Distribution network transmission resource trusted scheduling method, device and system assisted by large model

Abstract

The present invention relates to a method, device and system for trusted scheduling of distribution network computing resources assisted by a large model, belonging to the field of electric communication technology. The present invention jointly optimizes the global loss function of the large model and the total fine-tuning delay. By dynamically adjusting the optimization weight of the delay and learning the matching resource scheduling strategy, the trade-off between the delay and the convergence performance is achieved. In addition, by detecting FDI attacks, it prevents the parameters of the attacked large model from participating in global fine-tuning, improving the security of the large model and the credibility of resource scheduling. The resource scheduling strategy is learned through DQN, and a large model heuristic function is introduced to assist in the reconstruction of the Q function of DQN. The large model heuristic function can combine context information such as historical state information, actions, and rewards to score the newly generated resource scheduling actions, thereby inspiring the learning direction of the DQN network and supporting more accurate policy optimization. At the same time, the present invention uses a large model to predict unknown state information for resource scheduling optimization, improving the learning ability of the algorithm.

Description

Technical Field

[0001] This invention relates to a large-model-assisted reliable scheduling method, apparatus, and system for distribution network computing resources, belonging to the field of telecommunications technology. Background Technology

[0002] Currently, with the rapid development and increasing intelligence of power systems, distribution network operation and management are facing unprecedented challenges. To improve the accuracy and efficiency of distribution network control, it is essential to achieve digitalization and intelligentization of operation and management. In this process, generative artificial intelligence technology, especially the application of large-scale models, has become a key technological support. Large-scale models are deep learning models trained on massive amounts of data. By fine-tuning these pre-trained models, they can be adapted to the operating environment and task requirements in the distribution network field. The fine-tuning process first involves deploying power Internet of Things (IoT) terminals, which are responsible for collecting grid operation data. Subsequently, this data is used to fine-tune the large-scale model locally to adapt to the characteristics of the local grid. Finally, the results of the local fine-tuning are uploaded to the central system for global fine-tuning to achieve global optimization of model parameters. Large-scale model fine-tuning has strict requirements for convergence and latency. First, the dynamic nature of the distribution network requires the model to respond quickly to changes in grid status, meaning the fine-tuning process must be completed within a short time to maintain consistency between the model and the real-time grid operating status. Second, the model's convergence performance is directly related to the safety and economy of grid operation; a fast and accurate convergence process can reduce the risks and costs of grid operation. Therefore, fine-tuning of large models must not only ensure that model parameters can quickly adapt to changes in the power grid, but also ensure the stability and reliability of this process to meet the needs of efficient and safe operation of the distribution network.

[0003] To further improve the operational efficiency and response speed of distribution networks, flexible and intelligent scheduling of resources such as transmission and computing within the network is required. However, this process often takes place in untrusted environments, facing severe security challenges. Attackers may exploit various methods to attack distribution networks, one typical example being False Data Injection (FDI) attacks. FDI attacks inject false data into the power grid, affecting the fine-tuning process of large models, leading to model decision errors, and consequently severely impacting the safe and stable operation of the distribution network. The mechanism of FDI attacks lies in attackers intercepting or tampering with data during transmission, injecting carefully designed false information to mislead the learning and decision-making processes of large models. This type of attack may not only cause large models to fail to converge correctly but may also trigger instability in power grid operation, resulting in power grid failures. Therefore, fine-tuning of large distribution network models and optimization of transmission and computing resource scheduling in untrusted environments face the following challenges:

[0004] On the one hand, to improve the accuracy of distribution network management and control, it is essential to enhance the convergence performance of large-scale models, i.e., minimizing the loss function that reflects the deviation between the output values ​​of the large-scale model and the actual values. Simultaneously, fine-tuning of the large-scale model needs to be performed with low latency to maintain real-time consistency with the ever-growing knowledge of distribution network operation. Furthermore, due to the impact of FDI attacks, fine-tuning of large-scale models is often performed in untrusted environments, leading to difficulties in model convergence and even catastrophic forgetting. On the other hand, traditional machine learning-based resource scheduling optimization methods, such as Deep Q-Networks (DQNs), while capable of learning resource scheduling strategies even when optimization modeling and network environment information are unknown, have significant limitations in handling multi-dimensional resource allocation optimization problems. Specifically, the multi-dimensional resource scheduling problem involving channels (transmission) and datasets (computation) has high complexity. DQN networks need to learn within a large optimization space, resulting in significant uncertainty in the learning direction. Therefore, there is an urgent need to propose a large-scale model-assisted reliable scheduling method, device, and system for distribution network transmission and computational resources to achieve reliable scheduling of distribution network transmission and computational resources. Summary of the Invention

[0005] To address the aforementioned technical problems, the purpose of this invention is to provide a reliable scheduling method, apparatus, and system for distribution network computing resources assisted by a large model.

