Large-model-assisted distribution network calculation resource trusted scheduling method, device and system

Abstract

The invention relates to a large-model-assisted distribution network calculation resource trusted scheduling method, device and system, and belongs to the technical field of telecommunication. According to the method, the global loss function and the total fine tuning time delay of the large model are jointly optimized. By dynamically adjusting the optimization weight of the time delay and learning the matched resource scheduling strategy, the balance between the time delay and the convergence performance is realized. In addition, the attacked large model parameters are prevented from participating in global fine tuning through FDI attack detection, and the security of the large model and the resource scheduling credibility are improved. A resource scheduling strategy is learned through the DQN, and a large model heuristic function is introduced to assist Q function reconstruction of the DQN. The large model heuristic function can score newly generated resource scheduling actions in combination with context information such as historical state information, actions and rewards, so that the learning direction of the DQN network is heuristic, and more accurate strategy optimization is supported. Meanwhile, unknown state information is predicted for resource scheduling optimization by adopting a large model, and the learning ability of the algorithm is improved.

Description

Technical Field

[0001] The present invention relates to a large-model-assisted distribution network transmission and computing resource trusted scheduling method, device and system, belonging to the field of electrical communication technology. Background Art

[0002] At present, with the rapid development of power systems and the improvement of the level of intelligence, the operation and management of distribution networks are facing unprecedented challenges. In order to improve the accuracy and efficiency of distribution network management and control, it is necessary to realize the digitization and intelligence of operation and management. In this process, generative artificial intelligence technology, especially the application of large models, has become a key technical support. The large model is a deep learning model trained with large-scale data. By fine-tuning the pre-trained large model, it can be adapted to the operating environment and task requirements in the distribution network field. The fine-tuning process first involves the deployment of power Internet of Things terminals, which are responsible for collecting grid operation data. Subsequently, the large model is fine-tuned locally using this data to adapt to the characteristics of the local grid. Finally, the results of local fine-tuning are uploaded to the center for global fine-tuning to achieve global optimization of model parameters. Large model fine-tuning has strict requirements on convergence and latency. First, the dynamic nature of the distribution network requires the model to respond quickly to changes in the grid state, which means that the fine-tuning process must be completed in a short time to keep the model consistent with the real-time grid operation state. Secondly, the convergence performance of the model is directly related to the safety and economy of grid operation. A fast and accurate convergence process can reduce the risk and cost of grid operation. Therefore, large model fine-tuning must not only ensure that model parameters can quickly adapt to grid changes, but also ensure the stability and reliability of this process to meet the needs of efficient and safe operation of the distribution network.

[0003] In order to further improve the operating efficiency and response speed of the distribution network, it is necessary to flexibly and intelligently schedule the transmission, computing and other resources in the network. However, this process is often carried out in an untrusted environment and faces severe security challenges. Attackers may use various means to attack the distribution network. One typical attack is the false data injection (FDI) attack. The FDI attack affects the fine-tuning process of the large model by injecting false data into the power grid, resulting in model decision errors, which in turn has a serious impact on the safe and stable operation of the distribution network. The mechanism of the FDI attack is that the attacker intercepts or tampers with the data in the transmission process and injects carefully designed false information to mislead the learning and decision-making process of the large model. This attack may not only cause the large model to fail to converge correctly, but also cause instability in the operation of the power grid and cause power grid failures. Therefore, the fine-tuning of the large model of the distribution network and the optimization of the scheduling of transmission and computing resources in an untrusted environment face the following challenges:

[0004] On the one hand, in order to improve the accuracy of distribution network control, it is necessary to improve the convergence performance of the large model, that is, to minimize the loss function that reflects the deviation between the output value of the large model and the actual value. At the same time, the fine-tuning of the large model needs to be carried out under low latency to maintain its real-time consistency with the growing knowledge of distribution network operation. In addition, affected by FDI attacks, the fine-tuning of the large model is carried out in an untrusted environment, which makes it difficult for the model to converge or even catastrophic forgetting. On the other hand, although traditional machine learning-based resource scheduling optimization methods, such as deep Q networks (DQN), can learn resource scheduling strategies when information such as optimization modeling and network environment is unknown, they have great limitations in dealing with multi-dimensional resource allocation optimization problems. Specifically, the multi-dimensional resource scheduling problem of channels (transmission) and data sets (computation) has high complexity. The DQN network needs to learn in a large optimization space, which leads to the problem of large uncertainty in the learning direction of the algorithm. Therefore, it is urgent to propose a large model-assisted distribution network transmission and computing resource trusted scheduling method, device and system to realize the distribution network transmission and computing resource trusted scheduling. Summary of the invention

[0005] In order to solve the above technical problems, the purpose of the present invention is to provide a large-model-assisted distribution network computing resource trusted scheduling method, device and system.

