AI-driven key allocation optimization and fault warning system in quantum secure communication network

Abstract

The application belongs to the technical field of quantum secure communication of power systems, and discloses an AI-driven key distribution optimization and fault early warning system in a quantum secure communication network, which comprises a data acquisition module, a data preprocessing module, an AI prediction engine, a key pool dynamic scheduling module and a cooperative control interface; the data acquisition module is used for acquiring key consumption data, service traffic data and key pool state data; the data preprocessing module is used for preprocessing the collected data to construct a sample set; the AI prediction engine is internally provided with a time series prediction network and is used for predicting the key consumption in a future time window according to the preprocessed data and the current key pool inventory; the key pool dynamic scheduling module is used for dynamically generating a key distribution strategy according to the predicted key consumption in the future time window and the current key pool inventory, and triggering a key shortage early warning when a warning condition is met; and the cooperative control interface is used for issuing the key distribution strategy to a quantum key management system. In this way, the key resource utilization efficiency and the early warning accuracy are improved.

Description

Technical Field

[0001] This invention relates to the field of quantum secure communication technology in power systems, specifically to an AI-driven key distribution optimization and fault early warning system in a quantum secure communication network. Background Technology

[0002] With the development of the energy internet and smart grids, quantum secure communication technology has been gradually applied to core business operations such as power dispatch automation and source-grid-load coordinated control. In practical applications of power systems, quantum secure communication networks typically adopt a centralized key pool architecture. This involves generating quantum keys across multiple sites using quantum key generation terminals, storing them in a centralized key management system to form a key pool, and then having each business system request keys from the key pool for encrypted communication as needed. In current engineering practices, the key pool capacity is usually configured based on the number of links, peak traffic volume, and key update cycle, typically ranging from tens of megabits to gigabits in cache size. Some demonstration projects use megabit-level capacity to reduce deployment costs.

[0003] In actual operation, the existing system has the following technical problems:

[0004] First, the key scheduling strategy is statically fixed. Current key allocation in networks is typically based on fixed thresholds or simple polling mechanisms, failing to detect real-time changes in upper-layer power service traffic. For example, during grid failures or major events, dispatch automation service traffic can surge instantaneously, leading to a sharp increase in key consumption. However, existing systems cannot predict this in advance, potentially causing key pool depletion and forcing services to degrade or switch to backup traditional encryption channels.

[0005] Second, there is a lack of dynamic prediction capabilities for key consumption. Existing network management systems can only monitor the current key consumption rate and cannot predict the key demand in the future. This prevents the system from adjusting the key pool allocation strategy in advance, leaving it in a state of constant reactive response.

[0006] Therefore, there is an urgent need for an intelligent management method that can sense changes in business traffic, predict key consumption trends, and dynamically optimize key pool scheduling strategies accordingly. Summary of the Invention

[0007] This invention aims to overcome the technical defects of existing quantum secure communication networks in power systems, such as static and fixed key scheduling, lack of accurate dynamic prediction of key consumption, easy depletion of key pools, passive response to business needs, and cross-regional interaction that does not comply with power safety regulations. It provides a key allocation optimization and fault early warning method and system based on a key pool state gating mechanism. The invention uses artificial intelligence algorithms to predict the key consumption in quantum key distribution (QKD) networks and dynamically optimizes the centralized key pool scheduling strategy based on the prediction results.

[0008] This invention provides an AI-driven key allocation optimization and fault early warning system for quantum secure communication networks, comprising: a data acquisition module for acquiring data from a quantum key management system and a power business system, the data including key consumption data, business traffic data, and key pool status data; a data preprocessing module connected to the data acquisition module for preprocessing the acquired data to construct a sample set; an AI prediction engine connected to the data preprocessing module, with a built-in time-series prediction network for predicting key consumption within a future time window based on the preprocessed data and the current key pool inventory; a key pool dynamic scheduling module connected to the AI ​​prediction engine for dynamically generating a key allocation strategy based on the predicted key consumption within the future time window and the current key pool inventory, and triggering a key shortage warning when warning conditions are met; and a collaborative control interface connected to the key pool dynamic scheduling module for distributing the key allocation strategy to the quantum key management system.

[0009] Furthermore, the data acquisition module collects data from the production control area through a forward isolation device.