[0006] The present invention provides a large-model-assisted reliable scheduling method for distribution network computing resources, the specific scheduling steps of which are as follows:

[0007] Construct local large-scale model fine-tuning model and global large-scale model fine-tuning model, parameter transmission model, FDI attack model, privacy entropy model and total large-scale model fine-tuning latency model, and construct the trusted scheduling optimization problem of computing resources;

[0008] The above optimization problem is transformed and modeled as a Markov decision process;

[0009] By learning resource scheduling strategies through DQN and introducing a large model heuristic function to assist in the reconstruction of the Q function of DQN, the above Markov decision process is solved.

[0010] The above solutions are used to detect FDI attacks for trusted scheduling of computing resources. By discarding the large model parameters that are attacked, the adverse effects of false data on global fine-tuning are reduced, thus achieving trusted scheduling of computing resources.

[0011] Furthermore, the local large model fine-tuning model is as follows:

[0012] Define I terminals, represented as D = {d1, ..., d2}. i ,…,d I}; Define the terminal d for the t-th iteration. i The local large model is LLMi,t It is determined by the parameter w i,t Context c i,t and the prompt word q i,t Composition, represented as

[0013] LLM i,t =Γ(w i,t ,c i,t ,q i,t (1)

[0014] d i Receive global large model parameters w sent by the edge controller g,t-1 To synchronize its local large model parameters w i,t-1 Then, d i Fine-tuning w using dataset i,t-1 Define the dataset size set as A. i ={α min ,…,|C i |}, where |C i | represents d i The size of the dataset, α min Represents the minimum possible dataset size;

[0015] Define x k and y k Let x be the input label sequence and the target output, respectively, for the k-th sample in the dataset; k Includes context and question, y k The target probability distribution for the output label;

[0016] d i The cross-entropy function is

[0017]

[0018] In the formula, l CE (w i,t-1 ,x k ,y k () represents the cross-entropy function for each sample; based on the cross-entropy function, the local large model parameters w for the t-th iteration are updated using gradient descent. i,t ;

[0019] d i The local fine-tuning latency and fine-tuning power consumption are related to the floating-point operands γ of all tags in its parameters. i Parameter size |w i,t | and fine-tuning period e i The relevant parameters are: local fine-tuning latency and fine-tuning energy consumption, respectively expressed as follows:

[0020]

[0021] In the formula, μ i,t d i Available GPU resources, λ i Indicates the available GPU computing speed, κ i This indicates the available GPU computing efficiency.

[0022] Furthermore, the global large model fine-tuning model is as follows:

[0023] Global fine-tuning is represented as

[0024]

[0025] In the formula, Let ρ be the set of terminals accessing the channel in the t-th iteration; i d i Large model supports business priorities; define L g,t (w g,t ) is the global loss function, used to measure the convergence performance of the global model, expressed as:

[0026]

[0027] Furthermore, the parameter transmission model is as follows:

[0028] Define J channels, represented as N = {n1, ..., n} j ,…,n J Define the channel selection variable as β. i,j,t ∈{0,1}, β i,j,t =1 indicates that in the t-th iteration d i Select channel n j Transmission parameters; d i Through n j The transmission parameter rate is expressed as:

[0029]

[0030] In the formula, B i,j For d i medium channel n j The transmission bandwidth, P i,t For d i The transmission power, g i,j,t Let δ0 be the channel gain, δ0 be the Gaussian white noise power, and I0 be the channel gain. i,j,t For electromagnetic interference power; d i Through n j The latency and energy consumption of the transmission parameters are expressed as follows:

[0031]

[0032] Furthermore, the FDI attack model is as follows:

[0033] definition The FDI attack is modeled as injecting fake data.

[0034]

[0035] In the formula, θ i,t ∈{0,1} is the attack indicator variable, θ i,t =1 indicates d i The parameters are vulnerable to FDI attacks.

[0036] Furthermore, the privacy entropy up to the t-th iteration is expressed as:

[0037]

[0038] In the formula, This indicates that up to the t-th iteration, d... i Through n j The probability of transmitting parameters; when d i Privacy entropy is maximized when the probability of transmitting parameters through all channels is equal.

[0039] Furthermore, the overall large model fine-tuning delay model is as follows:

[0040] The total large-scale model fine-tuning latency consists of the total local fine-tuning latency and transmission latency of all terminals, as well as the global fine-tuning latency of the edge controller, and is expressed as follows:

[0041]

[0042] In the formula, τ g This is for global fine-tuning of latency.

[0043] Furthermore, the constructed reliable scheduling optimization problem for computing resources is as follows:

[0044] By coordinating the scheduling of power distribution network transmission and computing resources, the weighted sum of the global loss function and the overall model fine-tuning delay is minimized, specifically expressed as:

[0045]

[0046] In the formula, V τ The weights represent the overall model fine-tuning latency, used to unify the order of magnitude and balance the relationship between optimization objectives; as well as F represents the decision variables for dataset scheduling and channel scheduling, respectively; i,min For d i Tolerable lower bound of privacy entropy, E i,max For d iThe battery capacity; C1 represents the dataset scheduling constraint, C2, C3, and C4 represent the channel scheduling constraint, that is, each terminal can only choose one channel for parameter transmission; C5 and C6 represent the long-term energy consumption and privacy entropy constraints, respectively.