[0006] A large model-assisted distribution network computing resource trusted scheduling method of the present invention has the following specific scheduling steps:

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

[0008] Transform the above optimization problem and model it as a Markov decision process;

[0009] Through DQN learning resource scheduling strategy, and introducing large model heuristic function to assist DQN Q function reconstruction, solve the above Markov decision process;

[0010] The above solution results are used to detect FDI attacks for trusted scheduling of computing resources. By discarding the attacked large model parameters and reducing the adverse effects of false data on global fine-tuning, trusted scheduling of computing resources can be achieved.

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

[0012] Define I terminals, the set is represented by D = {d1,…,d i ,…,d I}; Define the tth iteration terminal d i The local large model is LLMi,t , which is determined by the parameter w i,t , context c i,t , and the prompt word q i,t Composition, expressed as

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

[0014] d i Receive the global large model parameter w sent by the edge controller g,t-1 To synchronize its local large model parameters w i,t-1 ; Then, d i Use the dataset to fine-tune w i,t-1 ; Define the dataset size set as A i ={α min ,…,|C i |}, where |C i | means d i The dataset size, α min Represents the minimum optional dataset size;

[0015] Define x k and k are the input label sequence and target output of the kth sample of the dataset respectively; x k Contains context and question, y k is the target probability distribution of 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 of each sample; based on the cross entropy function, the gradient descent method is used to update the local large model parameter w of the tth iteration i,t ;

[0019] d i The local fine-tuning delay and fine-tuning energy consumption are related to the number of floating-point operations of all labels in its parameters γ i 、Parameter size|w i,t |, and fine-tuning the cycle e i The local fine-tuning delay and fine-tuning energy consumption are respectively expressed as:

[0020]

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

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

[0023] Global fine-tuning is expressed as

[0024]

[0025] In the formula, is the set of terminals accessing the channel in the tth iteration; ρ i Indicates d i The business priorities supported by the large model; define L g,t (w g,t ) is the global loss function, which is used to measure the convergence performance of the global large model and is expressed as:

[0026]

[0027] Furthermore, the parameter transmission model is:

[0028] Define J channels, the set is represented by N = {n1,…,n j ,…,n J}, define the channel selection variable as β i,j,t ∈{0,1},β i,j,t =1 means at the tth iteration d i Select channel n j Transmission parameters; d i By j The rate of transmission parameters is expressed as

[0029]

[0030] In the formula, B i,j is d i Medium Channel n j The transmission bandwidth, P i,t is d i Transmission power, g i,j,t is the channel gain, δ0 is the Gaussian white noise power, I i,j,t is the electromagnetic interference power; d i By j The delay and energy consumption of transmission parameters are expressed as:

[0031]

[0032] Furthermore, the FDI attack model is:

[0033] definition The false data injected is the FDI attack model.

[0034]

[0035] In the formula, θ i,t ∈{0,1} is the attack indicator variable, θ i,t =1 means d i The parameters of are attacked by FDI.

[0036] Furthermore, the privacy entropy up to the tth iteration is expressed as:

[0037]

[0038] In the formula, Indicates that up to the tth iteration d i By j The probability of transmitting parameters; when d i The privacy entropy is maximized when the probability of transmitting parameters through all channels is equal.

[0039] Furthermore, the total large model fine-tuning delay model is:

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

[0041]

[0042] In the formula, τ g Fine-tune the latency globally.

[0043] Furthermore, the constructed trusted scheduling optimization problem of computing resources is:

[0044] Through the coordinated scheduling of distribution network transmission and computing resources, the weighted sum of the global loss function and the total large model fine-tuning delay is minimized, which is specifically expressed as:

[0045]

[0046] Where V τ The weight of the total model fine-tuning delay is used to unify the order of magnitude and weigh the relationship between optimization objectives; as well as They represent the decision variables for data set scheduling and channel scheduling respectively; F i,min is d i Tolerable privacy entropy lower bound, E i,max is d ibattery capacity; C1 represents the data set scheduling constraint, C2, C3, and C4 represent the channel scheduling constraints, that is, each terminal can only select one channel for parameter transmission; C5 and C6 represent the long-term energy consumption and privacy entropy constraints, respectively.

