Dynamic energy-saving cluster resource collaborative allocation method

Abstract

The invention discloses a dynamic energy-saving cluster resource collaborative allocation method. The method comprises the following steps: acquiring environment sensing data: acquiring environment data and node state data in real time through an unmanned aerial vehicle sensor; predicting a task demand: predicting a future task demand change condition by adopting a deep learning method according to historical data and environment change; optimizing the resource utilization rate: generating a resource allocation strategy, and dynamically adjusting calculation and communication resources of each node; making an energy-saving scheduling decision: generating an energy-saving scheduling strategy by using deep reinforcement learning, wherein the scheduling strategy considers two targets of task completion rate and energy consumption at the same time; feedback is collected, and self-adaptive adjustment is carried out: according to a task execution result and a node state feedback condition, an allocation strategy and energy-saving parameters are dynamically adjusted. By monitoring, predicting and optimizing resource allocation in real time and dynamically responding to environment changes, the task efficiency is improved, meanwhile, energy consumption is reduced, and the node operation time and service life are prolonged.

Description

Technical Field

[0001] This invention belongs to the field of unmanned aerial vehicle (UAV) control technology, and particularly relates to a dynamic and energy-saving cluster resource collaborative allocation method. Background Technology

[0002] In unmanned system clusters such as drone swarms and unmanned vehicle swarms, dynamic resource allocation and reducing drone energy consumption are crucial for mission completion and rapid response. Existing resource allocation methods are often based on static allocation or simple load balancing, which cannot effectively cope with environmental changes and fluctuations in mission requirements, resulting in low resource utilization, high energy consumption, and shortened node lifespan. Summary of the Invention

[0003] This application proposes a dynamic energy-saving cluster resource collaborative allocation method, which aims to improve task efficiency, reduce energy consumption, and extend node operating time and lifespan by real-time monitoring, prediction and optimization of resource allocation, dynamically responding to environmental changes.

[0004] To achieve the above objectives, this application discloses a dynamic energy-saving cluster resource collaborative allocation method, comprising the following steps:

[0005] S1 collects environmental perception data: Real-time collection of environmental data and node status data through the drone's sensors;

[0006] S2 predicts task requirements: Based on historical data and environmental changes, deep learning methods are used to predict future changes in task requirements;

[0007] S3 optimizes resource utilization: generates resource allocation strategies and dynamically adjusts the computing and communication resources of each node;

[0008] S4 makes energy-saving scheduling decisions: It uses deep reinforcement learning to generate energy-saving scheduling strategies, which simultaneously consider two objectives: task completion rate and energy consumption.

[0009] S5 collects feedback and makes adaptive adjustments: it dynamically adjusts allocation strategies and energy-saving parameters based on task execution results and node status feedback.

[0010] Preferably, in step S1, the collected data is denoised and normalized to generate a standardized dataset.

[0011] Preferably, step S2 uses a Long Short-Term Memory (LSTM) network model, inputting historical environmental data and historical task density to train the LSTM prediction model;

[0012] Then, the trained LSTM model is used to predict the task requirements for future time windows;

[0013] The prediction error function used is:

[0014]

[0015] E pred This is the prediction error, where w represents the length of the prediction time window. It is the task density predicted at time t+k. It represents the actual task density at time t+k.

[0016] Preferably, in step S3, the objective function considers both the task completion rate and resource utilization rate, and the objective function is calculated as follows:

[0017]

[0018] F alloc This represents the objective function value for resource allocation optimization, where w1, w2, and w4 are the task completion rate coefficient, resource utilization rate coefficient, and allocation balance coefficient, respectively; T comp (i,t) represents the amount of task completed by node i at time t. It is the predicted task density, R used (i,t) represents the amount of resources used by node i at time t, R total (i,t) represents the total resource amount of node i at time t, R alloc (i,t) represents the proportion of resources allocated to node i at time t. is the average resource allocation ratio of all nodes at time t, and N is the number of nodes.

[0019] Preferably, step S4 employs a deep reinforcement learning method, combining resource allocation ratios and node states to generate an energy-saving scheduling strategy; the reward function R in deep reinforcement learning... drl (t) Taking into account both task completion rate and energy consumption, the reward function is as follows:

[0020]

[0021] In the formula, R drl (t) represents the reward value at time t, w1, μ, v are the task completion rate coefficient, energy consumption coefficient, and load balancing coefficient, respectively, N is the number of nodes, and T is the reward value at time t. comp (i,t) represents the amount of task completed by node i at time t. It is the predicted task density, E cons (i,t) is the amount of electricity consumed by node i at time t, E node (i,t) is the charge of node i at time t, L comp (i,t) represents the computational load of node i at time t.

[0022] Preferably, the strategy effectiveness evaluation function in step S5 is as follows:

[0023]

[0024] S effect (t) is the strategy effectiveness evaluation value at time t, w1, μ, ω3 are the task completion rate coefficient, energy consumption coefficient, and node lifespan coefficient, respectively, N is the number of nodes, and T rate (t) is the task completion rate at time t, E total (t) represents the total energy consumption of the cluster at time t, E max It is the cluster's maximum energy capacity, L life (i,t) represents the predicted lifetime of node i at time L. max This is the maximum lifetime value of the node.

