Resource scheduling method based on risk awareness

Abstract

The invention relates to a resource scheduling method based on risk awareness, and belongs to the technical field of radio communication. The method comprises the following steps: respectively constructing communication models, time delay models and energy consumption models of a space-based network, an air-based network and a ground-based network in a space-air-ground integrated network; under the energy constraint of the unmanned aerial vehicle, the problem is expressed as an integer nonlinear optimization problem with uncertainty, and a conditional risk value CVaR is utilized to measure the check risk; the cost and the risk are evaluated by using double-critic networks, each network independently evaluates an action value function, and then comprehensive evaluation is carried out; through deep neural network parameterization and time sequence difference updating, an optimal task unloading and resource scheduling strategy is realized; the influence of inaccuracy of a single network on strategy updating is reduced, and meanwhile, the learning stability and convergence are improved; the resource scheduling method based on risk awareness provided by the invention not only can meet diversified requirements of users, but also can solve the problem of high risk caused by traditional decisions.

Description

Technical Field

[0001] The present invention belongs to the field of communication technology and relates to a resource scheduling method based on risk perception. Background Art

[0002] With the development of mobile communication technology, the applications running on user devices are becoming more diverse and complex, and their computing needs are becoming more frequent. These mobile applications often require a lot of computing resources and generate high energy consumption during operation. However, current user devices are usually constrained by their own limited computing resources and battery capacity, and it is difficult to meet the application's requirements for low latency and high computing power. Users can offload computing tasks to nearby ground base stations, which can not only reduce costs and reduce the latency of task execution, but also save users' own resources. However, simply relying on offloading tasks to ground base stations is difficult to ensure that task processing can achieve the ideal performance level. The ground network centered on base stations often faces the problem of overload, and may experience network congestion, slow processing, and reduced service quality. Drones and satellites are considered to be a powerful supplement to enhance the ground. Satellite communications are an effective means to provide global communication coverage. In order to reduce transmission latency and resource consumption and provide users with reliable services, Mobile Edge Computing (MEC) technology came into being. MEC servers can be deployed at (Low Earth Orbit Satellite, LEO) satellites to provide computing services to users. Drones have the characteristics of flexible deployment and agile management. They can be used as repeaters to transmit data or provide communication and computing resources on demand. The three are integrated to form an air-ground-space integrated network. The air-ground-space integrated network is a heterogeneous multi-dimensional network. The integration of multiple networks makes its network structure extremely complex, and network resources also show diversified characteristics. Due to the high dynamics and randomness of the arrival of computing tasks, traditional reinforcement learning usually focuses on maximizing cumulative rewards, but in some cases, only considering the maximization of rewards may lead to riskier decisions. Therefore, the scheduling of computing tasks in the air-ground-space integrated network still faces great challenges. Therefore, it is of great significance to design an effective resource allocation strategy that can not only meet the diverse needs of users, but also solve the high-risk problems caused by traditional decision-making. Summary of the invention

[0003] In order to achieve the above object, the present invention provides the following technical solutions:

[0004] A resource scheduling method based on risk perception includes the following steps:

[0005] S1: In the space-ground integrated network, the communication model, delay model and energy consumption model of the space-based network, air-based network and ground-based network are constructed respectively;

[0006] S2: Under the energy constraint of the UAV, the problem is formulated as an integer nonlinear optimization problem with uncertainty, and the conditional value at risk (CVaR) is used to measure the inspection risk.

[0007] S3: Use a dual critic network to evaluate costs and risks, and implement optimal task offloading and resource scheduling strategies through deep neural network (DNN) parameterization and temporal difference (TD) updates.

