Active sampling and constrained reinforcement learning edge collaboration method

Abstract

The invention provides an active sampling and constrained reinforcement learning edge collaboration method, and relates to the field of data processing. Modeling and sparse coding are carried out through sampling and event data collected by edge computing nodes, compressed coding data are generated, and transmission is optimized through a self-adaptive data flow scheduling algorithm. And verifying the authenticity of the data by utilizing a generative adversarial network, and optimizing an equipment scheduling scheme by adopting reinforcement learning according to a verification result. And summarizing the local models of the edge nodes through federal learning to generate a global model, and protecting data privacy by applying a differential privacy technology. And finally, based on the privacy protection model, in combination with meta-learning and online convex optimization technologies, executing an equipment scheduling scheme, generating an edge collaboration strategy, and optimizing carbon emission control and energy utilization efficiency of the green park. By implementing the technical scheme, the equipment can cooperate with each other through an intelligent technology, and the operation strategy is dynamically adjusted, so that the overall efficiency of the park system is improved.

Description

Technical Field

[0001] This application relates to the field of data processing technology, specifically to an active sampling and constrained reinforcement learning edge collaboration method. Background Technology

[0002] In green industrial parks, carbon emissions primarily originate from building heating, factory production, and transportation activities. Reducing carbon emissions is one of the core objectives of green industrial park management, and equipment scheduling and resource management are key aspects of improving overall energy efficiency and reducing carbon emissions.

[0003] Traditional equipment scheduling methods often focus on optimizing the operation of individual devices, neglecting the interdependencies and collaborations between them. This approach cannot effectively handle the interactive effects of multiple devices operating under different conditions, leading to low overall system efficiency and significant shortcomings in carbon emission control. Furthermore, since the data and control strategies of each device are typically optimized locally, lacking a global perspective, they are prone to local optima and cannot achieve global system optimization.

[0004] Therefore, the collaborative operation of multiple devices within a park depends not only on the operating status of individual devices but also on the mutual influence and collaborative optimization between them. How to leverage intelligent technologies to enable devices to cooperate and dynamically adjust their operating strategies, thereby improving the overall efficiency of the park system, has become a core technological challenge in carbon emission management for green parks. Summary of the Invention

[0005] This application provides an active sampling and constrained reinforcement learning edge collaboration method, which uses intelligent technology to enable devices to cooperate with each other and dynamically adjust their operating strategies, thereby improving the overall efficiency of the park system.

[0006] A first aspect of this application provides an active sampling and constrained reinforcement learning edge collaboration method, the method comprising: acquiring sampling data and event data sent by edge computing nodes targeting carbon emissions in green parks; modeling and sparsely encoding the sampling data and event data to obtain compressed coded data; processing the compressed coded data using an adaptive data stream scheduling algorithm to obtain optimized transmission data; performing multi-evidence verification on the optimized transmission data using a generative adversarial network to obtain verification results; optimizing the equipment scheduling for carbon emissions in green parks using a reinforcement learning algorithm based on the verification results to generate an optimal equipment scheduling scheme; aggregating the local models of each edge computing node using a federated learning algorithm to obtain a global model, and controlling the privacy level of data uploads in the global model using differential privacy technology to obtain a privacy-preserving model; and executing the optimal equipment scheduling scheme based on the privacy-preserving model using meta-learning and online convex optimization techniques to obtain an edge collaboration strategy.

[0007] A second aspect of this application provides an active sampling and constrained reinforcement learning edge collaboration device. The device includes an acquisition module and a processing module. The acquisition module is used to acquire sampling data and event data sent by edge computing nodes targeting carbon emissions in green parks. The processing module is used to model and sparsely encode the sampling data and event data to obtain compressed encoded data. The processing module is further used to process the compressed encoded data using an adaptive data flow scheduling algorithm to obtain optimized transmission data. The processing module is further used to perform multi-evidence verification on the optimized transmission data using a generative adversarial network to obtain verification results. The processing module is further used to optimize the equipment scheduling for carbon emissions in green parks using a reinforcement learning algorithm based on the verification results to generate an optimal equipment scheduling scheme. The processing module is further used to aggregate the local models of each edge computing node using a federated learning algorithm to obtain a global model, and control the privacy level of data uploads in the global model using differential privacy technology to obtain a privacy-preserving model. The processing module is further used to execute the optimal equipment scheduling scheme based on the privacy-preserving model using meta-learning and online convex optimization techniques to obtain an edge collaboration strategy.

[0008] A third aspect of this application provides an electronic device including a processor, a memory, a user interface, and a network interface. The memory is used to store instructions, and both the user interface and the network interface are used to communicate with other devices. The processor is used to execute the instructions stored in the memory to cause the electronic device to perform the method described above.

[0009] A fourth aspect of this application provides a computer-readable storage medium storing instructions that, when executed, perform the method described above.

[0010] In summary, one or more technical solutions provided in this application have at least the following technical effects or advantages: By introducing federated learning and differential privacy technologies, the privacy of local data on each edge computing node is ensured throughout the processing and model training process. Raw data is not directly transmitted to the central endpoint; only model parameters are uploaded, effectively avoiding the risk of data leakage and enabling optimization using distributed data without infringing on privacy. An adaptive data flow scheduling algorithm dynamically adjusts the data transmission process based on network bandwidth and data importance. This ensures efficient data transmission even with limited bandwidth, avoiding redundant information transmission and improving speed and reliability. Generative adversarial networks (GANs) are used for multi-evidence verification of optimized transmitted data, ensuring its authenticity and validity. Through multi-layered verification, the system can detect potential errors or biases in the data, ensuring that final decisions are based on accurate data, thereby improving system credibility. Reinforcement learning algorithms optimize equipment scheduling schemes, considering not only carbon emission and energy consumption targets but also dynamically responding to different operating states and environmental changes within the park, automatically generating optimal scheduling strategies. Furthermore, through meta-learning and online convex optimization techniques, the system can quickly adapt to changing environments and adjust equipment scheduling in real time, ensuring collaborative work between devices and improving the overall efficiency of the park system. Attached Figure Description

[0011] Figure 1 A flowchart illustrating an active sampling and constrained reinforcement learning edge collaboration method provided in an embodiment of this application; Figure 2 A schematic diagram of a module for an active sampling and constrained reinforcement learning edge collaboration device provided in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0012] Explanation of reference numerals in the attached figures: 21. Acquisition module; 22. Processing module; 31. Processor; 32. Communication bus; 33. User interface; 34. Network interface; 35. Memory. Detailed Implementation

[0013] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0014] In the description of the embodiments of this application, the words "for example" or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design that is described as "for example" or "for instance" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design options. Rather, the use of the words "for example" or "for instance" is intended to present the relevant concepts in a specific manner.