[0025] Preferably, in step S3, the objective function F is used. alloc Furthermore, by incorporating the dynamic adjustment mechanism of MARL, the resource allocation optimization problem is transformed into a multi-objective collaborative optimization problem.

[0026] The objective function is as follows:

[0027]

[0028] w1, w2, w4, w3, and η are the task completion rate coefficient, resource utilization coefficient, allocation balance coefficient, energy saving coefficient, and potential field penalty coefficient, respectively; N is the number of nodes; and T is the number of nodes. comp (i,t) represents the amount of task completed by node i at time t. R represents the predicted task density or total demand at time t. used (i,t) represents the amount of resources used by node i at time t, R total (i,t) describes the total available resources of node i at time t, R alloc (i,t) is the proportion of resources allocated to node i at time t. E is the average resource allocation ratio at time t. save (i,t) represents the energy saved by node i at time t, which is the energy reduction achieved through resource allocation optimization. P field (t) is the potential field penalty term at time t, used to measure the dynamic balance of resource allocation;

[0029] The potential field penalty items are as follows:

[0030]

[0031] in:

[0032]

[0033] F attr(i,t) represents the task demand attraction experienced by node i at time t, guiding resource allocation towards high-demand areas, F rep (i,t) represents the repulsive force of overloaded neighborhood resources experienced by node i at time t, to avoid excessive resource concentration. attr ,k rep As a moderating factor for attraction and repulsion, It is the predicted task demand value for a local region of node i, dist(i,D) center (t) is the distance from node i to the center point of the task requirement. Let i be the set of neighboring nodes of node i, and dist(i,j) be the distance between node i and neighboring node j.

[0034] The MARL reward function calculation formula is as follows:

[0035]

[0036] R MARL (i,t) is the reinforcement learning reward value of node i at time t, and w1, w2, w3, w4 are the task completion rate coefficient, resource utilization coefficient, energy saving coefficient, and allocation balance coefficient, respectively.

[0037] Preferably, step S4 generates an online scheduling strategy that balances energy saving and task completion rate, and dynamically adjusts node behavior based on predicted task requirements and real-time environmental changes. The specific implementation is as follows:

[0038] By dynamically adjusting the allocation ratio of computing, communication and storage resources, combined with hibernation / activation strategies, energy consumption caused by idle or inefficient operation can be reduced.

[0039] While saving energy, predictive models and real-time feedback are used to ensure that the task completion rate and response speed are not lower than the preset threshold.

[0040] Resources are allocated collaboratively among nodes to avoid single-point overload leading to a surge in energy consumption and a shortened hardware lifespan for individual nodes.

[0041] Each node acts as an intelligent agent, competing for limited resources while also collaborating to complete the overall task during resource allocation.

[0042] To achieve the multiple objectives of simultaneously improving task completion rate, optimizing energy consumption, and balancing load, the following reward function is used:

[0043]

[0044] w1,μ,ν,η,γ represent the task completion rate coefficient, energy consumption coefficient, load balancing coefficient, potential field penalty coefficient, and collaboration reward coefficient, respectively; N is the number of nodes; T comp(i,t) represents the amount of tasks or services completed by node i at time t. E is the total demand for the forecast task at time t; cons (i,t) represents the energy consumption or energy depletion of node i at time t; E node (i,t) represents the total available or remaining power of node i at time t; L comp (i,t) represents the load of node i at time t; U pot (i,t) is the potential energy of node i at time t, used to guide load balancing; C coop (t) represents the cooperative reward at time t, which is calculated according to the cooperative calculation of multi-agent game, where N is the total number of nodes;

[0045] Potential energy U pot (i,t) encourages tasks to migrate from high-load nodes to low-load nodes:

[0046]

[0047] k rep ,k attr >0 represents the repulsive and attractive coefficients, which control the strength of the potential field's influence; d ij The logical distance between nodes i and j; It is an indicator function; its value is 1 when the condition is true, and 0 otherwise; repulsion. Simulate the repulsive force of a high-load node on a low-load node; the greater the load difference and the closer the distance, the stronger the repulsive force. Gravity... Simulate the attractive force of a low-load node on a high-load node; the greater the load difference and the closer the distance, the stronger the attraction. comp (j,t) represents the load of node j at time t;

[0048] Collaboration Reward C coop (t) Encourage nodes to maximize the global task completion rate through resource competition:

[0049]

[0050] T mig (i,j,t) represents the amount of work required for node i to migrate to node j at time t. Attached Figure Description

[0051] Figure 1 A flowchart of a dynamic energy-saving cluster resource collaborative allocation method provided in an embodiment of this application;

[0052] Figure 2 A framework diagram of a dynamic energy-saving cluster resource collaborative allocation system provided in the embodiments of this application. Detailed Implementation

[0053] The present invention will be further described below with reference to the accompanying drawings, but this is not intended to limit the present invention in any way. Any modifications or substitutions made based on the teachings of the present invention shall fall within the protection scope of the present invention.

[0054] Artificial intelligence (AI) is the theory, methods, technology, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a way similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to possess the functions of perception, reasoning, and decision-making.