[0008] Further, the S1 comprises the following steps:

[0009] In ground-to-air communication, the channel gain between user k and drone m is defined as:

[0010]

[0011] Where α0 is the channel gain at reference distance d = 1m, q(i+1) is the position of drone m after it moves to the next time slot, and p k (i) is the position of user k in time slot i, H is the fixed flight altitude of drone m, ||q(i+1)-p k (i)|| 2 +H 2 represents the Euclidean distance between user k and drone m;

[0012] The data rate offloaded to UAV m is expressed as:

[0013]

[0014] Among them, B u is the available spectrum bandwidth from user k to drone m, η k is the bandwidth percentage allocated to user k, requiring ∑η k <1, P u is the user's transmit power on the uplink, g k (i) represents the channel gain, is the noise power, f k (i) Whether there is a blockage between user k and drone m. If there is a blockage, add P NLOS The noise power, P NLOS represents the additional noise power;

[0015] In ground / air-space communications, the channel gain from user k to the satellite is determined by the fixed distance D between the ground and the satellite and the satellite computing power f LEO The data rate from the ground or UAV m to the satellite is calculated as:

[0016]

[0017] in, is the data rate from the ground or UAV m to the satellite, B l represents the communication bandwidth to the satellite, P l is the transmission power on the link, is the noise power;

[0018] set up represents the partial uninstallation decision of user k, is the offloading ratio of user k to drone m, is the offloading ratio of user k to low-orbit satellites;

[0019] Set the partial uninstallation decision of user k, expressed as:

[0020]

[0021] in, is the offloading ratio of user k to drone m, is the offloading ratio of user k to low-orbit satellites.

[0022] Furthermore, the S1 further comprises the following steps:

[0023] By task size b k (i) The number of cycles required per bit s k (i) The computing power of the user f UE , the proportion of local computing Get the local computation delay of user k

[0024] The computing resources allocated by drone m to user k in each time slot are expressed as:

[0025]

[0026] Among them, κ1 is the energy consumption coefficient of UAV m, p k (i) is the computing resource allocated to user k by drone m, expressed as:

[0027] p k (i) = μ k P

[0028] Among them, μ k is the ratio of resource allocation, ∑μ k <1;

[0029] Computational tasks offloaded to drones The number of cycles required per bit s k (i) The computing power of the drone fUAV,k Calculate the computational delay D of the task offloaded to UAV m UAV,k (i);

[0030] The priority factor is expressed as:

[0031]

[0032] Among them, t k is the generation time of the task, t is the current time, the smaller the priority factor is, the higher the priority of the task is and the earlier it is unloaded and processed.

[0033] Furthermore, the S1 further comprises the following steps:

[0034] Energy consumption transmitted to UAV m and calculate energy consumption Respectively expressed as:

[0035]

[0036]

[0037] Depend on and The sum of the local computation delay of user k is obtained

[0038] The flight energy consumption E of UAV m fly (i), expressed as:

[0039] E fly (i)=φ||ν(i)|| 2

[0040] Among them, ν(i) contains the flight speed and flight angle of the drone. The flight speed is obtained by performing the modulo operation ||ν(i)||, which is in the range of [0,ν max ], ν max is the maximum value of the flight speed; β(i) represents the flight angle, β(i)∈[0,2π]; φ represents the relevant parameters of the UAV.

[0041] Further, the S2 comprises the following steps:

[0042] UAV m track It is expressed as:

[0043]

[0044] The unloading strategy Λ is expressed as:

[0045]

[0046] The resource allocation strategy Φ is expressed as:

[0047]

[0048] The delay D(i) is defined as the maximum delay of each part of the processing task, expressed as:

[0049]

[0050] The problem is modeled as:

[0051]

[0052]

[0053]

[0054]

[0055]

[0056]

[0057] Among them, C1 represents the value range of the offloading ratio; C2 represents that the resource allocation to all users does not exceed the total resources; C3 determines whether the delay tolerance of each user k is met; C4 represents that the movement range of the UAV does not exceed the predetermined range; C5 represents that the flight energy consumption, computing energy consumption and transmission energy consumption of the UAV in the time slot do not exceed the energy capacity constraint of the UAV, E b Represents the energy capacity of the drone;

[0058] Using Markov decision making, we get the cost c, expressed as:

[0059]

