific example of this application, the privacy-enhancing encryption module 130 includes: a pruning and quantization compression unit, used to prune and quantize the local anchor logic matrix to obtain a quantized logic matrix; a noise generation unit, used to calculate noise intensity parameters based on a preset privacy budget, and generate a random noise matrix with the same dimension as the quantized logic matrix based on the noise intensity parameters; a noise superposition and perturbation unit, used to superimpose the random noise matrix onto the quantized logic matrix to obtain a perturbation logic matrix; and a metadata encapsulation unit, used to append metadata header information containing a region ID and a quantization factor to the perturbation logic matrix to obtain an encrypted logic data packet.

[0031] Accordingly, the pruning and quantization compression unit is used to prune and quantize the local anchor logic matrix to obtain a quantized logic matrix. It should be noted that since the local anchor logic matrix is ​​usually composed of high-precision floating-point values, directly transmitting such a dense matrix in a cross-regional wide area network would consume valuable communication bandwidth resources. Furthermore, subsequent homomorphic encryption or differential privacy algorithms often require algebraic operations on an integer ring to avoid floating-point operation errors and improve computational efficiency. Based on this, the technical solution of this application further prunes and quantizes the local anchor logic matrix to obtain a quantized logic matrix, thereby eliminating redundant tail noise in the logic distribution and mapping the continuous floating-point probability space to a discrete integer space. Through the above processing, the data transmission volume can be effectively compressed, network communication overhead reduced, and a suitable and efficient data format provided for subsequent privacy protection protocols based on integer operations can be provided.

[0032] More specifically, in a concrete example of this application, the implementation process first performs a sparsity threshold pruning operation, traversing each probability element in the local anchor point logic matrix and setting non-critical probability terms with values ​​below a preset small threshold to zero, thereby removing background noise that contributes very little to power fault detection and improving the sparsity of the matrix. Subsequently, a fixed-point mapping operation is performed, setting a sufficiently large quantization scaling factor and multiplying this factor with the pruned retained probability values ​​to amplify the original decimal values ​​between zero and one to the integer range, ensuring sufficient effective digital precision is retained during the numerical conversion process. Finally, a rounding truncation operation is performed, rounding the amplified values ​​to the nearest integer or lower, converting them to a fixed-point integer format, thereby generating a compact quantization logic matrix that is convenient for subsequent encryption operations.

[0033] Accordingly, the noise generation unit and the noise superposition perturbation unit are used to calculate noise intensity parameters based on a preset privacy budget, and generate a random noise matrix with the same dimension as the quantization logic matrix based on the noise intensity parameters. This random noise matrix is ​​then superimposed onto the quantization logic matrix to obtain the perturbation logic matrix. It should be noted that although the quantized logic matrix reduces data precision, its internal numerical distribution still retains the discrimination characteristics of the local power model for specific load patterns. If this deterministic numerical matrix is ​​directly interacted with, attackers can still use member inference attacks or model inversion techniques to reconstruct sensitive local electricity consumption details or industrial production scheduling patterns based on the statistical characteristics of the probability distribution. Based on this, the technical solution of this application further calculates noise intensity parameters based on a preset privacy budget, and generates a random noise matrix with the same dimension as the quantization logic matrix based on the noise intensity parameters. This random noise matrix is ​​then superimposed onto the quantization logic matrix to obtain the perturbation logic matrix, thereby introducing statistical uncertainty interference. This allows the influence of a single data sample on the aggregation result to be masked by random noise, thus satisfying the indistinguishability requirement of the differential privacy mechanism. Through the above processing, the strong correlation between the output results and individual samples can be effectively severed at the mathematical level, ensuring that even if the data is intercepted during transmission, third parties cannot extract the original local load characteristics from the perturbed values. At the same time, the availability of global statistical regularities is preserved by utilizing the mutual cancellation properties of noise during large-scale aggregation.

[0034] More specifically, in a concrete example of this application, the implementation process first performs noise parameter calculation. Based on the system's requirement for a balance between privacy protection and data availability, a privacy budget parameter is set. Combining this with the maximum possible impact of a single anchor sample in the quantization logic matrix on the overall distribution—i.e., the global sensitivity—the intensity parameter controlling the noise dispersion is calculated using the mathematical definitions of the Laplace or Gaussian mechanism. Subsequently, a noise matrix of the same dimension is generated. A pseudo-random number generator is invoked to construct a matrix container with the same dimension as the quantization logic matrix. Based on the intensity parameter calculated above, a series of random values ​​are independently sampled from the Laplace or Gaussian distribution and filled into this container. Finally, an additive perturbation superposition operation is performed. Matrix addition is used to element-wise superimpose the generated random noise matrix onto the quantization logic matrix, generating a perturbation logic matrix where the values ​​are shifted but the statistical mean remains stable, thus completing the differential privacy processing of the local model knowledge.

[0035] Accordingly, the metadata encapsulation unit is used to append metadata header information containing the region ID and quantization factor to the perturbation logic matrix to obtain encrypted logic data packets. It should be noted that, in cross-regional collaborative scenarios, the central server needs to process parallel data streams from different geographical locations simultaneously, and the received perturbation logic matrix has already undergone fixed-point scaling, losing the original probability magnitude. Without necessary identification and restoration parameters, the server will be unable to correctly parse the data's origin and true value. Based on this, the technical solution of this application further appends metadata header information containing the region ID and quantization factor to the perturbation logic matrix to obtain encrypted logic data packets, thereby establishing a unique traceability index for the anonymized load data and transmitting the key scalar parameters required for reverse restoration. Through the above processing, it can be effectively ensured that the central server accurately matches the weight coefficients of each region in the subsequent aggregation process and recovers the true probability amplitude of the global logic distribution without loss based on the quantization factor.