[0055] Natural Language Processing (NLP) is an important field within computer science and artificial intelligence. It studies the theories and methods for enabling effective communication between humans and computers using natural language. NLP is a science that integrates linguistics, computer science, and mathematics. Therefore, research in this field involves natural language—the language people use in daily life—and thus it has a close relationship with linguistic research. NLP techniques typically include text processing, semantic understanding, machine translation, question answering, and knowledge graphs.

[0056] The technical solutions provided in this application involve technologies such as machine learning and natural language processing in artificial intelligence, which are specifically described and illustrated through the following embodiments.

[0057] refer to Figure 1 The dynamic resource allocation method proposed in this application is applied to the dynamic task allocation and energy management scenario of UAV swarms in disaster relief. First, real-time data is acquired through environmental perception; then, task demand prediction generates predicted values ​​for future demands based on the real-time data. Next, resource utilization optimization allocates resources based on the predicted values ​​and optimization strategies. Energy-saving scheduling decisions further optimize the scheduling scheme to reduce energy consumption. Finally, feedback is collected to adaptively adjust the strategy, forming a closed-loop system. The specific process mainly includes the following five steps:

[0058] S1 collects environmental perception data: It collects environmental data and node status data in real time through the drone's sensors. The collected environmental data includes wind speed, temperature, and mission requirements in the disaster area, while the node status data includes battery level and load.

[0059] S2 predicts task requirements: Based on historical data and environmental changes, it uses deep learning methods to predict future changes in task requirements.

[0060] S3 optimizes resource utilization: it generates resource allocation strategies and dynamically adjusts the computing and communication resources of each node.

[0061] S4 formulates energy-saving scheduling decisions: It uses deep reinforcement learning (DRL) to generate energy-saving scheduling strategies, which simultaneously consider two objectives: task completion rate and energy consumption.

[0062] S5 collects feedback and makes adaptive adjustments: it dynamically adjusts allocation strategies and energy-saving parameters based on task execution results and node status feedback.

[0063] In one embodiment, step S1 is responsible for real-time acquisition of the drone swarm's operating environment and node status data. Environmental data, including wind speed, temperature, and task density in the disaster area, is collected using sensors mounted on the drones. Simultaneously, node status data, including battery level, computational load, and communication load, is also collected. Task density is the task density at time t, measured in tasks per square kilometer. The computational load of node i at time t is measured in CPU cycles per second. The collected data is then denoised and normalized to generate a standardized dataset D. env (t) and D node (i,t).

[0064] In one embodiment, step S2 predicts changes in task requirements over a future period based on collected environmental data and historical task data, providing a basis for resource allocation. This embodiment employs a Long Short-Term Memory (LSTM) network model, with historical environmental data D as input. env (t―1:t―n) and historical task density D task (t―1:t―n), train the LSTM prediction model.

[0065] Then, the trained LSTM model is used to predict the task requirements in the future time window t+1:t+w.

[0066] The prediction error function used is:

[0067]

[0068] E pred is the prediction error, and w represents the prediction time window length. It is the task density predicted at time t+k. It represents the actual task density at time t+k.

[0069] This step, through the use of deep learning, anticipates fluctuations in mission requirements during disaster relief scenarios. This prevents resource allocation from lagging behind changes in mission needs, ensuring that drone swarms can allocate resources in advance during peak mission periods and reduce resource waste during off-peak periods, thereby accelerating response speed and improving mission completion rates.

[0070] In one embodiment, step S3 uses the optimization idea of ​​bird flock foraging behavior in intelligent algorithms to predict future task requirements and dynamically optimize the resource allocation of the drone swarm.

[0071] Considering both task completion rate and resource utilization rate, the objective function F alloc To calculate the resource allocation ratio R for each node i alloc (i,t).

[0072] The objective function is calculated using the following formula:

[0073]

[0074] F alloc This is the objective function value for resource allocation optimization. α, β, and γ are weight parameters that control the impact of task completion rate, resource utilization rate, and allocation balance, respectively, and their values ​​range from [0,1]. comp (i,t) represents the amount of task completed by node i at time t. R is the predicted task density. used (i,t) represents the amount of resources used by node i at time t. R total (i,t) represents the total resources of node i at time t. R alloc (i,t) is the proportion of resources allocated to node i at time t. It is the average resource allocation ratio of all nodes at time t.

[0075] The distribution balance penalty term in the above formula This is used to prevent excessive concentration of resources in a few nodes.

[0076] By simulating efficient resource allocation mechanisms in nature, we can ensure that drone swarms can collaborate efficiently in disaster relief. This translates predicted task requirements into specific resource allocation ratios, maximizing resource utilization and task completion rates. At the same time, by avoiding excessive resource concentration, we can improve the overall stability of the swarm.

[0077] In one embodiment, step S4 employs a deep reinforcement learning (DRL) method, combining resource allocation ratios and node states to generate an energy-saving scheduling strategy, thereby reducing overall energy consumption. The node state D... node (i,t), resource allocation ratio R alloc (i,t) and environmental data D env(t) serves as the input state for reinforcement learning.

[0078] In deep reinforcement learning, the action is the scheduling action A of each node. sched (i,t) includes adjustments to computing resources, adjustments to communication resources, and sleep / activation decisions.