[0047] Furthermore, the transformation of the optimization problem and its modeling as a Markov decision process specifically involves:

[0048] Construct a privacy entropy virtual queue and an energy consumption virtual queue corresponding to constraints C5 and C6; represented as:

[0049]

[0050] when and The queue is in a stable state, so C5 and C6 are considered satisfied; P1 is transformed into:

[0051]

[0052] To solve P2, it is modeled as a Markov decision process, whose key elements include state space, action space, and reward function, as detailed below:

[0053] 1) State Space: The large model can estimate information such as channel gain and electromagnetic interference based on historical information; the channel gain and electromagnetic interference estimated in the t-th iteration are expressed as... and d i The state space is filled by the privacy entropy virtual queue in the t-th iteration. Energy consumption virtual queue backlog in the t-th iteration Channel gain estimated by large model And the composition of electromagnetic interference power estimated by large models Represented as

[0054] 2) Action space: d i The action space consists of channel scheduling and dataset scheduling actions, represented as follows: in It is a Cartesian product;

[0055] 3) Reward Function: Since P2 is a minimization problem, the reward function Ω is... i,t Defined as the negative of the P2 optimization objective.

[0056] Furthermore, the specific method for solving the Markov decision process by learning resource scheduling strategies through DQN and introducing a large model heuristic function to assist in the reconstruction of the Q function of DQN is as follows:

[0057] The edge controller maintains a set of DQN neural networks for each terminal, i.e., one set of parameters is... The main network and parameters are The target network is defined as follows: At the beginning of each iteration, the edge controller optimizes the channel and dataset scheduling decisions based on the Q-value estimated by the main network of each terminal; at the end of each iteration, the loss function of DQN is calculated, and the parameters of the main network are updated with the assistance of the target network; the Q-function is reconstructed with the assistance of the large model; the reconstructed Q-function consists of two parts: a value function and a large model heuristic function; the value function reflects the expected cumulative reward obtained after taking a certain action in a certain state; the large model heuristic function combines contextual information such as historical state space, action space, and reward, and uses the current iteration state space and the action taken by the controller as the problem to evaluate the channel and dataset scheduling decisions of the current iteration, thereby inspiring the learning direction of the algorithm; the reconstructed Q-function is expressed as:

[0058]

[0059] In the formula, Represents the value function; The heuristic function of the large model is represented as:

[0060]

[0061] In the formula, This represents the heuristic context information of the input. This indicates a heuristic question. This represents the channel and dataset scheduling decision made by the controller in the t-th iteration; This represents the heuristic answer output by the large model. PARSE{.} represents the PARSE function, which parses the JSON-formatted answer from the large model into a numerical score. The channel-dataset collaborative scheduling steps based on large model-assisted learning are described below:

[0062] Step 1: Initialize the DQN network parameters for each terminal and Set the reward function Ω i,0 =0;

[0063] Step 2: At the start of each iteration, the edge controller selects the action with the highest Q-value for each terminal; when multiple terminals are assigned to the same channel, the controller assigns the channel to the terminal with the highest Q-value; terminals not assigned a channel select the action with the highest Q-value from the remaining actions; this process is repeated until all terminals are assigned a channel; each terminal executes its selected action.

[0064] Step 3: At the end of each iteration, the controller observes the performance of the global loss function, the total model fine-tuning latency, the privacy entropy virtual queue backlog, and the energy consumption virtual queue backlog, and calculates the reward; then, the controller updates the state space S of the (t+1)th iteration.i,t+1 The controller generates experience data. And store it in the experience replay pool;

[0065] Step 4: The controller randomly samples an experience dataset from the experience pool, and calculates the loss function of the DQN network based on the reconstructed Q function, expressed as:

[0066]

[0067] in For the sampled empirical dataset, The discount factor is used; based on the loss function, it is updated using gradient descent. Update every few iterations

[0068] Furthermore, the specific steps for detecting FDI attacks based on the solution results using trusted scheduling of computing resources are as follows:

[0069] Set the test dataset C Tes Based on C Tes Compare the local large model loss function of the terminal with the global large model loss function to verify whether the parameters are subject to FDI attacks; based on C Tes The local large model test loss function and the global test loss function are calculated as follows:

[0070]

[0071] when The loss function and w g,t-1 When the deviation between the loss functions exceeds a threshold ξ, that is... d i The transmitted parameters were subjected to an FDI attack, i.e. Here, The variable represents the attack detection variable; by discarding the large model parameters that have been attacked, the adverse effects of false data on global fine-tuning are reduced, and reliable scheduling of computing resources is achieved; Formula (7) is rewritten as:

[0072]

[0073] In the formula, For the t-th iteration, d i The size of the dataset used for local training; the edge controller performs global fine-tuning based on formula (19).

[0074] A large-model-assisted trusted scheduling device for power distribution network computing resources, deployed within an edge controller, includes:

[0075] Power module: Used to power the reliable scheduling device for distribution network computing resources assisted by the large model;

[0076] Communication module: Used to communicate with the terminal, send the terminal channel and dataset scheduling policies and the latest global large model parameters, and receive the local large model parameters uploaded by the terminal;

[0077] The trusted scheduling model construction module for distribution network computing resources is responsible for building local and global large-scale model fine-tuning models, parameter transmission models, FDI attack models, privacy entropy models, and overall large-scale model fine-tuning latency models, as well as constructing the trusted scheduling optimization problem for computing resources.

