A green electricity adaptive operation method and system of a containerized mobile computing power node
By using a green energy adaptive operation method with containerized mobile computing nodes, the power type and green energy ratio are identified in real time. Adaptive decision-making is carried out in combination with a large language model, which solves the problems of low green energy absorption rate and complex state coupling. This achieves efficient green energy priority scheduling and computing power service continuity, and reduces operating costs.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- JINGNENG DIGITAL IND CO LTD
- Filing Date
- 2026-04-03
- Publication Date
- 2026-06-19
AI Technical Summary
Existing containerized mobile computing nodes suffer from low green energy consumption rates, complex state coupling, and high operating costs, making it difficult to achieve priority scheduling of green energy and efficient computing services, especially in remote areas and emergency scenarios.
By monitoring mobility status, network connectivity status, and power supply status in real time, the system uses electrical feature fingerprinting and blockchain technology to identify power type and green electricity ratio. It then uses a lightweight large language model for adaptive decision-making and optimizes model parameters through federated learning and cloud collaboration to achieve multi-mode adaptive operation.
It has improved the green electricity consumption rate, reduced operating costs and carbon emissions, enhanced decision-making accuracy and system stability, expanded service coverage, and ensured the optimal allocation of computing resources and the continuous improvement of model capabilities.
Smart Images
Figure CN122247985A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of mobile computing technology, and in particular to a green electricity adaptive operation method and system for containerized mobile computing nodes. Background Technology
[0002] With the rapid development of 5G, the Internet of Things (IoT), and artificial intelligence (AI), the demand for computing power is experiencing explosive growth. Edge computing scenarios, in particular, place higher demands on the timeliness, reliability, and deployment flexibility of computing power. Traditional data centers, limited by fixed deployments, high energy consumption, and long construction cycles, struggle to meet the computing power needs of remote areas (such as wind farms, photovoltaic power stations, mines, and oil and gas fields) and emergency scenarios (such as natural disaster relief and temporary large-scale events). Containerized mobile computing nodes, due to their standardization, modularity, rapid deployment, and flexible migration capabilities, have become an effective technological solution to these problems.
[0003] Currently, containerized mobile computing nodes have achieved preliminary collaborative operation with renewable energy sources (wind power, photovoltaics) in engineering applications, i.e., "computing-power collaboration". However, existing technical solutions still have the following technical defects in practical applications: First, the power source is unknown, making it impossible to achieve true green electricity priority scheduling. Existing mobile computing nodes typically connect to multiple power sources (mains power, wind power, photovoltaics, energy storage, truck power), but the power monitoring module can only detect basic electrical parameters such as total power and total voltage, and cannot identify the type of current power supply and the contribution ratio of each power source in real time. Due to the lack of refined perception of the power source, the system cannot distinguish between green electricity (wind power, photovoltaics) and gray electricity (mains power, diesel generators), resulting in the inability to implement the "green electricity priority" scheduling strategy, low green electricity absorption rate, and high operating costs and carbon emissions. Second, the state coupling is complex, and existing rule-driven methods are difficult to cope with nonlinear coupling relationships. In actual operation, mobile computing nodes face a complex scenario where the mobile state, power state, and network state are highly coupled: movement leads to network switching, network quality affects data transmission power consumption, and insufficient power limits network transmission capacity. Existing technologies generally employ rule-driven methods such as threshold judgment, weighted averaging, and finite state machines for decision-making. These methods are prone to problems such as decision lag, frequent mode switching (oscillations), and low resource utilization when facing nonlinear, time-varying, and strongly coupled dynamic environments. While there have been attempts in recent years to introduce large language models into the field of energy management, such as patent application CN121364938A entitled "A Resource Management Method, System, and Equipment for Intelligent Computing Centers Based on Computing-Power Collaboration," this patent primarily utilizes large models for macro-level load forecasting rather than for local real-time decision-making for mobile computing nodes; and invention patent CN121094054B entitled "A Lightweight Large Language Model Fine-tuning Method for the Power Grid Field," this patent focuses on the model fine-tuning technology itself and does not address multi-mode adaptive operation control of mobile computing nodes.
[0004] Furthermore, existing technologies suffer from weak off-grid autonomy, rudimentary edge-cloud collaboration, and low green energy utilization. In remote areas without network coverage, nodes can only operate based on preset rules and cannot self-optimize according to environmental changes, making it difficult to guarantee the continuity of computing power services. When connected to the network, only data synchronization is performed, lacking a deep model co-evolution mechanism, and failing to leverage the global knowledge of large cloud models to improve local decision-making capabilities. Existing scheduling strategies fail to fully consider the volatility and predictability of renewable energy, resulting in low green energy consumption rates and high operating costs. Summary of the Invention
[0005] This invention provides a green energy adaptive operation method and system for containerized mobile computing nodes to solve the problems of low green energy absorption rate, complex state coupling, poor continuity of computing services, and high operating costs in the prior art.