[0079] Reward function R in deep reinforcement learning drl (t) is used to comprehensively consider task completion rate and energy consumption.

[0080] The reward function is as follows:

[0081]

[0082] In the formula, R drl (t) represents the reward value at time t, and λ, μ, v are the weighting parameters that control the impact of task completion rate, energy consumption, and load balancing, respectively. comp (i,t) represents the amount of task completed by node i at time t. This is the predicted task density. E cons (i,t) represents the amount of electricity consumed by node i at time t. node (i,t) represents the charge of node i at time t. comp (i,t) represents the computational load of node i at time t.

[0083] The load balancing reward item in the above formula A nonlinear excitation mechanism is used to ensure load balance and prevent node overload from increasing energy consumption.

[0084] This step uses deep reinforcement learning technology to dynamically generate scheduling strategies, balance task completion rate and energy consumption, optimize resource usage of each node based on resource allocation ratio and real-time status, switch between sleep or active operation modes, greatly reduce energy consumption, and at the same time, the reward function incentivizes load balancing, extends node running time, and extends the disaster relief work time of the drone swarm.

[0085] In one embodiment, step S5 dynamically adjusts resource allocation and energy-saving scheduling strategies based on task execution results and node status feedback to form a closed-loop optimization.

[0086] Collect task completion rate T rate (t), Energy consumption E total (t) and node lifetime prediction L life (i,t) is used to evaluate the effectiveness of the current policy. Then, based on the evaluation results, the weights α,β,γ of the objective function and the weights λ,μ,v of the reward function are adjusted and optimized to generate a new policy.

[0087] The strategy effectiveness evaluation function is as follows:

[0088]

[0089] S effect (t) is the policy effectiveness evaluation value at time t. ω1, ω2, and ω3 are the weight parameters for controlling the task completion rate, energy consumption, and node lifespan, respectively. rate (t) is the task completion rate at time t, E total (t) represents the total energy consumption of the cluster at time t. max This is the cluster's maximum energy capacity. L life (i,t) represents the predicted lifetime of node i at time t. max This is the maximum lifetime value of the node.

[0090] This step involves evaluating task execution results and node status to dynamically optimize resource allocation and energy-saving strategies, ensuring that the system can continuously adapt to environmental changes and fluctuations in task requirements, while maintaining high efficiency and energy saving.

[0091] In one embodiment, in step S3, the drone / unmanned system cluster is viewed as a "flock of birds," and the distribution of task requirements is simulated as a "food distribution." By integrating bird flock foraging behavior—that is, group cooperation and competition mechanisms—from intelligent algorithms, and treating each drone node as an intelligent agent, multiple agents collaboratively learn to generate an adaptive resource allocation strategy. This not only enables collaborative optimization and resource sharing among nodes but also allows for dynamic responses to environmental changes, such as fluctuations in task requirements and heterogeneity of node states. Thus, the predicted task requirements are transformed into the calculation and allocation of resource allocation ratio R for each node. alloc (i,t). The cluster can achieve self-organized resource allocation in a dynamic environment, while reducing overall energy consumption and extending node lifespan.

[0092] By modeling task requirements as an "attraction field," nodes are guided to allocate resources to high-demand areas. Resource overload or energy bottlenecks are modeled as a "repulsion field" to avoid excessive resource concentration, thereby forming a dynamic balance resource allocation mechanism.

[0093] In the MARL framework, each node acts as an agent, learning the optimal resource allocation strategy through interaction with the environment. MARL can not only capture changes in task requirements in dynamic environments, but also incentivize collaboration and energy-saving behaviors among nodes through reward function design, avoiding the limitations of traditional optimization methods in high-dimensional dynamic environments.

[0094] To achieve the above functionality, the objective function F is used. alloc Furthermore, by combining potential field theory with the dynamic adjustment mechanism of MARL, the resource allocation optimization problem is transformed into a multi-objective collaborative optimization problem.

[0095] The objective function is as follows:

[0096]

[0097] α,β,γ,δ,η∈[0,1] are weight parameters that control the impact of task completion rate, resource utilization rate, allocation balance, energy saving effect, and potential field penalty. comp (i,t) represents the amount of work completed by node i at time t. R represents the predicted task density or total demand at time t. used (i,t) represents the amount of resources used by node i at time t, including computing, communication, energy, etc. R total (i,t) describes the total available resources of node i at time t. R alloc (i,t) is the proportion of resources allocated to node i at time t. E is the average resource allocation ratio at time t. save (i,t) represents the node energy at time t, which is the energy consumption reduced through resource allocation optimization. field (t) is the potential field penalty term at time t, used to measure the dynamic balance of resource allocation.

[0098] The potential field penalty items are as follows:

[0099]

[0100] in:

[0101]

[0102] F attr (i,t) represents the task demand attraction experienced by node i at time t, guiding resource allocation towards high-demand regions. F rep (i,t) represents the repulsive force of overloaded neighborhood resources experienced by node i at time t, which aims to prevent excessive resource concentration. attr ,k rep This is the adjustment coefficient for attraction and repulsion. This is the predicted task demand value for the local region of node i. dist(i,D) center (t) is the distance from node i to the center point of the task requirement. Let i be the set of neighboring nodes of node i. Let dist(i,j) be the distance between node i and neighboring node j.