[0078] DQN network module: used to maintain a set of DQN neural networks and an experience replay pool for each terminal;

[0079] Q-function reconstruction module: Used to reconstruct the Q-function with the help of a large model. The reconstructed Q-function contains two parts: the value function and the large model heuristic function.

[0080] Trusted scheduling module for computing resources: used to optimize channel and dataset scheduling decisions based on the Q value estimated by the main network for each terminal;

[0081] FDI attack detection module: used to compare the local large model loss function and the global large model loss function of the terminal based on the test dataset to verify whether the parameters are subject to FDI attack;

[0082] Global Fine-Tuning Module: Used to fine-tune the global model based on the results of the FDI attack detection module;

[0083] Global large model module: It can estimate information such as channel gain and electromagnetic interference based on historical information and reconstruct the Q function, thereby providing assistance for reliable scheduling of computing resources.

[0084] A large-scale model-assisted trusted scheduling system for power distribution network computing resources.

[0085] It includes a local device layer, an edge layer, and a global large model layer;

[0086] In the local device layer, the power IoT terminal is deployed on the power distribution equipment. It fine-tunes the local large model according to the latest global large model parameters and channel and dataset scheduling strategies issued by the edge layer, and executes the corresponding scheduling strategies.

[0087] The edge layer includes a base station and an edge controller. The edge server is equipped with a large-model-assisted trusted scheduling device for distribution network computing resources. The power IoT terminal transmits local large-model parameters to the edge server via 4G / 5G / 6G channels. In addition, the edge layer can also provide attack detection for the uploaded parameters.

[0088] The global large model layer can also estimate channel gain and electromagnetic interference information, and reconstruct the Q function to assist in resource allocation optimization; the global large model layer is maintained by the edge controller and is responsible for performing global large model fine-tuning.

[0089] By means of the above-described solution, the present invention has at least the following advantages:

[0090] 1. This invention jointly optimizes the global loss function and total fine-tuning latency of a large model. By dynamically adjusting the optimization weights of latency and learning a matching resource scheduling strategy, a trade-off between latency and convergence performance is achieved. Furthermore, this invention prevents attacked large model parameters from participating in global fine-tuning through FDI attack detection, improving the security of the large model and the reliability of resource scheduling.

[0091] 2. This invention proposes a large-model-assisted reliable resource scheduling algorithm for distribution network computing. It learns resource scheduling strategies through Distributed Quad-N (DQN) and introduces a large-model heuristic function to assist in the reconstruction of the DQN's Q-function. The large-model heuristic function can combine historical state information, actions, rewards, and other contextual information to score newly generated resource scheduling actions, thereby inspiring the learning direction of the DQN network and supporting more accurate strategy optimization. Simultaneously, this invention uses a large model to predict unknown state information for resource scheduling optimization, such as channel gain and electromagnetic interference power, improving the algorithm's learning ability.

[0092] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings. Attached Figure Description

[0093] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show a certain embodiment of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0094] Figure 1 This is a flowchart of the reliable scheduling method for distribution network computing resources assisted by the large model of this invention;

[0095] Figure 2 This is a diagram of the reliable scheduling device for distribution network computing resources assisted by the large model of this invention;

[0096] Figure 3 This is a diagram of the reliable scheduling system for distribution network computing resources assisted by the large model of this invention. Detailed Implementation

[0097] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0098] This invention designs a large-model-assisted reliable scheduling method for power distribution network computing resources, the process of which is as follows: Figure 1 As shown.

[0099] 1.1. Construction of a Reliable Scheduling Model for Distribution Network Computing Resources

[0100] The optimization time is divided into T iterations, denoted as T = {1,…,t,…,T}. In each iteration, the edge controller distributes the terminal's channel and dataset scheduling policies, as well as the latest global large model parameters. Then, the terminal fine-tunes its local large model parameters based on the scheduled dataset and transmits these parameters to the edge controller via the allocated 4G / 5G / 6G channels. Finally, the edge controller performs attack detection on the uploaded parameters and aggregates the local large model parameters based on the detection results to fine-tune the global large model.

[0101] 1.1.1 Local Large Model Fine-tuning

[0102] Large-scale model fine-tuning ensures that the pre-trained model parameters remain unchanged, saving computational resources while preventing catastrophic forgetting. Consider I terminals, represented as D = {d1, ..., d...} i ,…,d I Define the terminal d for the t-th iteration. i The local large model is LLM i,t It is determined by the parameter w i,t Context c i,t and the prompt word q i,t Composition, specifically represented as

[0103] LLM i,t =Γ(w i,t ,c i,t ,q i,t (1)

[0104] d i Receive global large model parameters w sent by the edge controller g,t-1 To synchronize its local large model parameters w i,t-1 Then, d i Fine-tuning w using dataset i,t-1 Define the dataset size as A. i ={α min ,…,|C i |}, where |C i | represents d i The size of the dataset, α min This represents the minimum possible dataset size.

