Power network data security encryption method and system based on deep learning

Abstract

The invention relates to the technical field of network security, in particular to a power network data security encryption method and system based on deep learning. The method comprises the following steps: acquiring power network data, and performing structured processing to generate a standardized power data unit; constructing a power time sequence fragment according to the standardized power data unit; inputting the power time sequence fragment into a pre-constructed deep learning model, and outputting a high-dimensional feature vector; generating a dynamic encryption parameter by using the high-dimensional feature vector; performing hierarchical encryption processing on the power network data based on the dynamic encryption parameters, and recording encrypted power data; evaluating an integrity check state according to the encrypted power data; and when the integrity check state is abnormal check, isolating a communication link in the encrypted power data. According to the method, adaptive hierarchical encryption and security protection of the power network data are realized based on a network security technology, so that the security reliability and the tampering attack recognition rate of power network data encryption are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of network security technology, and in particular to a method and system for secure encryption of power network data based on deep learning. Background Technology

[0002] In existing power network data security protection and encryption technologies, static key configuration, fixed encryption algorithms, or rule-based access control mechanisms are typically used to protect power network data. For example, a preset symmetric key or public key system is used to encrypt the transmission of collected power data, or encryption strategies are pre-configured according to the communication link type and service level, and simple integrity checks are used to determine whether the data has been tampered with. These methods generally suffer from the following shortcomings: First, the encryption parameters are mostly static or periodically updated, making it difficult to reflect the characteristics of power system operation status, load changes, and dynamic adjustments in communication topology. The encryption mechanism is disconnected from the actual operating status of the power network, resulting in insufficient adaptive security protection capabilities. Second, existing methods mostly focus on data encryption and verification for a single session or link, lacking system modeling of communication behavior and link relationships in multi-device collaborative scenarios. They have limited ability to identify complex security threats caused by communication link hijacking, replay, or partial tampering, and are prone to missed detections or misjudgments. Third, data integrity verification and anomaly handling are usually independent of the encryption process. There is a lack of a mechanism to jointly analyze the integrity verification results with communication session, link status, and key usage, making it difficult to achieve accurate isolation and re-encryption of abnormal communication links and meet the application requirements of power networks under conditions of large-scale access, high real-time performance, and high security. Summary of the Invention

[0003] Therefore, the present invention needs to provide a method and system for secure encryption of power network data based on deep learning to solve at least one of the above-mentioned technical problems.

[0004] To achieve the above objectives, a deep learning-based method for secure encryption of power network data includes the following steps: Step S1: Acquire power network data, process it in a structured manner, and generate standardized power data units; construct power time series segments based on the standardized power data units; Step S2: Input the power time series segment into the pre-built deep learning model and output a high-dimensional feature vector; use the high-dimensional feature vector to generate dynamic encryption parameters; Step S3: Perform hierarchical encryption processing on the power network data based on dynamic encryption parameters, and record the encrypted power data; evaluate the integrity verification status based on the encrypted power data; Step S4: When the integrity verification status is abnormal, isolate the communication link in the encrypted power data; use the communication link to perform the re-encryption task.

[0005] Preferably, this specification also provides a deep learning-based power network data security encryption system for executing the above-described deep learning-based power network data security encryption method. The deep learning-based power network data security encryption system includes: The structured processing module is used to acquire power network data, process it in a structured manner, and generate standardized power data units; and to construct power time series segments based on the standardized power data units. The deep learning module is used to input power time-series segments into a pre-built deep learning model and output high-dimensional feature vectors; dynamic encryption parameters are generated using the high-dimensional feature vectors. The hierarchical encryption module is used to perform hierarchical encryption processing on power network data based on dynamic encryption parameters, record encrypted power data, and evaluate the integrity verification status based on the encrypted power data. The re-encryption module is used to isolate the communication link in the encrypted power data when the integrity verification status is abnormal; and to perform the re-encryption task using the communication link.

[0006] The beneficial effects of this invention are as follows: (1) By introducing a deep learning model based on power time series data, high-dimensional feature modeling of power network data in terms of time evolution, load change and communication behavior is carried out, and dynamic encryption parameters are generated accordingly. This enables the encryption process to be adaptively associated with the power system operation status and communication session characteristics, avoiding the problem of insufficient security of traditional static keys or fixed strategies in complex operating environments, and improving the dynamic adaptability and overall security strength of power network data encryption.

[0007] (2) In the process of hierarchical encryption and session management of power network data, a session-level encryption seed is generated based on the high-dimensional feature vector combined with the communication link identifier and the basic key identifier. The key is dynamically evolved in a fine-grained manner by introducing perturbation combination and time factor, so that the data encryption process under different communication links and different business scenarios has significant distinguishability, effectively reducing the risk of key reuse and associated attacks, and improving the security and anti-attack capability of power network data transmission in multi-device collaborative scenarios.

[0008] (3) In the data integrity verification and anomaly handling stage, by jointly analyzing the integrity verification results with the encrypted data, communication links and key usage status, the tampering behavior and abnormal communication sessions can be accurately identified. When an anomaly is detected, the corresponding communication link is isolated and re-encrypted, forming a closed-loop security mechanism that coordinates encryption, anti-tampering and anomaly handling, thereby improving the data security guarantee capability and engineering application reliability of the power network under large-scale access and high real-time operation conditions. Attached Figure Description

[0009] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a flowchart illustrating the steps of a deep learning-based power network data security encryption method according to the present invention. Figure 2 This is a schematic diagram illustrating the execution flow of a deep learning-based power network data security encryption method according to the present invention. Figure 3 This is a schematic diagram of the architecture of the deep learning model in this invention; The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0010] The technical method of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this invention.