[0103] The MARL reward function is used to dynamically learn resource allocation.

[0104]

[0105] R MARL(i,t): The reinforcement learning reward value of node i at time t. w1,w2,w3,w4: The weights of each part in the reward function, used to balance task completion, resource utilization, energy saving and allocation balance.

[0106] The energy-saving term E in the objective function save (i,t) and potential field penalty term P field (t) enables resource allocation to not only focus on task completion and balance, but also directly optimize energy consumption and extend node lifespan.

[0107] This method supports adaptive resource reallocation and collaborative optimization in dynamic environments, improving the overall performance of UAV swarms while extending node runtime and lifespan.

[0108] The specific steps are as follows:

[0109] Based on the distribution of task requirements and the status of nodes, construct a task attraction field and a resource repulsion field.

[0110] Initialize the policy network for each node, and optimize the resource allocation ratio R through interactive learning. alloc (i,t), reward function R MARL (i,t) guides collaboration and energy conservation.

[0111] Real-time calculation of F alloc Resource allocation is dynamically adjusted based on weight parameters and potential field penalties.

[0112] Update the MARL strategy and weight parameters based on task completion status and energy consumption feedback:

[0113] Real-time network environments can experience various changes, such as fluctuating communication latency, sudden surges in task load, and unstable energy supply. In one embodiment, step S4 generates an online scheduling strategy that balances energy saving and task completion rate, enabling dynamic adjustment of node behavior based on predicted task demands and real-time environmental changes. Specifically, this is implemented as follows:

[0114] By dynamically adjusting the allocation ratio of computing, communication and storage resources, combined with hibernation / activation strategies, energy consumption caused by idle or inefficient operation can be reduced.

[0115] While saving energy, predictive models and real-time feedback are used to ensure that the task completion rate and response speed are not lower than the preset threshold.

[0116] Resources are allocated collaboratively among nodes to avoid single-point overload leading to a surge in energy consumption and a shortened hardware lifespan for individual nodes.

[0117] Each node is treated as an intelligent agent, competing for limited resources such as bandwidth and energy while cooperating to complete the overall task. By introducing the concept of Nash equilibrium and adding cooperation and competition penalty terms to the reward function, nodes can find a balance between local energy conservation and global task completion.

[0118] State space S(t): The state consists of three parts, comprehensively representing the dynamics of the system and the environment:

[0119] Node state D node (i,t)={C comp (i,t),C comm (i,t),E rem (i,t),L curr (i,t)}, where C comp (i,t) represents the resource utilization rate, and C is used to calculate the resource utilization rate. comm (i,t) represents the communication resource utilization rate, E rem (i,t) represents the remaining energy, L curr (i,t) represents the current load; resource allocation ratio R alloc (i,t)={R comp (i,t),R comm ... in To predict task requirements, E supply (t) represents the energy supply status, N delay (t) represents the network latency, W cond (t) represents external environmental conditions such as the effect of temperature and humidity on energy consumption. Action space A sched (i,t) defines multidimensional continuous / discrete actions for each node, including: computational resource adjustment A comp (i,t)∈[0,1], representing the scaling ratio of CPU / GPU frequency or task concurrency; communication resource adjustment A comm (i,t)={B alloc (i,t),L sel (i,t),R rate (i,t)}, where B alloc (i,t) represents the bandwidth allocation ratio, L sel (i,t) represents the link selection, R rate (i,t) represents the transmission rate; sleep / activation decision A sleep (i,t)∈{0,1}, where 0 represents dormancy and 1 represents activation; Task migration decision A mig (i,t)={T mig (i,j,t),P mig(i,j,t)} represents the task quantity and priority of migrating from node i to node j; resource pre-allocation A pre (i,t)∈[0,1] represents the pre-allocation ratio of resources based on prediction.

[0120] To achieve the multiple objectives of simultaneously improving task completion rate, optimizing energy consumption, and balancing load, the following reward function is used, which includes potential energy and game-theoretic cooperation:

[0121]

[0122] λ,μ,v,η,γ∈[0,1] are weight parameters representing the relative importance of the control task completion rate, energy consumption penalty, load balancing reward, potential field energy reward, and cooperation reward, respectively; T comp (i,t) represents the amount of tasks or services completed by node i at time t. E is the total predicted task demand at time t, used to normalize the task completion rate; cons (i,t) represents the energy consumption or energy depletion of node i at time t; E node (i,t) represents the total available or remaining power of node i at time t, used to normalize the energy consumption effect; L comp (i,t) represents the load on node i at time t, including utilization, queue length, or latency metrics; N is the total number of nodes; U pot (i,t) is the potential energy of node i at time t, used to guide load balancing; C coop (t) represents the cooperative reward at time t, calculated based on the cooperative nature of multi-agent games.