[0006] This invention is achieved through the following technical solution, providing a green energy adaptive operation method for containerized mobile computing nodes, comprising the following steps: S1, Status Awareness and Quantitative Assessment, real-time monitoring of the mobile status, network connection status and power supply status of containerized mobile computing nodes, and calculation of network status score and power status score; the monitoring of power supply status includes: collecting electrical parameters of each power input interface, using electrical feature fingerprint recognition technology to identify the currently connected power type, determining each power type and its contribution ratio, and calculating the green electricity ratio based on the contribution ratio. S2, State Encoding and Large Language Model Inference, encodes the mobility status, network connection status, and power supply status monitored by S1, as well as the calculated network status score and power status score, into natural language text according to a preset format. It is input into a lightweight large language model deployed locally on a containerized mobile computing node and outputs the target operating mode and operating parameters. S3, multi-mode adaptive operation, executes mobile mode, off-network mode, weak network mode or network mode on containerized mobile computing nodes according to the target operation mode output by S2; S4, Local closed-loop feedback and online learning, monitor the actual effect after the decision is executed, combine the effect data with the original state text to form a new training sample, and fine-tune the local lightweight large language model online; S5, Edge-Cloud Collaboration Model Optimization: When the network connection status meets preset conditions, the preset conditions are the network status score. When the value is greater than or equal to the first preset threshold, the parameter gradient of the local lightweight large language model is uploaded to the cloud-based large language model collaboration platform; the gradient uploaded by multiple containerized mobile computing power nodes is aggregated in the cloud, the global model parameters are updated, and the optimized parameters are sent to the local lightweight large language model. S6, Network Adaptive Model Working State Switching: Based on the network state score, the local lightweight large language model's working mode is automatically switched, including network collaborative working state, weak network compressed inference state, off-network autonomous inference state, and mobile low-power inference state.
[0007] Specifically, green electricity in S1 refers to the electricity generated by wind power and photovoltaic power generation; The calculation of the green electricity ratio described in S1 specifically includes the following steps: S11, Electrical parameter acquisition, acquires voltage and current waveforms of each power input interface at a sampling rate of not less than 1kHz; S12, Feature extraction and matching: Wavelet packet decomposition is used to extract time-frequency features, including fundamental amplitude, harmonic content, frequency fluctuation standard deviation, and transient response features. Feature vectors are constructed and input into a support vector machine classifier to match with a pre-stored electrical feature fingerprint database to identify the power supply type. S13, Blind Source Separation: When multiple power sources supply power simultaneously, the independent component analysis algorithm is used to separate the independent components of each power source from the mixed signal and calculate the contribution ratio of each power source. S14, Green Electricity Certification, uses blockchain technology to query green electricity transaction records and renewable energy certificates for the corresponding time period and location to verify the authenticity of the green electricity ratio; S15, the real-time contribution power of each power source is obtained through electrical feature fingerprint matching and blind source separation algorithm, and is denoted as wind power. Photovoltaic power Mains power Truck power supply and energy storage discharge power The energy storage discharge power may include a portion of electricity from green sources, which needs to be traced back based on the energy storage's charging history: if the proportion of green electricity absorbed by the energy storage system during charging is... Then the portion of the energy storage discharge power originating from green electricity is If it is not possible to trace accurately, The green electricity ratio G is calculated using the following formula:
[0008] Total power supply .
[0009] Specifically, the monitoring of the movement status described in S1 includes: Continuous location data is acquired using GPS sensors to calculate the moving speed v. When v>0 and the duration exceeds the first preset threshold, it is determined to be in a moving state; When v=0 and the duration exceeds the second preset threshold, it is determined to be a fixed deployment state.
[0010] The network state score mentioned in S1 is calculated according to the following formula:
[0011] in, Rate the network status; , This is the bandwidth normalized value. Preset desired bandwidth; , For delayed normalized values, The preset maximum allowable delay; , This is the normalized value for packet loss rate. The maximum allowable packet loss rate is preset; , This is the jitter normalization value. Set the maximum permissible jitter; , This is the normalized value of the signal strength; in, To monitor network bandwidth, For delay, For packet loss rate, For shaking, Signal strength; The corresponding weight coefficients, and satisfying .
[0012] Specifically, the power state score mentioned in S1 is calculated according to the following formula:
[0013] in, To score the power status, To score power availability, To score power quality, For power stability scoring, Rate available capacity; The corresponding weight coefficients, and satisfying ; If mains power is available, then A = 1.0; otherwise...
[0014] in, For wind power output power, For photovoltaic output power, This is the maximum power that the energy storage system can continuously and stably output under the current state of charge (SOC). Its value is taken as the smaller of the rated maximum continuous discharge power of the energy storage system and the SOC limit power. This represents the current load power requirement. Power quality rating Calculate using the following formula:
[0015] Wherein, THD represents harmonic content. To preset the maximum permissible harmonic content, For frequency deviation, The preset maximum allowable frequency deviation; For renewable energy supply Calculate using the following formula: ; in The variance of renewable energy power generation. This is the adjustment coefficient; Available capacity rating Calculate according to the following formula: ; SOC stands for State of Charge of the energy storage system.
[0016] Specifically, the online fine-tuning described in S4 includes... The performance data and the original state text are combined to form new training samples. An online gradient descent algorithm is used to fine-tune the model, with the learning rate dynamically adjusted based on the actual green energy consumption rate. The adjustment formula is as follows:
[0017] in, For learning rate, The initial learning rate, To adjust the coefficient, For green electricity consumption rate, To achieve the target green electricity consumption rate, .