[0011] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0012] It should be understood that although the terms "first," "second," etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are used merely to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0013] To achieve the above objectives, please refer to Figures 1 to 3 This invention provides a deep learning-based method for secure encryption of power network data, the method comprising the following steps: Step S1: Acquire power network data, process it in a structured manner, and generate standardized power data units; construct power time series segments based on the standardized power data units; In one embodiment, raw power network data is acquired through acquisition terminals deployed at power network nodes. This power network data includes, but is not limited to, voltage values, current values, power factor, frequency, load status identifiers, and communication timestamps. The acquired raw data undergoes structured processing, including field alignment, outlier removal, and missing value imputation, and is uniformly encapsulated into standardized power data units. Each standardized power data unit contains at least a data identifier field, a data content field, and a timestamp field. Furthermore, continuous standardized power data units are aggregated using a sliding aggregation method according to a preset time window length to construct power time series segments, used to characterize the operational status changes of the power network over continuous time periods.

[0014] In another embodiment, it is assumed that voltage, current, active power, and reactive power data are collected from the substation side with a sampling period of 10ms, and each standardized power data unit has an 8-dimensional data dimension. In structured processing, outlier sampling points are eliminated using the three-standard-deviation method, and missing data of no more than two consecutive sampling points are supplemented using linear interpolation. With a time window length of 1s and a step size of 0.5s, each power time series segment contains 100 standardized power data units, forming a time series input sample with 800-dimensional features, providing a unified data foundation for subsequent deep learning feature extraction.

[0015] Step S2: Input the power time series segment into the pre-built deep learning model and output a high-dimensional feature vector; use the high-dimensional feature vector to generate dynamic encryption parameters; In one embodiment, the constructed power time series segments are input into a pre-trained deep learning model. The deep learning model comprises a structure combining a multi-layer convolutional neural network (CNN) and a time series modeling network, used to automatically learn multi-scale time series correlation features and operating state features from the power time series segments. The deep learning model outputs a high-dimensional feature vector corresponding to each power time series segment. Based on this high-dimensional feature vector, dynamic encryption parameters are generated through a nonlinear mapping function. These dynamic encryption parameters include at least a dynamic key seed, a perturbation factor, and round control parameters.

[0016] In another embodiment, the deep learning model is assumed to consist of three one-dimensional convolutional layers (with kernel sizes of 3, 5, and 7, corresponding to 64, 128, and 256 channels, respectively) and one bidirectional LSTM layer (with 128 hidden units). After passing through a fully connected layer, a 512-dimensional high-dimensional feature vector is output. Based on this 512-dimensional feature vector, a dynamic key seed of 256 bits is generated using hash mapping and modular arithmetic. Simultaneously, a perturbation factor ε (ranging from 0.1 to 0.5) is generated based on the mean and variance of the feature vector, and the encryption round parameters are dynamically determined to be 8 to 16 rounds, achieving adaptive changes in encryption parameters according to the operating status of the power network.

[0017] Step S3: Perform hierarchical encryption processing on the power network data based on dynamic encryption parameters, and record the encrypted power data; evaluate the integrity verification status based on the encrypted power data; In one embodiment, power network data is subjected to hierarchical encryption based on dynamic encryption parameters. Data fields with different security levels are processed using encryption algorithms of different strengths. For example, core operational data is encrypted with high strength, while auxiliary status data is encrypted with lightweight encryption, generating corresponding encrypted power data. Integrity verification parameters are generated synchronously during the encryption process. Digest calculation is performed on the encrypted power data, and the integrity verification status is evaluated based on the digest comparison results to determine whether the data has been tampered with during transmission or storage.

[0018] In another embodiment, it is assumed that key fields such as voltage and current are encrypted using the AES-256 algorithm, and the status identifier field is encrypted using the AES-128 algorithm. A 256-bit checksum calculated based on the SHA-256 algorithm is appended to each encrypted power data packet. The receiving end recalculates the checksum and compares it with the appended checksum. If the proportion of discrepancies exceeds two out of five consecutive data packets, the integrity check status of the current encrypted power data is determined to be abnormal; otherwise, it is determined to be normal.

[0019] Of particular importance, step S3, which assesses the integrity verification status based on the encrypted power data, includes: Extract integrity verification parameters from encrypted power data; calculate verification digest based on integrity verification parameters; detect tampering attacks based on verification digest and generate tampering indication identifier; In one embodiment, after receiving encrypted power data, the system first parses and extracts integrity verification parameters from the data frame structure of the encrypted power data. These integrity verification parameters include a data block identifier, data length information, a timestamp field, and a corresponding encryption verification code field. Subsequently, based on the extracted integrity verification parameters, the system performs a hash operation on the encrypted power data according to a preset digest calculation rule to generate a verification digest uniquely corresponding to the current data frame. The system then compares the verification digest with the original verification value carried in the data frame. If the two are inconsistent, a potential tampering attack is identified, and a corresponding tampering indication identifier is generated to mark the abnormal data state.

[0020] In another embodiment, it is assumed that the encrypted power data is divided into 32 data blocks, each 1024 bytes in length. The integrity verification parameters include the block number, CRC32 checksum, and 64-bit timestamp. The system calculates a checksum digest for each data block using the SHA-256 algorithm, generating a 256-bit digest value, and compares it block by block with the original checksum. When the digest comparison results of three consecutive data blocks are inconsistent, the system determines that the data stream is at risk of tampering and generates a tampering indication flag Flag=1; if all data blocks are consistent, a tampering indication flag Flag=0 is generated for subsequent integrity status assessment.