[0047] Furthermore, the optimization problem is transformed and modeled as a Markov decision process as follows:

[0048] Construct the privacy entropy virtual queue and energy consumption virtual queue corresponding to constraints C5 and C6; expressed as:

[0049]

[0050] when and The queue is in a stable state, 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 follows:

[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 at the tth iteration are expressed as and d i The state space of is backlogged by the privacy entropy virtual queue of the tth iteration Energy consumption virtual queue backlog at the tth iteration Channel gain estimated by large model And the electromagnetic interference power composition estimated by the large model Expressed as

[0054] 2) Action space: d i The action space of is composed of channel scheduling and data set scheduling actions, expressed as in is the Cartesian product;

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

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

[0057] The edge controller maintains a set of DQN neural networks for each terminal, that is, a parameter The main network and parameters are target network; at the beginning of each iteration, the edge controller optimizes the channel and data set 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 main network parameters 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 contains 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 the historical state space, action space, reward and other contextual information, takes the current iteration state space and the action made by the controller as the problem, and evaluates the channel and data set scheduling decisions of the current iteration, thereby inspiring the learning direction of the algorithm; the Q function reconstruction is expressed as:

[0058]

[0059] In the formula, represents the value function; represents the large model heuristic function, expressed as

[0060]

[0061] In the formula, represents the heuristic context information of the input, Indicates the inspiration question, represents the channel and data set scheduling decision made by the controller at the tth iteration; represents the heuristic answer output by the big model, PARSE{.} represents the PARSE function, which parses the JSON-formatted answer of the big model into a numerical score. The steps of channel-dataset collaborative scheduling based on big model-assisted learning are introduced as follows:

[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 beginning 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 to the channel selects the action with the largest Q value from the remaining actions; this process is repeated until all terminals are assigned to the channel; each terminal executes the selected action

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

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

[0066]

[0067] in is the sampled empirical data set, is the discount factor; based on the loss function, update according to the gradient descent method Update every several iterations

[0068] Furthermore, the specific steps of performing FDI attack detection on the solution results for trusted scheduling of computing resources are as follows:

[0069] Set the test data set 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 attacked by FDI; 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 is related to w g,t-1 When the deviation between the loss functions exceeds a threshold ξ, that is, Indicates d i The transmitted parameters are attacked by FDI, i.e. Here, represents the attack detection variable; by discarding the large model parameters under attack, the adverse impact of false data on global fine-tuning is reduced, and the trusted scheduling of computing resources is achieved; formula (7) is rewritten as:

[0072]

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

[0074] A large model-assisted distribution network computing resource trusted scheduling device, deployed in an edge controller, includes:

[0075] Power module: used to power the trusted scheduling device of the distribution network computing resources assisted by the large model;

[0076] Communication module: used to communicate with the terminal, send the terminal's channel and data set scheduling strategy 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 model fine-tuning models, parameter transmission models, FDI attack models, privacy entropy models, and total large model fine-tuning delay models, and building trusted scheduling optimization problems for computing resources.

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

[0079] Q function reconstruction module: used to reconstruct the Q function with the help of the big model. The reconstructed Q function includes two parts: the value function and the big model heuristic function.

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

[0081] FDI attack detection module: used to compare the local large model loss function of the terminal with the global large model loss function based on the test data set to verify whether the parameters are attacked by FDI;

[0082] Global fine-tuning module: used to fine-tune the global large model based on the results of the FDI attack detection module;

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

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

[0085] Includes local device layer, edge layer, and global large model layer;

[0086] In the local device layer, the power IoT terminal is deployed on the power distribution equipment, fine-tunes the local big model according to the latest global big model parameters and channel and data set 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. A large-model-assisted distribution network computing resource trusted scheduling device is deployed in the edge server. The power Internet of Things terminal transmits the local large-model parameters to the edge server through the 4G / 5G / 6G channel. 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, electromagnetic interference information, and reconstruct the Q function, thereby assisting in resource allocation optimization; the global large model layer is maintained by the edge controller, which is responsible for performing global large model fine-tuning.

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

[0090] 1. The present invention jointly optimizes the global loss function and total fine-tuning delay of the large model. By dynamically adjusting the optimization weight of the delay and learning the matching resource scheduling strategy, a trade-off between delay and convergence performance is achieved. In addition, the present invention prevents the attacked large model parameters from participating in global fine-tuning through FDI attack detection, thereby improving the security of the large model and the credibility of resource scheduling.

[0091] 2. The present invention proposes a trusted scheduling algorithm for distribution network computing resources assisted by a large model. The resource scheduling strategy is learned through DQN, and a large model heuristic function is introduced to assist the reconstruction of the Q function of DQN. The large model heuristic function can score the newly generated resource scheduling actions in combination with historical state information, actions, rewards and other contextual information, thereby inspiring the learning direction of the DQN network and supporting more accurate strategy optimization. At the same time, the present invention uses a large model to optimize resource scheduling to predict unknown state information, such as channel gain and electromagnetic interference power, to improve the learning ability of the algorithm.