[0018] Specifically, the cloud aggregation described in S5 uses a federated learning algorithm to update the global model parameters according to the following formula:
[0019] in, These are global model parameters. These are the local model parameters for the i-th node. This represents the number of nodes participating in the aggregation.
[0020] Specifically, the switching of the network adaptive model's working state in S6 includes: When the movement speed v>0, the local model enters a low-power inference state, and the inference frequency is reduced to once every 10 minutes. Priority is given to urgent tasks, and non-critical tasks are suspended. When v=0 and When the value is greater than or equal to the first preset threshold, switch to network collaborative working mode, and the local model and the cloud model work together to receive optimization parameters sent from the cloud in real time; When v=0 and the second preset threshold ≤ When the first preset threshold is reached, switch to weak network compressed inference state. The local model uses dynamic quantization technology to convert the model weights from floating point to integer, reducing the amount of inference computation and reducing the inference frequency to once every 5 minutes. When v=0 and 0≤ When the second preset threshold is reached, switch to off-network autonomous inference mode. The local model is fully autonomous, and all decision logs and effect data are stored locally and uploaded after the network is restored.
[0021] Specifically, this also includes S7, seamless mode switching. During the switching of operating modes, checkpoint saving is performed for computing tasks, state transition is performed for containerized tasks, and breakpoint resumption is performed for data transmission tasks to reduce task interruptions. This is achieved through the following steps: S71, micro-task-level checkpoints, save checkpoints once periodically for computationally intensive tasks; S72, task-level state migration, for containerized tasks, uses CRIU technology to save container state and a pre-copy strategy to migrate memory pages; S73, data-level breakpoint resume, for data transmission tasks, divides the file into blocks, confirms the transmission of each block, records the index of the transmitted block, and resumes transmission from the breakpoint after switching.
[0022] It also includes a green energy adaptive operation system for containerized mobile computing nodes, used to execute the above methods, including: The physical infrastructure layer includes wind power generation units, photovoltaic power generation units, energy storage systems, truck power supply interfaces, satellite communication equipment, GPS modules, and environmental control units, which provide power supply, communication, positioning, and operating environment support for containerized mobile computing nodes; the hardware resource layer includes GPU servers, CPU servers, and storage devices, which provide computing and storage resources required for model inference, task processing, and data storage. The software control layer, which is deployed on the hardware resource layer, includes a state awareness and power source deconstruction module, a local lightweight large language model decision engine, a multi-mode execution module, an edge-cloud collaboration module, and a container orchestration module. The status perception and power source deconstruction module is used to collect mobile status, network connection status and power supply status, and identify power source type and calculate green electricity ratio; The local lightweight large language model decision engine is used to generate target operating modes and operating parameters based on the state information and scoring results output by the state perception and power source deconstruction module. The multi-mode execution module is used to perform control operations corresponding to mobile mode, offline mode, weak network mode or network mode according to the target operating mode; The edge-cloud collaboration module is used to upload local model update parameters to the cloud and receive optimization parameters sent from the cloud when the network connection status meets preset conditions. The container orchestration module is used to perform task checkpoint saving, container state migration, and data breakpoint resume during the operation mode switching process. The cloud collaboration layer includes a cloud-based large language model collaboration platform, which receives model update parameters uploaded by multiple containerized mobile computing power nodes, aggregates and processes them using a federated learning algorithm, generates optimized global model parameters, and distributes them to the corresponding nodes. The data layer includes local storage, cloud storage, and a pre-trained data module specifically for green electricity. The local storage is used to save local decision logs, performance data, and offline running data, while the cloud storage is used to save global model data and collaborative training data. The green electricity-specific pre-training data module is used to provide green electricity-related training data and knowledge support for the local lightweight large language model decision engine and the cloud-based large language model collaborative platform.
[0023] Compared with the prior art, the present invention has the following beneficial effects: This invention utilizes electrical fingerprint recognition and blockchain-based green electricity traceability technology to achieve real-time identification of power source types and reliable measurement of green electricity proportions in mixed power supply scenarios. This method accurately distinguishes power source types and their contribution ratios, and verifies the authenticity of green electricity through blockchain. This technical solution enables mobile computing nodes to possess refined perception capabilities, thereby solving the problem in existing technologies where priority scheduling of green electricity cannot be implemented due to the unknown source of electricity. This effectively improves the green electricity consumption rate and reduces operating costs and carbon emissions.
[0024] This invention introduces a local lightweight large language model, encoding the coupled states of mobility, power, and network into natural language text input for model inference. Utilizing the semantic understanding capabilities of this local lightweight large language model, it learns complex nonlinear coupling relationships and outputs the optimal operating mode and related operating parameters. Simultaneously, the local lightweight large language model possesses network adaptive capabilities: it automatically enables dynamic quantization compression and reduces inference frequency in weak network conditions; it collaborates with the cloud when connected to the network; and it enters low-power autonomous operation in offline and mobile states. This technical solution replaces the traditional rule-driven threshold judgment method, solving the problems of nonlinear coupling and frequent pattern oscillations that existing technologies struggle to handle, significantly improving decision-making accuracy and system stability, and expanding service coverage in edge scenarios.