[0015] In the description of the embodiments of this application, the term "multiple" means two or more. For example, multiple systems means two or more systems, and multiple screen terminals means two or more screen terminals. Furthermore, 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 indicated technical features. Thus, a feature defined with "first" or "second" may explicitly or implicitly include one or more of that feature. The terms "comprising," "including," "having," and variations thereof all mean "including but not limited to," unless otherwise specifically emphasized.

[0016] To address the aforementioned technical problems, this application provides a collaborative method for active sampling and constrained reinforcement learning at the edge, referring to... Figure 1 , Figure 1 This is a flowchart illustrating an active sampling and constrained reinforcement learning edge collaboration method provided in an embodiment of this application. The method is applied to a server and includes steps S110 to S170, as follows:

[0017] S110. Obtain sampling data and event data sent by edge computing nodes targeting carbon emissions in green parks.

[0018] Specifically, a server refers to a computer system or hardware device used to process, store, and manage data. In a green park carbon emission management system, the server acts as a centralized processing platform, responsible for receiving data from multiple edge computing nodes, analyzing the data, and making decisions. Servers possess strong computing and storage capabilities, supporting large-scale data processing and system operation. For example, assuming there are multiple monitoring points within the park, the server is responsible for receiving real-time data transmitted from these monitoring points, performing centralized processing and analysis. A green park refers to an industrial park or residential area with sustainable development as its core objective, committed to reducing carbon emissions, improving energy efficiency, optimizing resource utilization, and protecting the environment. Green parks integrate various environmental protection technologies and intelligent management systems, implementing measures such as smart buildings, clean energy, and energy conservation and emission reduction. For example, a green park may include multiple solar power generation systems, smart buildings, and efficient wastewater treatment facilities, all aimed at reducing the park's carbon emissions and energy consumption.

[0019] Carbon emissions refer to the release of greenhouse gases such as carbon dioxide into the atmosphere due to human activities, including the burning of fossil fuels and industrial production. Carbon emissions are a significant factor in global climate change. For example, in green parks, carbon emissions primarily originate from building heating, factory production, and transportation activities within the park, and reducing carbon emissions is one of the core objectives of green park management. Edge computing nodes are computing devices deployed close to data sources or users, responsible for data collection, preliminary processing, and storage. Compared to traditional cloud computing, edge computing distributes data processing tasks to the "edge" of the network, reducing latency and improving processing efficiency. Edge computing nodes can be various devices, such as sensors, gateways, and smart devices. For example, environmental sensors in green parks, such as carbon dioxide concentration sensors and temperature and humidity sensors, are edge computing nodes responsible for real-time data collection and transmitting the data to a central server for further processing.

[0020] Sampling data refers to information related to the operation of an environment or equipment, collected periodically through sensors or monitoring devices. In green parks, sampling data may include temperature, humidity, carbon dioxide concentration, and energy consumption. Sampling data is collected at regular intervals or under specific conditions for subsequent analysis and decision-making. For example, temperature and humidity sensors record temperature and humidity data hourly and transmit the data to edge computing nodes. Event data refers to data related to specific events occurring in the system or environment; it is immediate, sudden, or abnormal. Event data reflects changes in system status or abnormal equipment behavior. For example, event data in green parks may include equipment malfunction warnings, peak energy consumption, and excessive carbon emissions. This data is crucial for real-time monitoring and emergency response. For example, when the operation of a piece of equipment exceeds a set carbon emission threshold, the system generates event data to notify management for intervention.

[0021] S120. Model and sparsely encode the sampled data and event data to obtain compressed coded data.

[0022] Specifically, modeling refers to constructing mathematical models that reflect the characteristics, patterns, or system behavior of data through mathematical methods or algorithms. In carbon emission management in green industrial parks, modeling can help understand and predict changes in equipment and environmental data within the park. For example, based on collected data such as temperature, humidity, and electricity consumption, a regression model can be built to predict trends in carbon emissions, or a classification model can be established to identify different energy use patterns. Modeling is a crucial step in data analysis and decision support, transforming raw data into a structure that is easy to understand and analyze.

[0023] Sparse coding is a technique that represents data as a linear combination of a set of sparse coefficients and dictionary elements. The goal of sparse coding is to represent data more compactly by selecting a small number of dictionary elements and removing redundant information. In practical applications, sparse coding can effectively reduce data storage and computational overhead. In green parks, sparse coding is used to represent high-dimensional data collected from sensors, compressing it into a sparse vector while retaining the most critical information and removing redundancy. For example, sparse coding can transform a set of high-dimensional environmental data into a few key features without needing to store all data points.

[0024] Compressed data is a data representation that has undergone processing using techniques such as sparse coding, resulting in a significant reduction in data size or storage requirements. By removing redundant and repetitive information, compressed coding makes data more compact without losing key features, facilitating storage and transmission. For example, by sparsely coding environmental monitoring data, multi-dimensional data that would otherwise require storage, such as minute-by-minute temperature and humidity values, can be compressed into smaller vector representations, thereby reducing the bandwidth requirements for data transmission. For instance, in a green park, sparsely coded sensor data can be transmitted to a central server via a low-bandwidth network, reducing data transmission latency and saving storage space.

[0025] Furthermore, firstly, the goal of the modeling process is to process the sampled and event data using mathematical models, capture the temporal characteristics of the data, and generate multidimensional modeling data. Assume that the environmental monitoring data collected at different time points in the green park are as follows: ,in Indicates a point in time. It is a multidimensional data vector, representing time... All environmental data (e.g., temperature, humidity, carbon dioxide concentration) were collected at all times. A time series model was selected, and model parameters were estimated using historical data. For the sampled data... This can be represented using an autoregressive model:

[0026] in: These are the autoregressive coefficients of the model; At any moment The collected data; This is the error term, reflecting the portion that failed to be predicted by the model. At this point, It represents a time series data vector, and through model training and parameter estimation, it captures the trend and change pattern of time series data.