[0060] Among them, ζ is the penalty for exceeding the delay constraint, o k (i) It is used to determine whether the task delay tolerance of each user k is met;

[0061] The current scheduling problem is described as a constrained Markov decision process, expressed as:

[0062]

[0063] stC1,C2,C3,C4

[0064]

[0065] The Conditional Value at Risk (CVaR) is used as a metric to test risk, which is used to measure the average loss in the worst case scenario under a certain confidence level;

[0066] Let X be the random variable of the drone's energy consumption exceeding the energy capacity limit:

[0067]

[0068] in, is the energy actually consumed by the drone, E b is the energy capacity of the drone; when X>0, it means that the drone energy consumption exceeds the limit;

[0069] Set the confidence level α = 0.05; for a given α, the value at risk (VaR) needs to satisfy P(X≤-VaR) = α; find all data points X that satisfy X>-VaR j (j=1,2,...,m), calculate CVaR, expressed as:

[0070]

[0071] Take CVaR as the main component of the risk function; let the risk function R(X) = CVaR α (X).

[0072] Further, the S3 comprises the following steps:

[0073] The value function of the expected long-term discounted cost V(s) is expressed as:

[0074]

[0075] in, represents the discount factor at the current time slot, which is used to measure the importance of future rewards relative to current rewards;

[0076] The Q-value function is defined to further evaluate the expected long-term discounted cost obtained under the current state-action pair, expressed as:

[0077]

[0078] With the minimum cost as the goal, select the minimum Q value function Q * (s,a)=minQ(s,a) as the optimal Q value; take action a at the same time * Get the optimal Q value; the Q value function is updated according to the temporal difference (TD) equation, expressed as:

[0079]

[0080] Among them, η represents the learning rate, which is used to determine the magnitude of each update;

[0081] Using deep neural networks (DNN) for parameterization, the loss function is expressed as:

[0082]

[0083] in, represents the Q-value function based on DNN, is the TD error;

[0084] The risk function for the part exceeding the energy capacity of the drone is expressed as:

[0085]

[0086] Among them, Ω is a set of defined error states, indicating that the energy consumption of the drone exceeds its energy capacity.

[0087] Furthermore, the S3 further comprises the following steps:

[0088] The expected long-term discounted risk value function It is expressed as:

[0089]

[0090] The expected long-term discounted risk estimated by the Q-value function is expressed as:

[0091]

[0092] Execute greedy action To obtain the optimal Q value

[0093] Assume that the optimal Q values ​​obtained under two different objectives are Q(s i ,a i )and A dynamically updated weight coefficient δ is added between the two to balance them and obtain a new Q value function, which is expressed as:

[0094]

[0095] Furthermore, the S3 further comprises the following steps:

[0096] Initialize the parameterized cost network Risk Network and the action network μ(s|θ μ ); Initialization parameters Experience Replay Weight δ and step size Δ; according to μ(s|θ μ )Select a i , storage experience (s i ,a i ,c i ,r i ,s i+1 )arrive middle;

[0097] According to the target value y j Update, expressed as:

[0098]

[0099] Calculate the network loss L c , expressed as:

[0100]

[0101] Get the best strategy action It is expressed as:

[0102]

[0103] The weights are updated while the parameters are soft updated, expressed as:

[0104]

[0105] if Then δ←δ+Δ, otherwise δ←δ-Δ.

[0106] Furthermore, the maximum value λ of the unloading ratio max is 0.5, and the unloading ratio is constrained, which is expressed as:

[0107]

[0108] The beneficial effects of the present invention are:

[0109] Since the availability of links and the arrival of tasks are highly dynamic, the focus is on minimizing the average processing delay of all tasks; in the traditional integrated air-ground-space network, the communication model, delay model and energy consumption model of the three-layer network are constructed respectively according to the respective characteristics of the three-layer network; in order to solve the instability caused by the dynamic arrival of tasks and optimize the deviation problem of the estimation result caused by a single goal, an integer nonlinear optimization problem with uncertainty is proposed under the energy constraint of the UAV, and the conditional risk value CVaR is introduced to measure the inspection risk; a dual critic network structure is designed to balance cost and risk, and a resource scheduling method based on risk perception is proposed; the method provided by the present invention can achieve a more significant improvement in service quality, and a good trade-off between cost and risk is achieved.