[0021] The tampered data block is located based on the tampering indicator; the tampered data block is written into the tampering record table; and the integrity verification status of the encrypted power data is evaluated based on the tampering record table.

[0022] In one embodiment, after a tampering indication is generated, the system performs further analysis on the encrypted power data based on the tampering indication to locate the abnormal data block. For each data block determined to be abnormal, the system records its corresponding data block number, timestamp, and anomaly type information, and writes it into a preset tampering record table. Subsequently, based on the number, distribution location, and frequency of abnormal data blocks recorded in the tampering record table, the system comprehensively evaluates the overall integrity verification status of the current encrypted power data, thereby determining whether the data meets the conditions for trusted transmission or subsequent decryption processing.

[0023] In another embodiment, assuming the tampering record table is used to record the integrity detection results of the most recent 100 data blocks, it includes data block numbers, anomaly flags, and anomaly count statistics. If a total of 8 abnormal data blocks are detected within a detection cycle, and these abnormal blocks are concentrated between numbers 15 and 22, the system determines that there is concentrated tampering behavior in this interval and assesses the integrity verification status as "high risk." If the number of abnormal data blocks is less than 3 and they are discretely distributed, the integrity verification status is assessed as "low risk." This assessment result can serve as the basis for decisions regarding power data discarding, retransmission, or alarm triggering.

[0024] Step S4: When the integrity verification status is abnormal, isolate the communication link in the encrypted power data; use the communication link to perform the re-encryption task.

[0025] In one embodiment, when an integrity verification status of abnormality is detected, the control module immediately performs logical isolation on the corresponding communication link to block the abnormal data from continuing to propagate to the upstream node; at the same time, a re-encryption task is triggered to regenerate dynamic encryption parameters for the power network data in the isolated link and perform encryption processing. After confirming that the re-encrypted data passes the integrity verification, the data interaction of the communication link is restored.

[0026] In another embodiment, assuming a communication link detects three consecutive verification anomalies within 10 seconds, the system marks the communication link as a high-risk link and puts it into isolation. Subsequently, a dynamic key seed is regenerated based on the latest batch of power time-series segments, and the encryption rounds are increased to 16 rounds, re-encrypting the 200 cached power network data entries. After recalculating the checksum, if 20 consecutive data entries pass the checksum verification, the isolation state is lifted and communication is restored; otherwise, isolation is maintained and a security alarm is reported.

[0027] Of particular importance, in step S4, when the integrity verification status is abnormal, the communication links in the isolated encrypted power data include: Extract the message verification code based on the integrity verification status; if the message verification code differs from the preset message verification code by any number of bits, it is determined to be an abnormal communication session. In one embodiment, after completing the integrity verification of encrypted power data, the system extracts a message verification code field from the corresponding data session based on the obtained integrity verification status. The message verification code characterizes the consistency and trustworthiness of the current communication session during transmission. The system compares the extracted message verification code bit by bit with a preset message verification code template. When any inconsistency is detected, the current communication session is determined to be an abnormal communication session, and the session is marked as abnormal to prevent abnormal data from participating in subsequent business processing or decryption procedures, thereby improving the security of the power data communication process.

[0028] In another embodiment, assuming the message verification code uses a 128-bit HMAC value, the system receives message verification codes from 20 communication sessions within a detection cycle. The preset message verification code is generated by key K and stored in the security module. The system performs a bit-by-bit XOR comparison on the message verification code for each session. When at least one bit difference is detected in a session, that session is determined to be an abnormal communication session. For example, if session 7 has a difference between bits 45 and 89, then that session is marked as abnormal; the remaining 19 sessions all match successfully and are determined to be normal communication sessions for subsequent normal data processing.

[0029] Extract the abnormal communication link identifier from the abnormal communication session; generate a link isolation instruction based on the abnormal communication link identifier; and use the link isolation instruction to isolate the communication link in the encrypted power data.

[0030] In one embodiment, upon detecting an abnormal communication session, the system extracts the corresponding abnormal communication link identifier based on the source node identifier, destination node identifier, and link path information carried in the abnormal communication session. The system generates a link isolation command based on the abnormal communication link identifier and sends the command to the network control module to perform logical or physical isolation operations on the abnormal communication link, thereby blocking the continued transmission of encrypted power data in the abnormal link and preventing the abnormal communication from affecting the overall operational security of the power network.

[0031] In another embodiment, it is assumed that the abnormal communication link identifier consists of "source node ID - relay node ID - destination node ID", for example, L12 = (N03 – N08 – N15). When the system detects that three abnormal communication sessions are all associated with link L12, it generates a corresponding link isolation command, sets the forwarding priority of the link to 0, and closes the corresponding port connection at the switching node, with the isolation duration set to 300 seconds. Meanwhile, the remaining communication links not involved in the abnormality continue to operate normally, thereby achieving precise isolation of abnormal links and ensuring the security of power data communication.

[0032] It should be noted that you should refer to [link / reference]. Figure 2 Starting with data acquisition and preprocessing, the power network data is first processed into standardized data units and power time series segments are generated. Then, a 1024-dimensional high-dimensional feature vector is extracted using a deep learning model that includes a time series coding layer and a feature aggregation layer, and dynamic encryption parameters are generated. Next, based on the power impact range, symmetric encryption of a single device or hybrid encryption of multiple devices is performed on the data to generate encrypted power data and perform integrity verification. When the verification is abnormal, the abnormal handling and re-encryption process is triggered, the communication link of the abnormal data is isolated and the re-encryption operation is performed, and finally the secure transmission and storage of power data is guaranteed.