[0027] Next, sparse coding is implemented using sparse representation techniques to process these multidimensional modeling data. This is transformed into a more compact sparse representation. The goal of sparse coding is to use a dictionary... Data To represent, dictionary For one The matrix, where It is a data dimension. This represents the number of base elements in the dictionary. The goal of sparse coding is to find a sparse coefficient matrix. This makes the data Able to use a dictionary and sparsity coefficient Represent it as a linear combination, i.e.:

[0028] in: The original data matrix, For the dimensions of the data, This represents the number of data samples. It is a sparse dictionary matrix, where each column of the dictionary is a basic element; It is a sparse coefficient matrix, representing the linear combination coefficients of each sample across the dictionary elements. To find the optimal dictionary... sparse coefficient matrix The following optimization problem can be solved:

[0029] in: It is the reconstruction error, which measures the error when representing data using a dictionary; It is a sparsity constraint term; the L1 norm forces a sparse coefficient matrix. Only a few non-zero elements; This is the regularization parameter, used to control the balance between sparsity and reconstruction error. By optimizing this objective function, the optimal dictionary is obtained. sparse coefficient matrix ,in This is the final compressed and encoded data. Sparse coefficient matrix. It retains the key information of the original data, while making the data more compact through sparse representation, thus making it more efficient in storage and transmission.

[0030] By modeling sampled data and event data, the system can capture the temporal characteristics of the data and generate multidimensional modeling data. Subsequently, sparse representation technology is used to encode these data, so that the data is compressed in a sparse form, thereby reducing redundancy and improving the efficiency of data processing.

[0031] S130. An adaptive data stream scheduling algorithm is used to process the compressed encoded data to obtain optimized transmission data.

[0032] Specifically, adaptive data flow scheduling algorithms refer to algorithms that dynamically adjust data flow processing and transmission methods based on real-time network conditions and data demand. Its core objective is to maximize data transmission efficiency by flexibly adjusting data transmission priority, transmission rate, and data batch size when network bandwidth is limited or unstable. Adaptive algorithms can sense network load in real time and dynamically adjust transmission strategies based on changes in network bandwidth and latency. For example, when network bandwidth is high, the system can transmit more data simultaneously; while when network bandwidth is low, the system will reduce the amount of data transmitted each time to avoid network congestion. For instance, suppose there are multiple sensor nodes in a green park that transmit monitoring data to the central system in real time. If the system detects low network bandwidth at a certain time, the data flow scheduling algorithm will automatically adjust, prioritizing the transmission of higher-priority carbon emission data while delaying the upload of less urgent data.

[0033] Optimized data transmission refers to the efficient transmission of data after a series of algorithmic processing steps. The optimization process includes not only data compression but also data sorting and priority adjustment, ensuring that the most important data is transmitted and processed in the shortest possible time. The core purpose of optimized data transmission is to minimize data loss and improve network resource utilization under bandwidth constraints or network instability, thereby guaranteeing the real-time performance and accuracy of data transmission. For example, when network bandwidth is sufficient, devices within a green space can quickly transmit large amounts of data. However, when network bandwidth is insufficient, the system will optimize transmission strategies, prioritizing the uploading of critical information while delaying or merging the transmission of other information. In this way, the system can ensure the timely delivery of important information and reduce latency or packet loss during data transmission.

[0034] Furthermore, firstly, the communication parameters of the current network environment are determined, and network conditions are obtained based on these parameters. The system needs to monitor the current network status in real time to adjust the data flow transmission strategy according to different network conditions. Network environment communication parameters include bandwidth, latency, packet loss rate, network congestion, and other information, all of which are key factors affecting data transmission efficiency and reliability. By monitoring these communication parameters in real time, the system can assess the current network quality. For example, the server can periodically or dynamically measure network bandwidth. ,Delay Packet loss rate Parameters such as bandwidth allow for large-volume, high-speed data transmission; however, when latency is high or packet loss is significant, the system needs to make corresponding adjustments to avoid affecting data accuracy and timeliness. Specific network conditions are also crucial. This can be expressed through the following relationship:

[0035] in: It's bandwidth. It's a delay. It's the packet loss rate; It is a function that represents how communication parameters collectively affect network conditions.

[0036] Next, based on the timeliness and importance of different data under varying network conditions, the priority of compressed and encoded data is determined, and a data stream is generated. Once the current network conditions are obtained, the system needs to determine the priority of each batch of data based on its timeliness and importance. Data with high timeliness, such as urgent equipment failure information or carbon emission exceeding warnings, needs to be uploaded first to ensure the park management system can respond promptly; while data with low timeliness, such as historical environmental data or periodically collected status information, can be transmitted later. The system assigns a priority to each data packet by setting a priority strategy. For example, high-priority data can be given lower transmission delays, while low-priority data can be scheduled using a delay mechanism. A data stream is generated based on timeliness and importance. Each data stream contains multiple data packets, and the priority of each data packet is determined based on its importance and timeliness. The formula is as follows:

[0037] in: Indicates the first One data packet; Indicates data packet The lower the value, the higher the priority.

[0038] Then, an adaptive data flow scheduling algorithm is used to schedule the data flow in real time, automatically adjusting the data transmission frequency, batch size, and order. After acquiring the data flow and determining the priority of each data packet, the system needs to use the adaptive data flow scheduling algorithm to schedule the data flow in real time. The core of this algorithm is to dynamically adjust the data transmission frequency, batch size, and order based on the current network conditions. For example, if the network bandwidth is high, large-batch data transmission can be selected; if the network bandwidth is limited, it should be adjusted to small-batch, high-frequency data transmission. Through this mechanism, the system can ensure that high-priority data is transmitted first, avoiding the loss of critical data when the network is congested. Assume that at a certain moment, the system dynamically adjusts the transmission order of data packets based on the current network conditions and priorities. And adjust the transmission batch size according to network conditions. , can be represented as:

[0039] in: This refers to the current network conditions, including bandwidth and latency. It refers to the priority of the data packets; This is the adjusted data packet transmission order.