[0010] Furthermore, the data preprocessing module performs operations including: missing value handling, data normalization, and sliding window construction.

[0011] Furthermore, the time-series prediction network built into the AI ​​prediction engine is the Power Business Adaptive Time-Series Prediction Network PAT-Net, which includes: a power business feature adaptation layer, used to convert general business traffic data into feature representations with power industry semantics; a multi-scale convolutional feature extraction layer, connected to the power business feature adaptation layer, used to extract features at different time scales in parallel through multiple convolutional kernels with different receptive fields; an attention mechanism layer, connected to the multi-scale convolutional feature extraction layer, used to dynamically weight the extracted multi-scale features according to the channel dimension; a key pool state-gated recurrent unit (GRU) layer, connected to the attention mechanism layer, used to use the current key pool inventory as a gating signal to modulate the hidden state update of the GRU; and an output layer, connected to the key pool state-gated GRU layer, used to map the output of the GRU to the final predicted key consumption value.

[0012] Furthermore, the power business feature adaptation layer is specifically used for: extracting the daily periodic features of timestamp t and performing sine coding to obtain the daily periodic code sin(2πt / 1440); extracting the weekly periodic features of timestamp t and performing sine coding to obtain the weekly periodic code sin(2πt / 10080); performing one-hot coding on abnormal events including fault triggering, load peaks, new energy output fluctuations, and maintenance plans; and splicing and fusing the abnormal event features obtained from the daily periodic code, weekly periodic code, and one-hot coding with the original business traffic data collected from the power business system to form a power business enhanced feature vector.

[0013] Furthermore, the multi-scale convolutional feature extraction layer adopts a parallel multi-branch convolutional structure, specifically including: a first branch, using a convolutional kernel with a kernel size of k=2, used to capture rapid burst changes at the minute level; a second branch, using a convolutional kernel with a kernel size of k=4, used to capture local trend changes; and a third branch, using a convolutional kernel with a kernel size of k=6, used to capture medium-term trend changes; the outputs of the first branch, the second branch, and the third branch are concatenated and fused to form a multi-scale feature vector.

[0014] Furthermore, the attention mechanism layer is a Squeeze-Excitation attention mechanism layer with a compression ratio of 16. It is used to calculate the weights of each feature channel through global average pooling, fully connected layers, and activation functions, and then weights the calculated weights with the input multi-scale features channel by channel.

[0015] Furthermore, the key pool state gating GRU layer is used to perform the following gating modulation operation: calculating the key pool state gating function. ,in This represents the current key pool balance. For learnable weights, For bias; according to the key pool state gating function The hidden state update of the modulated GRU is calculated using the following formula: ,in, This is the update gate for the GRU at time t. This represents the GRU hidden state at the previous time step t-1. Let be the candidate hidden state of GRU at time t; is the hidden state of GRU after update at time t; ⊙ is the Hadamard product.

[0016] Furthermore, the output layer consists of two fully connected layers, which are used to map the output of the key pool state gated GRU layer to the predicted key consumption value within the next 30 minutes.

[0017] Furthermore, the scheduling logic executed by the key pool dynamic scheduling module includes: triggering a key shortage warning when the total key consumption within the predicted future time window exceeds the current key pool inventory multiplied by a preset threshold; dynamically adjusting the key pool allocation to each link according to the predicted key consumption within the future time window of each service link, with the scheduling algorithm being: ;in, The key amount allocated to the i-th link, This represents the current key pool balance. , These represent the key consumption amounts predicted by the AI ​​prediction engine for the i-th and j-th service links within future time windows, respectively, where N is the total number of service links. When the total key consumption amount within the predicted future time window exceeds the current key pool inventory, priority is given to ensuring the key supply for high-priority service links. The priority order of the high-priority service links is: scheduling automation service, source-grid-load control service, and power distribution and consumption service.

[0018] Compared with the prior art, the present invention has the following beneficial effects:

[0019] First, this invention employs a time-series prediction network optimized for power QKD scenarios, achieving resource-aware adaptive prediction through a key pool state gating mechanism, thereby improving key resource utilization efficiency. Simulation results show that, compared to traditional fixed-threshold scheduling strategies, the key shortage event incidence rate is reduced by approximately 40%, and key pool resource utilization is improved by approximately 15%.