[0033] It should be noted that you should refer to [link / reference]. Figure 3This is a deep learning model architecture diagram, illustrating the complete process from input to output: The input data contains power time series segments (batch size 40, sequence length 1000, dimension 40×1000×1). After outlier removal, missing value compensation, and normalization to [0,1], it enters the temporal coding layer, which contains three 1D convolutional layers (convolutional layer 1: kernel 3, channels 32 using ReLU and stride 2; convolutional layer 2: kernel 5, channels 64 using ReLU and stride 2; convolutional layer 3: kernel 7, channels 128 using ReLU and stride 2), and then enters the feature aggregation layer, which contains three different pooling operations (pooling... Pooling 1: Window 2, short scale; Pooling 2: Window 4, medium scale; Pooling 3: Window 8, long scale). Channel dimensions are concatenated through feature concatenation, and then weighted fusion is performed to obtain the load aggregation feature (256 dimensions). Then it enters the feature mapping layer, which contains two fully connected layers (fully connected layer 1: 256→512 using Dropout (p=0.3) and BatchNorm; fully connected layer 2: 512→1024 using Dropout (p=0.3) and BatchNorm). Finally, a high-dimensional feature vector of 1024 dimensions is output as the feature representation of the model output.

[0034] Preferably, in step S2, inputting the power time series segment into a pre-built deep learning model and outputting a high-dimensional feature vector includes: The power time series segments are input into a pre-built deep learning model, which includes a time series encoding layer, a feature aggregation layer, and a feature mapping layer. In one embodiment, continuous power operation data is acquired from a power network acquisition system, and the power operation data is sliced ​​according to a preset time window to form power time series segments. After outlier removal, missing value compensation, and amplitude normalization are performed on the power time series segments, they are used as input data for the model. The processed power time series segments are input into a pre-constructed deep learning model, which includes a time series encoding layer, a feature aggregation layer, and a feature mapping layer, for hierarchical feature modeling of power load variation patterns, thereby providing a unified data input foundation for subsequent feature extraction and analysis.

[0035] In another embodiment, assuming the original power data sampling frequency is 200Hz and the time window length is 5s, each power time series segment contains 1000 sampling points; 40 consecutive power time series segments are extracted to form an input batch with a data dimension of 40×1000×1. Further assuming that each sampling point is subjected to min-max normalization, mapping the amplitude to the [0,1] interval, the normalized batch data is then input into the deep learning model to ensure that data from different time periods and different power nodes have a consistent numerical scale during the model input stage.

[0036] The temporal coding layer is used to perform temporal convolution on power time-series segments to generate temporal features; In one embodiment, the temporal coding layer is used to perform temporal convolution operations on the input power time series segments. By sliding multiple one-dimensional convolution kernels along the time axis, local dynamic features of power load changes over time are extracted. At the same time, a nonlinear activation function is combined to enhance the features of the convolution results, thereby encoding the original power time series segments into a temporal feature representation that can reflect load fluctuation trends, abrupt change characteristics and periodic characteristics, providing basic input for subsequent feature aggregation.

[0037] In another embodiment, the temporal coding layer is assumed to consist of three one-dimensional convolutional networks, where the first convolutional layer has a kernel size of 3 and 32 channels, the second convolutional layer has a kernel size of 5 and 64 channels, and the third convolutional layer has a kernel size of 7 and 128 channels. Each convolutional layer is followed by a ReLU activation function, and a downsampling operation with a stride of 2 is performed. For an input temporal segment of length 1000, after the three temporal convolutional layers, a temporal feature matrix of approximately length 125 and feature dimension 128 is finally obtained.

[0038] The feature aggregation layer is used to perform multi-scale aggregation and feature fusion of temporal features to generate load aggregation features; In one embodiment, the feature aggregation layer is used to perform multi-scale aggregation processing on the time-series features output by the time-series coding layer. By fusing the feature responses at different time scales, the overall characteristics of power load in short-term fluctuations and medium-to-long-term changes are comprehensively characterized. The feature aggregation layer can uniformly represent the dispersed time-series features and generate stable and discriminative load aggregation features to reduce the information bias caused by single time-scale features.

[0039] In another embodiment, it is assumed that the feature aggregation layer uses three different scale pooling operations with pooling window sizes of 2, 4 and 8, respectively, to process the temporal feature matrix in parallel and obtain three sets of feature vectors with different time resolutions. The three sets of feature vectors are then concatenated according to the channel dimension to form a fused feature vector with a dimension of 256. The contribution ratio of each scale feature is adjusted by weighted fusion, and the load aggregation feature is finally output.

[0040] The feature mapping layer is used to map the aggregated features of the load to a high-dimensional feature space of a preset dimension, and output a high-dimensional feature vector.

[0041] In one embodiment, the feature mapping layer is used to map load aggregation features to a high-dimensional feature space of a preset dimension. The nonlinear mapping enhances the distinguishability and expressive power between features, making different power operation states more separable in the high-dimensional space. The high-dimensional feature vector can be used as the basic feature representation for subsequent encryption parameter generation, anomaly detection, or security analysis.

[0042] In another embodiment, the feature mapping layer is assumed to consist of two fully connected network layers. The first layer maps the 256-dimensional load aggregated features to a 512-dimensional feature space, and the second layer further maps them to a 1024-dimensional high-dimensional feature space. Batch Normalization and Dropout with a ratio of 0.3 are introduced between the fully connected layers to improve model stability. The final output 1024-dimensional high-dimensional feature vector serves as the feature representation of this power time series segment.