[0123] Potential energy U pot (i,t) represents the potential energy state of the load distribution for each node, encouraging tasks to migrate from high-load nodes to low-load nodes:

[0124]

[0125] k rep ,k attr >0 represents the repulsive and attractive coefficients, which control the strength of the potential field's influence; d ij The logical distance between nodes i and j, including physical distance or communication delay, etc. Indicator function, value is 1 when the condition is true, otherwise it is 0; repulsion force Simulate the repulsive force of a high-load node on a low-load node; the greater the load difference and the closer the distance, the stronger the repulsive force. Gravity... The attraction between low-load nodes and high-load nodes is simulated; the greater the load difference and the closer the distance, the stronger the attraction.

[0126] The above formula uses a physics-inspired "force field" mechanism to make load balancing more natural and avoids the policy oscillations that may be caused by traditional load variance minimization methods.

[0127] Collaboration Reward C coop (t) Encourage nodes to maximize the global task completion rate through resource competition:

[0128]

[0129] T mig (i,j,t) represents the amount of task that node i migrates to node j at time t. This parameter encourages task migration behavior through a logarithmic function, while weighting it with the proportion of task completion contributed by each node to ensure that collaborative behavior makes a positive contribution to the overall task completion. comp (j,t) represents the amount of tasks or services completed by node j at time t.

[0130] Through the aforementioned state / action space design and reward function, fine-grained management of computing, communication, and storage resources is achieved through multi-dimensional actions such as frequency adjustment, task migration, and resource pre-allocation, reducing idle energy consumption and redundant overhead; through load balancing and energy distribution optimization, single-point overload and energy bottlenecks are avoided, greatly extending the lifespan of node hardware.

[0131] Preferably, DRL algorithms that support continuous action spaces, such as Proximal Policy Optimization (PPO) or Soft Actor-Critic (SAC), are employed to adapt to the high complexity of multi-dimensional action spaces. Simultaneously, experience replay buffers and target networks are used to improve sample utilization and training stability; the reward function is decomposed into short-term and long-term objectives to alleviate conflicts in multi-objective optimization; and minimum service guarantees, such as a lower bound on task completion rate and minimum remaining battery power, are introduced to avoid catastrophic decisions during the learning process.

[0132] This application improves the naturalness and efficiency of load balancing, and also balances the goals of local energy saving and global task completion through a game theory mechanism.

[0133] refer to Figure 2 The dynamic energy-saving cluster resource collaborative allocation system disclosed in this application includes the following modules:

[0134] Environmental perception module: Collects environmental data and node status data in real time through the drone's sensors;

[0135] Prediction Task Module: Based on historical data and environmental changes, it uses deep learning methods to predict future changes in task requirements;

[0136] Resource utilization optimization module: Generates resource allocation strategies and dynamically adjusts the computing and communication resources of each node;

[0137] Energy-saving scheduling module: Utilizes deep reinforcement learning to generate energy-saving scheduling strategies, which simultaneously consider both task completion rate and energy consumption objectives;

[0138] Feedback and Adjustment Module: Dynamically adjusts allocation strategies and energy-saving parameters based on task execution results and node status feedback.

[0139] To verify the effectiveness of the proposed dynamic energy-saving cluster resource collaborative allocation method in unmanned system clusters, this application conducted a series of simulation experiments and some real-world scenario tests to evaluate the performance of the method in terms of task completion rate, energy consumption, load balancing, and node lifetime extension. The experiments primarily used unmanned aerial vehicle (UAV) clusters in disaster relief scenarios as the main application background. The experiments mainly used simulation environment-generated data, and based on the open-source UAV cluster simulation platforms AirSim and ROS (Robot Operating System), a dataset of task demand distribution and environmental changes in disaster relief scenarios was constructed. The simulation data includes:

[0140] Wind speed, temperature, and mission density are simulated using Gaussian distributions and sinusoidal fluctuations to model the dynamic changes in the disaster area. The drone node's power E... node (i,t)∈[0,5000]mAh, computational load L comp (i,t)∈[0,10 9 Communication load L comm (i,t)∈[0,10 6 The initial value is randomly distributed and dynamically changed based on task execution and energy consumption models. Finally, simulation data lasting 24 hours is generated, with a time step of 1 minute and a total of 1440 time points, each containing data from 50 drone nodes.

[0141] Real-world drone flight and mission data were extracted from the publicly available UAV123 dataset and disaster relief case data to verify the reliability of the simulation results. The real-world data includes drone battery level changes, mission distribution, and environmental parameters, spanning a 2-hour period and involving 10 drone nodes.

[0142] Synthetic and real data are mixed to generate test data, which is then used to train the LSTM prediction model.

[0143] 1.2 Experimental Environment and Hardware Configuration

[0144] The hardware configuration is an Intel Core i9-10900K CPU, an NVIDIA RTX 3090 GPU, and 64GB of RAM. The simulation environment is AirSim+ROS, running on Ubuntu 20.04.

[0145] Then, a real-world test was conducted using 10 DJI Matrice 300RTK drones in a 1-square-kilometer outdoor test area, equipped with sensors to collect environmental data and node status. The simulation experiment used 50 drone nodes, while the real test used 10. The task requirement prediction time window was 10 minutes. Environmental data and node status data were collected in real-time within the AirSim simulation environment. The data was then denoised and Min-Max normalized to the range [0,1].