[0020] Secondly, this invention achieves a shift from passive response to proactive early warning. Through AI prediction, the risk of key resource shortage can be predicted 30 minutes in advance, with an accuracy rate of approximately 85%, providing ample response time for operations and maintenance personnel.

[0021] Third, the model of this invention is lightweight and efficient, meeting engineering deployment requirements. The model has approximately 200,000 parameters, and a single inference takes approximately milliseconds, enabling real-time inference on existing power grid management servers and meeting the second-level response requirements within a 5-minute sampling period.

[0022] Fourth, it differs from traditional static allocation schemes. This invention introduces power business feature encoding and key pool state gating mechanism into the key consumption prediction model, and realizes prediction-driven dynamic scheduling of the key pool through key resource state feedback mechanism. This breaks through the traditional static allocation mode of key pool based on fixed thresholds or simple polling, and forms a dedicated solution for power QKD scenarios. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the system architecture of an AI-driven key distribution optimization and fault early warning system in a quantum secure communication network according to the present invention.

[0024] Figure 2 This is a schematic diagram of the AI ​​prediction engine (PAT-Net) structure according to an embodiment of the present invention.

[0025] Figure 3 This is a schematic diagram of the entire quantum key distribution process design structure according to an embodiment of the present invention. Detailed Implementation

[0026] The following detailed implementation of an AI-driven key allocation optimization and fault early warning system in a quantum secure communication network according to the present invention is provided in conjunction with specific embodiments. The embodiments of the present invention are only used to explain the present invention and are not intended to limit the scope of protection of the present invention.

[0027] Figure 1 This invention relates to an AI-driven key distribution optimization and fault early warning system in a quantum secure communication network, as described in an embodiment of the present invention. Figure 1 As shown, the system architecture includes: a data acquisition module 110, which conforms to the security zoning requirements of the power monitoring system and acquires data from the production control area through a forward isolation device; a data preprocessing module 120, connected to the data acquisition module, used for cleaning, normalizing, and constructing sliding windows for multi-source time-series data; an AI prediction engine 130, connected to the data preprocessing module, which incorporates a time-series prediction network optimized for power QKD scenarios to predict key consumption within future time windows; a key pool dynamic scheduling module 140, connected to the AI ​​prediction engine 130, used for dynamically adjusting the key allocation quota of the key pool for each business link based on the prediction results; and a collaborative control interface 150, connected to the key pool dynamic scheduling module 140, which interacts with the quantum key management system through a standardized interface and issues scheduling instructions. Specifically:

[0028] The data acquisition module 110 is used to acquire data from the quantum key management system and the power business system, including key consumption data, business traffic data and key pool status data.

[0029] Specifically, the data acquisition module 110 adopts an acquisition architecture that meets the security zoning requirements of the power monitoring system, and acquires the following data from the production control area through a forward isolation device:

[0030] First, key consumption data: the key consumption rate and total key consumption of each business link are collected from the northbound interface of the quantum key management machine. The unit of measurement is key bits (unit: Mbit), and the collection period is 5 minutes.

[0031] Second, business traffic data, which is collected from the southbound interface of the power business system in real time, including the number of business traffic and messages. The data interface complies with the power industry standard DL / T 634.5101-2002 or DL / T 634.5104-2009.

[0032] Third, key pool status data, which is collected from the quantum key management system to show the current stock (unit: Mbit) of the centralized key pool and the key allocation quota for each business link.

[0033] The data preprocessing module 120 is connected to the data acquisition module and is used to preprocess the acquired data to construct a sample set;

[0034] Specifically, the data preprocessing module 120 performs the following operations:

[0035] For missing values, linear interpolation is used to fill in temporary data gaps caused by network jitter.

[0036] Data normalization is performed by using the max-min normalization method to map all feature data to the [0,1] interval;

[0037] A sliding window is constructed using a fixed-length sliding window to build the sample set. The window length L is determined by five-fold cross-validation. The window traverses L∈[6,12,18,24,30] and selects the L value that minimizes the RMSE of the validation set. Simulation results show that L=12 (corresponding to 60 minutes) is optimal.