[0036] More specifically, in a concrete example of this application, the implementation process first performs metadata extraction and structured definition operations. A pre-allocated unique region identifier is read from the secure storage area of ​​the local configuration file to characterize the geographic or logical identity of the computing node. Simultaneously, the numerical scaling factor used in the preceding quantization step is obtained as the quantization factor. Subsequently, data packet assembly and serialization operations are performed. A header field containing a version number, region identifier, and quantization factor is constructed according to a predefined communication protocol specification. This header field is concatenated with the perturbation logic matrix, which serves as the payload. A serialization algorithm is used to convert the combined structured object into a binary byte stream suitable for network transmission. Finally, a transport layer security encapsulation operation is performed. The transport layer security protocol interface is called to encrypt and encapsulate the aforementioned binary byte stream, generating an encrypted logical data packet with tamper-proof and eavesdropping-proof capabilities, ready to be sent to the central coordination server via a wide area network link.

[0037] In the aforementioned cross-regional power data collaborative analysis system 100 based on privacy computing, the global knowledge aggregation module controls the central server to receive encrypted logical data packets from multiple different regions and performs correlation-based global knowledge aggregation on these encrypted logical data packets to obtain a global logical matrix containing the power pattern characteristics of the entire network. It should be noted that, given the inherent significant statistical heterogeneity of cross-regional power data, traditional average aggregation strategies are prone to misclassifying high-value long-tail specific patterns, such as heavy industrial impact loads, as outliers and suppressing them. Furthermore, the differential privacy noise contained in the encrypted data packets uploaded by each region requires large-scale statistical aggregation to eliminate its negative impact on model accuracy. Based on this, the technical solution of this application further controls the central server to receive encrypted logical data packets from multiple different regions and performs correlation-based global knowledge aggregation on these encrypted logical data packets to obtain a global logical matrix containing the power pattern characteristics of the entire network. This utilizes a correlation evaluation mechanism to decouple data quality and knowledge scarcity, dynamically assigning higher aggregation weights to nodes with high confidence and high complementarity, while simultaneously using the law of large numbers to mutually cancel out random noise interference introduced by each node during the aggregation process. Through the above processing, robust features that are both universally applicable and region-specific can be effectively extracted from massive heterogeneous data while protecting privacy, and a global consensus logic that can cope with extreme working conditions can be constructed.

[0038] Figure 4 This is a block diagram of the global knowledge aggregation module in a cross-regional power data collaborative analysis system based on privacy computing, according to an embodiment of this application. Figure 4 As shown, the global knowledge aggregation module 140 includes: an unpacking and verification unit 141, used to control the central server to unpack and verify the integrity of encrypted logical data packets from multiple regions to obtain a set of perturbed logical matrices; a dynamic weight estimation unit 142, used to perform correlation-based dynamic weight estimation on the set of perturbed logical matrices to obtain a regional aggregation weight vector; and a global weighted aggregation unit 143, used to perform global weighted aggregation and noise cancellation on the set of perturbed logical matrices based on the regional aggregation weight vector to obtain a global logical matrix.

[0039] In the aforementioned cross-regional power data collaborative analysis system 100 based on privacy computing, the unpacking and verification unit 141 is used to control the central server to unpack and verify the integrity of encrypted logic data packets from multiple regions to obtain a set of perturbation logic matrices. It should be noted that, since the cross-regional power collaborative network is built on a complex wide area network environment, long-distance data transmission links face objective security risks such as packet loss, latency jitter, and man-in-the-middle tampering. Furthermore, encrypted data packets from nodes in different regions are mixed in with massive amounts of network traffic. Without strict identity authentication and content integrity checks, if tampered or forged malicious data packets are mixed into the subsequent aggregation calculation process, it will cause the final generated global model logic to deviate or even completely fail. Based on this, the technical solution of this application further controls the central server to unpack and verify the integrity of encrypted logic data packets from multiple different regions to obtain a set of perturbation logic matrices. This constructs a robust security filtering barrier at the data aggregation layer, automatically eliminating abnormal data with transmission errors or unknown sources, and extracting the effective payload for subsequent core calculations from the communication encapsulation format. Through the above processing, it can be effectively ensured that every piece of data participating in global aggregation originates from an authorized and legitimate power area node and retains its original state at the sending end, laying a reliable and pure data foundation for subsequent high-precision knowledge fusion.

[0040] More specifically, in a concrete example of this application, the implementation process first performs network reception and protocol parsing operations. It utilizes the high-concurrency network interface of the central server to listen on a designated service port, capturing data streams uploaded from various distributed power area nodes in real time. It parses the header control information of the data packets according to predefined communication protocol specifications, checks the timestamp field to determine the timeliness of the data packets and prevent replay attacks, and simultaneously reads the area identifier to confirm the sender's legitimacy. Subsequently, it performs digital signature verification and integrity screening operations, retrieving the public key certificates or shared keys of each area node pre-stored in the system, performing cryptographic verification on the digital signature segment in the data packet, and comparing the calculated hash digest with the digest value carried in the packet to ensure strict consistency. If a signature mismatch or hash value discrepancy is found, the risky or corrupted data packet is immediately discarded. Finally, the payload stripping and set construction operations are performed. For all valid data packets that pass the verification, the outer transport layer encryption encapsulation and additional metadata header information are removed, the internally encapsulated perturbation logic matrix is ​​extracted, and the matrix objects from all valid nodes are aggregated into a unified memory buffer to form a set of perturbation logic matrices to be processed, ready to enter the next stage of the aggregation calculation process.