[0092] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention and implement it according to the contents of the specification, the following is a detailed description of the preferred embodiments of the present invention in conjunction with the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0093] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show a certain embodiment of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0094] Figure 1 It is a flow chart of the reliable scheduling method of distribution network computing resources assisted by the large model of the present invention;

[0095] Figure 2 It is a diagram of a trusted scheduling device for distribution network computing resources assisted by a large model of the present invention;

[0096] Figure 3 It is a diagram of a trusted scheduling system for distribution network computing resources assisted by a large model of the present invention. DETAILED DESCRIPTION

[0097] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0098] The present invention designs a reliable scheduling method for distribution network computing resources assisted by a large model, the process is as follows Figure 1 shown.

[0099] 1.1. Construction of a trusted scheduling model for distribution network computing resources

[0100] The optimization time is divided into T iterations, and the set is represented as T = {1,…,t,…,T}. In each iteration, the edge controller sends the terminal's channel and data set scheduling strategy and the latest global large model parameters. Then, the terminal fine-tunes the local large model parameters according to the scheduled data set and transmits the large model parameters to the edge controller through the allocated 4G / 5G / 6G channel. 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 Fine-tuning the local large model

[0102] Fine-tuning the large model can ensure that the parameters of the pre-trained model remain unchanged, saving computing resources while preventing catastrophic forgetting. Consider I terminals, the set is represented by D = {d1,…,d i ,…,d I}. Define the tth iteration terminal d i The local large model is LLM i,t , which is determined by the parameter w i,t , context c i,t , and the prompt word q i,t Composition, specifically expressed as

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

[0104] d i Receive the global large model parameter w sent by the edge controller g,t-1 To synchronize its local large model parameters w i,t-1 Then, d i Use the dataset to fine-tune w i,t-1 The dataset size set is defined as A i ={α min ,…,|C i |}, where |C i | means d i The dataset size, α min Represents the minimum optional dataset size.

[0105] The cross entropy function is used to measure the convergence performance of the large model. The smaller the cross entropy function, the better the convergence performance of the large model. Define x k and k are the input label sequence and target output of the kth sample of the dataset respectively. k Contains context and question, y k is the target probability distribution of the output label.

[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 of each sample. Based on the cross entropy function, the gradient descent method is used to update the local large model parameter w of the tth iteration. i,t .

[0109] d i The local fine-tuning delay and fine-tuning energy consumption are related to the number of floating-point operations of all labels in its parameters γ i 、Parameter size|w i,t |, and fine-tuning the cycle e i The local fine-tuning delay and fine-tuning energy consumption are expressed as

[0110]

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

[0112] 1.1.2 Parameter Transmission Model

[0113] After fine-tuning the local large model, the terminal transmits the parameters to the edge controller through the channel. Consider J channels, the set is 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 means at the tth iteration d i Select channel n j Transmission parameters. i By j The rate of transmission parameters is expressed as

[0114]

[0115] In the formula, B i,j is d i Medium Channel n j The transmission bandwidth, P i,t is d i Transmission power, g i,j,t is the channel gain, δ0 is the Gaussian white noise power, I i,j,t is the electromagnetic interference power. i By j The delay and energy consumption of transmission parameters are expressed as

[0116]

[0117] 1.1.3FDI attack model

[0118] FDI attacks intercept large model parameters during transmission and inject false data to destroy model convergence performance. The false data injected is the FDI attack model.

[0119]

[0120] In the formula, θ i,t ∈{0,1} is the attack indicator variable, θ i,t =1 means d i The parameters of are attacked by FDI.

[0121] 1.1.4 Fine-tuning 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 expressed as

[0123]

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

[0125]

[0126] 1.1.5 Privacy Entropy Model

[0127] Privacy entropy is defined as the degree of confusion of the terminal transmission parameters. The greater the degree of confusion of the local large model parameters being distributed to different channels for transmission, the greater the privacy entropy. The privacy entropy up to the tth iteration is expressed as

[0128]

[0129] In the formula, Indicates that up to the tth iteration d i By j The probability of transmitting parameters. i The privacy entropy is maximized when the probability of transmitting parameters through all channels is equal.

[0130] 1.1.5 Fine-tuning the delay model for the overall large model

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

[0132]

[0133] In the formula, τ g Fine-tune the latency globally.