[0038] AI prediction engine 130, connected to the data preprocessing module, has a built-in time series prediction network for predicting key consumption in future time windows based on the preprocessed data and the current key pool inventory.

[0039] Specifically, such as Figure 2The diagram shows the structure of the AI ​​prediction engine (PAT-Net) according to an embodiment of the present invention. The AI ​​prediction engine 130 incorporates a temporal prediction network optimized for the power QKD scenario, named Power-Adaptive Temporal Prediction Network (PAT-Net). Its structure includes a power service feature adaptation layer, a multi-scale convolutional feature extraction layer, a Squeeze-Excitation attention mechanism layer, a key pool state-gated GRU layer, and a fully connected output layer. The system comprises the following components: a power business feature adaptation layer, used to convert general business traffic data into feature representations with power industry semantics; a multi-scale convolutional feature extraction layer, connected to the power business feature adaptation layer, used to extract features at different time scales in parallel using multiple convolutional kernels with different receptive fields; an attention mechanism layer, connected to the multi-scale convolutional feature extraction layer, used to dynamically weight the extracted multi-scale features along the channel dimension; a key pool state-gated recurrent unit (GRU) layer, connected to the attention mechanism layer, used to modulate the hidden state update of the GRU using the current key pool inventory as a gating signal; and an output layer, connected to the key pool state-gated GRU layer, used to map the output of the GRU to the final predicted key consumption value. Specifically:

[0040] 1. Power Business Characteristics Adaptation Layer

[0041] The power business feature adaptation layer is specifically used for: extracting the daily periodic features of timestamp t and performing sine coding to obtain the daily periodic code; extracting the weekly periodic features of timestamp t and performing sine coding to obtain the weekly periodic code; performing one-hot coding on abnormal events including fault triggering, load peaks, new energy output fluctuations, and maintenance plans; and concatenating and fusing the abnormal event features obtained from the daily periodic code, weekly periodic code, and one-hot coding with the original business traffic data collected from the power business system to form an enhanced feature vector for power business.

[0042] Specifically, the power business feature adaptation layer transforms general business traffic data into feature expressions with power industry semantics, including:

[0043] (1) Load feature coding: Extract the daily and weekly periodic features of the power load. Let the daily periodic code corresponding to the timestamp t be sin(2πt / 1440) and the weekly periodic code be sin(2πt / 10080), as additional feature inputs;

[0044] (2) Event feature coding: Abnormal events such as power grid faults and anti-accident drills are coded using a single hot code. The event categories include fault triggering, load peak, new energy output fluctuation, and maintenance plan, totaling 4 categories.

[0045] (3) Feature fusion: The above-mentioned encoded features are concatenated with the original business traffic data features to form an enhanced feature vector for power business. The feature dimension after concatenation is the original feature dimension plus 6.

[0046] 2. Multi-scale convolutional feature extraction layer

[0047] The multi-scale convolutional feature extraction layer employs a parallel multi-branch convolutional structure, simultaneously using three convolutional kernels with different receptive fields to perform parallel convolution operations on the input sequence. The sampling interval is 5 minutes, and the window length is 12 (60 minutes).

[0048] (1) The first branch has a kernel size of k=2, covering 10 minutes, capturing rapid bursts of change at the minute level, and the number of kernels is 64;

[0049] (2) The second branch has a kernel size of k=4, covers 20 minutes, captures local trend changes, and has 128 kernels.

[0050] (3) The third branch has a kernel size of k=6, covers 30 minutes, captures mid-term trend changes, and has 256 kernels.

[0051] The outputs of the first, second, and third branches are concatenated and fused to form a multi-scale feature vector, resulting in 448 channels after fusion.

[0052] 3. Squeeze-Excitation attention mechanism layer

[0053] The attention mechanism layer is a Squeeze-Excitation attention mechanism layer with a compression ratio of 16. It is used to calculate the weights of each feature channel through global average pooling, fully connected layers, and activation functions. The calculated weights are then weighted channel-by-channel with the input multi-scale features. The calculation process is as follows:

[0054]

[0055] in, The input features have a dimension of . z represents the result of global average pooling, with dimension 1. ; Dimensions , Dimensions ReLU is the linear rectified activation function; Sigmoid is the sigmoid activation function. For channel weights, the dimension is... The weighted output has the following dimensions: ; B is the time step of the input features; B is the batch size of the input features. For the first The eigenvalues ​​of each channel at time step t; For the first The global average pooling result for each channel.