[0041] In the aforementioned privacy-preserving computing-based cross-regional power data collaborative analysis system 100, the dynamic weight estimation unit 142 is used to perform correlation-based dynamic weight estimation on the perturbation logic matrix set to obtain a regional aggregated weight vector. It should be noted that, given that in cross-regional power data collaboration scenarios, traditional similarity calculation paradigms often assume that reliability equates to similarity, they are prone to misinterpreting data that deviates from the mainstream distribution but contains key specific knowledge, such as heavy industrial loads exhibiting square wave patterns or regional data from new energy reverse power transmission with strong randomness, as noise or anomalies and suppressing them, thus making the aggregated model unable to cope with extreme operating conditions. Based on this, the technical solution of this application further performs correlation-based dynamic weight estimation on the perturbation logic matrix set to obtain a regional aggregated weight vector, including uncertainty assessment of the perturbation logic matrix set to obtain a regional confidence vector, orthogonal projection of the global feature space to obtain a positive interactive complement score vector, and complementary fusion of the regional confidence vector and the positive interactive complement score vector under confidence constraints. This approach decouples data credibility and knowledge scarcity, using information entropy theory to quantify the determinism of local model prediction distributions to eliminate low-quality noise. Simultaneously, orthogonal projection is used to mine complementary knowledge that cannot be linearly represented by the mainstream consensus across the network. Through this process, accurate identification and protection of high-value heterogeneous data can be effectively achieved. This ensures that the global model, while eliminating over 90% of low-quality interference data, can losslessly capture key long-tail scenario features, thereby improving the accuracy and robustness of fault diagnosis in power grid systems under complex and changing environments.

[0042] Figure 5 This is a block diagram of the dynamic weight estimation unit in a cross-regional power data collaborative analysis system based on privacy computing, according to an embodiment of this application. Figure 5 As shown, the dynamic weight estimation unit 142 includes: an uncertainty assessment subunit 1421, used to perform uncertainty assessment on the set of perturbation logic matrices to obtain a regional confidence vector; an orthogonal projection subunit 1422, used to perform orthogonal projection of the set of perturbation logic matrices onto the global feature space to obtain a positive interactive complement score vector; and a complementary fusion subunit 1423, used to perform complementary fusion of the regional confidence vector and the positive interactive complement score vector under confidence constraints to obtain a regional aggregate weight vector.

[0043] In the aforementioned cross-regional power data collaborative analysis system 100 based on privacy computing, the uncertainty assessment subunit 1421 is used to perform uncertainty assessment on the set of perturbation logic matrices to obtain a regional confidence vector. It should be noted that, given that in a cross-regional power collaborative network, the low relevance of data uploaded by the local model may stem from poor data quality, such as white noise caused by sensor failures, or from the high specificity of the data itself, simply relying on relevance screening is prone to mistakenly eliminating high-value heterogeneous nodes. Therefore, the technical solution of this application further performs uncertainty assessment on the set of perturbation logic matrices to obtain a regional confidence vector, thereby using information entropy theory to quantify the certainty of the local model's predicted distribution and accurately distinguish between low relevance caused by the model's failure to learn effective discrimination logic and high specificity caused by special operating conditions. Through the above processing, truly low-quality noise nodes can be effectively eliminated from the source, clearing obstacles for the subsequent introduction of heterogeneous data and ensuring the purity of the aggregation basis.

[0044] More specifically, in a particular example of this application, the implementation process first performs an entropy calculation operation on the probability distribution, for the probability distribution from the first... The perturbation logic matrix received from each region is used to extract its predicted probability distribution for the common anchor point data. If the prediction distribution exhibits high entropy (i.e., the predicted probability is flat and ambiguous), it indicates that the model itself has not learned effective discriminative logic, and its output is considered unreliable. Conversely, if the entropy value is low and the distribution is sharp, it indicates that the model has high internal logical consistency. Subsequently, a confidence mapping operation based on the entropy value is performed to calculate its average information entropy. and the derived confidence coefficient The specific calculation logic follows the following mathematical expression: in, The number of anchor point samples. The total number of categories, The normalized prediction probability. For the first The average entropy value of each region, The threshold value is the average entropy of the entire network. To adjust the hyperparameters of sensitivity, To prevent the use of tiny constants that are meaningless in logarithm, These are the elements in the region confidence vector. The closer a value is to 1, the more confident the model is. Taking a specific power fault monitoring scenario as an example, suppose a hardware failure occurs in the measuring equipment in a certain area, causing the collected voltage data to be filled with random white noise. A local model trained based on this noisy data cannot make a clear fault classification judgment when faced with a common anchor point waveform. Its output prediction probability is evenly distributed across all categories, and the calculated average information entropy... This will be significantly higher than the average threshold for the entire network. This leads to the confidence coefficient after mapping through the Sigmoid function. When the confidence level approaches 0, the system will determine that the node is an unreliable noise source based on this low confidence level and suppress it to prevent it from polluting the global model. Meanwhile, high-quality heterogeneous nodes with distinctive data distributions but decisive fault judgments and low entropy values ​​will be retained, thereby achieving accurate denoising in the dimension of uncertainty.