[0134] 1.1.6 Modeling and transformation of trusted scheduling optimization problems for computing resources

[0135] The present invention aims to solve the problem of trusted scheduling of transmission and computing resources in distribution networks under FDI attacks. First, increasing the size of the data set used by the terminal can make local fine-tuning more sufficient, but this will increase the local fine-tuning delay and fine-tuning cost. Second, changing the channel scheduling strategy will affect the terminal parameter transmission delay, thereby changing the fine-tuning cost. Therefore, the optimization goal is to minimize the weighted sum of the global loss function and the total large model fine-tuning delay through the coordinated scheduling of distribution network transmission and computing resources, which is specifically expressed as

[0136]

[0137] Where V τ Represents the weight of the total model fine-tuning delay, which is used to unify the order of magnitude and weigh the relationship between optimization objectives. as well as They represent the decision variables for data set scheduling and channel scheduling respectively. i,min is d i Tolerable privacy entropy lower bound, E i,max is d i The battery capacity of the UE is 1000M. C1 represents the data set scheduling constraint, C2, C3, and C4 represent the channel scheduling constraints, that is, each terminal can only select one channel for parameter transmission. C5 and C6 represent the long-term energy consumption and privacy entropy constraints, respectively.

[0138] Due to the coupling between the short-term resource allocation strategy optimization and the long-term privacy entropy and energy consumption constraints, the present invention constructs a privacy entropy virtual queue and an energy consumption virtual queue corresponding to constraints C5 and C6. It is expressed as

[0139]

[0140] when and In the stable state of the queue, C5 and C6 can be regarded as satisfied. P1 can be transformed into

[0141]

[0142] 1.2 Design of a trusted scheduling algorithm for distribution network computing resources assisted by a large model

[0143] The present invention designs a large-model-assisted distribution network computing resource trusted scheduling algorithm. First, the channel and data set scheduling strategy is solved based on large-model-assisted learning. Then, the edge controller uses the test data set to perform FDI attack detection, realize the trusted scheduling of computing resources, and prevent false data from causing adverse effects on global fine-tuning.

[0144] 1.2.1 Channel-dataset collaborative scheduling based on large model-assisted learning

[0145] To solve P2, the present invention models it as a Markov decision process, whose key elements include state space, action space and reward function, which are specifically introduced as follows:

[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 at the tth iteration are expressed as and d i The state space of is backlogged by the privacy entropy virtual queue of the tth iteration Energy consumption virtual queue backlog at the tth iteration Channel gain estimated by large model And the electromagnetic interference power composition estimated by the large model Expressed as

[0147] 2) Action space: d i The action space of is composed of channel scheduling and data set scheduling actions, expressed as in is the Cartesian product.

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

[0149] DQN provides a model-free solution to the Markov decision process mentioned above. The edge controller maintains a set of DQN neural networks for each terminal, that is, a parameter The main network and parameters are target network. At the beginning of each iteration, the edge controller optimizes the channel and data set 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 main network parameters are updated with the assistance of the target network. However, due to the large optimization space of P2, the DQN algorithm is prone to large uncertainty in the learning direction. To solve this problem, the present invention reconstructs the Q function with the assistance of a large model. The reconstructed Q function includes 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 rewards, and takes the current iteration state space and the actions made by the controller as problems to evaluate the channel and data set scheduling decisions of the current iteration, thereby inspiring the learning direction of the algorithm. The Q function reconstruction is expressed as

[0150]

[0151] In the formula, Represents the value function. represents the large model heuristic function, expressed as

[0152]

[0153] In the formula, represents the heuristic context information of the input, Indicates the inspiration question, represents the channel and data set scheduling decision made by the controller at the tth iteration. represents the heuristic answer output by the big model, and PARSE{.} represents the PARSE function, which parses the JSON-formatted answer of the big model into a numerical score. The steps of channel-dataset collaborative scheduling based on big model-assisted learning are introduced as follows.

[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 beginning 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 terminals that are not assigned to the channel select the action with the largest Q value from the remaining actions. This process is repeated until all terminals are assigned to the channel. Each terminal executes the selected action.

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

[0157] Step 4: The controller randomly samples an experience data set from the experience pool and calculates the loss function of the DQN network based on the reconstructed Q function, which is expressed as

[0158]

[0159] in is the sampled empirical data set, is the discount factor. Based on the loss function, update according to the gradient descent method Update every several iterations

[0160] 1.2.2 FDI attack detection for trusted scheduling of computing resources

[0161] The present invention designs an FDI attack detection mechanism for trusted scheduling of computing resources. The principle of attack detection is to give a test data set 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 attacked by FDI. Tes The local large model test loss function and the global test loss function are calculated as

[0162]

[0163] when The loss function is related to w g,t-1 When the deviation between the loss functions exceeds a threshold ξ, that is, Indicates d i The transmitted parameters are attacked by FDI, i.e. Here, represents the attack detection variable. By discarding the large model parameters that are attacked, the adverse impact of false data on global fine-tuning is reduced, and the trusted scheduling of computing resources is achieved. Formula (7) can be rewritten as

[0164]

[0165] In the formula, is d in the tth iteration iThe dataset size used for local training. The edge controller performs global fine-tuning based on formula (19).