[0025] This invention employs a federated learning mechanism to aggregate model parameters across multiple nodes while protecting data privacy. It leverages the global knowledge of a large cloud-based model to optimize local models, forming a collaborative optimization capability of "local continuous learning + cloud-based global evolution." Simultaneously, a three-level state preservation mechanism enables seamless task-level switching during various operating mode transitions. This technical solution addresses the problems of inefficient cloud-mid-device collaboration, weak off-network autonomy, and task interruptions caused by mode switching in existing technologies. While ensuring business continuity, it achieves globally optimized allocation of computing resources and continuous improvement of model capabilities. Attached Figure Description
[0026] Figure 1 This is a flowchart of the green electricity adaptive operation method for containerized mobile computing nodes according to the present invention; Figure 2 This is a flowchart of the power source identification method for the green electricity adaptive operation method of the containerized mobile computing node of the present invention. Figure 3 This is a local lightweight large language model and online learning flowchart for the green electricity adaptive operation method of the containerized mobile computing node of the present invention. Figure 4 This is a schematic diagram of the optimization of the edge-cloud collaborative model for the green electricity adaptive operation method of the containerized mobile computing node of the present invention. Figure 5 This is an architecture diagram of the containerized mobile computing node green electricity adaptive operation system of the present invention. Detailed Implementation
[0027] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0028] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature; in the description of this application, unless otherwise stated, "multiple" means two or more.
[0029] Please see Figures 1-5 This invention provides a green energy adaptive operation method for containerized mobile computing nodes, comprising the following steps: S1: Status awareness and quantitative assessment, real-time monitoring of the mobility status, network connection status and power supply status of containerized mobile computing nodes, and calculation of network status score and power status score; In this embodiment, the monitoring of movement status includes: Continuous location data is acquired using GPS sensors to calculate the moving speed v, in km / h. When v>0 and the duration exceeds the first preset threshold T1, it is determined to be in a moving state. Preferably, in this embodiment, T1=10 seconds. When v=0 and the duration exceeds the second preset threshold T2, it is determined to be a fixed deployment state. Preferably, in this embodiment, T2=5 minutes.
[0030] In this embodiment, the network state score is calculated according to the following formula:
[0031] in, Rate the network status; , This is the bandwidth normalized value. To preset the desired bandwidth, 100Mbps is preferred; , For delayed normalized values, The maximum allowable delay is preset to 500ms; , This is the normalized value for packet loss rate. To preset the maximum allowable packet loss rate, a value of 10% is preferred. , This is the jitter normalization value. To preset the maximum allowable jitter, 100ms is preferred; , The signal strength is a normalized value, preferably ranging from -110dBm to -50dBm; To monitor network bandwidth (in Mbps). The delay is measured in milliseconds (ms). Packet loss rate (in %) Jitter (in milliseconds) Signal strength (in dBm); The corresponding weight coefficients, and satisfying .
[0032] Preferred, , , , , .
[0033] In this embodiment, the power status score is calculated according to the following formula:
[0034] in, To score the power status, To score power availability, To score power quality, For power stability scoring, Rate available capacity; The corresponding weight coefficients, and satisfying Preferred , , , .
[0035] If the mains power is available and of good quality, then A = 1.0; otherwise...
[0036] in, Wind power output (unit: kW). Photovoltaic output power (unit: kW). This is the maximum power that the energy storage system can continuously and stably output under the current state of charge (SOC). Its value is taken as the smaller of the rated maximum continuous discharge power of the energy storage system and the SOC limit power. Power required by the current load Power quality rating Calculate according to the following formula:
[0037] Wherein, THD is the harmonic content (unit: %). To preset the maximum permissible harmonic content, a value of 10% is preferred. Frequency deviation (in Hz) The maximum permissible frequency deviation is preset, preferably 1Hz; For renewable energy supply, power stability score Calculate according to the following formula: ; in The variance of renewable energy power generation. The adjustment coefficient is preferably set to 0.1; for mains power supply, S = 1.0. Available capacity rating Calculate according to the following formula. ; SOC stands for State of Charge of Energy Storage (in %).
[0038] In this embodiment, the monitoring of power supply status includes: collecting electrical parameters of each power input interface, using electrical feature fingerprinting technology to identify the currently connected power type, determining each power type and its contribution ratio, and calculating the green electricity ratio based on the contribution ratio.
[0039] In this invention, green electricity refers to the electrical energy generated by wind power generation and photovoltaic power generation.
[0040] Electrical fingerprinting technology is used to identify power supply types and determine the contribution ratio of each power supply when multiple power sources are supplying power simultaneously. Figure 2 As shown, the calculation of the green electricity ratio in S1 specifically includes the following steps: S11: Electrical parameter acquisition: Acquire the voltage and current waveforms of each power input interface at a sampling rate of not less than 1kHz; S12: Feature Extraction and Matching: Wavelet packet decomposition is used to extract time-frequency features, including fundamental amplitude, harmonic content (2nd-15th), frequency fluctuation standard deviation, and transient response features. A 128-dimensional feature vector is constructed and input into a support vector machine (SVM) classifier. The vector is then matched with a pre-stored electrical feature fingerprint database to identify the power source type (at least one of mains power, wind power, photovoltaic, energy storage, and truck power). S13: Blind source separation: When multiple power sources supply power simultaneously, the independent component analysis (ICA) algorithm is used to separate the independent components of each power source from the mixed signal and calculate the contribution ratio of each power source. S14: Green Electricity Certification: Based on blockchain technology, query green electricity transaction records and renewable energy certificates (RECs) for the corresponding time period and location to verify the authenticity of the green electricity ratio.