[0045] In the aforementioned cross-regional power data collaborative analysis system 100 based on privacy computing, the orthogonal projection subunit 1422 is used to perform global feature space orthogonal projection on the set of perturbation logic matrices to obtain positive interactive complementary score vectors. It should be noted that, based on confirming the reliability of the data, it is necessary to further explore who can provide new knowledge; simply superimposing homogeneous data is insufficient to improve the generalization boundary of the model. Based on this, the technical solution of this application further performs global feature space orthogonal projection on the set of perturbation logic matrices to obtain positive interactive complementary score vectors. By constructing the mainstream consensus vector (mean vector) of the entire network and utilizing the Gram-Schmidt orthogonalization idea, the data vectors of each region are decomposed into two components parallel to and perpendicular to the consensus. Subsequently, the magnitude of the vertical component (i.e., orthogonal projection) is calculated to quantify the unique information increment contained in that region that cannot be linearly represented by other nodes in the entire network. Through the above processing, extremely valuable complementary knowledge is introduced into the system, breaking the redundant accumulation of homogeneous information and thus achieving a substantial expansion of the knowledge boundary.

[0046] More specifically, in a particular example of this application, the implementation process first performs a consensus vector construction operation, averaging the feature vectors uploaded from all regions in the perturbation logic matrix set to obtain the mainstream consensus vector representing the universal mode across the entire network. Subsequently, orthogonal decomposition and complementarity scoring operations are performed. For each participating region node, the projection component of its eigenvector onto the consensus vector direction is calculated using the vector projection formula. Then, the orthogonal component perpendicular to the consensus direction is separated by vector subtraction, and the squared Euclidean norm of this orthogonal component is calculated as the complementarity score. The mathematical expression of this step is as follows: Let... Given the mainstream consensus vector (the mean of features across all regions), calculate the... Feature vectors of each region Positive interactive complementary score : in, Represents the vector dot product. The norm of a vector is the square of its magnitude. Representing vectors exist The projection vector on, The larger the value, the greater the difference between the features provided by the region and the mainstream consensus, i.e., the stronger the complementarity. Taking a specific cross-regional power grid coordination scenario as an example, assuming that most of the regions participating in the coordination are urban residential loads, their uploaded feature vectors all show typical daily load curves with morning and evening peaks. The consensus vector formed by the aggregation of these vectors... It also exhibits a smooth bimodal shape. At this point, if there exists an industrial region node containing a large electrolytic aluminum plant, its feature vector... It manifests as a sustained, stable, high-energy-consuming straight line or a specific square-wave impact pattern. When performing orthogonal projection calculations, the projection component of the characteristic vector of this industrial area onto the consensus vector of the bimodal shape is small, while the magnitude of the orthogonal component perpendicular to the consensus vector is extremely large, resulting in a calculated complementary score. The weight is very high. Based on this, the system recognizes that although the node is different from the mainstream distribution, it provides industrial load characteristics that are lacking in the entire network. Therefore, it assigns a high weight to it in subsequent aggregation to prevent this key specific pattern from being overwhelmed by the majority of residential loads.

[0047] In the aforementioned cross-regional power data collaborative analysis system 100 based on privacy computing, the complementary fusion subunit 1423 is used to perform complementary fusion of the regional confidence vector and the positive interactive complementarity score vector under confidence constraints to obtain a regional aggregate weight vector. It should be noted that, given that a single-dimensional evaluation is insufficient to comprehensively measure node value, focusing solely on complementarity may lead to high-entropy noise data being mistakenly assigned high weights due to its uniqueness, thus polluting the global model; focusing solely on confidence may overlook key heterogeneous nodes that, although possessing high confidence, differ significantly from the mainstream distribution due to sample scarcity. Based on this, the technical solution of this application further performs complementary fusion of the regional confidence vector and the positive interactive complementarity score vector under confidence constraints to obtain a regional aggregate weight vector, thereby constructing an incentive mechanism with hard constraints: if a node has low confidence, regardless of its specificity, it is considered untrusted noise and suppressed (weight tends to 0); only when a node simultaneously possesses high confidence and high complementarity (i.e., a confident and unique expert node) is the system given a significant weight reward. Through the above processing, a balance is effectively found between denoising and fidelity preservation, ensuring that the global model is not contaminated by poor data while absorbing unique experiences from different regions. In the end, a robust and intelligent global optimization model is generated, which improves the robustness and global generalization accuracy of the cross-regional power collaborative analysis model in complex power grid environments.

[0048] More specifically, in a concrete example of this application, the implementation process first performs a nonlinear fusion calculation operation, reading the regional confidence coefficients and positive interaction complement scores generated in the preceding steps. A hyperbolic tangent function is used to saturate and map the complementary scores to prevent numerical overflow. Then, a specific reward factor is introduced to adjust the incentive intensity for heterogeneous knowledge. The confidence coefficients are used as multiplicative gating factors and weighted together with the mapped complementary scores to calculate the unnormalized final score. Subsequently, a global normalization operation based on Softmax is performed, introducing a temperature coefficient to control the entropy of the weight distribution. The final score is mapped into a probabilistic aggregated weight vector, which directly determines the contribution ratio of the perturbation logic matrix uploaded from each region in building global consensus. The mathematical expression and symbol interpretation of this step are as follows: in, For the first Integration score for each region The confidence coefficient is... To compensate for the positive interaction score, It is a specific reward factor used to regulate the incentive intensity for heterogeneous knowledge. The hyperbolic tangent function is used to limit the reward cap and prevent weight explosion. The temperature coefficient is Softmax. The final generated improved aggregate weights incorporate considerations of quality and complementarity. This represents the total number of regions participating in the aggregation. Taking a specific cross-regional power grid collaboration scenario as an example, suppose there is a regional node in the system that generates random white noise due to a sensor malfunction. Although its data distribution is drastically different from the mainstream of the entire network, resulting in a complementary score... Extremely high, but the confidence coefficient is low due to its excessively high predictive entropy value. Approaching 0, the final score calculated according to the above formula will be forcibly reset to zero, thus effectively suppressing the pollution of the global model by more than 90% of low-quality interference data; at the same time, for a key long-tail scenario node containing industrial heavy load or new energy backfeeding characteristics, because its waveform is unique and its model discrimination logic is clear (high... And high The system will assign it a high weight, thereby achieving lossless capture and fusion of key long-tail scenario features such as heavy industrial loads and new energy reverse power transmission.