[0110] Other advantages, objectives and features of the present invention will be explained in part in the following description, and in part will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0111] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:

[0112] Figure 1 A network system framework diagram of an embodiment of the present invention;

[0113] Figure 2 The steps of the risk-aware resource scheduling method according to an embodiment of the present invention;

[0114] Figure 3 The present invention is a flowchart of a risk-aware resource scheduling method for solving a problem. DETAILED DESCRIPTION

[0115] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0116] See also Figure 1 , which is a network system framework diagram of an embodiment of the present invention; please refer to Figure 2, which are steps of the risk-aware resource scheduling method according to an embodiment of the present invention; please refer to Figure 3 , is a flow chart of a problem solving method for resource scheduling based on risk perception according to an embodiment of the present invention. Figures 1 to 3 The present invention specifically describes a risk-aware resource scheduling method, which specifically includes the following steps:

[0117] S1: In the integrated space-ground network, the communication model, delay model and energy consumption model of the three-layer network are constructed according to their respective characteristics;

[0118] S1.1: The system considered consists of K users, M drones and a LEO satellite, where user k = {1,2,...,K} and drone m = {1,2,...,M}. Assuming that the satellite always covers the area within a period T, and the time is divided into I time slots, then T = {1,2,...,I}. The positions of user k and drone m in each time slot are denoted by p k (i), q(i). The task generated by the user in time slot i is represented as a tuple where b k (i) represents the task size generated by user k in time slot i, represents the delay tolerance of all tasks generated by the same user k, s k (i) represents the number of cycles per bit required to process the task.

[0119] S1.2: The channel gain between user k and drone m is defined as:

[0120]

[0121] Where α0 is the channel gain at reference distance d = 1m, q(i+1) is the position of drone m after it moves to the next time slot, and p k (i) is the position of user k in time slot i, H is the fixed flight altitude of drone m, ||q(i+1)-p k (i)|| 2 +H 2 represents the Euclidean distance between user k and drone m.

[0122] Data rate offloaded to UAV m It is expressed as:

[0123]

[0124] Among them, B u is the available spectrum bandwidth from user k to drone m, η k is the bandwidth percentage allocated to user k, requiring ∑η k <1, Pu is the user's transmit power on the uplink, g k (i) represents the channel gain, is the noise power, f k (i) is used to indicate whether there is a blockage between user k and drone m. If there is a blockage, P is added NLOS The noise power, P NLOS represents the additional noise power.

[0125] Ground / air-space communication: The channel gain from user k to the satellite can be determined by the fixed distance D between the ground and the satellite and the satellite computing power f LEO The data rate offloaded to LEO is calculated as:

[0126]

[0127] in, is the data rate from the ground or UAV m to the satellite, B l represents the communication bandwidth to the satellite, P l is the transmission power on the link, is the noise power.

[0128] S1.3: Set the partial offloading decision of user k, expressed as:

[0129]

[0130] in, is the offloading ratio of user k to drone m, is the offloading ratio of user k to low-orbit satellites. The maximum offloading ratio is set to 0.5, expressed as:

[0131]

[0132] S1.4: Local computation delay of user k By task size b k (i) The number of cycles required per bit s k (i) The computing power of the user f UE , the proportion of local computing Calculated, expressed as:

[0133]

[0134] The computing resources allocated by drone m to user k in each time slot are expressed as:

[0135]

[0136] Among them, κ1 is the energy consumption coefficient of UAV m, p k(i) is the computing resource allocated to user k by drone m, expressed as:

[0137] p k (i) = μ k P

[0138] Among them, μ k is the ratio of resource allocation, which is required to satisfy ∑μ k <1.