[0041] S15: The real-time contribution power of each power source is obtained through electrical feature fingerprint matching and blind source separation algorithm, and is denoted as wind power. Photovoltaic power Mains power Truck power supply and energy storage discharge power .
[0042] The energy storage discharge power may include a portion of the electricity from green electricity, which needs to be traced back based on the charging history of the energy storage: If the proportion of green electricity in the electricity absorbed by the energy storage system during charging is , The portion of the energy storage discharge power originating from green electricity is:
[0043] If it is not possible to trace accurately,
[0044] The green electricity ratio G is calculated using the following formula:
[0045] Total power supply .
[0046] The green electricity ratio G provides input for subsequent local lightweight large language model decisions, enabling the system to optimize its operating mode and task scheduling strategy based on the real-time green electricity ratio.
[0047] S2: Status encoding and large language model inference. The mobile status, network connection status and power supply status monitored by S1, as well as the calculated network status score and power status score, are encoded into natural language text according to a preset format. The input is a lightweight large language model deployed locally on the containerized mobile computing node, which can be called a local LLM. The output is the target operating mode and operating parameters. In this embodiment, the encoding format is: "Moving speed: [v] km / h, Network score: [S_net], Bandwidth: [B] Mbps, Latency: [D] ms, Packet loss rate: [L]%, Power score: [S_power], Mains power: [Available / Unavailable], Wind power: [P_wind] kW, Photovoltaic: [P_pv] kW, Energy storage SOC: [SOC]%, Truck power supply: [Yes / No], Green electricity ratio: [G]%." The local LLM is a compressed and quantized lightweight model (e.g., a GPT-2 series model with 150 million parameters, or a lightweight model fine-tuned by LoRA). Its pre-training data includes historical state data and corresponding optimal decisions. The pre-training objective is to learn the mapping relationship between the power-network-mobility coupling state and the optimal operating mode. The local LLM output text format is: "Recommended mode: [Mobile / Offline / Weak Network / Online], Load ratio: [X]%, Data compression: [On / Off], Synchronization frequency: [Y] minutes / times". The system parses the output text and extracts the target operating mode and related operating parameters.
[0048] S3: Multi-mode adaptive operation, which executes the corresponding operation in mobile mode, offline mode, weak network mode or network mode according to the target operation mode output by S2; S31, the operation in mobile mode includes: Trigger condition: The movement state is determined to be moving; Power supply method: Connect to the truck's power supply system and activate the energy storage system as a backup power source to achieve seamless power supply switching; Network connectivity: Switch to 4G / 5G mobile network and use QUIC protocol to optimize transmission efficiency; Task scheduling: Dynamically allocate computing resources based on task priority, reduce computing load to a preset ratio (preferably 70%), prioritize the execution of urgent tasks, and suspend non-critical tasks; Data management: Enable local storage, pause data upload, enable data compression, and adopt a breakpoint resume mechanism; Location-aware pre-scheduling: Combining GPS trajectory and destination information, it predicts the task requirements after arrival and loads relevant models and data in advance during transportation.
[0049] S32, off-network mode operation includes: Triggering conditions: Fixed deployment status, no network connection (0≤ <30); Power supply method: Switch to local renewable energy supply (wind power / solar power), with energy storage system as backup; Task scheduling: Start the local task scheduler and optimize task scheduling based on renewable energy generation forecasts; Data management: Store the data generated by task processing in the local storage system and generate data fingerprint identifiers.
[0050] S33, weak network mode operation includes: Triggering conditions: fixed deployment status, 30≤ <80; Power supply method: Automatic switching between mains power / wind power / photovoltaic / energy storage; Network connectivity: Enable data compression transmission; the compression rate is dynamically adjusted based on network quality. Task scheduling: Optimize task scheduling strategies to reduce data transmission requirements; Data Management: Enable breakpoint resume mechanism, implement priority transmission according to data importance level, and reduce data synchronization frequency.
[0051] S34, the operation in network mode includes: Triggering condition: Fixed deployment status. ≥80 and ≥60; Power supply method: Municipal power priority, supplemented by renewable energy; Network connectivity: Restore normal data transmission mechanism, disable data compression and resume interrupted transmission; Task scheduling: Register with the global coordinator in the cloud, participate in global load balancing, and receive and execute tasks assigned by the global coordinator; Data management: Synchronize local data stored during off-network mode, weak network mode, and mobile mode to the cloud storage system.
[0052] S4: As Figure 3As shown, local closed-loop feedback and online learning monitor the actual effect after the decision is implemented, combine the effect data with the original state text to form new training samples, and fine-tune the local lightweight large language model online. Monitoring the actual effects of decision implementation includes: Green electricity consumption rate Task completion time Number of mode switches ; Data transmission success rate .