[0049] In the aforementioned cross-regional power data collaborative analysis system 100 based on privacy computing, the global weighted aggregation unit 143 is used to perform global weighted aggregation and noise cancellation on the perturbation logic matrix set based on the regional aggregation weight vector to obtain a global logic matrix. It should be noted that, because random noise is artificially injected into the logic matrix at the local end of each region to meet differential privacy protection requirements, the data uploaded from a single region has numerical deviations. Furthermore, different regions occupy different importance in the global knowledge system due to differences in data quality and heterogeneity. If they are directly merged without weighted processing, the noise cannot be effectively canceled, and low-quality data will dilute high-value knowledge. Based on this, the technical solution of this application further performs global weighted aggregation and noise cancellation on the perturbation logic matrix set based on the regional aggregation weight vector to obtain a global logic matrix. This utilizes the law of large numbers to allow independent random noise following a specific distribution to cancel each other out during the weighted summation process. Simultaneously, based on the dynamic weights calculated in the preceding steps, the contributions of high-confidence and highly complementary nodes are accurately amplified. Through the above processing, a high-fidelity consensus logic for the entire network can be effectively restored from noisy distributed data, and a global knowledge carrier that has both broad generalization capabilities and retains key specific features can be constructed.

[0050] More specifically, in a concrete example of this application, the implementation process first performs a weighted matrix calculation operation. It iterates through each matrix object in the perturbation logic matrix set, reads the corresponding regional aggregation weight scalar generated in the previous steps, and uses scalar multiplication to apply the weight coefficients element-wise to the perturbation logic matrix, generating a weighted matrix reflecting the proportion of that region's global contribution. Subsequently, noise cancellation and tensor aggregation operations are performed, summing the weighted matrices calculated for all regions element-wise. During this process, since the noise injected into each region follows a zero-mean distribution and is independent, as the number of nodes participating in the aggregation increases, the expected value of the superimposed noise approaches zero, thereby achieving the enhancement of the effective signal and the self-cancellation of random noise. Finally, the results are encapsulated and normalized for confirmation. The accumulated aggregation result is used as the global logic matrix, which mathematically approximates the noise-free weighted true value distribution, establishing a consensus on power fault mode discrimination across the entire network.

[0051] In the aforementioned cross-regional power data collaborative analysis system 100 based on privacy computing, the global data decryption and verification module 150 is used to receive the global logical matrix broadcast and distributed back by the central server, and to decrypt and verify the global logical matrix in conjunction with the local decryption private key to obtain a global plaintext soft label for model guidance. It should be noted that although the global logical matrix generated by the central server contains consensus knowledge of the entire network's power patterns, it still resides in the cloud and is in a quantized and compressed state. If local computing nodes want to use this consensus knowledge to correct their own judgment biases, they must securely obtain this data through an unreliable wide area network link and restore it to a floating-point probability distribution that the model can understand. Based on this, the technical solution of this application further distributes the global logical matrix as downlink data payload through a central server to obtain the received global logical matrix. The received global logical matrix is ​​then verified and unpacked using a pre-set key to obtain a verified global logical matrix. Finally, a quantized inverse mapping is performed on the verified global logical matrix to obtain a global plaintext soft tag. This constructs a secure data backhaul channel from the cloud to the edge, ensuring the authenticity of the source and the integrity of the guidance signal delivered locally. Through the above processing, data tampering caused by man-in-the-middle attacks or network transmission errors can be effectively prevented, and the compressed integer matrix is ​​losslessly restored to a high-precision probability distribution containing rich dark knowledge, providing accurate global supervision signals for subsequent high-fidelity knowledge distillation by the local model.

[0052] More specifically, in a specific example of this application, the global data decryption and verification module 150 includes: a data broadcasting and distribution unit, used to broadcast and distribute the global logical matrix as downlink data payload through a central server to obtain the received global logical matrix; an integrity authentication and unpacking unit, used to perform data integrity authentication and unpacking on the received global logical matrix based on a preset key to obtain the verified global logical matrix; and a quantization inverse mapping unit, used to perform quantization inverse mapping on the verified global logical matrix to obtain a global plaintext soft tag.

[0053] More specifically, the process first involves broadcasting and distributing data. Utilizing the high-throughput network interface of the central server, the aggregated global logical matrix is ​​encapsulated into standardized downlink communication messages. Employing IP multicast technology or a publish-subscribe message queue model, these messages are synchronously sent to all participating power data nodes distributed across different geographical regions, ensuring that each local node can obtain the latest network-wide consensus status almost in real time. Subsequently, integrity authentication and unpacking operations are performed. Upon detecting the downlink message, the local node does not directly read the content through the application layer. Instead, it first calls the cryptographic component to read the digital signature and hash digest fields in the message header. The signature is verified using the pre-distributed and stored public key certificate of the central server. The hash value of the message payload is recalculated and compared with the received digest. If they match perfectly, it confirms that the data has not been tampered with and was indeed sent by a legitimate central server. The transmission protocol header is then stripped to extract the internally encapsulated and verified global logical matrix. Finally, a quantization inverse mapping operation is performed. The local node performs a division operation on each integer element in the verified global logic matrix according to the quantization scaling factor agreed in the communication protocol or carried in the message metadata, inversely mapping the discrete integer values ​​back to the continuous floating-point probability space, thereby reconstructing a global plaintext soft tag that can delicately represent the network's logic for judging the waveform of the common anchor point.