[0043] Preferably, the feature aggregation layer is used to perform multi-scale aggregation and feature fusion of temporal features to generate load aggregation features, including: The feature aggregation layer is used to extract multi-scale subsequences of temporal features under different time windows; the multi-scale subsequences are then hierarchically segmented to generate multi-scale temporal feature groups. In one embodiment, the feature aggregation layer first slides and truncates the input power time-series features according to different time window lengths to form multi-scale time-series subsequences covering short-term fluctuations, medium-term changes, and long-term trends. Then, the multi-scale subsequences are hierarchically segmented according to time scales, ensuring that the time-series features corresponding to each time scale remain structurally relatively independent, thereby generating multi-scale time-series feature groups. In this way, both transient change features and overall evolutionary features of power load can be preserved within the same model framework, providing a foundational input for subsequent cross-scale feature modeling.

[0044] In another embodiment, assuming the input time-series feature length is 1024, and the time window lengths are set to 16, 32, 64, and 128, corresponding to step sizes of 8, 16, 32, and 64, respectively, 128, 64, 32, and 16 subsequences can be obtained. The subsequences are further segmented hierarchically according to the window length, forming four groups of multi-scale time-series feature groups. Each group of subsequences has dimensions of 16×C, 32×C, 64×C, and 128×C, where C is the number of channels, used to characterize load variation features at different time scales.

[0045] One-dimensional dilated convolution is applied to the multi-scale temporal feature set to capture local load features; cross-scale correlation weights are calculated based on the local load features; weighted fusion is performed using the cross-scale correlation weights to generate scale fusion features; In one embodiment, a one-dimensional dilated convolution operation is applied to each of the generated multi-scale time series feature groups to capture the change pattern of power load in a local time range by expanding the receptive field of the convolution. Based on this, cross-scale correlation weights are calculated according to the response similarity and change trend between features of different scales, and the correlation weights are used to perform weighted fusion of features of each scale to generate scale fusion features that take into account both local details and overall trends.

[0046] In another embodiment, it is assumed that one-dimensional convolutions with dilation rates of 1, 2, 4, and 8 are applied to four sets of multi-scale features, with a uniform kernel size of 3 and 64 output channels for each set. Then, relevance weights are calculated based on the global average response of each scale feature, resulting in a weight vector such as [0.15, 0.25, 0.35, 0.25]. These weights are then used to perform a weighted summation of the four sets of features, ultimately generating a scale fusion feature with dimension 64, used to uniformly represent load information at different time scales.

[0047] Channel-level splicing is performed based on scale fusion features, and aggregated feature vectors are recorded; load aggregation features are determined based on aggregated feature vectors.

[0048] In one embodiment, based on the distribution characteristics of scale fusion features in the channel dimension, a channel-level splicing operation is performed on the fused multi-scale features to form a unified aggregated feature vector; then, based on the aggregated feature vector, the overall power load change state is characterized to determine the load aggregated features that can reflect the current power operation characteristics, thereby providing highly stable feature inputs for subsequent encryption parameter generation or security analysis.

[0049] In another embodiment, assuming the scale fusion feature is composed of four sets of features, each with a dimension of 64, the channel-level concatenation results in an aggregated feature vector with a dimension of 256. Further linear transformation and normalization are then applied to map this aggregated feature vector to a 128-dimensional load aggregated feature. This load aggregated feature can simultaneously reflect short-term disturbances, medium-term load changes, and long-term operating trends, making it suitable as a core feature representation for power network data security processing.

[0050] Preferably, step S2, which uses high-dimensional feature vectors to generate dynamic encryption parameters, includes: Power state feature codes are generated based on high-dimensional feature vectors; target base key identifiers are determined using preset key indexing rules and power state feature codes. In one embodiment, the system first extracts a high-dimensional feature vector reflecting the power operating status from encrypted power data. This high-dimensional feature vector includes at least voltage amplitude features, phase offset features, load fluctuation features, and sampling time series features. The high-dimensional feature vector is then normalized and concatenated to generate a power status feature code. Subsequently, according to a pre-defined key indexing rule, the power status feature code is mapped to a key index space. Index matching is used to determine the target base key identifier corresponding to the current power status, thereby achieving dynamic binding between the base key and the power operating status.

[0051] In another embodiment, assuming the high-dimensional feature vector has 256 dimensions, with voltage features occupying 64 dimensions, current features occupying 64 dimensions, power factor features occupying 32 dimensions, and the remainder being time-series statistical features, a 256-bit power state feature code is generated after normalization. The key indexing rule is set to use the first 12 bits of the encoded hash value as the index field, allowing for rapid location of the target basic key identifier among 4096 basic key identifiers. For example, when the index value is "0xA3F", the system selects the corresponding basic key number as a candidate basic key for the current session.

[0052] The communication link identifier is extracted from the high-dimensional feature vector; perturbation combination is performed based on the communication link identifier and the target basic key identifier to generate a session-level encryption seed; dynamic encryption parameters are generated based on the session-level encryption seed.

[0053] In one embodiment, the system further extracts a communication link identifier based on the high-dimensional feature vector. This communication link identifier characterizes the link features of power data during transmission, including source node identifier, destination node identifier, transmission delay characteristics, and link stability characteristics. The communication link identifier is then perturbed and combined with the target base key identifier to generate a session-level encryption seed. This perturbation combination includes XOR operations, shift transformations, or nonlinear mapping operations. Encryption parameters are dynamically generated based on the session-level encryption seed for data encryption processing in the current communication session.