[0056] 4. Key Pool State Gated GRU Layer

[0057] The key pool state-gated GRU layer performs the following gating and modulation operations: calculating the key pool state gating function; and modulating the hidden state update of the GRU according to the key pool state gating function. The specific implementation process is as follows:

[0058] The key pool state-gated GRU layer uses the key pool state as a gating signal to modulate the hidden state update of the GRU.

[0059] The hidden state update of the traditional GRU is as follows:

[0060]

[0061] in, This is the update gate for the GRU at time t, used to control the hidden state from the previous time step. Bring to the current hidden state The degree; This is the GRU's reset gate at time t, used to control the hidden state of the previous time step. Write to the current candidate hidden state The degree; Let t be the hidden state of the GRU at time t-1; The GRU input features at the current time t; Update the gate weight matrix for GRU; Reset the gate weight matrix for the GRU; Let be the weight matrix of the GRU candidate hidden states; Let be the candidate hidden state of GRU at time t; represents the hidden state of the GRU at time t; ⊙ represents the Hadamard product, indicating the multiplication of corresponding elements.

[0062] This invention introduces a key pool state gating mechanism, defined as follows:

[0063]

[0064] in, Current key pool inventory (normalized, unit: Mbit); These are learnable weights with a dimension of 1×1; The bias is used; Sigmoid is the activation function. This is the output of the key pool state gating function.

[0065] The hidden state after gating modulation is updated as follows:

[0066]

[0067] When the key pool has a low inventory Approaching 0, gating factor When the value approaches 1, the model maintains normal predictive ability; when the key pool has sufficient reserves, The gating factor approaches 1, and the gating factor approaches 2, enhancing the contribution of candidate states to the current hidden state. To avoid gradient instability, LayerNorm normalization is applied to the hidden state.

[0068] By introducing key pool reserves Gating function This invention enables dynamic modulation of GRU hidden state updates.

[0069] 5. Output layer and model training settings

[0070] The output layer consists of two fully connected layers with 128 and 1 neurons respectively. It maps the output of the key pool state-gated GRU layer to the final predicted value, which is the key consumption amount (unit: Mbit) within the first preset time window in the future. The optimal time window is 30 minutes after five-fold cross-validation.

[0071] The model training settings are as follows:

[0072] (1) The loss function is the mean squared error (MSE);

[0073] (2) The optimizer is Adam, the initial learning rate is 0.001, and the weight decay is 1e-5;

[0074] (3) Batch size is 32;

[0075] (4) The training rounds are 100 rounds. The early stopping method is adopted. The training is stopped when the validation set loss does not decrease for 10 consecutive rounds.

[0076] (5) Regularization uses a Dropout layer with a dropout rate of 0.2, placed between the GRU layer and the fully connected layer;

[0077] (6) The training dataset uses historical running data of the QKD network for at least 3 months, with a sampling interval of 5 minutes. The training set and the validation set are divided in a 7:3 ratio. The dataset contains at least 100 complete key consumption event cycles, covering normal load, fault events, and maintenance plan scenarios.

[0078] (7) After the model training is completed, the model weight parameters are fixed in the inference engine. The inference deployment uses a CPU server or a GPU server and adopts the ONNX Runtime or TensorRT inference framework.

[0079] The key pool dynamic scheduling module 140 is connected to the AI ​​prediction engine 130. It is used to dynamically generate a key allocation strategy based on the predicted key consumption in the future time window and the current key pool inventory, and to trigger a key shortage warning when the warning conditions are met.

[0080] Specifically, the key pool dynamic scheduling module 140 receives the output of the AI ​​prediction engine 130 and executes the following scheduling logic:

[0081] First, a key shortage warning. If the predicted key consumption within a future time window exceeds a preset threshold T of the current key pool inventory, a warning is triggered. The preset threshold T is determined to be 80% through simulation experiments. The warning information includes the expected shortage time, the shortage amount, and suggested actions.