[0053] In this embodiment, online fine-tuning specifically includes: The performance data and the original state text are combined to form new training samples. An online gradient descent algorithm is used to fine-tune the model, with the learning rate dynamically adjusted based on the actual green energy consumption rate. The adjustment formula is as follows:
[0054] in, For learning rate, The initial learning rate, To adjust the coefficient, For green electricity consumption rate, The target is the green energy absorption rate. Performance data should include at least one or more of the following: green energy absorption rate, task completion time, number of switching operations, or transmission success rate.
[0055] S5: Optimization of the edge-cloud collaboration model. When the network connection status meets the preset conditions, the preset conditions are: when the network status score is... In this embodiment, when the threshold value is greater than or equal to the first preset threshold, When the value is ≥80, the parameter gradients (not the original data) of the local lightweight large language model are uploaded to the cloud-based large language model collaboration platform. The cloud aggregates the gradients of multiple nodes, updates the global model parameters, and sends the optimized parameters to the local lightweight large language model. The node decrypts and updates the local model. At the same time, the node retains some underlying parameters unchanged to adapt to the local green electricity characteristics.
[0056] In this embodiment, as Figure 4 As shown, cloud aggregation uses the FedAvg federated learning algorithm to aggregate the gradients of multiple nodes and updates the global model parameters according to the following formula:
[0057] in, These are global model parameters. Here are the local model parameters for the i-th node, and N is the number of nodes participating in the aggregation.
[0058] S6: Network adaptive model working state switching. Based on the network state score, the local lightweight large language model working mode is automatically switched, including network collaborative working state, weak network compressed inference state, off-network autonomous inference state, and mobile low-power inference state.
[0059] The switching of the working state of the network adaptive model in S6 includes: When the movement speed v > 0, the local model enters a low-power inference state, reducing the inference frequency to once every 10 minutes, prioritizing urgent tasks and suspending non-critical tasks; when v = 0 and When the value is greater than or equal to the first preset threshold, the system switches to network collaborative working mode, where the local model and the cloud model work together, and the system receives optimization parameters from the cloud in real time. When v=0 and the second preset threshold ≤ When the first preset threshold is reached, switch to weak network compressed inference state. The local model uses dynamic quantization technology to convert the model weights from floating point to integer, reducing the amount of inference computation and reducing the inference frequency to once every 5 minutes. When v=0 and 0≤ When the second preset threshold is reached, the system switches to off-network autonomous inference mode. The local model is fully autonomous, and all decision logs and effect data are stored locally until the network is restored. S7, Seamless Mode Switching: During the mode switching process, checkpoint saving is performed on computational tasks, state migration is performed on containerized tasks, and breakpoint resumption is performed on data transmission tasks to reduce task interruptions. This includes the following steps: S71, Microtask-level checkpoints: For computationally intensive tasks, a checkpoint is saved every 30 seconds or after a certain number of iterations. The checkpoints include CPU / GPU register status, memory page status, and task execution progress.
[0060] S72, Task-level State Transition: For containerized tasks, CRIU (Checkpoint / Restore In Userspace) technology is used to save the container state, and a pre-copy strategy is used to migrate memory pages.
[0061] S73, Data-level Resume Transmission: For data transmission tasks, the file is divided into blocks, each block is confirmed after transmission, the index of the transmitted block is recorded, and transmission continues from the breakpoint after switching.
[0062] S74, time adaptation strategy: power switching time <1ms (hardware redundancy), network switching time <100ms (multiple links maintained simultaneously), task migration interruption time <100ms (pre-copy technology).
[0063] like Figure 5 As shown, the present invention also includes a green energy adaptive operation system for containerized mobile computing nodes, used to implement the above method, comprising: The physical infrastructure layer includes wind power generation units, photovoltaic power generation units, energy storage systems, truck power supply interfaces, satellite communication equipment, GPS modules, and environmental control units, which provide power supply, communication, positioning, and operating environment support for containerized mobile computing nodes. The hardware resource layer includes GPU servers, CPU servers, and storage devices, which provide the computing and storage resources required for model inference, task processing, and data storage. The software control layer is deployed on the hardware resource layer and includes a state awareness and power source deconstruction module, a local lightweight large language model decision engine, a multi-mode execution module, an edge-cloud collaboration module, and a container orchestration module. The status awareness and power source deconstruction module is used to collect mobile status, network connection status and power supply status, and identify power source type and calculate green electricity ratio; The local lightweight large language model decision engine is used to generate the target operating mode and operating parameters based on the state information and scoring results output by the state perception and power source deconstruction module; The multi-mode execution module is used to execute control operations corresponding to mobile mode, offline mode, weak network mode or network mode according to the target operating mode; The edge-cloud collaboration module is used to upload local model update parameters to the cloud and receive optimization parameters sent from the cloud when the network connection status meets preset conditions. The container orchestration module is used to perform task checkpoint saving, container state migration, and data breakpoint resumption during run mode switching. The cloud collaboration layer includes a cloud-based large language model collaboration platform, which receives model update parameters uploaded by multiple containerized mobile computing power nodes, aggregates and processes them using federated learning algorithms, generates optimized global model parameters, and distributes them to the corresponding nodes. The data layer includes local storage, cloud storage, and a pre-trained data module specifically for green electricity. Local storage is used to save local decision logs, performance data, and offline running data. Cloud storage is used to save global model data and collaborative training data. The green electricity-specific pre-training data module is used to provide green electricity-related training data and knowledge support for the local lightweight large language model decision engine and the cloud-based large language model collaborative platform.