[0054] In another embodiment, it is assumed that the communication link identifier consists of a source node ID (32 bits), a destination node ID (32 bits), and a link delay feature (16 bits), with a total length of 80 bits. This communication link identifier is XORed with the target base key identifier and then left-shifted by 3 bits to generate a 128-bit session-level encryption seed. Based on this encryption seed, dynamic encryption parameters are generated using a linear congruential algorithm, including an initialization vector IV (96 bits) and a round key sequence (10 rounds, 128 bits per round), to achieve session-level, link-aware dynamic encryption control.

[0055] Preferably, the process of generating a session-level encryption seed by performing a perturbation combination based on the communication link identifier and the target base key identifier includes: The communication link identifier and the target basic key identifier are concatenated at the bit level to generate an identifier combination sequence; the clock count register value is obtained; perturbation processing is performed based on the identifier combination sequence and the clock count register value to generate a perturbation result; Perform truncation or expansion processing based on the perturbation results to determine the fixed session length; generate a session-level encryption seed based on the fixed session length.

[0056] Preferably, perturbation processing is performed based on the identifier combination sequence and the clock count register value, and the perturbation result includes: A time perturbation factor is generated based on the clock count register value; the time perturbation factor is used to perform a sequence rearrangement on the identifier combination sequence to generate a rearranged identifier sequence; a nonlinear mapping is performed based on the rearranged identifier sequence to form an initial perturbation sequence; The offset amplitude parameter is determined based on the clock count register value; the initial perturbation sequence is shifted using the offset amplitude parameter, and out-of-range sequence units are recorded; the out-of-range sequence units are wrapped around to generate the perturbation result.

[0057] Preferably, in step S3, hierarchical encryption processing is performed on the power network data based on dynamic encryption parameters, and the encrypted power data is recorded as follows: The power impact range is calculated based on power network data; if the power impact range is a single device, then symmetric encryption is performed on the power network data using dynamic encryption parameters, and the symmetric encryption parameters are recorded. In one embodiment, the system first obtains the communication link identifier corresponding to the current power data communication session and the identified target basic key identifier, and then concatenates them sequentially according to a preset bit-level concatenation rule to generate an identifier combination sequence for session perturbation. Subsequently, it reads a real-time updated clock count register value from an encryption device or security chip; this clock count register value reflects the time state of the current session. The identifier combination sequence and the clock count register value are used together as perturbation input, and perturbation processing is performed through XOR, cyclic shift, or nonlinear mapping to obtain a perturbation result associated with the current communication link and time state, thereby enhancing session randomness.

[0058] In another embodiment, assuming the communication link identifier is 64 bits long and the target base key identifier is 128 bits long, the system performs bit-level concatenation according to the rule of "link identifier first, key identifier last," generating a 192-bit identifier combination sequence. The clock count register value is assumed to be a 32-bit incrementing count value, for example, the current value is 0x00AF23C1. The system expands and aligns the 192-bit identifier combination sequence with the 32-bit clock count register value, performs a bitwise XOR operation, and then performs a left circular shift of 5 bits on the result, finally generating a 192-bit perturbation result for subsequent session-level encryption parameter generation.

[0059] If the power impact involves multiple devices working together, then the power network data is encrypted using dynamic encryption parameters, and the hybrid encryption parameters are recorded; the symmetric encryption parameters and the hybrid encryption parameters are then encapsulated into encrypted power data.

[0060] In one embodiment, after obtaining the perturbation result, the system performs length normalization processing on the perturbation result according to preset session security level requirements. When the length of the perturbation result is greater than the target session length, its high or low bits are truncated; when the length of the perturbation result is less than the target session length, it is expanded through repeated padding, hash expansion, or pseudo-random padding to determine a uniform fixed session length. Based on the data sequence of this fixed session length, a session-level encryption seed corresponding to the current communication session is generated for use in subsequent dynamic encryption parameter or round key derivation.

[0061] In another embodiment, assuming the system sets the fixed session length to 128 bits, and the generated perturbation result is 192 bits long, the system truncates the middle 128 bits according to a preset rule as the fixed session length data; if the perturbation result is only 96 bits, it is padded to 128 bits through a single SHA expansion. The final generated 128-bit data serves as the session-level encryption seed, for example, represented as... This session-level encryption seed remains fixed within a single communication session and expires after the session ends, thereby achieving session isolation and enhanced security of encryption parameters.

[0062] Preferably, symmetric encryption is performed on power network data using dynamic encryption parameters, and the symmetric encryption parameters include: The system generates a wheel key using dynamic encryption parameters; performs group processing on power network data to generate several power network data blocks; and performs iterative symmetric encryption operations on the several power network data blocks using the wheel key to generate encrypted data blocks. In one embodiment, the system first generates a wheel key corresponding to the current communication session based on the dynamic encryption parameters generated in the previous stage, using a key derivation function. The wheel key is used to limit the key input for single-round or multi-round symmetric encryption operations. Subsequently, the power network data to be processed is grouped according to a preset data block length, forming several power network data blocks to adapt to the block encryption mechanism. For each data block, the system uses the wheel key as input to perform multi-round iterative symmetric encryption operations, ensuring that different data blocks within the same session have consistent security constraints during the encryption process, thereby generating a corresponding encrypted data block.

[0063] In another embodiment, assuming the dynamic encryption parameter length is 256 bits, a 128-bit wheel key is generated by deriving from HKDF. The total length of the power network data stream is 8192 bytes. The system groups the data into 32 blocks of 256 bytes each. For each data block, the AES algorithm is used to perform 10 rounds of iterative symmetric encryption, with each round using a round key. The round function calculations ultimately generate 32 encrypted data blocks. This method ensures consistent encryption of different data blocks under the same round key constraint.