[0166] 2. A trusted dispatching device for distribution network computing resources assisted by a large model

[0167] The present invention proposes a large model-assisted distribution network computing resource trusted scheduling device such as Figure 2 As shown, it is deployed in the edge controller, including power module, communication module, distribution network computing resource trusted scheduling model construction module, DQN network module, Q function reconstruction module, computing resource trusted scheduling module, FDI attack detection module, global fine-tuning module, and global large model module. The specific introduction is as follows.

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

[0169] Communication module: responsible for communicating with the terminal, sending the terminal's channel and data set scheduling strategy and the latest global large model parameters, and receiving the local large model parameters uploaded by the terminal.

[0170] Distribution network computing resource trusted scheduling model construction module: The distribution network computing resource trusted scheduling model construction module is responsible for constructing local and global large model fine-tuning models, parameter transmission models, FDI attack models, privacy entropy models and total large model fine-tuning delay models, and constructing the trusted scheduling optimization problem of 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 help of the big model. The reconstructed Q function contains two parts: the value function and the big model heuristic function.

[0173] Trusted scheduling module for computing resources: optimizes channel and data set scheduling decisions based on the Q value estimated by the main network of each terminal.

[0174] FDI attack detection module: compares the local large model loss function of the terminal with the global large model loss function based on the test data set to verify whether the parameters are attacked by FDI.

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

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

[0177] 3. A trusted dispatching system for distribution network computing resources assisted by a large model

[0178] The large model-assisted distribution network computing resource trusted scheduling system considered by the present invention is as follows Figure 3 As shown, it includes the local device layer, edge layer, and global large model layer.

[0179] In the local device layer, the power IoT terminal is deployed on the distribution equipment, which fine-tunes the local big model according to the latest global big model parameters and channel and data set scheduling strategies issued by the edge layer, and executes the corresponding scheduling strategies.

[0180] The edge layer includes a base station and an edge controller. The edge server is deployed with a large model-assisted distribution network computing resource trusted scheduling device. The power IoT terminal transmits the local large model parameters to the edge server through the 4G / 5G / 6G channel. In addition, the edge layer can also 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 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.

[0182] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. It should be pointed out that a person skilled in the art can make several improvements and modifications without departing from the technical principles of the present invention, and these improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A large model-assisted distribution network computing resource trusted scheduling method, characterized in that The specific scheduling steps are: Construct local large model fine-tuning model and global large model fine-tuning model, parameter transmission model, FDI attack model, privacy entropy model and total large model fine-tuning delay model, and construct the trusted scheduling optimization problem of transmission and computing resources; Transform the above optimization problem and model it as a Markov decision process; Through DQN learning resource scheduling strategy, and introducing large model heuristic function to assist DQN Q function reconstruction, solve the above Markov decision process; The above solution results are used to detect FDI attacks for trusted scheduling of computing resources. By discarding the attacked large model parameters and reducing the adverse effects of false data on global fine-tuning, trusted scheduling of computing resources can be achieved.

2. According to claim 1, a large model-assisted distribution network computing resource trusted scheduling method is characterized by: The local large model fine-tuning model is: Define I terminals, the set is represented by D = {d1,…,d i ,…,d I }; Define the tth iteration terminal d i The local large model is LLM i,t , which is determined by the parameter w i,t , context c i,t , and the prompt word q i,t Composition, expressed as LLM i,t =Γ(w i,t ,c i,t ,q i,t ) (1) d i Receive the global large model parameter w sent by the edge controller g,t-1 To synchronize its local large model parameters w i,t-1 ; Then, d i Use the dataset to fine-tune w i,t-1 ; Define the dataset size set as A i ={α min ,…,|C i |}, where |C i | means d i The dataset size, α min Represents the minimum optional dataset size; Define x k and k are the input label sequence and target output of the kth sample of the dataset respectively; x k Contains context and question, y k is the target probability distribution of the output label; d i The cross entropy function is In the formula, l CE (w i,t-1 ,x k ,y k ) represents the cross entropy function of each sample; based on the cross entropy function, the gradient descent method is used to update the local large model parameter w of the tth iteration i,t ; d i The local fine-tuning delay and fine-tuning energy consumption are related to the number of floating-point operations of all labels in its parameters γ i 、Parameter size|w i,t |, and fine-tuning the cycle e i The local fine-tuning delay and fine-tuning energy consumption are respectively expressed as: In the formula, μ i,t Indicates d i Available GPU resources, λ i represents the available GPU computing speed, κ i Indicates the available GPU computing efficiency.