[0040] Next, data transmission efficiency is adjusted through a data flow control mechanism to optimize data transmission. To further optimize data transmission, the system uses a data flow control mechanism for adjustment. This mechanism ensures balanced bandwidth usage during data transmission, preventing network congestion caused by excessive simultaneous data transmission. The flow control mechanism can dynamically adjust the data transmission rate of each batch. This ensures the rational use of network resources. Specifically, the system can dynamically adjust the data transmission rate for each transmission based on real-time measured network bandwidth and data priority through a feedback mechanism. For example, when bandwidth is low, the system will reduce the transmission rate, decreasing the amount of data uploaded each time; while under better network conditions, the system can increase the transmission rate, thereby improving transmission efficiency. The formula is expressed as:

[0041] in: It refers to the data transmission rate; This refers to the current network conditions; These are flow control parameters that reflect the maximum allowable transmission rate under current network conditions.

[0042] Finally, through the above adjustments, the final result is optimized data transmission. This optimized data transmission features efficient transmission order, adaptable transmission frequency, and appropriate data batch size, ensuring efficient and real-time data transmission under different network conditions. Optimized data transmission maximizes bandwidth utilization while ensuring data integrity, avoiding network congestion or data loss.

[0043] Through the above steps, the system can adjust its data transmission strategy in real time according to the network environment, thereby achieving efficient and stable data transmission. The implementation of the adaptive data flow scheduling algorithm enables the system to flexibly respond to different data demands under varying network conditions, achieving efficient data transmission in green park carbon emission management.

[0044] S140. The optimized transmission data is verified using a generative adversarial network to obtain the verification results.

[0045] Specifically, Generative Adversarial Networks (GANs) are deep learning models comprising two neural networks: a generator and a discriminator. The generator aims to produce fake data that resembles real data, while the discriminator distinguishes between the generated fake data and real data. Through this adversarial training, the generator progressively improves its generation capabilities, making the generated fake data increasingly closer to real data. GANs are primarily used for tasks such as generative inference and data augmentation. In this scheme, the role of the GAN is to verify the optimized transmitted data, ensuring its authenticity and validity. For example, in a carbon emission monitoring system in a green park, a GAN could generate a set of potential hypothetical event data, reflecting different environmental changes or equipment states that the system might experience, such as equipment failure. The discriminator verifies the correctness of the transmitted data by comparing this generated data with the actual sensor data collected.

[0046] Multi-evidence verification refers to verifying the same data through multiple sources or methods to ensure its authenticity and reliability. In carbon emission management in green industrial parks, the generated optimized transmission data may contain errors or inaccuracies, thus requiring multiple verifications using different methods. Generative adversarial networks (GANs) can serve as a verification tool, validating data by comparing it with actual data. For example, the latent hypothetical event data generated by a GAN is compared with actual sensor data, and the discriminator uses this comparison to assess the reliability of the optimized transmission data. If the generated data is highly consistent with the actual data, the optimized transmission data is considered to have high validity and credibility; if there are significant differences, the system will detect potential errors or inconsistencies and take appropriate corrective measures. The verification result refers to the conclusion obtained through the multi-evidence verification process, indicating whether the data is reliable and valid. In this scheme, the verification result reflects the authenticity of the optimized transmission data, confirming whether the data conforms to the actual situation. By verifying the optimized transmission data, the system can ensure that decisions and optimizations based on this data are reliable.

[0047] Furthermore, firstly, the generator in the generative adversarial network generates potential hypothetical event data based on optimized transmission data. The generator's goal is to generate data based on the optimized transmission data input. Generate hypothetical event data that matches it. These hypothetical event data can represent possible anomalies or environmental changes. For example, hypothesizing optimized data transmission. This represents carbon emission data from certain equipment within the park. The generator will generate hypothetical event data for possible equipment malfunctions or excessive carbon emissions. The generator's output can be represented as:

[0048] in: It optimizes data transmission, including actual carbon emission data of equipment within the park; It is the input noise of the generator, representing possible changes or abnormal situations; It is a generator function used to generate data from input data. and noise Generate potential hypothetical event data .

[0049] Next, the discriminator in the generative adversarial network compares the generated latent hypothetical event data with the optimized transmission data, calculating their similarity to determine whether the optimized transmission data is genuine. The discriminator's goal is to evaluate the generated data. With optimized data transmission The discriminator measures the difference between the generated and actual data. If the generated data is highly consistent with the actual data, the discriminator outputs a higher "true" signal; otherwise, it outputs a "false" signal. The discriminator calculates a similarity score. This is used to evaluate the degree of matching between the generated data and the optimized data. This can be achieved through the discriminator's loss function. To indicate:

[0050] in: It is the discriminator that judges the data The result of the judgment is output as a value between [0,1], indicating whether the data is "real" or "fake". It is the distribution of real data; It is the distribution of noise variables; This refers to the potential hypothetical event data generated by the generator. Discriminator The system will evaluate the authenticity of the data based on this loss function and generate a similarity score. This score indicates the degree of similarity between the generated data and the actual data.

[0051] If the verification results indicate a high degree of consistency between the latent hypothetical event data and the optimized transmission data, then the optimized transmission data is determined to be the real data. If the discriminator's similarity score... Higher than the preset threshold If the generated potential hypothetical event data is highly consistent with the optimized transmission data, and the optimized transmission data is considered reliable, then the system will classify the optimized transmission data as "real data," indicating that the data has not been interfered with or erroneously affected, and can be used for subsequent decision-making and operations. If Then optimize data transmission. It is considered to be real data.

[0052] If the verification results indicate a high degree of inconsistency between the latent hypothesis event data and the optimized transmission data, then the optimized transmission data is determined to be biased data. If the discriminator's similarity score... Below the set threshold If the generated potential hypothetical event data differs significantly from the optimized transmission data, the optimized transmission data may be biased. In this case, the system will mark the optimized transmission data as "biased data" and may initiate a data correction mechanism or a process of re-collecting data. Then optimize data transmission. It is considered biased data.

[0053] By using generative adversarial networks (GANs) to perform multi-evidence verification on optimized transmission data, the system can comprehensively check the authenticity of the data. GANs confirm the reliability of the data by generating potential hypothetical event data and comparing it with the optimized transmission data. If the verification results show a high degree of consistency, the optimized transmission data is considered credible; if the verification results show a significant difference, the data is marked as biased data, preventing decisions based on erroneous data. In this way, the green park carbon emission management system ensures that it makes accurate decisions based on real data.

[0054] S150. Based on the verification results, a reinforcement learning algorithm is used to optimize the equipment scheduling for carbon emissions in the green park, generating the optimal equipment scheduling scheme.