[0139] Computational tasks offloaded to drones The number of cycles required per bit s k (i) The computing power of the drone f UAV,k Calculate the computational delay D of the task offloaded to UAV m UAV,k (i), expressed as:

[0140]

[0141] The task transmission delay from user k to drone m is expressed as The computational delay of the task that user k offloads to the satellite is expressed as

[0142] Since the distance from the ground / air to the satellite is long, the data propagation delay d must also be considered during the transmission process. s Consider the task's tolerance delay, task generation time, and priority factor

[0143] S1.5: Energy consumption transmitted to drone m and calculate energy consumption Respectively expressed as:

[0144]

[0145]

[0146] Energy consumption E of user k k (i), by and Add and calculate the total.

[0147] Energy consumption for transmission to satellite Transmitting power from the ground to the satellite P l , the energy consumption per CPU cycle of the user is e loc Calculated.

[0148] The flight energy consumption of UAV m is expressed as:

[0149] E fly (i)=φ||ν(i)|| 2

[0150] Among them, ν(i) contains the flight speed and flight angle of the drone. The flight speed is obtained by performing the modulo operation ||ν(i)||, which is in the range of [0,ν max ], ν max is the maximum value of the flight speed; β(i) represents the flight angle, β(i)∈[0,2π]; φ represents the relevant parameters of the UAV, which is determined by the payload M of the UAV and the fixed flight time t fly Calculated, expressed as:

[0151] φ=0.5M UAV t fly

[0152] S2: Under the energy constraint of the UAV, an integer nonlinear optimization problem with uncertainty is proposed, and the conditional value at risk (CVaR) is used to measure the inspection risk;

[0153] S2.1: The trajectory of UAV m is The uninstall policy is Resource allocation strategy The delay D(i) is defined as the maximum delay of each part of the processing task, expressed as:

[0154]

[0155] The problem can be modeled as follows:

[0156]

[0157] Where C1 represents the value range of the offloading ratio; C2 represents that the resource allocation to all users cannot exceed the total resource amount; C3 is used to determine whether the task delay tolerance of each user k is met; C4 represents that the movement range of the UAV cannot exceed the predetermined range; C5 represents that the flight energy consumption, computing energy consumption and transmission energy consumption of the UAV in all time slots cannot exceed the energy capacity constraint of the UAV, E b Represents the energy capacity of the drone.

[0158] S2.2: Define a Markov decision MDP as a tuple At time slot i, the description of the system state s i , expressed as:

[0159] s i ={E[i],Q[i],P[i],D remain [i],f K [i]}

[0160] At time slot i, take action a according to the current statei , expressed as:

[0161] a i =(β(i),ν(i),λ(i))

[0162] Among them, β(i), ν(i), and λ(i) represent the flight direction of the UAV, the moving speed of the UAV, and the unloading ratio, respectively.

[0163] Considering the penalty for exceeding the task delay requirement, the cost c is expressed as:

[0164]

[0165] Among them, ζ is the penalty for exceeding the delay constraint, o k (i) It is used to determine whether the task delay tolerance of each user k is met.

[0166] The task scheduling problem of minimizing delay based on MDP is expressed as:

[0167]

[0168] stC1,C2,C3,C4

[0169]

[0170] Among them, problem P2 represents the expected average cost. Energy consumption is not a component of the cost function of problem P2. It is necessary to design a method suitable for handling the problem of Constrained Markov Decision Process (CMDP). The Conditional Value at Risk (CVaR) is introduced as a measure of risk. It is used to measure the average loss in the worst case under a certain confidence level.

[0171] First, let X be the random variable of the drone's energy consumption exceeding the energy capacity limit:

[0172]

[0173] in, is the energy actually consumed by the drone, E b is the energy capacity of the drone. When X>0, it means that the drone energy consumption exceeds the limit.

[0174] According to the acceptable risk level of the drone mission, select an appropriate confidence level α. It is hoped that the drone can work normally in 95% of cases without energy overspending, so set α = 0.05.