[0105] The cross-entropy function is used to measure the convergence performance of large models. A smaller cross-entropy function indicates better convergence performance for large models. Define x... k and y k Let x be the input label sequence and the target output, respectively, for the k-th sample of the dataset. k Includes context and question, y k This is the target probability distribution for the output labels.

[0106] d i The cross-entropy function is expressed as

[0107]

[0108] In the formula, l CE (w i,t-1 ,x k ,y k () represents the cross-entropy function for each sample. Based on the cross-entropy function, the local large model parameters w for the t-th iteration are updated using gradient descent. i,t .

[0109] d i The local fine-tuning latency and fine-tuning power consumption are related to the floating-point operands γ of all tags in its parameters. i Parameter size |w i,t | and fine-tuning period e i Related. Local fine-tuning latency and fine-tuning power consumption are expressed as follows:

[0110]

[0111] In the formula, μ i,t d i Available GPU resources, λ i Indicates the available GPU computing speed, κ i This indicates the available GPU computing efficiency.

[0112] 1.1.2 Parameter Transfer Model

[0113] After fine-tuning the local large-scale model, the terminal transmits the parameters to the edge controller via channels. Consider J channels, represented as N = {n1, ..., n}. j ,…,n J Define the channel selection variable as β. i,j,t ∈{0,1}, β i,j,t =1 indicates that in the t-th iteration d i Select channel n j Transmission parameters. d i Through n j The transmission parameter rate is expressed as:

[0114]

[0115] In the formula, B i,j For d i medium channel n j The transmission bandwidth, P i,t For d i The transmission power, g i,j,t Let δ0 be the channel gain, δ0 be the Gaussian white noise power, and I0 be the channel gain. i,j,t This refers to electromagnetic interference power. (d) i Through n j The latency and energy consumption of the transmission parameters are expressed as follows:

[0116]

[0117] 1.1.3 FDI Attack Model

[0118] FDI attacks disrupt model convergence performance by intercepting large model parameters during transmission and injecting spoofed data. (Definition) The FDI attack is modeled as injecting fake data.

[0119]

[0120] In the formula, θ i,t ∈{0,1} is the attack indicator variable, θ i,t =1 indicates d i The parameters are vulnerable to FDI attacks.

[0121] 1.1.4 Fine-tuning of the global large model

[0122] After receiving the transmitted local large model parameters, the edge controller performs global fine-tuning to update the global large model parameters. Global fine-tuning is represented as...

[0123]

[0124] In the formula, Let ρ be the set of terminals accessing the channel in the t-th iteration. i d i The business priorities supported by the large model. Define L g,t (w g,t ) is the global loss function, used to measure the convergence performance of the global model, expressed as:

[0125]

[0126] 1.1.5 Privacy Entropy Model

[0127] Privacy entropy is defined as the degree of disorder in the terminal's transmitted parameters. The greater the disorder in the local large model parameters being distributed across different channels for transmission, the greater the privacy entropy. The privacy entropy up to the t-th iteration is expressed as...

[0128]

[0129] In the formula, This indicates that up to the t-th iteration, d... i Through n j The probability of transmitting parameters. When d i Privacy entropy is maximized when the probability of transmitting parameters through all channels is equal.

[0130] 1.1.5 Overall Model Fine-tuning Delay Model

[0131] The total large-scale model fine-tuning latency consists of the total local fine-tuning latency and transmission latency of all terminals, as well as the global fine-tuning latency of the edge controller, and is expressed as follows:

[0132]

[0133] In the formula, τ g This is for global fine-tuning of latency.

[0134] 1.1.6 Modeling and Transformation of the Reliable Scheduling Optimization Problem for Computational Resources

[0135] This invention aims to solve the problem of reliable scheduling of power distribution network transmission and computing resources under FDI attacks. First, increasing the size of the dataset used by the terminal allows for more thorough local fine-tuning, but this increases local fine-tuning latency and cost. Second, changing the channel scheduling strategy affects the terminal parameter transmission latency, thus altering the fine-tuning cost. Therefore, the optimization objective is to minimize the weighted sum of the global loss function and the overall model fine-tuning latency through coordinated scheduling of power distribution network transmission and computing resources, specifically expressed as:

[0136]

[0137] In the formula, V τ The weights represent the overall model fine-tuning delay, used to unify the order of magnitude and balance the relationship between optimization objectives. as well as F represents the decision variables for dataset scheduling and channel scheduling, respectively. i,min For d i Tolerable lower bound of privacy entropy, E i,max For d i The battery capacity. C1 represents the dataset scheduling constraint, and C2, C3, and C4 represent the channel scheduling constraint, meaning that each terminal can only choose one channel for parameter transmission. C5 and C6 represent long-term energy consumption and privacy entropy constraints, respectively.

[0138] Because short-term resource allocation strategy optimization is coupled with long-term privacy entropy and energy consumption constraints, this invention constructs a privacy entropy virtual queue and an energy consumption virtual queue corresponding to constraints C5 and C6. (Represented as...)