[0146] The experimental results are as follows: the average task completion rate was 92.5%, with a peak of 98.2% and a trough of 85.3%. Total energy consumption was 1.85 × 10⁵ mAh, a 23.6% reduction compared to the baseline method. The average load variance was 1.2 × 10⁸ CPU cycles / second, a 31.4% reduction compared to the baseline. The predicted average node lifetime was 36.7 hours, an extension of 18.9%.

[0147] This application achieves an 11% improvement in task completion rate compared to static allocation methods, a 17% reduction in energy consumption compared to load balancing methods, and a 13.5% improvement in load balancing compared to single reinforcement learning, demonstrating that the combination of ecological optimization and DRL significantly improves performance.

[0148] As used herein, the term "preferred" is meant as an example, illustration, or illustration. Any aspect or design described herein as "preferred" need not be construed as being more advantageous than other aspects or designs. Rather, the use of the term "preferred" is intended to present the concept in a specific manner. As used in this application, the term "or" is intended to mean an inclusive "or" rather than an exclusionary "or." That is, unless otherwise specified or clear from the context, "X uses A or B" naturally includes either of the permutations. That is, if X uses A; X uses B; or X uses both A and B, then "X uses A or B" is satisfied in any of the foregoing examples.

[0149] Furthermore, although this disclosure has been shown and described with respect to one or more implementations, equivalent variations and modifications will occur to those skilled in the art based on a reading and understanding of this specification and the accompanying drawings. This disclosure includes all such modifications and variations and is limited only by the scope of the appended claims. In particular, with respect to the various functions performed by the aforementioned components (e.g., elements, etc.), the terminology used to describe such components is intended to correspond to any component (unless otherwise indicated) that performs the specified function of said component, even if structurally not equivalent to the disclosed structure performing the functions in the exemplary implementations of this disclosure shown herein. Moreover, although a particular feature of this disclosure has been disclosed with respect to only one of several implementations, such feature may be combined with one or more features of other implementations that may be desirable and advantageous for a given or particular application. Furthermore, with regard to the use of the terms “comprising,” “having,” “containing,” or variations thereof in the Detailed Description or claims, such terms are intended to be included in a manner similar to the term “including.”

[0150] The functional units in this invention embodiment can be integrated into a processing module, or each unit can exist physically separately, or multiple units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. The aforementioned devices or systems can execute the storage methods in the corresponding method embodiments.

[0151] In summary, the above embodiments are one implementation of the present invention, but the implementation of the present invention is not limited to the embodiments described above. Any changes, modifications, substitutions, combinations, or simplifications made that deviate from the spirit and principle of the present invention should be considered equivalent substitutions and are included within the protection scope of the present invention.

Claims

1. A dynamic energy-saving cluster resource collaborative allocation method, characterized in that, Includes the following steps: S1 collects environmental perception data: Real-time collection of environmental data and node status data through the drone's sensors; S2 predicts task requirements: Based on historical data and environmental changes, deep learning methods are used to predict future changes in task requirements; S3 optimizes resource utilization: generates resource allocation strategies and dynamically adjusts the computing and communication resources of each node; S4 makes energy-saving scheduling decisions: It uses deep reinforcement learning to generate energy-saving scheduling strategies, which simultaneously consider two objectives: task completion rate and energy consumption. S5 collects feedback and makes adaptive adjustments: it dynamically adjusts allocation strategies and energy-saving parameters based on task execution results and node status feedback.

2. The dynamic energy-saving cluster resource collaborative allocation method according to claim 1, characterized in that, In step S1, the collected data is denoised and normalized to generate a standardized dataset.

3. The dynamic energy-saving cluster resource collaborative allocation method according to claim 1, characterized in that, Step S2 uses a Long Short-Term Memory (LSTM) network model, inputting historical environmental data and historical task density to train the LSTM prediction model; Then, the trained LSTM model is used to predict the task requirements for future time windows; The prediction error function used is: E pred This is the prediction error, where w represents the length of the prediction time window. It is the task density predicted at time t+k. It represents the actual task density at time t+k.

4. The dynamic energy-saving cluster resource collaborative allocation method according to claim 1, characterized in that, In step S3, the objective function considers both task completion rate and resource utilization rate. The formula for calculating the objective function is as follows: F alloc This represents the objective function value for resource allocation optimization, where w1, w2, and w4 are the task completion rate coefficient, resource utilization rate coefficient, and allocation balance coefficient, respectively; T comp (i,t) represents the amount of task completed by node i at time t. It is the predicted task density, P used (i,t) represents the amount of resources used by node i at time t, R total (i,t) represents the total resource amount of node i at time t, R alloc (i,t) represents the proportion of resources allocated to node i at time t. is the average resource allocation ratio of all nodes at time t, and N is the number of nodes.

5. The dynamic energy-saving cluster resource collaborative allocation method according to claim 1, characterized in that, Step S4 employs a deep reinforcement learning method, combining resource allocation ratios and node states to generate an energy-saving scheduling strategy; the reward function R in deep reinforcement learning... drl (t) Taking into account both task completion rate and energy consumption, the reward function is as follows: In the formula, R drl (t) represents the reward value at time t, w1, μ, and ν are the task completion rate coefficient, energy consumption coefficient, and load balancing coefficient, respectively, N is the number of nodes, and T comp (i,t) represents the amount of task completed by node i at time t. It is the predicted task density, E cons (i,t) is the amount of electricity consumed by node i at time t, E node (i,t) is the charge of node i at time t, L comp (i,t) represents the computational load of node i at time t.