[0055] Specifically, reinforcement learning is a machine learning method that optimizes decision-making strategies through interaction with the environment. In carbon emission management of green parks, reinforcement learning algorithms are used to optimize equipment scheduling. The algorithm continuously adjusts equipment scheduling strategies by simulating environmental feedback to minimize carbon emissions and optimize energy use. A key characteristic of reinforcement learning is trial-and-error learning, which improves decision-making strategies through continuous experience accumulation. For example, a green park may contain multiple energy devices, such as air conditioners, lighting equipment, and air compressors, and the operation of each device affects carbon emissions. Reinforcement learning algorithms automatically adjust the start-up, shutdown, and power regulation of equipment based on the operating status and carbon emission requirements of different devices, thereby optimizing the overall carbon emission level of the park.

[0056] Equipment scheduling refers to the rational arrangement of equipment operating times and modes based on actual needs and objectives to achieve optimized resource allocation and system efficiency. In carbon emission management within green industrial parks, the goal of equipment scheduling is to reduce energy consumption and carbon emissions by optimizing equipment operating strategies. Equipment scheduling includes, but is not limited to, decisions regarding equipment start-up and shutdown, operating mode switching, and power adjustment. For example, multiple air conditioning systems within a park may experience periodic load variations. The equipment scheduling system needs to dynamically adjust the start-up and shutdown times of the air conditioning systems based on climate change and energy demand, thereby reducing energy waste and carbon emissions.

[0057] Optimal equipment scheduling schemes are decision strategies generated through reinforcement learning algorithms or other optimization methods, aiming to achieve the best scheduling effect under given constraints. This scheme is the result of comprehensively considering factors such as equipment power consumption, carbon emission standards, and equipment health status. Generating an optimal equipment scheduling scheme requires considering multiple factors, including the overall energy demand of the park, equipment efficiency, and environmental changes. For example, assuming that in a green park, the electricity consumption and carbon emissions of the air conditioning system are the main objectives of scheduling strategy optimization, the optimal equipment scheduling scheme might reduce the operation of certain air conditioning units during peak electricity demand periods or adjust the equipment's temperature control strategies based on weather forecasts, thereby reducing carbon emissions while ensuring comfort.

[0058] Furthermore, firstly, the system acquires the current status and environmental conditions of each device within the green park, and then generates a state space for device scheduling based on these conditions. The system needs to collect current status information and environmental conditions for each device within the green park to form the state space for device scheduling. Device status includes device power, device health status, and device operating mode. For example, device power represents the current workload of the device, health status indicates whether the device is operating normally, and operating mode indicates whether the device is in a working, standby, or off state. Current environmental conditions include temperature, humidity, and carbon emission levels, which directly affect the operating efficiency and carbon emissions of the device. By collecting this information, the system generates a state space. Each state It includes all relevant variables. For example, for device i, the state... It can be represented as:

[0059] in: Is device i at time The power; Is device i at time Health condition; Is device i at time Operating mode; It's the temperature in the park; It's the humidity level in the park; It refers to the carbon emission level of the park.

[0060] Next, based on the state space, the actions corresponding to the device scheduling strategy are determined. After obtaining the current state of the device and the environmental conditions, the system will determine the actions corresponding to the device scheduling strategy based on the state space. Device scheduling actions include starting the device, adjusting the device power, and switching the device operating mode. Each action may have different execution effects under different states. In order to select the optimal action, the system needs to consider the current state space. Evaluate the possible actions. Assume the action set is... Then the equipment scheduling strategy Corresponding to state and actions The mapping relationship. Each device is based on its current state. Choose an action , can be represented as:

[0061] in: This indicates an action that the device can perform, such as starting the device, adjusting the power, or switching modes. It depends on the state The generated scheduling strategy. For example, if the carbon emission level of the park is too high, the equipment scheduling strategy may choose to reduce the power of some high-energy-consuming equipment or switch the equipment to a low-carbon emission mode.

[0062] Then, after the control device executes the action corresponding to the device scheduling strategy, it obtains feedback and calculates a reward signal. Once the device scheduling strategy is determined and the corresponding action is executed, the system obtains feedback information based on the device's operating status and environmental changes. This feedback information is an evaluation of the effectiveness of the current action, often expressed as a reward signal. The reward signal reflects the contribution of a particular action to a goal, such as reducing carbon emissions or improving energy efficiency. The calculation of the reward signal depends on the performance of the device and its impact on the overall system objective. The reward signal can be represented by a function. This indicates that it depends on the current state. and the actions performed For example, if adjusting equipment power reduces carbon emissions, a reward signal is given. It is a positive value; if a device malfunction occurs, a reward signal is given. It is a negative value. Specific reward signals. It can be defined as:

[0063] in: It represents the change in carbon emissions; a decrease is a negative value, and an increase is a positive value. It represents the change in electrical energy consumption; a decrease is a negative value, and an increase is a positive value. It represents changes in the health status of the equipment; a positive value indicates improvement, while a negative value indicates deterioration. These are the corresponding weighting coefficients, which determine the priority of different objectives.

[0064] Finally, based on the reward signal, the equipment scheduling strategy is progressively optimized using a reinforcement learning algorithm to generate the optimal equipment scheduling scheme. Based on the reward signal received from feedback, the system optimizes the equipment scheduling strategy using a reinforcement learning algorithm. The goal of the reinforcement learning algorithm is to maximize the cumulative reward, thereby optimizing the equipment scheduling scheme. By continuously adjusting the scheduling strategy and obtaining feedback, the reinforcement learning algorithm can progressively improve the equipment scheduling strategy to minimize carbon emissions and maximize energy efficiency. The reinforcement learning algorithm uses Q-learning or the policy gradient method for learning. The Q-learning method selects the optimal policy by updating the Q-value function, while the policy gradient method directly optimizes the policy function. For each state... Reinforcement learning algorithms are based on reward signals Update the scheduling strategy. Ultimately, the system will obtain an optimal equipment scheduling scheme. This scheme can dynamically adjust equipment operation at various points in time based on the actual needs of the park and environmental changes to achieve the best overall carbon emission and energy consumption results. Optimal Equipment Scheduling Scheme It is a strategy obtained by reinforcement learning algorithms through multiple iterations and optimizations, satisfying the following optimization objectives:

[0065] in: It is a discount factor used to balance the importance of current and future rewards; This represents the expected value, indicating the value achieved by executing the scheduling policy. The cumulative rewards obtained under different states. Ultimately, the optimal equipment scheduling scheme generated by the system will achieve efficient coordination of equipment and minimize carbon emissions within the green park.