[0139]

[0140] when and Given a stable queue state, C5 and C6 can be considered satisfied. P1 can be transformed into...

[0141]

[0142] 1.2 Design of a Reliable Scheduling Algorithm for Distribution Network Computing Resources Assisted by a Large Model

[0143] This invention designs a large-model-assisted trusted scheduling algorithm for power distribution network computing resources. First, a channel and dataset scheduling strategy is solved based on large-model-assisted learning. Then, the edge controller uses a test dataset to detect FDI attacks, achieving trusted scheduling of computing resources and preventing false data from adversely affecting global fine-tuning.

[0144] 1.2.1 Channel-Dataset Cooperative Scheduling Based on Large Model-Assisted Learning

[0145] To solve P2, this invention models it as a Markov decision process, whose key elements include state space, action space, and reward function, as detailed below:

[0146] 1) State Space: The large model can estimate information such as channel gain and electromagnetic interference based on historical information. The channel gain and electromagnetic interference estimated in the t-th iteration are expressed as... and d i The state space is filled by the privacy entropy virtual queue in the t-th iteration. Energy consumption virtual queue backlog in the t-th iteration Channel gain estimated by large model And the composition of electromagnetic interference power estimated by large models Represented as

[0147] 2) Action space: d i The action space consists of channel scheduling and dataset scheduling actions, represented as follows: in It is a Cartesian product.

[0148] 3) Reward Function: Since P2 is a minimization problem, the reward function Ω is... i,t Defined as the negative of the P2 optimization objective.

[0149] DQN provides a model-free solution to the aforementioned Markov decision process. The edge controller maintains a set of DQN neural networks for each terminal, i.e., one set of parameters... The main network and parameters are The target network is the edge controller. At the beginning of each iteration, the edge controller optimizes the channel and dataset scheduling decisions based on the Q-value estimated by the main network of each terminal. At the end of each iteration, the loss function of DQN is calculated, and the parameters of the main network are updated with the assistance of the target network. However, due to the large optimization space of P2, the DQN algorithm is prone to the problem of large uncertainty in the learning direction. To solve this problem, this invention reconstructs the Q-function with the assistance of a large model. The reconstructed Q-function consists of two parts: a value function and a large model heuristic function. The value function reflects the expected cumulative reward obtained after taking a certain action in a certain state. The large model heuristic function combines contextual information such as historical state space, action space, and reward, and uses the current iteration state space and the action taken by the controller as the problem to evaluate the channel and dataset scheduling decisions of the current iteration, thereby inspiring the learning direction of the algorithm. The reconstructed Q-function is expressed as follows:

[0150]

[0151] In the formula, This represents the value function. The heuristic function of the large model is represented as:

[0152]

[0153] In the formula, This represents the heuristic context information of the input. This indicates a heuristic question. This represents the channel and dataset scheduling decision made by the controller in the t-th iteration. This represents the heuristic answer output by the large model, and PARSE{.} represents the PARSE function, which parses the JSON-formatted answer from the large model into a numerical score. The steps of channel-dataset collaborative scheduling based on large model-assisted learning are described below.

[0154] Step 1: Initialize the DQN network parameters for each terminal and Set the reward function Ω i,0 =0.

[0155] Step 2: At the start of each iteration, the edge controller selects the action with the highest Q-value for each terminal. When multiple terminals are assigned to the same channel, the controller allocates the channel to the terminal with the highest Q-value. Terminals not assigned a channel select the action with the highest Q-value from the remaining actions. This process is repeated until all terminals are assigned a channel. Each terminal then executes its selected action.

[0156] Step 3: At the end of each iteration, the controller observes the performance of the global loss function, the total model fine-tuning latency, the privacy entropy virtual queue backlog, and the energy consumption virtual queue backlog, and calculates the reward. Then, the controller updates the state space S for the (t+1)th iteration. i,t+1 The controller generates empirical data. And store it in the experience replay pool.

[0157] Step 4: The controller randomly samples an experience dataset from the experience pool, calculates the loss function of the DQN network based on the reconstructed Q function, and expresses it as follows:

[0158]

[0159] in For the sampled empirical dataset, This is the discount factor. Based on the loss function, it is updated using gradient descent. Update every few iterations

[0160] 1.2.2 FDI Attack Detection for Trusted Scheduling of Computing Resources

[0161] This invention designs an FDI attack detection mechanism for trusted scheduling of computing resources. The principle of attack detection is based on a given test dataset C. Tes Based on C Tes Compare the local large model loss function and the global large model loss function on the terminal to verify whether the parameters are susceptible to FDI attacks. Based on C... Tes The local large model test loss function and the global test loss function are calculated as follows:

[0162]

[0163] when The loss function and w g,t-1 When the deviation between the loss functions exceeds a threshold ξ, that is... d i The transmitted parameters were subjected to an FDI attack, i.e. Here, This represents the attack detection variable. By discarding the large model parameters that have been attacked, the adverse effects of false data on global fine-tuning are reduced, thus achieving reliable scheduling of computing resources. Formula (7) is rewritten as...