[0175] For a given confidence level α, the value at risk (VaR) needs to satisfy P(X≤-VaR) = α. Suppose there are n historical or simulated energy consumption data X1, X2, ..., X n , sort the energy consumption data, and according to the confidence level α, VaR is approximately expressed as Represents the maximum possible loss boundary of the UAV energy consumption exceeding the capacity limit under the confidence level α.

[0176] After determining VaR, calculate CVaR to represent the average of these losses under the condition that the losses exceed VaR. Find all data points X that satisfy X>-VaR j (j=1,2,...,m), calculate CVaR, and get:

[0177]

[0178] Take CVaR as the main component of the risk function. Suppose the risk function R(X) = CVaR α (X).

[0179] S3: A dual-critic network structure is designed and a risk-aware resource scheduling method is proposed.

[0180] The discounted cost model can be used to consider the consequences of long-term decisions and is used to calculate state value functions or action value functions. Let V(s) represent the value function of the expected long-term discounted cost, expressed as:

[0181]

[0182] in, Represents the discount factor at the current time slot, which is used to measure the importance of future rewards relative to current rewards.

[0183] Based on the above formula, the Q-value function is defined to further evaluate the expected long-term discounted cost obtained under the current state-action pair:

[0184]

[0185] With the minimum cost as the goal, select the minimum Q value function Q * (s,a)=minQ(s,a) as the optimal Q value. At the same time, take action a * Get the optimal Q value. The Q value function is updated according to the temporal difference (TD) equation:

[0186]

[0187] Among them, η represents the learning rate, which determines the magnitude of each update.

[0188] Using deep neural networks (DNN) for parameterization, the loss function is expressed as:

[0189]

[0190] in, represents the Q-value function based on DNN, is the TD error.

[0191] The risk function for the part exceeding the energy capacity of the drone is expressed as:

[0192]

[0193] Among them, Ω is the set of defined error states, which means that the energy consumption of the drone exceeds its energy capacity.

[0194] Similarly, we can obtain the value function of expected long-term discounted risk It is expressed as:

[0195]

[0196] The expected long-term discounted risk estimated by the Q-value function is expressed as:

[0197]

[0198] Execute greedy action To obtain the optimal Q value

[0199] Assume that the optimal Q values ​​obtained under two different objectives are Q(s i ,a i )and A dynamically updated weight coefficient δ is added between the two to balance them and obtain a new Q value function, which is expressed as:

[0200]

[0201] Initialize the parameterized cost network Risk Network and the action network μ(s|θ μ ), initialization parameters Experience Replay Weight δ and step size Δ. According to μ(s|θ μ )Select a i , storage experience (s i ,a i ,c i,r i ,s i+1 )arrive middle.

[0202] According to the target value y j Update, expressed as:

[0203]

[0204] Calculate the network loss L c , expressed as:

[0205]

[0206] Get the best strategy action It is expressed as:

[0207]

[0208] The weights are updated while the parameters are soft updated, expressed as:

[0209]

[0210] if Then δ←δ+Δ, otherwise δ←δ-Δ. Thus, the optimal task offloading and resource scheduling method is obtained. The structure of the dual critic network introduced in the present invention is learned, each network independently evaluates the action value function, and then a comprehensive evaluation is performed. In this way, the impact of the inaccuracy of a single network on the policy update can be reduced, and the problem of over-estimation can be reduced. At the same time, the stability and convergence of learning are improved.

[0211] The present invention designs a resource scheduling method based on risk perception. First, in the traditional integrated air-ground-space network, the communication model, delay model and energy consumption model of the three-layer network (space-based network, air-based network and ground-based network) are constructed respectively according to the respective characteristics of the three-layer network. In order to solve the instability caused by the dynamic arrival of tasks and optimize the deviation problem caused by a single goal to the estimation result, an integer nonlinear optimization problem with uncertainty is proposed under the energy constraint of the UAV, and the conditional value at risk (CVaR) is used to measure the inspection risk. A dual critic network structure is designed to balance cost and risk, and a resource scheduling method based on risk perception is proposed. The method provided by the present invention can achieve a more significant improvement in service quality, and a good balance is achieved in terms of cost and risk.