[0082] Second, dynamic allocation strategy generation. Based on the predicted key demand of each business link, the allocation quota of the key pool to each link is dynamically adjusted proportionally. The scheduling period is aligned with the prediction time window (30 minutes). The scheduling algorithm is as follows:

[0083]

[0084] in, The amount of key allocated to the link (in Mbit). Current key pool inventory (unit: Mbit); The key consumption (in Mbit) of the i-th business link predicted by AI in the next 30 minutes. This represents the total number of business links. When the total predicted consumption exceeds... At the same time, the system prioritizes the supply of keys for high-priority service links, with the priority order being: scheduling automation services, source-grid-load control services, and power distribution and consumption services.

[0085] The collaborative control interface 150 is connected to the key pool dynamic scheduling module 140 and is used to distribute the key allocation strategy to the quantum key management system.

[0086] Specifically, the collaborative control interface 150 uses a standardized interface to interact with the underlying devices:

[0087] (1) The key management interface follows the RESTful API provided by the QKD device manufacturer and is used to issue key allocation strategies to the quantum key management system;

[0088] (2) In the security partition adaptation, the AI ​​prediction engine is deployed in the production control area II, and the scheduling instructions are sent to the quantum key management system in the management information area III through the forward isolation device, which complies with the "Regulations on Security Protection of Power Monitoring System".

[0089] The AI-driven key allocation optimization and fault early warning system in the quantum secure communication network according to the above embodiments of the present invention constructs a dedicated time-series prediction mechanism for the key pool stock-gated modulation GRU, integrating the key pool resource status into the prediction model. This breaks through the static allocation mode of traditional key pools with fixed thresholds and simple polling, solving the industry pain points of conventional predictions being unable to perceive the key pool surplus and having insufficient prediction accuracy. It realizes prediction-driven dynamic scheduling of the key pool and proactive early warning of key shortages, predicting the risk of key resource depletion in advance, avoiding forced degradation of core businesses such as power dispatch automation due to insufficient keys, and completing the mode transformation from passive response to proactive prevention and control. It is adapted to the daily / weekly cycle characteristics of power business and the power grid. The system leverages multi-scale feature extraction and channel attention weighting to enhance prediction adaptability by identifying abnormal events such as faults and load peaks, optimizing key resource allocation, and improving the overall utilization efficiency of the key pool. A high-priority power business key protection mechanism is established to prioritize core business supply during periods of key resource scarcity, thereby improving the operational reliability of the power quantum secure communication network. Strict adherence to power monitoring system security zoning standards is ensured through forward isolation devices for secure cross-regional data exchange and command issuance, guaranteeing compliant system deployment and engineering feasibility. Finally, a lightweight and efficient prediction and scheduling solution is provided, adaptable to existing power grid management hardware, achieving real-time inference and second-level response, balancing prediction accuracy with engineering practicality.

[0090] The technical solution of the present invention will be described in detail below with reference to a specific embodiment in the general power QKD network scenario.

[0091] Implementation Scenario: A provincial power company's quantum secure communication network adopts a centralized key pool architecture. A quantum key management system is deployed at the main station, establishing quantum keys with multiple substations via point-to-point QKD links. The typical key generation rate of the QKD device is 20 kbit / s (25 km fiber). The key pool capacity is configured based on the number of links, peak traffic, and key update cycle. This embodiment uses 4 Mbit as an example capacity to demonstrate the algorithm's effectiveness. Each business system requests keys from the key pool for encrypted communication. In one-time pad or high-frequency key update modes, key consumption is approximately positively correlated with the amount of business data.

[0092] like Figure 3The diagram illustrates the complete quantum key distribution process of this invention. The complete quantum key distribution process includes key generation, key management, and key distribution. Key generation: Quantum key generation terminal B and quantum key generation terminal A achieve quantum state interaction without IP data, including photon emission and reception. Quantum key generation terminals B and A are simultaneously connected to a key management server and a key management machine via classical switches. The classical switches negotiate key measurement signals and complete the necessary IP data interaction to generate the original key. The generated key then enters the key management stage: the key management server and the key management machine engage in bidirectional data interaction. In the key distribution stage, each quantum VPN (i.e., key distributor) establishes a secure connection through a quantum VPN tunnel. The quantum VPN obtains keys from quantum key generation terminal A and the key management machine, achieving secure key distribution and application.