[0066] Through the steps described above, the system can optimize equipment scheduling based on reinforcement learning algorithms after acquiring real-time data on equipment status and the environment within the park. Feedback and reward signals after each scheduling operation will drive the algorithm to progressively improve the scheduling strategy, ultimately generating the optimal equipment scheduling plan to ensure the park achieves the lowest carbon emissions and highest energy efficiency.

[0067] S160. The local models of each edge computing node are aggregated through a federated learning algorithm to obtain a global model. Differential privacy technology is used to control the privacy level of data uploads in the global model to obtain a privacy-protected model.

[0068] Specifically, a local model refers to a machine learning model trained on each edge computing node, based on locally collected data. Because edge computing nodes process local data, each node's local model reflects the local environment or device's operational status. These models are used for tasks such as real-time decision-making, data analysis, and prediction. For example, a device might train a local model based on data from a temperature and humidity sensor to predict trends in carbon emissions or assess the device's health. Each edge node trains and updates its own model to make optimal decisions based on real-time data.

[0069] Federated learning is a distributed machine learning approach that allows multiple computing nodes to collaboratively train a global model without exchanging local data. Each edge computing node only needs to share the parameters of its local model, such as weights and biases, without transmitting the raw data. This approach protects data privacy while integrating the learning results of each node to obtain a global model. For example, individual sensors within a green park, such as temperature and humidity sensors and carbon dioxide sensors, can train their own models locally to predict the operating status of different devices. Through federated learning algorithms, the parameters of these local models are aggregated into a global model to more accurately predict the overall carbon emission level of the park.

[0070] A global model is a model that aggregates local models from multiple edge computing nodes using a federated learning algorithm. By combining the learning outcomes from each edge node, the global model better reflects the characteristics or behavior of the entire system. Computed on a central server, the global model can be used for decision-making or prediction tasks across multiple devices. For example, a global model in a green park might combine sensor data from different devices, such as temperature, humidity, and carbon emissions, to more comprehensively predict the overall carbon emission trend of the park and provide optimization suggestions.

[0071] Differential privacy is a data privacy protection technique designed to ensure that individual data cannot be leaked or used for inference. By adding noise during data distribution and usage, differential privacy minimizes the impact of any single data point on the overall data analysis results, thus preventing the leakage of sensitive information. In federated learning, differential privacy is used to protect the privacy of local models, preventing the inference of private data from a particular edge computing node. For example, when sensor nodes in a green park upload parameters from their local models, differential privacy adds noise to the uploaded model parameters, ensuring that even if someone obtains these parameters, they cannot infer the specific data of a particular sensor or the private information of any individual device.

[0072] A privacy-preserving model is a model obtained after differential privacy protection. This model ensures that all privacy-related data is protected during training and inference. While maintaining data utilization efficiency, the privacy-preserving model avoids disclosing any data details about individuals. Federated learning algorithms combined with differential privacy technology can ensure that no sensitive information is leaked even when model parameters are shared and aggregated across multiple edge computing nodes. For example, in a carbon emission management system in a green park, a privacy-preserving model can train a global model using model parameters provided by each edge computing node without exposing specific device data, while ensuring the data privacy of each device is not violated.

[0073] Furthermore, firstly, each edge computing node trains a local model based on its local environmental data and device status data. In this stage, each edge computing node trains its model using local environmental and device status data. Specifically, the edge computing node collects environmental data related to carbon emission management, such as temperature, humidity, and carbon dioxide concentration, as well as device status data, such as device power, operating mode, and health status, and trains a local model based on this data. These local models include regression or time series models related to carbon emission prediction, as well as device scheduling strategy models. For example, suppose the edge computing node uses environmental data and device operating status data to build a device scheduling strategy model to predict carbon emission changes under different scheduling strategies. The goal of this model is to predict the impact of different device scheduling schemes by analyzing the current state. The process of training the local model includes data preprocessing, such as denoising, standardization, feature extraction, and model training. The local model for each device can be trained through the following process:

[0074] in: It is a local model for edge computing nodes; It is local environmental data, including information such as temperature and humidity; It is local device status data, including information such as power and operating mode; It is a model function built based on data and device status.

[0075] Next, the parameters of the local models on each edge computing node are uploaded using a federated learning algorithm and aggregated through a weighted average to obtain the global model. Once the local model on each edge computing node is trained, the parameters of each local model are uploaded to the central server using the federated learning algorithm. It is important to note that the edge computing nodes do not upload local data, but only the parameters of their local models, such as weights and biases. On the server side, these local model parameters are aggregated through a weighted average to generate the global model. Assume that the local model parameters of each edge computing node are... Then the parameters of the global model The weighted average can be calculated using the following formula:

[0076] in: These are the parameters of the global model; It is the first Local model parameters for each edge computing node; It is the first The weight of each edge computing node is related to the amount of data or computing power of the node; This represents the total number of edge computing nodes. Through a weighted average, the central server aggregates the learning outcomes of all edge computing nodes into a global model. This model encompasses global information about all devices and the environment, thus providing comprehensive decision support for the park's carbon emission management.

[0077] Then, differential privacy technology is used to inject noise into the parameters of the global model, resulting in a privacy-preserving model. After the global model is generated, the system uses differential privacy technology to process the model parameters to ensure data privacy is protected. Differential privacy technology ensures that individual data cannot be inferred by injecting noise into the parameters of the global model. The purpose of noise injection is to ensure that changes in any single data point will not significantly affect the results of the global model, thereby preventing potential data leakage. Specifically, the implementation process of differential privacy technology includes the following steps: Noise injection: According to a preset privacy budget, the system adds noise to the parameters of the global model. This noise follows a Laplace or Gaussian distribution to ensure that each uploaded parameter does not leak individual private information; Privacy budget: The privacy budget is a parameter that controls the strength of data privacy protection and determines the strength of noise injection. The smaller the privacy budget, the stronger the noise injection and the better the data protection, but it may affect the accuracy of the model. Assume... These are the original parameters of the global model. Differential privacy technology injects noise in the following ways. :

[0078] in: These are the global model parameters after applying differential privacy; The noise is sampled from a specific distribution, such as a Laplace or Gaussian distribution, to ensure the privacy of the model. In this way, differential privacy technology can effectively prevent the data of a certain edge computing node from being back-engineered, thereby ensuring the data privacy of every device and sensor in the park is protected.