[0212] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims

1. A resource scheduling method based on risk perception, characterized in that: The following steps are involved: S1: In the space-ground integrated network, the communication model, delay model and energy consumption model of the space-based network, air-based network and ground-based network are constructed respectively; S2: Under the energy constraint of the UAV, the problem is formulated as an integer nonlinear optimization problem with uncertainty, and the conditional value at risk (CVaR) is used to measure the inspection risk; S3: Use a dual critic network to evaluate costs and risks, and implement optimal task offloading and resource scheduling strategies through deep neural network (DNN) parameterization and temporal difference (TD) updates.

2. The risk-aware resource scheduling method according to claim 1, characterized in that: The S1 The following steps are involved: In ground-to-air communication, the channel gain between user k and drone m is defined as: Where α0 is the channel gain at reference distance d = 1m, q(i+1) is the position of drone m after it moves to the next time slot, and p k (i) is the position of user k in time slot i, H is the fixed flight altitude of drone m, ||q(i+1)-p k (i)|| 2 +H 2 represents the Euclidean distance between user k and drone m; The data rate offloaded to UAV m is expressed as: Among them, B u is the available spectrum bandwidth from user k to drone m, η k is the bandwidth percentage allocated to user k, requiring ∑η k <1, P u is the user's transmit power on the uplink, g k (i) represents the channel gain, is the noise power, f k (i) Whether there is a blockage between user k and drone m. If there is a blockage, add P NLOS The noise power, P NLOS represents the additional noise power; In ground / air-space communications, the channel gain from user k to the satellite is determined by the fixed distance D between the ground and the satellite and the satellite computing power f LEO The data rate from the ground or UAV m to the satellite is calculated as: in, is the data rate from the ground or UAV m to the satellite, B l represents the communication bandwidth to the satellite, P l is the transmission power on the link, is the noise power; set up represents the partial uninstallation decision of user k, is the offloading ratio of user k to drone m, is the offloading ratio of user k to low-orbit satellites; Set the partial uninstallation decision of user k, expressed as: in, is the offloading ratio of user k to drone m, is the offloading ratio of user k to low-orbit satellites.

3. The risk-aware resource scheduling method according to claim 2, characterized in that: The S1 further comprises the following steps: By task size b k (i) The number of cycles required per bit s k (i) The computing power of the user f UE , the proportion of local computing Get the local computation delay of user k The computing resources allocated by drone m to user k in each time slot are expressed as: Among them, κ1 is the energy consumption coefficient of UAV m, p k (i) is the computing resource allocated to user k by drone m, expressed as: p k (i)=μ k P Among them, μ k is the ratio of resource allocation, ∑μ k <1; Computational tasks offloaded to drones The number of cycles required per bit s k (i) The computing power of the drone f UAV,k Calculate the computational delay D of the task offloaded to UAV m UAV,k (i); The priority factor is expressed as: Among them, t k is the generation time of the task, t is the current time, and the smaller the priority factor is, the higher the priority of the task is and the earlier it can be unloaded and processed.

4. The risk-aware resource scheduling method according to claim 3 is characterized in that: The S1 further comprises the following steps: Energy consumption transmitted to UAV m and calculate energy consumption Respectively expressed as: Depend on and The sum of the local computation delay of user k is obtained The flight energy consumption E of UAV m fly (i), expressed as: E fly (i)=φ||ν(i)|| 2 Among them, ν(i) contains the flight speed and flight angle of the drone. The flight speed is obtained by performing the modulo operation ||ν(i)||, which is in the range of [0,ν max ], ν max is the maximum value of the flight speed; β(i) represents the flight angle, β(i)∈[0,2π]; φ represents the relevant parameters of the UAV.