6. The dynamic energy-saving cluster resource collaborative allocation method according to claim 1, characterized in that, The strategy effectiveness evaluation function in step S5 is as follows: S effect (t) is the strategy effectiveness evaluation value at time t, w1, μ, ω3 are the task completion rate coefficient, energy consumption coefficient, and node lifespan coefficient, respectively, N is the number of nodes, and T rate (t) is the task completion rate at time t, E total (t) represents the total energy consumption of the cluster at time t, E max It is the cluster's maximum energy capacity, L life (i,t) represents the predicted lifetime of node i at time L. max This is the maximum lifetime value of the node.

7. The dynamic energy-saving cluster resource collaborative allocation method according to claim 1, characterized in that, In step S3, the objective function F is used. alloc Furthermore, by incorporating the dynamic adjustment mechanism of MARL, the resource allocation optimization problem is transformed into a multi-objective collaborative optimization problem. The objective function is as follows: w1, w2, w4, w3, and η are the task completion rate coefficient, resource utilization coefficient, allocation balance coefficient, energy saving coefficient, and potential field penalty coefficient, respectively; N is the number of nodes; and T is the number of nodes. comp (i,t) represents the amount of task completed by node i at time t. R represents the predicted task density or total demand at time t. used (i,t) represents the amount of resources used by node i at time t, R total (i,t) describes the total available resources of node i at time t, R alloc (i,t) is the proportion of resources allocated to node i at time t. E is the average resource allocation ratio at time t. save (i,t) represents the energy saved by node i at time t, which is the energy reduction achieved through resource allocation optimization. P field (t) is the potential field penalty term at time t, used to measure the dynamic balance of resource allocation; The potential field penalty items are as follows: in: F attr (i,t) represents the task demand attraction experienced by node i at time t, guiding resource allocation towards high-demand areas, F rep (i,t) represents the repulsive force of overloaded neighborhood resources experienced by node i at time t, to avoid excessive resource concentration. attr ,k rep As a moderating factor for attraction and repulsion, It is the predicted task demand value for a local region of node i, dist(i,D) center (t) is the distance from node i to the center point of the task requirement. Let i be the set of neighboring nodes of node i, and dist(i,j) be the distance between node i and neighboring node j. The MARL reward function calculation formula is as follows: R MARL (i,t) is the reinforcement learning reward value of node i at time t, and w1, w2, w3, w4 are the task completion rate coefficient, resource utilization coefficient, energy saving coefficient, and allocation balance coefficient, respectively.

8. The dynamic energy-saving cluster resource collaborative allocation method according to claim 1, characterized in that, Step S4 implements an online scheduling strategy that balances energy saving and task completion rate, dynamically adjusting node behavior based on predicted task requirements and real-time environmental changes. The specific implementation is as follows: By dynamically adjusting the allocation ratio of computing, communication and storage resources, combined with hibernation / activation strategies, energy consumption caused by idle or inefficient operation can be reduced. While saving energy, predictive models and real-time feedback are used to ensure that the task completion rate and response speed are not lower than the preset threshold. Resources are allocated collaboratively among nodes to avoid single-point overload leading to a surge in energy consumption and a shortened hardware lifespan for individual nodes. Each node acts as an intelligent agent, competing for limited resources while also collaborating to complete the overall task during resource allocation. To achieve the multiple objectives of simultaneously improving task completion rate, optimizing energy consumption, and balancing load, the following reward function is used: w1,μ,v,η,γ represent the task completion rate coefficient, energy consumption coefficient, load balancing coefficient, potential field penalty coefficient, and collaboration reward coefficient, respectively; T comp (i,t) represents the amount of tasks or services completed by node i at time t. E is the total demand for the forecast task at time t; cons (i,t) represents the energy consumption or energy depletion of node i at time t; E node (i,t) represents the total available or remaining power of node i at time t; L comp (i,t) represents the load of node i at time t; U pot (i,t) is the potential energy of node i at time t, used to guide load balancing; C coop (t) represents the cooperative reward at time t, which is calculated according to the cooperative calculation of multi-agent game, where N is the total number of nodes; Potential energy U pot (i,t) encourages tasks to migrate from high-load nodes to low-load nodes: k rep ,k attr >0 represents the repulsive and attractive coefficients, which control the strength of the potential field's influence; d ij The logical distance between nodes i and j; It is an indicator function; its value is 1 when the condition is true, and 0 otherwise; repulsion. Simulate the repulsive force of a high-load node on a low-load node; the greater the load difference and the closer the distance, the stronger the repulsive force. Gravity... Simulate the attractive force of a low-load node on a high-load node; the greater the load difference and the closer the distance, the stronger the attraction. comp (j,t) represents the load of node j at time t; Collaboration Reward C coop (t) Encourage nodes to maximize the global task completion rate through resource competition: T mig (i,j,t) represents the amount of task that node i needs to migrate to node j at time t, where T is the amount of task. comp (j,t) represents the amount of work completed by node j at time t.

Images