[0064] The wheel key identifier is identified based on the encrypted data block; the key usage status is determined based on the wheel key identifier, and a data block encryption index is generated based on the key usage status; symmetric encryption parameters are generated based on the data block encryption index.

[0065] In one embodiment, after obtaining the encrypted data block, the system extracts a wheel key identifier corresponding to the wheel key from the additional fields or implicit structure of each encrypted data block to identify the key instance used by the current data block. Based on the wheel key identifier, the system determines the usage status of the corresponding wheel key, including whether it is being used for the first time, whether it has reached its usage limit, or whether it is about to expire. Subsequently, a corresponding data block encryption index is generated for each encrypted data block according to the key usage status, and the symmetric encryption parameters are dynamically adjusted based on the index to achieve linkage control between the key usage status and the encryption parameters.

[0066] In another embodiment, it is assumed that an 8-bit wheel key identifier is embedded in each encrypted data block. This is used to identify the current wheel key instance. The system statistics for this... The system classifies usage counts as follows: less than 100 uses are considered "normal usage"; 100 to 120 uses are considered "critical usage". For each of these states, the system generates different data block encryption index values, for example... , Based on this, the initial vector of the round function and the round shift factor in the symmetric encryption parameters are adjusted to improve the dynamism and security of the encryption process.

[0067] Preferably, hybrid encryption is performed on power network data using dynamic encryption parameters, and the recorded hybrid encryption parameters include: Generate session-level symmetric and asymmetric keys based on dynamic encryption parameters; perform symmetric encryption on power network data based on the session-level symmetric key to generate key ciphertext; In one embodiment, after acquiring dynamic encryption parameters, the system first generates a session-level symmetric key and a session-level asymmetric key corresponding to the current communication session using a key derivation algorithm. The session-level symmetric key is used for high-speed encryption of power network data, while the session-level asymmetric key is used for subsequent key protection and encapsulation operations. Subsequently, the system uses the session-level symmetric key as encryption input to perform symmetric encryption operations on the power network data to be sent, obtaining the key ciphertext bound to that session. By binding power network data to the session-level key, cross-session data replay or illegal reuse can be effectively prevented, improving the confidentiality and integrity of power data transmission.

[0068] In another embodiment, assuming the dynamic encryption parameter length is 256 bits, the system generates a session-level symmetric key of length 128 bits based on this parameter. and a 2048-bit session-level asymmetric key pair (public key) Private key The total length of the power network data is 4096 bytes. The AES algorithm is used to perform block symmetric encryption on the data, with each block being 256 bytes long, generating 16 key ciphertexts. Each key ciphertext is uniquely associated with its corresponding session and becomes invalid after the session ends, thereby reducing the security risks associated with long-term key leakage.

[0069] Asymmetric encryption of power network data is performed using a session-level asymmetric key to generate ciphertext; the key ciphertext and key ciphertext are encapsulated as hybrid encryption parameters.

[0070] In one embodiment, after generating the key ciphertext, the system further utilizes the session-level asymmetric key to perform asymmetric encryption on the power network data or its key digest information, generating corresponding data ciphertext. Subsequently, the system encapsulates the key ciphertext and the data ciphertext together to form a hybrid encryption parameter for secure transmission within the power network. By introducing a hybrid encryption mechanism combining symmetric and asymmetric encryption, the system ensures both the processing efficiency of large-scale power data transmission and enhances the security level during key distribution and data protection.

[0071] In another embodiment, assuming the system uses the RSA algorithm as a session-level asymmetric encryption method, utilizing a public key... The hash digest (256 bits) of the power network data is encrypted to generate a 2048-bit ciphertext. Sixteen sets of key ciphertext are combined with this ciphertext according to a preset encapsulation format to form a hybrid encryption parameter structure, which includes a key identifier, encryption algorithm identifier, and timestamp information. This hybrid encryption parameter can be transmitted along with the power network data for rapid decryption and verification at the receiving end.

[0072] Preferably, this specification also provides a deep learning-based power network data security encryption system for executing the above-described deep learning-based power network data security encryption method. The deep learning-based power network data security encryption system includes: The structured processing module is used to acquire power network data, process it in a structured manner, and generate standardized power data units; and to construct power time series segments based on the standardized power data units. The deep learning module is used to input power time-series segments into a pre-built deep learning model and output high-dimensional feature vectors; dynamic encryption parameters are generated using the high-dimensional feature vectors. The hierarchical encryption module is used to perform hierarchical encryption processing on power network data based on dynamic encryption parameters, record encrypted power data, and evaluate the integrity verification status based on the encrypted power data. The re-encryption module is used to isolate the communication link in the encrypted power data when the integrity verification status is abnormal; and to perform the re-encryption task using the communication link.

[0073] Therefore, the embodiments should be considered as exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of the equivalents of the application are intended to be included within the invention.

[0074] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A deep learning-based method for secure encryption of power network data, characterized in that, Includes the following steps: Step S1: Acquire power network data, process it in a structured manner, and generate standardized power data units; Power time-series segments are constructed based on standardized power data units; Step S2: Input the power time series segment into the pre-built deep learning model and output a high-dimensional feature vector; use the high-dimensional feature vector to generate dynamic encryption parameters; Step S3: Perform hierarchical encryption processing on the power network data based on dynamic encryption parameters, and record the encrypted power data; The integrity verification status is assessed based on the encrypted power data. Step S4: When the integrity verification status is abnormal, isolate the communication link in the encrypted power data; use the communication link to perform the re-encryption task.