[0064] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit it. This application is not limited to the exact structures described above and illustrated in the accompanying drawings, and it should not be considered that the specific implementation of this application is limited to these descriptions. For those skilled in the art, various changes and modifications made without departing from the concept of this application should be considered to fall within the protection scope of this application.
Claims
1. A green energy adaptive operation method for containerized mobile computing nodes, characterized in that, Includes the following steps: S1, Status Awareness and Quantitative Assessment, monitors the mobility status, network connectivity status, and power supply status of containerized mobile computing nodes in real time, and calculates network status scores and power status scores. The monitoring of the power supply status includes: collecting electrical parameters of each power input interface, using electrical feature fingerprinting technology to identify the currently connected power source type, determining each power source type and its contribution ratio, and calculating the green electricity ratio based on the contribution ratio. S2, State Encoding and Large Language Model Inference, encodes the mobility status, network connection status, and power supply status monitored by S1, as well as the calculated network status score and power status score, into natural language text according to a preset format. It is input into a lightweight large language model deployed locally on a containerized mobile computing node and outputs the target operating mode and operating parameters. S3, multi-mode adaptive operation, executes mobile mode, off-network mode, weak network mode or network mode on containerized mobile computing nodes according to the target operation mode output by S2; S4, Local closed-loop feedback and online learning, monitor the actual effect after the decision is executed, combine the effect data with the original state text to form a new training sample, and fine-tune the local lightweight large language model online; S5, Edge-Cloud Collaboration Model Optimization: When the network connection status meets preset conditions, the preset conditions are the network status score. When the value is greater than or equal to the first preset threshold, the parameter gradient of the local lightweight large language model is uploaded to the cloud-based large language model collaboration platform; the gradient uploaded by multiple containerized mobile computing power nodes is aggregated in the cloud, the global model parameters are updated, and the optimized parameters are sent to the local lightweight large language model. S6, Network Adaptive Model Working State Switching: Based on the network state score, the local lightweight large language model's working mode is automatically switched, including network collaborative working state, weak network compressed inference state, off-network autonomous inference state, and mobile low-power inference state.
2. The green energy adaptive operation method for containerized mobile computing nodes according to claim 1, characterized in that, In S1, green electricity refers to the electrical energy generated by wind power and photovoltaic power generation. The calculation of the green electricity ratio described in S1 specifically includes the following steps: S11, Electrical parameter acquisition, acquires voltage and current waveforms of each power input interface at a sampling rate of not less than 1kHz; S12, Feature extraction and matching: Wavelet packet decomposition is used to extract time-frequency features, including fundamental amplitude, harmonic content, frequency fluctuation standard deviation, and transient response features. Feature vectors are constructed and input into a support vector machine classifier to match with a pre-stored electrical feature fingerprint database to identify the power supply type. S13, Blind Source Separation: When multiple power sources supply power simultaneously, the independent component analysis algorithm is used to separate the independent components of each power source from the mixed signal and calculate the contribution ratio of each power source. S14, Green Electricity Certification, uses blockchain technology to query green electricity transaction records and renewable energy certificates for the corresponding time period and location to verify the authenticity of the green electricity ratio; S15, the real-time contribution power of each power source is obtained through electrical feature fingerprint matching and blind source separation algorithm, and is denoted as wind power. Photovoltaic power Mains power Truck power supply and energy storage discharge power The energy storage discharge power may include a portion of electricity from green sources, which needs to be traced back based on the energy storage's charging history: if the proportion of green electricity absorbed by the energy storage system during charging is... Then the portion of the energy storage discharge power originating from green electricity is ,in If it is not possible to trace accurately, The green electricity ratio G is calculated using the following formula: Total power supply .
3. The green energy adaptive operation method for containerized mobile computing nodes according to claim 1, characterized in that, The monitoring of the movement status described in S1 includes: Continuous location data is acquired using GPS sensors to calculate the moving speed v. When v>0 and the duration exceeds the first preset threshold, it is determined to be in a moving state; When v=0 and the duration exceeds the second preset threshold, it is determined to be a fixed deployment state.
4. The green energy adaptive operation method for containerized mobile computing nodes according to claim 1, wherein the network status score in S1 is calculated according to the following formula: in, Rate the network status; , This is the bandwidth normalized value. Preset desired bandwidth; , For delayed normalized values, The preset maximum allowable delay; , This is the normalized value for packet loss rate. The maximum allowable packet loss rate is preset; , This is the jitter normalization value. Set the maximum allowable jitter; , This is the normalized value of the signal strength; in, To monitor network bandwidth, For delay, For packet loss rate, For shaking, Signal strength; The corresponding weight coefficients, and satisfying .