[0054] In the aforementioned cross-regional power data collaborative analysis system 100 based on privacy computing, the knowledge distillation optimization module 160 is used to perform knowledge distillation closed-loop optimization on the pre-trained local deep network model based on local power cleaning data, common anchor waveform data, and global plaintext soft labels to obtain a regional optimized power model. It should be noted that although the returned global plaintext soft labels encapsulate the generalization consensus of the entire network's power patterns, blindly fitting this global knowledge to the local model can easily lead to catastrophic forgetting of local region-specific load characteristics, such as specific industrial impact patterns. Furthermore, continuing to rely solely on local data for training will fail to effectively absorb complementary experiences from external regions regarding rare faults, resulting in insufficient generalization ability of the model when facing unseen operating conditions. Based on this, the technical solution of this application further optimizes the pre-trained local deep network model by performing knowledge distillation closed-loop optimization based on local power cleaning data, common anchor waveform data, and global plaintext soft labels to obtain a regional optimized power model. Specifically, it constructs a hybrid objective function that includes cross-entropy and KL divergence constraints, uses local real labels to calculate task loss to lock the fidelity of local features, and uses global soft labels to calculate distillation loss to force the model's response distribution at common anchors to converge with the network consensus. This achieves a dynamic balance between maintaining local personalized features and absorbing global general knowledge at the mathematical level. Through the above processing, the dual optimization effect of local data parameter brushing and global knowledge logic repair can be effectively achieved, ensuring that the final generated regional optimized power model can accurately identify the bi-peak characteristics of local residential or commercial loads and has strong robustness to cope with extreme fault scenarios across the entire network, thereby solving the problem of collaborative failure caused by the non-independent and identically distributed cross-regional power data.

[0055] Figure 6 This is a block diagram of the knowledge distillation optimization module in a cross-regional power data collaborative analysis system based on privacy computing, according to an embodiment of this application. Figure 6 As shown, the knowledge distillation optimization module 160 includes: a local feature fidelity calculation unit 161, used to switch the pre-trained local deep network model back to training mode, and use real label samples in local power cleaning data to calculate the local feature fidelity of the pre-trained local deep network model to obtain the local task loss value; a global knowledge alignment calculation unit 162, used to calculate the global knowledge alignment of the pre-trained local deep network model based on common anchor waveform data and global plaintext soft labels to obtain the global distillation loss value; a loss weighted fusion unit 163, used to introduce a balancing hyperparameter to weightedly fuse the local task loss value and the global distillation loss value to obtain the overall objective function; and a model weight update unit 164, used to calculate the gradient based on the overall objective function using the stochastic gradient descent algorithm and update the model weights of the pre-trained local deep network model until the model converges to obtain the regional optimized power model.

[0056] In the aforementioned cross-regional power data collaborative analysis system 100 based on privacy computing, the local feature fidelity calculation unit 161 is used to switch the pre-trained local deep network model back to training mode, and to calculate the local feature fidelity of the pre-trained local deep network model using real label samples from local power cleaning data to obtain the local task loss value. It should be noted that although the returned global plaintext soft labels encapsulate the generalized consensus of the entire network's power patterns, they are essentially statistical averages of multi-source heterogeneous data. If the local model blindly and singularly fits this global signal in subsequent optimization, it is highly likely to cause a catastrophic forgetting of local region-specific load characteristics, such as specific industrial impact patterns, thereby leading to a decrease in the model's accuracy in discriminating local basic services. Based on this, the technical solution of this application further switches the pre-trained local deep network model back to training mode, and uses real labeled samples from local power cleaning data to calculate the local feature fidelity of the pre-trained local deep network model to obtain the local task loss value. This introduces a hard constraint mechanism of supervised learning, forcing the model to minimize the prediction error of local historical data while absorbing external complementary generalization knowledge through distillation. Through the above processing, a dynamic balance can be effectively achieved between maintaining local personalized features and integrating global general knowledge, ensuring that the model obtains network-wide robustness while strictly adhering to the diagnostic baseline for local specific fault types.

[0057] More specifically, in a concrete example of this application, the implementation process first performs a reverse switching operation of the model's running state. Through the control interface of the underlying computing framework, the pre-trained local deep network model is reset from the previous frozen inference state to an updatable training state, releasing the lock on the gradient of the weight parameters and restoring the dynamic statistical function of the randomly deactivated neuron layer and the batch normalization layer, thereby reconstructing the complete backpropagation computation graph. Subsequently, the fidelity verification and loss measurement operations of local features are performed. Training samples containing multi-dimensional time-series waveforms of voltage, current, and active power, along with manually diagnosed fault labels, are read from the local power cleaning dataset. These samples are input into the model for forward propagation to generate the class prediction probability in the current state. Then, the log-likelihood difference between the predicted probability distribution and the local true label is calculated using the cross-entropy loss function. By quantifying this difference, a local task loss value is generated. This loss value directly reflects the degree of fit of the model's current parameter distribution to the local inherent power characteristics, providing a key localized constraint term for the subsequent construction of the hybrid objective function.