2. The deep learning-based power network data security encryption method according to claim 1, characterized in that, In step S2, the power time series segments are input into a pre-built deep learning model, and the high-dimensional feature vectors output include: The power time series segments are input into a pre-built deep learning model, which includes a time series encoding layer, a feature aggregation layer, and a feature mapping layer. The temporal coding layer is used to perform temporal convolution on power time-series segments to generate temporal features; The feature aggregation layer is used to perform multi-scale aggregation and feature fusion of temporal features to generate load aggregation features; The feature mapping layer is used to map the aggregated features of the load to a high-dimensional feature space of a preset dimension, and output a high-dimensional feature vector.

3. The deep learning-based power network data security encryption method according to claim 2, characterized in that, The feature aggregation layer is used to perform multi-scale aggregation and feature fusion of temporal features, generating load aggregation features including: The feature aggregation layer is used to extract multi-scale subsequences of temporal features under different time windows; the multi-scale subsequences are then hierarchically segmented to generate multi-scale temporal feature groups. One-dimensional dilated convolution is applied to the multi-scale temporal feature set to capture local load features; cross-scale correlation weights are calculated based on the local load features; weighted fusion is performed using the cross-scale correlation weights to generate scale fusion features; Channel-level splicing is performed based on scale fusion features, and aggregated feature vectors are recorded; load aggregation features are determined based on aggregated feature vectors.

4. The deep learning-based power network data security encryption method according to claim 1, characterized in that, Step S2, which uses high-dimensional feature vectors to generate dynamic encryption parameters, includes: Power state feature codes are generated based on high-dimensional feature vectors; target base key identifiers are determined using preset key indexing rules and power state feature codes. The communication link identifier is extracted from the high-dimensional feature vector; perturbation combination is performed based on the communication link identifier and the target basic key identifier to generate a session-level encryption seed; dynamic encryption parameters are generated based on the session-level encryption seed.

5. The deep learning-based power network data security encryption method according to claim 4, characterized in that, Based on the communication link identifier and the target base key identifier, a perturbation combination is performed to generate a session-level encryption seed, including: The communication link identifier and the target basic key identifier are concatenated bit-wise to generate an identifier combination sequence; the clock count register value is obtained; perturbation processing is performed based on the identifier combination sequence and the clock count register value to generate a perturbation result; Perform truncation or expansion processing based on the perturbation results to determine the fixed session length; generate a session-level encryption seed based on the fixed session length.

6. The deep learning-based power network data security encryption method according to claim 5, characterized in that, Perturbation processing is performed based on the identifier combination sequence and the clock count register value, generating the following perturbation results: A time perturbation factor is generated based on the clock count register value; the time perturbation factor is used to perform a sequence rearrangement on the identifier combination sequence to generate a rearranged identifier sequence; a nonlinear mapping is performed based on the rearranged identifier sequence to form an initial perturbation sequence; The offset amplitude parameter is determined based on the clock count register value; the initial perturbation sequence is shifted using the offset amplitude parameter, and out-of-range sequence units are recorded; the out-of-range sequence units are wrapped around to generate the perturbation result.

7. The deep learning-based power network data security encryption method according to claim 1, characterized in that, In step S3, hierarchical encryption processing is performed on the power network data based on dynamic encryption parameters, and the encrypted power data is recorded as follows: The power impact range is calculated based on power network data; if the power impact range is a single device, then symmetric encryption is performed on the power network data using dynamic encryption parameters, and the symmetric encryption parameters are recorded. If the power impact involves multiple devices working together, then the power network data is encrypted using dynamic encryption parameters, and the hybrid encryption parameters are recorded; the symmetric encryption parameters and the hybrid encryption parameters are then encapsulated into encrypted power data.

8. The deep learning-based power network data security encryption method according to claim 7, characterized in that, Symmetric encryption is performed on power network data using dynamic encryption parameters. The symmetric encryption parameters include: The system generates a wheel key using dynamic encryption parameters; performs group processing on power network data to generate several power network data blocks; and performs iterative symmetric encryption operations on the several power network data blocks using the wheel key to generate encrypted data blocks. The wheel key identifier is identified based on the encrypted data block; the key usage status is determined based on the wheel key identifier, and a data block encryption index is generated based on the key usage status; symmetric encryption parameters are generated based on the data block encryption index.

9. The deep learning-based power network data security encryption method according to claim 7, characterized in that, Hybrid encryption is performed on power network data using dynamic encryption parameters. The recorded hybrid encryption parameters include: Generate session-level symmetric and asymmetric keys based on dynamic encryption parameters; perform symmetric encryption on power network data using the session-level symmetric key to generate key ciphertext; Asymmetric encryption of power network data is performed using a session-level asymmetric key to generate ciphertext; the key ciphertext and the key ciphertext are encapsulated as hybrid encryption parameters.

10. A deep learning-based power grid data security encryption system, characterized in that, The deep learning-based power grid data security encryption method for performing the deep learning-based power grid data security encryption system as described in claim 1 includes: The structured processing module is used to acquire power network data, process it in a structured manner, and generate standardized power data units; and to construct power time series segments based on the standardized power data units. The deep learning module is used to input power time-series segments into a pre-built deep learning model and output high-dimensional feature vectors; dynamic encryption parameters are generated using the high-dimensional feature vectors. The hierarchical encryption module is used to perform hierarchical encryption processing on power network data based on dynamic encryption parameters, record encrypted power data, and evaluate the integrity verification status based on the encrypted power data. The re-encryption module is used to isolate the communication link in the encrypted power data when the integrity verification status is abnormal; and to perform the re-encryption task using the communication link.

Images