[0093] System deployment and operation process:

[0094] The first step is data acquisition. An agent program is deployed in the quantum key management system. Through a forward isolation device, key consumption data is pushed to the AI ​​prediction server in Production Control Zone II every 5 minutes. The unit for key consumption data is Mbit. The scheduling automation system transmits business traffic data to the AI ​​prediction server in real time via the IEC 60870-5-104 protocol. The unit for traffic data is Mbit.

[0095] The second step is data preprocessing. Linear interpolation is used to fill missing values ​​in the collected data, and the window length L=12 (60 minutes) is determined through five-fold cross-validation to construct the sample.

[0096] The third step is AI model prediction. Inputs include the key consumption sequence of the past 60 minutes, the business traffic sequence, load feature encoding, event feature encoding, and the current key pool inventory. The power business feature adaptation layer extracts the load cycle features at the current time point, identifies it as the daily load peak period, and enhances the weight of traffic features. A multi-scale convolutional layer uses three parallel convolutions to extract features at 10-minute, 20-minute, and 30-minute scales. A Squeeze-Excitation attention mechanism dynamically weights the 448-channel features. A key pool state-gated GRU layer combines the current key pool inventory to gate and update the hidden state of the modulation GRU. The output is the predicted key consumption value for the next 30 minutes.

[0097] Step 4: Dynamic scheduling of the key pool. The dynamic scheduling module receives the prediction results: In the next 30 minutes, Link A (scheduling automation) will consume 0.35 Mbit, Link B (source-network-load control) will consume 0.18 Mbit, and Link C (power distribution and consumption services) will consume 0.12 Mbit; the current key pool stock is 1.2 Mbit, and the predicted total consumption is 0.65 Mbit, which does not exceed the threshold (1.2 × 80% = 0.96 Mbit), so no warning is triggered; Dynamic allocation strategy is generated: The 0.65 Mbit in the key pool is allocated proportionally, with Link A receiving 0.35 Mbit, Link B receiving 0.18 Mbit, and Link C receiving 0.12 Mbit.

[0098] The fifth step is to issue scheduling instructions. The allocation instructions are sent to the quantum key management system via a RESTful API. The system then updates the key allocation quotas for each business link, achieving a closed-loop scheduling mechanism.

[0099] The parameters are determined based on the following table 1:

[0100] Table 1

[0101]

[0102] The technical solution of this invention was tested using quantum key generation rate. The test results showed that the hourly average key generation rate during the test period was 19.868 kbps, which meets the requirement of a key generation rate ≥ 2 kbps (hourly average key generation rate under 13 dB total path attenuation). It was confirmed with NARI's debugging personnel that the NARI VPN device could acquire quantum keys and perform encryption and decryption normally. This is consistent with the expected results.

[0103] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0104] The above embodiments are merely illustrative examples and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. An AI-driven key distribution optimization and fault early warning system for quantum secure communication networks, characterized in that, include: The data acquisition module is used to acquire data from the quantum key management system and the power business system. The data includes key consumption data, business traffic data and key pool status data. A data preprocessing module, connected to the data acquisition module, is used to preprocess the acquired data to construct a sample set; The AI ​​prediction engine is connected to the data preprocessing module and has a built-in time-series prediction network to predict the key consumption in future time windows based on the preprocessed data and the current key pool inventory. The key pool dynamic scheduling module is connected to the AI ​​prediction engine and is used to dynamically generate a key allocation strategy based on the predicted key consumption in the future time window and the current key pool inventory, and trigger a key shortage warning when the warning conditions are met. The collaborative control interface is connected to the key pool dynamic scheduling module and is used to distribute the key allocation strategy to the quantum key management system.

2. The AI-driven key distribution optimization and fault early warning system in a quantum secure communication network according to claim 1, characterized in that, The data acquisition module collects data from the production control area through a forward isolation device.

3. The AI-driven key distribution optimization and fault early warning system in a quantum secure communication network according to claim 1, characterized in that, The data preprocessing module performs the following operations: missing value handling, data normalization, and sliding window construction.