[0058] In the aforementioned cross-regional power data collaborative analysis system 100 based on privacy computing, the global knowledge alignment calculation unit 162 is used to calculate the global knowledge alignment of the pre-trained local deep network model based on common anchor waveform data and global plaintext soft labels to obtain the global distillation loss value. It should be noted that, due to the inherent non-independent and identically distributed characteristics of power data in different regions caused by differences in geographical environment and industrial structure, relying solely on local data for closed-loop training leads to a lack of awareness of rare fault patterns occurring in different locations. Furthermore, simple hard label supervision cannot convey the potential correlation information between categories, i.e., dark knowledge, thus limiting the model's generalization ability when facing complex power grid conditions. Based on this, the technical solution of this application further calculates the global knowledge alignment of the pre-trained local deep network model based on local power cleaning data, common anchor waveform data, and global plaintext soft labels to obtain the global distillation loss value. This quantifies the cognitive bias between the local model's predicted distribution on the standard common anchor dataset and the weighted consensus distribution of the entire network, and uses this bias to construct a soft constraint signal to force the local model's discrimination logic to move closer to the global optimal solution. Through the above processing, the knowledge barriers caused by data silos can be effectively broken down. Without directly interacting with the original private data, the general discrimination experience of the entire network for specific power waveforms can be seamlessly transferred to the local model, thereby improving the regional model's ability to provide early warning and identification of unknown fault types.

[0059] More specifically, in a specific example of this application, the implementation process first performs re-inference and softening operations on the anchor data, calls the pre-set common anchor waveform data in memory and inputs it again into the local deep network model of the current training batch for forward propagation calculation, obtains the non-normalized log probability vector output by the model for the standard dataset, and reuses the distillation temperature coefficient determined in the previous step to perform division scaling and exponential normalization processing on the vector to generate the local anchor softening probability distribution in the current state. Subsequently, a difference metric calculation based on relative entropy is performed. The decrypted and verified global plaintext soft tag is read as the teacher signal. This soft tag carries the weighted aggregation of the network consensus probability distribution. The Kullback-Leibler divergence formula is used to calculate the asymmetric difference value between the global plaintext soft tag and the local anchor softening probability distribution. This difference value is multiplied by the square of the temperature coefficient to compensate for the gradient magnitude. Finally, the global distillation loss value is obtained by summing or averaging the difference values ​​of all anchor samples. This value accurately represents the degree of alignment between the local model and the global consensus in the feature space, providing a key global guiding gradient for subsequent multi-objective optimization.

[0060] In the aforementioned privacy-preserving computation-based cross-regional power data collaborative analysis system 100, the loss weighted fusion unit 163 and the model weight update unit 164 are used to introduce balanced hyperparameters to weightedly fuse the local task loss value and the global distillation loss value to obtain the overall objective function. Based on this overall objective function, the gradient is calculated using the stochastic gradient descent algorithm, and the model weights of the pre-trained local deep network model are updated until the model converges to obtain the regional optimized power model. It should be noted that since there is a natural non-convex competitive relationship between local feature fidelity and global knowledge alignment in the optimization direction, focusing solely on the local task loss will lead to the model falling into an overfitting trap and losing the ability to generalize and identify common cross-regional faults. On the other hand, over-reliance on the global distillation loss will smooth out the unique key load features of the local area due to the strong regularization effect. Based on this, the technical solution of this application further introduces a balancing hyperparameter to weight and fuse the local task loss value and the global distillation loss value to obtain the overall objective function. Based on this overall objective function, the gradient is calculated using the stochastic gradient descent algorithm, and the model weights of the pre-trained local deep network model are updated until the model converges to obtain a regional optimized power model. This mathematically constructs an adjustable joint optimization plane, precisely controlling the trade-off between retaining the ability to distinguish specific local industrial or residential power consumption patterns and absorbing general power fault logic experience from the entire network by adjusting the balancing hyperparameter. Through the above processing, catastrophic forgetting can be effectively avoided, ensuring that the generated regional optimized model can maintain high-precision diagnosis of existing local services while leveraging complementary knowledge from the entire network to improve its robustness against interference from unseen abnormal operating conditions.

[0061] More specifically, in a concrete example of this application, the implementation process first performs a weighted construction operation of a multi-objective loss function. A scalar with a value between zero and one is set as a balancing hyperparameter. This parameter is used to assign retention weights to the local task loss value and corresponding transfer weights to the global distillation loss value. The two are then weighted and summed to generate the overall objective function, which defines the comprehensive error potential surface of the model in the current iteration. Subsequently, gradient calculation and parameter update operations based on backpropagation are performed. The automatic differentiation engine is used to calculate the partial derivatives of the overall objective function with respect to the weight parameters of each layer of the model using the chain rule, generating a gradient vector pointing towards the direction of the minimum error. The stochastic gradient descent optimizer is then invoked to fine-tune the convolutional kernel weights and bias terms in the local deep network model along the negative gradient direction according to a preset learning rate, thereby driving the model parameter distribution to move towards the Pareto optimal solution. Finally, the iterative convergence determination and model solidification operation are performed. The change trend of the total objective function value is continuously monitored over multiple training cycles. Once the loss value decreases below the preset minimum threshold or the validation set accuracy no longer improves, the model is determined to have reached convergence. The training process is then terminated and the parameter structure at the current moment is solidified and saved, outputting a regional optimized power model with dual characteristics.

[0062] In summary, the privacy-preserving cross-regional power data collaborative analysis system based on the embodiments of this application is explained. It generates a logical matrix containing implicit features through inference of anchor point data using a local model, and uploads it after encryption with enhanced privacy, effectively avoiding the direct parameter averaging failure problem caused by heterogeneous data distribution. In the global aggregation stage, the system does not mechanically perform mean calculation, but dynamically adjusts the aggregation weights by performing a two-dimensional evaluation of the entropy weight confidence and positive complementarity of the feature space of the uploaded data from each region. This removes low-quality noise while retaining key load patterns with high specificity. The finally generated global soft label containing network-wide consensus is distributed back to the local system to guide the local model in closed-loop distillation optimization. Thus, while ensuring data privacy, it solves the problem of balancing universality and specificity in cross-regional collaboration, improving the robustness of the model under extreme conditions.