3. According to the large model-assisted distribution network computing resource trusted scheduling method of claim 1, it is characterized by: The global large model fine-tuning model is: Global fine-tuning is expressed as Where D t a is the set of terminals accessing the channel in the tth iteration; ρ i Indicates d i The business priorities supported by the large model; define L g,t (w g,t ) is the global loss function, which is used to measure the convergence performance of the global large model and is expressed as:

4. According to the large model-assisted distribution network computing resource trusted scheduling method of claim 1, it is characterized by: The parameter transmission model is: Define J channels, the set is represented by N = {n1,…,n j ,…,n J }, define the channel selection variable as β i,j,t ∈{0,1},β i,j,t =1 means at the tth iteration d i Select channel n j Transmission parameters; d i By j The rate at which the parameters are transmitted is expressed as In the formula, B i,j is d i Medium channel n j The transmission bandwidth, P i,t is d i Transmission power, g i,j,t is the channel gain, δ0 is the Gaussian white noise power, I i,j,t is the electromagnetic interference power; d i By j The delay and energy consumption of transmission parameters are expressed as 5. According to the large model-assisted distribution network computing resource trusted scheduling method of claim 1, it is characterized by: The FDI attack model is: definition The false data injected is the FDI attack model. In the formula, θ i,t ∈{0,1} is the attack indicator variable, θ i,t =1 means d i The parameters of are attacked by FDI.

6. According to a large model-assisted distribution network computing resource trusted scheduling method according to claim 1, it is characterized by: The privacy entropy up to the tth iteration is expressed as: In the formula, Indicates that up to the tth iteration d i By j The probability of transmitting parameters; when d i The privacy entropy is maximized when the probability of transmitting parameters through all channels is equal.

7. According to claim 1, a large model-assisted distribution network computing resource trusted scheduling method is characterized by: The total large model fine-tuning delay model is: The total large model fine-tuning delay consists of the total local fine-tuning delay and transmission delay of all terminals, as well as the global fine-tuning delay of the edge controller, which is expressed as In the formula, τ g Fine-tune the latency globally.

8. The method for reliable scheduling of distribution network computing resources assisted by a large model according to claim 1 is characterized by: The constructed trusted scheduling optimization problem of computing resources is: Through the coordinated scheduling of distribution network transmission and computing resources, the weighted sum of the global loss function and the total large model fine-tuning delay is minimized, which is specifically expressed as: Where V τ The weight of the total model fine-tuning delay is used to unify the order of magnitude and weigh the relationship between optimization objectives; as well as They represent the decision variables for data set scheduling and channel scheduling respectively; F i,min is d i Tolerable privacy entropy lower bound, E i,max is d i battery capacity; C1 represents the data set scheduling constraint, C2, C3, and C4 represent the channel scheduling constraints, that is, each terminal can only select one channel for parameter transmission; C5 and C6 represent the long-term energy consumption and privacy entropy constraints, respectively.

9. The method for reliable scheduling of distribution network computing resources assisted by a large model according to claim 1 is characterized by: The transformation of the optimization problem and modeling it as a Markov decision process is specifically as follows: Construct the privacy entropy virtual queue and energy consumption virtual queue corresponding to constraints C5 and C6; expressed as: when and The queue is in a stable state, C5 and C6 are considered satisfied; P1 is transformed into: To solve P2, it is modeled as a Markov decision process, whose key elements include state space, action space and reward function, as follows: 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 at the tth iteration are expressed as and d i The state space of is backlogged by the privacy entropy virtual queue of the tth iteration Energy consumption virtual queue backlog at the tth iteration Channel gain estimated by large model And the electromagnetic interference power composition estimated by the large model Expressed as 2) Action space: d i The action space of is composed of channel scheduling and data set scheduling actions, expressed as in is the Cartesian product; 3) Reward function: Since P2 is a minimization problem, the reward function Ω i,t Defined as the negative of the P2 optimization objective.