[0079] Through the steps described above, the system first trains local models on each edge computing node, and then aggregates the parameters of these local models using a federated learning algorithm to generate a global model. Subsequently, differential privacy technology is applied to inject noise into the parameters of the global model to ensure data privacy is protected. This process effectively combines the advantages of data privacy protection and distributed learning, enabling the optimization of the decision-making capabilities of the green park carbon emission management system while ensuring privacy.

[0080] S170. Based on the privacy protection model, the optimal device scheduling scheme is executed through meta-learning and online convex optimization techniques to obtain the edge collaboration strategy.

[0081] Specifically, meta-learning is a machine learning method that aims to learn more effectively by learning patterns and transferability between tasks. In meta-learning, algorithms acquire learning experience from multiple tasks, enabling them to quickly adapt to new tasks. The goal of meta-learning is to allow systems to rapidly optimize the learning process with less data and time when facing new tasks or environmental conditions. Suppose that the operating modes of equipment in a green park need to be quickly adjusted according to environmental changes. Through meta-learning, the system can learn from past equipment scheduling tasks and quickly adjust strategies to cope with new environmental conditions, such as temperature changes and carbon emission requirements. This rapid adaptability plays a crucial role in optimizing equipment scheduling within the park.

[0082] Online convex optimization is an optimization method used to solve situations involving online decision-making problems, especially when these problems change over time or with varying environmental conditions. Unlike traditional batch optimization methods, online convex optimization can progressively adjust optimization decisions without requiring complete data. It updates the optimization strategy in real time to adapt to new data inputs and environmental changes. The "convex" in convex optimization means that the objective function and constraints satisfy certain mathematical properties, allowing the optimization problem to be solved using efficient algorithms. For example, in a green park, equipment operation strategies, such as starting, stopping, and power adjustment, need to be adjusted based on real-time environmental data, such as temperature and humidity. Online convex optimization can optimize equipment operation strategies in real time, thereby quickly finding the optimal scheduling scheme under unstable network or environmental conditions.

[0083] Edge collaboration strategy refers to maximizing overall system efficiency by coordinating device scheduling decisions among multiple edge computing nodes, and balancing tasks, resources, and power consumption among multiple devices. The implementation of edge collaboration strategy relies on a global optimization model or multi-agent system to ensure that devices work collaboratively to achieve overall goals without directly exchanging data. For example, a green park may have multiple air conditioners, lighting, and other devices that need to coordinate their operation to optimize energy use. In this case, the edge collaboration strategy can adjust the operating mode of each device according to the park's real-time needs, ensuring that all devices operate with minimal power consumption and carbon emissions while maintaining comfort.

[0084] Furthermore, firstly, based on the privacy-preserving model, a meta-learning model is initialized and trained using historical environmental change data. The meta-learning model is trained using historical environmental change data, such as temperature variations, humidity, and carbon emission levels. The goal of meta-learning is to enable the system to quickly adapt and optimize equipment scheduling strategies when facing new tasks. During initialization, the system uses environmental change and equipment scheduling data from historical data for training. Through this data, the meta-learning model can learn how to quickly adjust equipment scheduling strategies under different environmental conditions. For example, during training, historical data... This includes environmental data, such as temperature, humidity, and carbon emissions, as well as equipment scheduling data, such as equipment start-up and shutdown, and power adjustments. The trained model... This can be represented by the following relationship:

[0085] in: It is the initialized meta-learning model; It includes historical environmental change data and equipment scheduling data; It is the training function, which indicates how the meta-learning model adjusts its parameters based on the training data.

[0086] Next, the device scheduling task is decomposed, and a meta-learning model is used to train the scheduling strategies for each device, resulting in candidate device scheduling strategies. The device scheduling task is broken down into multiple sub-tasks, each corresponding to a device's scheduling objective. Based on the initialized meta-learning model, the system trains the scheduling strategy for each device, generating multiple candidate device scheduling strategies. Each candidate strategy corresponds to an operation that a device might perform under different environmental conditions, such as starting the device, adjusting device power, or changing the operating mode. This is achieved through the meta-learning model. The system generates a candidate device scheduling strategy. It can be represented as:

[0087] in: It is equipment Scheduling strategy; It is a device scheduling strategy learned by the meta-learning model; It is equipment The current status, including equipment power, equipment health status, etc.; It contains environmental data, including information such as temperature, humidity, and carbon emissions. Through a meta-learning model, the system can generate multiple candidate device scheduling strategies under different environmental conditions. These strategies are generated based on the learning results of historical data.

[0088] Then, in each round of optimization, the candidate device scheduling strategy is dynamically adjusted using online convex optimization technology to obtain the current device scheduling strategy. During each round of optimization, the system uses online convex optimization technology to dynamically adjust the candidate device scheduling strategy. Online convex optimization technology can dynamically adjust the device scheduling strategy based on real-time feedback to ensure that device scheduling can be optimized in a timely manner and achieve the best results under different network conditions or device states. For example, suppose the current candidate device scheduling strategy is... Online convex optimization technology is based on real-time feedback This strategy is dynamically adjusted based on factors such as carbon emissions and electricity consumption. The goal of the online optimization process is to minimize carbon emissions or maximize energy efficiency. The current equipment scheduling strategy is adjusted in real time. This can be expressed as an optimization function:

[0089] in: This is the adjusted equipment scheduling strategy; It is a loss function, representing the device scheduling strategy. Real-time feedback The degree of optimization; It serves as a feedback signal, reflecting the impact of equipment scheduling strategies on the park's carbon emissions or energy use. Through online convex optimization, the system can dynamically adjust equipment scheduling strategies based on real-time data and feedback, ensuring that equipment scheduling meets optimal requirements at every moment.