Claims

1. A cross-regional power data collaborative analysis system based on privacy computing, characterized in that, include: The local model pre-training module is used to perform the first round of pre-training on the initialized local deep network model based on local power cleaning data to obtain the pre-trained local deep network model. The local anchor inference module is used to input the acquired public anchor waveform data into the pre-trained local deep network model for forward inference to obtain the local anchor logic matrix. The privacy-enhanced encryption module is used to perform privacy-enhanced encryption on the local anchor point logic matrix to obtain encrypted logic data packets; The global knowledge aggregation module is used to control the central server to receive encrypted logical data packets from multiple different regions, and to perform global knowledge aggregation based on correlation on multiple encrypted logical data packets to obtain a global logical matrix containing the power pattern characteristics of the entire network. The global data decryption and verification module is used to receive the global logic matrix broadcast and distributed back by the central server, and to decrypt and verify the global logic matrix in combination with the local decryption private key to obtain the global plaintext soft tag used for model guidance. The knowledge distillation optimization module is used to perform knowledge distillation closed-loop optimization on the pre-trained local deep network model based on local power cleaning data, common anchor waveform data, and global plaintext soft labels to obtain a regional optimized power model.

2. The cross-regional power data collaborative analysis system based on privacy computing according to claim 1, characterized in that, The local raw power dataset includes multi-dimensional time-series features of voltage, current, and active power. The local model pre-training module includes: The multidimensional data cleaning and standardization unit is used to perform multidimensional power data cleaning and standardization on the local raw power dataset to obtain local power cleaned data. The local model pre-training unit is used to divide the local power cleaning data into training and validation sets and input them into the initialized local deep network model for forward propagation calculation. It also performs backpropagation update based on the deviation between the predicted value and the true label until the loss function converges, so as to obtain the pre-trained local deep network model.

3. The cross-regional power data collaborative analysis system based on privacy computing according to claim 1, characterized in that, The local anchor inference module includes: The mode switching and freeze unit is used to perform inference mode switching and gradient freezing on the pre-trained local deep network model to obtain the inference state local model. Anchor point forward mapping unit is used to input common anchor point waveform data into the local model of inference state in batches for forward mapping to obtain the original output vector of the model containing the original confidence score; The temperature scaling and softening unit is used to perform temperature scaling and probability distribution softening on the original output vector of the model based on a preset distillation temperature coefficient to obtain the local anchor logic matrix.

4. The cross-regional power data collaborative analysis system based on privacy computing according to claim 1, characterized in that, Privacy-enhancing encryption modules include: The trimming and quantization compression unit is used to trim and quantize the local anchor logic matrix to obtain the quantized logic matrix. The noise generation unit is used to calculate the noise intensity parameters based on a preset privacy budget, and generate a random noise matrix with the same dimension as the quantization logic matrix based on the noise intensity parameters. The noise superposition perturbation unit is used to superimpose the random noise matrix onto the quantization logic matrix to obtain the perturbation logic matrix; Metadata encapsulation unit is used to append metadata header information containing region ID and quantization factor to the perturbation logic matrix to obtain encrypted logic data packets.

5. The cross-regional power data collaborative analysis system based on privacy computing according to claim 1, characterized in that, The global knowledge aggregation module includes: The unpacking and verification unit is used to control the central server to unpack and verify the integrity of encrypted logical data packets from multiple regions to obtain a set of perturbation logic matrices. The dynamic weight estimation unit is used to perform correlation-based dynamic weight estimation on the set of perturbation logic matrices to obtain the regional aggregated weight vector; The global weighted aggregation unit is used to perform global weighted aggregation and noise cancellation on the set of perturbation logic matrices based on the regional aggregation weight vector to obtain the global logic matrix.

6. The cross-regional power data collaborative analysis system based on privacy computing according to claim 1, characterized in that, The global data decryption and verification module includes: The data broadcasting and distribution unit is used to broadcast and distribute the global logical matrix as downlink data payload through the central server in order to obtain the received global logical matrix. The integrity authentication and unpacking unit is used to perform data integrity authentication and unpacking on the received global logic matrix based on a preset key to obtain the verified global logic matrix; The quantization inverse mapping unit is used to perform quantization inverse mapping on the verified global logic matrix to obtain global plaintext soft tags.

7. The cross-regional power data collaborative analysis system based on privacy computing according to claim 1, characterized in that, The knowledge distillation optimization module includes: The local feature fidelity calculation unit is used to switch the pre-trained local deep network model back to training mode and use real label samples in the local power cleaning data to calculate the local feature fidelity of the pre-trained local deep network model to obtain the local task loss value. The global knowledge alignment calculation unit is used to calculate the global knowledge alignment of the pre-trained local deep network model based on common anchor waveform data and global plaintext soft labels to obtain the global distillation loss value. The loss-weighted fusion unit is used to introduce a balancing hyperparameter to weight and fuse the local task loss value and the global distillation loss value to obtain the overall objective function. The model weight update unit is used to calculate the gradient based on the overall objective function using the stochastic gradient descent algorithm and update the model weights of the pre-trained local deep network model until the model converges to obtain the regional optimized power model.

8. The cross-regional power data collaborative analysis system based on privacy computing according to claim 5, characterized in that, The dynamic weight estimation unit includes: The uncertainty assessment subunit is used to perform uncertainty assessment on the set of perturbation logic matrices to obtain the regional confidence vector. The orthogonal projection sub-unit is used to perform global orthogonal projection of the perturbation logic matrix set into the feature space to obtain the positive interactive complement score vector. The complementary fusion subunit is used to perform complementary fusion of the regional confidence vector and the positive interactive complementary score vector under confidence constraints to obtain the regional aggregate weight vector.

Images