10. The method for reliable scheduling of distribution network computing resources assisted by a large model according to claim 1 is characterized by: The specific method of solving the Markov decision process by learning resource scheduling strategy through DQN and introducing large model heuristic function to assist DQN Q function reconstruction is as follows: The edge controller maintains a set of DQN neural networks for each terminal, that is, a parameter The main network and parameters are target network; At the beginning of each iteration, the edge controller optimizes the channel and data set 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 main network parameters 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: the value function and the large model heuristic function; The value function reflects the expected cumulative reward after taking a certain action in a certain state; the large model heuristic function combines historical state space, action space, reward and other contextual information, takes the current iteration state space and the action taken by the controller as the problem, evaluates the channel and data set scheduling decision of the current iteration, and thus inspires the learning direction of the algorithm; the Q function reconstruction is expressed as: In the formula, represents the value function; represents the large model heuristic function, expressed as In the formula, represents the heuristic context information of the input, Indicates the inspiration question, represents the channel and data set scheduling decision made by the controller at the tth iteration; represents the heuristic answer output by the big model, PARSE{.} represents the PARSE function, which parses the JSON-formatted answer of the big model into a numerical score. The steps of channel-dataset collaborative scheduling based on big model-assisted learning are introduced as follows: Step 1: Initialize the DQN network parameters for each terminal and Set the reward function Ω i,0 =0; Step 2: At the beginning 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 to the channel selects the action with the largest Q value from the remaining actions. This process is repeated until all terminals are assigned to the channel. Each terminal performs the selected action Step 3: At the end of each iteration, the controller observes the global loss function, the total large model fine-tuning delay, the privacy entropy virtual queue backlog, the energy consumption virtual queue backlog and other performance, and calculates the reward; then, the controller updates the state space S of the t+1th iteration i,t+1 ; The controller generates experience data And store it in the experience replay pool; Step 4: The controller randomly samples an experience data set from the experience pool and calculates the loss function of the DQN network based on the reconstructed Q function, which is expressed as: in is the sampled empirical data set, is the discount factor; based on the loss function, update according to the gradient descent method Update every several iterations 11. The method for reliable scheduling of distribution network computing resources assisted by a large model according to claim 1 is characterized by: The specific steps of performing FDI attack detection on the solution results for trusted scheduling of computing resources are as follows: Set the test data set 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 attacked by FDI; based on C Tes The local large model test loss function and the global test loss function are calculated as follows: when The loss function is related to w g,t-1 When the deviation between the loss functions exceeds a threshold ξ, that is, Indicates d i The transmitted parameters are attacked by FDI, i.e. Here, represents the attack detection variable; by discarding the large model parameters under attack, the adverse impact of false data on global fine-tuning is reduced, and the trusted scheduling of computing resources is achieved; formula (7) is rewritten as: In the formula, is d in the tth iteration i The dataset size used for local training; the edge controller performs global fine-tuning based on formula (19).

12. A large model-assisted distribution network computing resource trusted scheduling device, characterized in that: Deployed in the edge controller, including: Power module: used to power the trusted scheduling device of the distribution network computing resources assisted by the large model; Communication module: used to communicate with the terminal, send the terminal's channel and data set scheduling strategy and the latest global large model parameters, and receive the local large model parameters uploaded by the terminal; The trusted scheduling model construction module for distribution network computing resources is responsible for building local and global large model fine-tuning models, parameter transmission models, FDI attack models, privacy entropy models, and total large model fine-tuning delay models, and building trusted scheduling optimization problems for computing resources. DQN network module: used to maintain a set of DQN neural networks and experience replay pools for each terminal; Q function reconstruction module: used to reconstruct the Q function with the help of the big model. The reconstructed Q function includes two parts: the value function and the big model heuristic function. Trusted scheduling module for computing resources: used to optimize channel and data set scheduling decisions based on the Q value estimated by the main network of each terminal; FDI attack detection module: used to compare the local large model loss function of the terminal with the global large model loss function based on the test data set to verify whether the parameters are attacked by FDI; Global fine-tuning module: used to fine-tune the global large model based on the results of the FDI attack detection module; Global large model module: It can estimate channel gain, electromagnetic interference and other information based on historical information, and reconstruct the Q function, thereby providing assistance for the trusted scheduling of transmission and computing resources.

13. A large model-assisted distribution network computing resource trusted scheduling system, characterized by: Includes local device layer, edge layer, and global large model layer; In the local device layer, the power IoT terminal is deployed on the power distribution equipment, fine-tunes the local big model according to the latest global big model parameters and channel and data set scheduling strategies issued by the edge layer, and executes the corresponding scheduling strategies; The edge layer includes a base station and an edge controller. A large-model-assisted distribution network computing resource trusted scheduling device is deployed in the edge server. The power Internet of Things terminal transmits the local large-model parameters to the edge server through the 4G / 5G / 6G channel. In addition, the edge layer can also provide attack detection for the uploaded parameters. The global large model layer can also estimate channel gain, electromagnetic interference information, and reconstruct the Q function, thereby assisting in resource allocation optimization; the global large model layer is maintained by the edge controller, which is responsible for performing global large model fine-tuning.

Images