5. The green energy adaptive operation method for containerized mobile computing nodes according to claim 4, characterized in that, The power status score mentioned in S1 is calculated according to the following formula: in, To score the power status, To score power availability, To score power quality, For power stability scoring, Rate the available capacity; The corresponding weight coefficients, and satisfying ; If mains power is available, then A = 1.0; otherwise... in, For wind power output power, For photovoltaic output power, This is the maximum power that the energy storage system can continuously and stably output under the current state of charge (SOC). Its value is taken as the smaller of the rated maximum continuous discharge power of the energy storage system and the SOC limit power. Current load power demand; power quality score Calculate using the following formula: Wherein, THD represents harmonic content. To preset the maximum permissible harmonic content, For frequency deviation, The preset maximum allowable frequency deviation; For renewable energy supply Calculate using the following formula: ； in The variance of renewable energy power generation fluctuations. This is the adjustment coefficient; Available capacity rating Calculate according to the following formula: ； SOC stands for State of Charge of the energy storage system.
6. The green energy adaptive operation method for containerized mobile computing nodes according to claim 1, characterized in that, The online fine-tuning described in S4 specifically includes: The performance data and the original state text are combined to form new training samples. An online gradient descent algorithm is used to fine-tune the model, with the learning rate dynamically adjusted based on the actual green energy consumption rate. The adjustment formula is as follows: in, For learning rate, The initial learning rate, To adjust the coefficient, For green electricity consumption rate, To achieve the target green electricity consumption rate, .
7. The green energy adaptive operation method for containerized mobile computing nodes according to claim 1, characterized in that, The cloud aggregation described in S5 uses a federated learning algorithm to update the global model parameters according to the following formula: in, These are global model parameters. Here are the local model parameters for the i-th node, and N is the number of nodes participating in the aggregation.
8. The green energy adaptive operation method for containerized mobile computing nodes according to claim 1, characterized in that, The switching of the working state of the network adaptive model in S6 includes: When the movement speed v > 0, the local model enters a low-power inference state, reducing the inference frequency to once every 10 minutes, prioritizing urgent tasks and suspending non-critical tasks; when v = 0 and When v=0 and the second preset threshold is less than or equal to the first preset threshold, the system switches to network collaborative working mode, where the local model and the cloud model work together, receiving optimization parameters from the cloud in real time; when v=0 and the second preset threshold is less than or equal to the second preset threshold, the system switches to network collaborative working mode, where the local model and the cloud model work together, receiving optimization parameters from the cloud in real time. When the threshold is less than the first preset threshold, switch to weak network compressed inference mode. The local model uses dynamic quantization technology to convert model weights from floating-point to integer, reducing the amount of inference computation and reducing the inference frequency to once every 5 minutes; when v=0 and 0≤ When the second preset threshold is reached, the system switches to off-network autonomous inference mode, where the local model is fully autonomous, and all decision logs and effect data are stored locally and uploaded after the network is restored.
9. The green energy adaptive operation method for containerized mobile computing nodes according to claim 1, characterized in that, It also includes S7, seamless switching, which performs checkpoint saving for computing tasks, state transition for containerized tasks, and breakpoint resumption for data transmission tasks during the operation mode switching process to reduce task interruption. Specifically, it is implemented through the following steps: S71, micro-task level checkpoint, for computing-intensive tasks, checkpoints are saved once periodically. S72, task-level state migration, for containerized tasks, uses CRIU technology to save container state and a pre-copy strategy to migrate memory pages; S73, data-level breakpoint resume, for data transmission tasks, divides the file into blocks, confirms the transmission of each block, records the index of the transmitted block, and resumes transmission from the breakpoint after switching.
10. A green energy adaptive operation system for containerized mobile computing nodes, used to execute the method as described in any one of claims 1-9, characterized in that, include: The physical infrastructure layer includes wind power generation units, photovoltaic power generation units, energy storage systems, truck power supply interfaces, satellite communication equipment, GPS modules, and environmental control units, which are used to provide power supply, communication, positioning, and operating environment support for containerized mobile computing nodes. The hardware resource layer includes GPU servers, CPU servers, and storage devices, which are used to provide computing and storage resources required for model inference, task processing, and data storage. The software control layer, deployed on the hardware resource layer, includes a state awareness and power source deconstruction module, a local lightweight large language model decision engine, a multi-mode execution module, an edge-cloud collaboration module, and a container orchestration module. The state awareness and power source deconstruction module is used to collect mobile status, network connection status, and power supply status, and identify power type and calculate green electricity ratio. The local lightweight large language model decision engine is used to generate target operating modes and operating parameters based on the state information and scoring results output by the state perception and power source deconstruction module. The multi-mode execution module is used to perform control operations corresponding to mobile mode, offline mode, weak network mode or network mode according to the target operating mode; The edge-cloud collaboration module is used to upload local model update parameters to the cloud and receive optimization parameters sent from the cloud when the network connection status meets preset conditions. The container orchestration module is used to perform task checkpoint saving, container state migration, and data breakpoint resume during the operation mode switching process. The cloud collaboration layer includes a cloud-based large language model collaboration platform, which receives model update parameters uploaded by multiple containerized mobile computing power nodes, aggregates and processes them using a federated learning algorithm, generates optimized global model parameters, and distributes them to the corresponding nodes. The data layer includes local storage, cloud storage, and a pre-trained data module specifically for green electricity. The local storage is used to save local decision logs, performance data, and offline running data, while the cloud storage is used to save global model data and collaborative training data. The green electricity-specific pre-training data module is used to provide green electricity-related training data and knowledge support for the local lightweight large language model decision engine and the cloud-based large language model collaborative platform.