[0164]

[0165] In the formula, For the t-th iteration, d iThe size of the dataset used for local training. The edge controller performs global fine-tuning based on equation (19).

[0166] 2. Large-scale model-assisted reliable scheduling device for power distribution network computing resources

[0167] This invention proposes a large-model-assisted reliable scheduling device for distribution network computing resources, such as... Figure 2 As shown, the module deployed within the edge controller includes a power module, a communication module, a trusted scheduling model construction module for power distribution network computing resources, a DQN network module, a Q function refactoring module, a trusted scheduling module for computing resources, an FDI attack detection module, a global fine-tuning module, and a global large model module, which are described in detail below.

[0168] Power module: Responsible for supplying power to the reliable scheduling device for distribution network computing resources assisted by the large model.

[0169] Communication module: Responsible for communicating with the terminal, sending the terminal channel and dataset scheduling policies and the latest global large model parameters, and receiving the local large model parameters uploaded by the terminal.

[0170] The trusted scheduling model construction module for distribution network computing resources is responsible for constructing local and global large-scale model fine-tuning models, parameter transmission models, FDI attack models, privacy entropy models, and overall large-scale model fine-tuning latency models, as well as constructing the trusted scheduling optimization problem for computing resources.

[0171] DQN network module: responsible for maintaining a set of DQN neural networks and experience replay pools for each terminal.

[0172] Q-function Reconstruction Module: Reconstructs the Q-function with the aid of a large model. The reconstructed Q-function consists of two parts: a value function and a large model heuristic function.

[0173] Trusted scheduling module for computing resources: Optimizes channel and dataset scheduling decisions based on the Q-value estimated by the main network for each terminal.

[0174] FDI attack detection module: Based on the test dataset, compare the local large model loss function of the terminal with the global large model loss function to verify whether the parameters are subject to FDI attack.

[0175] Global Fine-tuning Module: Fine-tunes the global model based on the results of the FDI attack detection module.

[0176] Global large model module: It can estimate information such as channel gain and electromagnetic interference based on historical information and reconstruct the Q function, thereby providing assistance for reliable scheduling of computing resources.

[0177] 3. Large-scale model-assisted trusted scheduling system for power distribution network computing resources

[0178] The large-model-assisted reliable scheduling system for distribution network computing resources considered in this invention is as follows: Figure 3 As shown, it includes a local device layer, an edge layer, and a global large model layer.

[0179] In the local device layer, power IoT terminals are deployed on power distribution equipment. They fine-tune the local large model based on the latest global large model parameters issued by the edge layer, as well as the channel and dataset scheduling strategies, and execute the corresponding scheduling strategies.

[0180] The edge layer comprises a base station and an edge controller, with a large-model-assisted trusted scheduling device for distribution network computing resources deployed in the edge server. Power IoT terminals transmit local large-model parameters to the edge server via 4G / 5G / 6G channels. Furthermore, the edge layer can provide attack detection for the uploaded parameters.

[0181] The global large model layer can also estimate information such as channel gain and electromagnetic interference, and reconstruct the Q function, thereby assisting in resource allocation optimization. The global large model layer is maintained by the edge controller and is responsible for performing global large model fine-tuning.

[0182] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A reliable scheduling method for power distribution network computing resources assisted by a large model, characterized in that... The specific scheduling steps are as follows: Construct local large-scale model fine-tuning model, global large-scale model fine-tuning model, parameter transmission model, FDI attack model, privacy entropy model and total large-scale model fine-tuning latency model, and construct the trusted scheduling optimization problem of computing resources; The above optimization problem is transformed and modeled as a Markov decision process; By learning resource scheduling strategies through DQN and introducing a large model heuristic function to assist in the reconstruction of the Q function of DQN, the above Markov decision process is solved. The above solution results are subjected to FDI attack detection for trusted scheduling of computing resources. By discarding the large model parameters that are attacked, the adverse effects of false data on global fine-tuning are reduced, and trusted scheduling of computing resources is achieved. The local large-model fine-tuning model is as follows: definition A set of terminals, represented as ; Define the first Second iteration terminal The local large model is It is composed of parameters Context and prompt words Composition, represented as （1） Receive global large model parameters sent by the edge controller To synchronize its local large model parameters ;Then, Fine-tuning using datasets ; Define the dataset size set as ,in express The size of the dataset, Represents the minimum possible dataset size; definition and The datasets are respectively the first and second. Each sampled input label sequence and target output; Includes context and question; The cross-entropy function is （2） In the formula, This represents the cross-entropy function for each sample; based on the cross-entropy function, the gradient descent method is used to update the... Local large model parameters in the next iteration ; The local fine-tuning latency and fine-tuning power consumption are related to the floating-point operands of all tags in its parameters. Parameter size and fine-tuning cycle The relevant parameters are: local fine-tuning latency and fine-tuning energy consumption, respectively expressed as follows: （3） In the formula, express Available GPU resources Indicates the available GPU computing speed. Indicates the available GPU computing efficiency; The global large model fine-tuning model is as follows: （7） In the formula, For the first The set of terminals accessing the channel in the next iteration; express The business priorities supported by the large model; definition Here, is the global loss function, used to measure the convergence performance of the global model, expressed as: （8）。 2. The reliable scheduling method for distribution network computing resources assisted by a large model as described in claim 1, characterized in that: The parameter transmission model is as follows: definition There are channels, and the set is represented as... Define the channel selection variable as , Indicates the first iteration Select Channel Transmission parameters; pass The transmission parameter rate is expressed as: （4） In the formula, for Mid-channel transmission bandwidth, for Transmission power, For channel gain, The power of Gaussian white noise. Electromagnetic interference power; pass The latency and energy consumption of the transmission parameters are expressed as follows: （5）。 3. The reliable scheduling method for distribution network computing resources assisted by a large model, as described in claim 2, is characterized in that: The FDI attack model is as follows: definition The FDI attack is modeled as injecting fake data. （6） In the formula, As an attack indicator variable, express The parameters are vulnerable to FDI attacks.