5. The resource scheduling method based on risk perception according to claim 1 is characterized in that: The S2 comprises the following steps: UAV m track It is expressed as: The unloading strategy Λ is expressed as: The resource allocation strategy Φ is expressed as: The delay D(i) is defined as the maximum delay of each part of the processing task, expressed as: The problem is modeled as: Among them, C1 represents the value range of the offloading ratio; C2 represents that the resource allocation to all users does not exceed the total resources; C3 determines whether the delay tolerance of each user k is met; C4 represents that the movement range of the UAV does not exceed the predetermined range; C5 represents that the flight energy consumption, computing energy consumption and transmission energy consumption of the UAV in the time slot do not exceed the energy capacity constraint of the UAV, E b Represents the energy capacity of the drone; Using Markov decision making, we get the cost c, expressed as: Among them, ζ is the penalty for exceeding the delay constraint, o k (i) It is used to determine whether the task delay tolerance of each user k is met; The current scheduling problem is described as a constrained Markov decision process, expressed as: stC1,C2,C3,C4 The risk conditional value CVaR is used as a measure of risk inspection to measure the average loss in the worst case under a certain confidence level; Let X be the random variable of the drone's energy consumption exceeding the energy capacity limit: in, is the energy actually consumed by the drone, E b is the energy capacity of the drone; when X>0, it means that the drone energy consumption exceeds the limit; Set the confidence level α = 0.05; for a given α, the value at risk (VaR) needs to satisfy P(X≤-VaR) = α; find all data points X that satisfy X>-VaR j (j=1,2,...,m), calculate CVaR, expressed as: Take CVaR as the main component of the risk function; let the risk function R(X) = CVaR α (X).

6. A risk-aware resource scheduling method according to claim 1, characterized in that: The S3 comprises the following steps: The value function of the expected long-term discounted cost V(s) is expressed as: in, represents the discount factor at the current time slot, which is used to measure the importance of future rewards relative to current rewards; The Q-value function is defined to further evaluate the expected long-term discounted cost obtained under the current state-action pair, expressed as: With the minimum cost as the goal, select the minimum Q value function Q * (s,a)=minQ(s,a) as the optimal Q value; take action a at the same time * Get the optimal Q value; the Q value function is updated according to the temporal difference (TD) equation, expressed as: Among them, η represents the learning rate, which is used to determine the magnitude of each update; Using deep neural networks (DNN) for parameterization, the loss function is expressed as: in, represents the Q-value function based on DNN, is the TD error; The risk function for the part exceeding the energy capacity of the drone is expressed as: Among them, Ω is a set of defined error states, indicating that the energy consumption of the drone exceeds its energy capacity.

7. A risk-aware resource scheduling method according to claim 6, characterized in that: The S3 further comprises the following steps: The expected long-term discounted risk value function It is expressed as: The expected long-term discounted risk estimated by the Q-value function is expressed as: Execute greedy action To obtain the optimal Q value Assume that the optimal Q values ​​obtained under two different objectives are Q(s i ,a i )and A dynamically updated weight coefficient δ is added between the two to balance them and obtain a new Q value function, which is expressed as:

8. The risk-aware resource scheduling method according to claim 7, characterized in that: The S3 further comprises the following steps: Initialize the parameterized cost network Risk Network and the action network μ(s|θ μ ); Initialization parameters θ μ' ←θ μ , Experience Replay Weight δ and step size Δ; according to μ(s|θ μ )Select a i , storage experience (s i ,a i ,c i ,r i ,s i+1 )arrive middle; According to the target value y j Update, expressed as: Calculate the network loss L c , expressed as: Get the best strategy action It is expressed as: The weights are updated while the parameters are soft updated, expressed as: if Then δ←δ+Δ, otherwise δ←δ-Δ.

9. The risk-aware resource scheduling method according to claim 2, characterized in that: The maximum value of the unloading ratio λ max is 0.5, and the unloading ratio is constrained, which is expressed as:

Images