4. The AI-driven key distribution optimization and fault early warning system in a quantum secure communication network according to claim 1, characterized in that, The AI ​​prediction engine's built-in time-series prediction network is the Power Business Adaptive Time-Series Prediction Network (PAT-Net), which includes: The power business feature adaptation layer is used to convert general business traffic data into feature expressions with power industry semantics. A multi-scale convolutional feature extraction layer, connected to the power business feature adaptation layer, is used to extract features at different time scales in parallel through multiple convolutional kernels with different receptive fields. The attention mechanism layer, connected to the multi-scale convolutional feature extraction layer, is used to dynamically weight the extracted multi-scale features according to the channel dimension. The Key Pool State Gated Recurrent Unit (GRU) layer, connected to the attention mechanism layer, is used to modulate the hidden state update of the GRU by using the current key pool inventory as a gating signal. The output layer, connected to the key pool state-gated GRU layer, is used to map the output of the GRU to the final predicted key consumption value.

5. The AI-driven key distribution optimization and fault early warning system in a quantum secure communication network according to claim 4, characterized in that, The power service feature adaptation layer is specifically used for: Extract the daily periodicity features of the timestamp t and perform sine coding to obtain the daily periodic code sin(2πt / 1440); Extract the periodicity feature of the timestamp t and perform sine coding to obtain the periodic code sin(2πt / 10080); Abnormal events, including fault triggering, peak load, fluctuations in new energy output, and maintenance plans, are encoded using unique thermal codes. The abnormal event features obtained from the daily cycle encoding, weekly cycle encoding, and unique hot encoding are spliced ​​and fused with the original business traffic data collected from the power business system to form an enhanced feature vector for power business.

6. The AI-driven key distribution optimization and fault early warning system in a quantum secure communication network according to claim 4, characterized in that, The multi-scale convolutional feature extraction layer adopts a parallel multi-branch convolutional structure, specifically including: The first branch uses a convolution kernel with a kernel size of k=2 to capture rapid, sudden changes on a minute-by-minute scale. The second branch uses a convolution kernel with a kernel size of k=4 to capture local trend changes; The third branch uses a convolution kernel with a kernel size of k=6 to capture medium-term trend changes; The outputs of the first branch, the second branch, and the third branch are concatenated and fused to form a multi-scale feature vector.

7. The AI-driven key distribution optimization and fault early warning system in a quantum secure communication network according to claim 4, characterized in that, The attention mechanism layer is a Squeeze-Excitation attention mechanism layer with a compression ratio of 16. It is used to calculate the weights of each feature channel through global average pooling, fully connected layers and activation functions, and then weight the calculated weights with the input multi-scale features channel by channel.

8. The AI-driven key distribution optimization and fault early warning system in a quantum secure communication network according to claim 4, characterized in that, The key pool state-gated GRU layer is used to perform the following gating modulation operations: Calculate the key pool state gating function ,in This represents the current key pool balance. For learnable weights, For bias; According to the key pool state gating function The hidden state update of the modulated GRU is calculated using the following formula: ,in, This is the update gate for the GRU at time t. This represents the GRU hidden state at the previous time step t-1. Let be the candidate hidden state of GRU at time t; is the hidden state of GRU after update at time t; ⊙ is the Hadamard product.

9. The AI-driven key distribution optimization and fault early warning system in a quantum secure communication network according to claim 4, characterized in that, The output layer consists of two fully connected layers, which are used to map the output of the key pool state gated GRU layer to the predicted key consumption value within a preset time window.

10. The AI-driven key distribution optimization and fault early warning system in a quantum secure communication network according to claim 1, characterized in that, The scheduling logic executed by the key pool dynamic scheduling module includes: When the total key consumption within the predicted future time window exceeds the current key pool inventory multiplied by a preset threshold, a key shortage warning is triggered. Based on the predicted key consumption of each business link within the future time window, the allocation of the key pool to each link is dynamically adjusted proportionally. The scheduling algorithm is as follows: ; in, The key amount allocated to the i-th link, This represents the current key pool balance. , These represent the key consumption amounts predicted by the AI ​​prediction engine for the i-th and j-th business links within future time windows, respectively, where N is the total number of business links. When the total key consumption within the predicted future time window exceeds the current key pool inventory, priority will be given to ensuring the key supply for high-priority service links. The priority order of the high-priority service links is: scheduling automation service, source-grid-load control service, and power distribution and consumption service.

Images