4. The reliable scheduling method for distribution network computing resources assisted by a large model as described in claim 3, characterized in that: up to the The privacy entropy of the next iteration is expressed as: （9） In the formula, , indicating up to the iteration pass The probability of transmitting parameters; when Privacy entropy is maximized when the probability of transmitting parameters through all channels is equal.

5. The reliable scheduling method for distribution network computing resources assisted by a large model, as described in claim 4, is characterized in that: The overall model fine-tuning delay model is as follows: The total large-scale model fine-tuning latency consists of the total local fine-tuning latency and transmission latency of all terminals, as well as the global fine-tuning latency of the edge controller, and is expressed as follows: （10） In the formula, This is for global fine-tuning of latency.

6. The reliable scheduling method for distribution network computing resources assisted by a large model, as described in claim 5, is characterized in that: The specific method for learning resource scheduling strategies through DQN and introducing a large model heuristic function to assist in the reconstruction of the Q function of DQN to solve the Markov decision process is as follows: The edge controller maintains a set of DQN neural networks for each terminal, i.e., one set of parameters is... The main network and parameters are The target network is defined as follows: at the beginning of each iteration, the edge controller optimizes the channel and dataset scheduling decisions based on the Q-value estimated by the main network of each terminal; at the end of each iteration, the loss function of DQN is calculated, and the parameters of the main network are updated with the assistance of the target network; the Q-function is reconstructed with the assistance of the large model; the reconstructed Q-function includes two parts: the value function and the large model heuristic function. The value function reflects the expected cumulative reward obtained after taking a certain action in a certain state; The large model heuristic function combines historical state space, action space, and reward context information. Using the current iteration's state space and the controller's actions as the problem, it evaluates the channel and dataset scheduling decisions for the current iteration, thus inspiring the algorithm's learning direction. The channel-dataset collaborative scheduling steps based on large model-assisted learning are described below: Step 1: Initialize the DQN network parameters for each terminal and Set reward function = 0; Step 2: At the start of each iteration, the edge controller selects the action with the largest Q value for each terminal; when multiple terminals are assigned to the same channel, the controller assigns the channel to the terminal with the largest Q value; the terminal that is not assigned a channel selects the action with the largest Q value from the remaining actions; this process is repeated until all terminals are assigned a channel. Step 3: At the end of each iteration, the controller observes the global loss function, the total model fine-tuning latency, the privacy entropy virtual queue backlog, and the energy consumption virtual queue backlog performance, and calculates the reward; then, the controller updates the... The state space of the next iteration The controller generates experience data and stores it in the experience replay pool. Step 4: The controller randomly samples an experience dataset from the experience pool and calculates the loss function of the DQN network based on the reconstructed Q function.

7. A large-model-assisted reliable scheduling device for distribution network computing resources, used to run the method of any one of claims 1-6, characterized in that: Deployed within the edge controller, including: Power module: Used to power the reliable scheduling device for distribution network computing resources assisted by the large model; Communication module: Used to communicate with the terminal, send the terminal channel and dataset scheduling policies and the latest global large model parameters, and receive the local large model parameters uploaded by the terminal; The module for building a trusted scheduling model for power distribution network computing resources is responsible for constructing local and global large-scale model fine-tuning models, parameter transmission models, FDI attack models, privacy entropy models, and overall large-scale model fine-tuning latency models, as well as constructing a trusted scheduling optimization problem for computing resources. DQN network module: used to maintain a set of DQN neural networks and an experience replay pool for each terminal; Q-function reconstruction module: Used to reconstruct the Q-function with the help of a large model. The reconstructed Q-function contains two parts: the value function and the large model heuristic function. Trusted scheduling module for computing resources: used to optimize channel and dataset scheduling decisions based on the Q value estimated by the main network for each terminal; FDI attack detection module: used to compare the local large model loss function and the global large model loss function of the terminal based on the test dataset to verify whether the parameters are subject to FDI attack; Global Fine-Tuning Module: Used to fine-tune the global model based on the results of the FDI attack detection module; Global large model module: Based on historical information, it estimates channel gain and electromagnetic interference information and reconstructs the Q function, thereby providing assistance for reliable scheduling of computing resources.

Images