[0090] Next, the current device scheduling strategy is corrected based on real-time feedback signals. Gradient calculation is used to optimize the current device scheduling strategy, resulting in the optimal device scheduling scheme. Based on real-time feedback signals, the system corrects the current device scheduling strategy, gradually optimizing the device scheduling scheme. Gradient calculation is used to determine the direction and magnitude of the adjustment of the current device scheduling strategy, thereby optimizing system performance. By calculating the gradient of the current device scheduling strategy relative to the feedback signal, the system can evaluate the effectiveness of the strategy and make adjustments. Assume... It comes from the device. Real-time feedback signals, equipment scheduling strategies gradient It can be represented as:

[0091] in: Equipment scheduling strategy The gradient; It is a loss function that represents the degree to which the equipment scheduling strategy optimizes the feedback signal; This refers to gradient calculation, which represents the rate of change of the loss function relative to the device scheduling strategy. By calculating the gradient and adjusting the device scheduling strategy, the system gradually optimizes the device scheduling strategy to obtain the optimal device scheduling scheme. This minimizes carbon emissions or maximizes energy efficiency.

[0092] Finally, based on the optimal device scheduling scheme, the system generates an edge collaboration strategy. The purpose of this strategy is to ensure coordinated device scheduling across multiple edge computing nodes within the park, optimizing overall carbon emission control and energy efficiency. By coordinating the scheduling of different devices, the system avoids conflicts and resource waste, ensuring that all devices work collaboratively within a globally optimized framework. For example, the scheduling of multiple air conditioning units within the park needs to be coordinated based on real-time environmental changes to prevent multiple units from starting simultaneously during high-temperature periods, which would lead to excessive energy consumption. The edge collaboration strategy coordinates the operation of these devices to ensure optimal energy allocation and minimal carbon emissions. Ultimately, the edge collaboration strategy can be represented by the following relationship:

[0093] in: It is the optimal scheduling strategy for each device; It is the edge collaboration strategy generation function, which represents how to coordinate the operation of various devices to achieve the global optimal goal.

[0094] Through meta-learning and online convex optimization techniques, the system can progressively optimize equipment scheduling strategies based on historical data and real-time feedback. It then refines these strategies through gradient calculation to ultimately generate the optimal equipment scheduling scheme. Based on this scheme, the system further generates edge collaboration strategies to ensure coordinated equipment scheduling among multiple edge computing nodes, thereby optimizing carbon emission control and energy efficiency in the green park.

[0095] This application also provides an active sampling and constrained reinforcement learning edge collaboration device, referring to... Figure 2 , Figure 2This is a schematic diagram of a module for an active sampling and constrained reinforcement learning edge collaboration device provided in an embodiment of this application. The device is a server, comprising an acquisition module 21 and a processing module 22. The acquisition module 21 acquires sampling data and event data sent by edge computing nodes targeting carbon emissions in green parks. The processing module 22 models and sparsely encodes the sampling data and event data to obtain compressed encoded data. The processing module 21 further processes the compressed encoded data using an adaptive data flow scheduling algorithm to obtain optimized transmission data. The processing module 22 performs multi-evidence verification on the optimized transmission data using a generative adversarial network to obtain verification results. Based on the verification results, the processing module 22 optimizes the equipment scheduling for carbon emissions in green parks using a reinforcement learning algorithm to generate an optimal equipment scheduling scheme. The processing module 22 also aggregates the local models of each edge computing node using a federated learning algorithm to obtain a global model, and controls the privacy level of data uploads in the global model using differential privacy technology to obtain a privacy-preserving model. Finally, based on the privacy-preserving model, the processing module 22 executes the optimal equipment scheduling scheme using meta-learning and online convex optimization techniques to obtain an edge collaboration strategy.

[0096] It should be noted that the above embodiments of the apparatus are only illustrated by the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the apparatus and method embodiments provided in the above embodiments belong to the same concept, and the specific implementation process can be found in the method embodiments, which will not be repeated here.

[0097] This application also provides an electronic device, with reference to... Figure 3 , Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device may include: at least one processor 31, at least one network interface 34, a user interface 33, a memory 35, and at least one communication bus 32.

[0098] The communication bus 32 is used to enable communication between these components.

[0099] The user interface 33 may include a display screen and a camera. Optionally, the user interface 33 may also include a standard wired interface and a wireless interface.

[0100] The network interface 34 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface).

[0101] The processor 31 may include one or more processing cores. The processor 31 connects to various parts of the server via various interfaces and lines, executing instructions, programs, code sets, or instruction sets stored in the memory 35, and calling data stored in the memory 35 to perform various server functions and process data. Optionally, the processor 31 may be implemented using at least one hardware form of Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), or Programmable Logic Array (PLA). The processor 31 may integrate one or a combination of several of the following: Central Processing Unit (CPU), Graphics Processing Unit (GPU), and modem. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the content to be displayed on the screen; and the modem handles wireless communication. It is understood that the modem may also not be integrated into the processor 31 and may be implemented as a separate chip.

[0102] The memory 35 may include random access memory (RAM) or read-only memory. Optionally, the memory 35 may include a non-transitory computer-readable storage medium. The memory 35 can be used to store instructions, programs, code, code sets, or instruction sets. The memory 35 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as touch function, sound playback function, image playback function, etc.), instructions for implementing the above-described method embodiments, etc.; the data storage area may store data involved in the above-described method embodiments, etc. Optionally, the memory 35 may also be at least one storage device located remotely from the aforementioned processor 31. Figure 3 As shown, the memory 35, which serves as a computer storage medium, may include an operating system, a network communication module, a user interface module, and an application program for an active sampling and constrained reinforcement learning edge collaboration method.

[0103] exist Figure 3In the electronic device shown, the user interface 33 is mainly used to provide an interface for users to input data and obtain user input data; while the processor 31 can be used to call an application stored in the memory 35 that is an active sampling and constrained reinforcement learning edge collaboration method. When executed by one or more processors, the electronic device executes one or more methods as described in the above embodiments.

[0104] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0105] This application also provides a computer-readable storage medium storing instructions. When executed by one or more processors, these instructions cause an electronic device to perform one or more of the methods described in the above embodiments.

[0106] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0107] In the several embodiments provided in this application, it should be understood that the disclosed apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the shown or discussed mutual couplings or direct couplings or communication connections may be through some service interfaces; indirect couplings or communication connections between apparatuses or units may be electrical or other forms.

[0108] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0109] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0110] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned memory includes various media capable of storing program code, such as USB flash drives, portable hard drives, magnetic disks, or optical disks.

[0111] The foregoing description is merely an exemplary embodiment of this disclosure and should not be construed as limiting the scope of this disclosure. Any equivalent changes and modifications made in accordance with the teachings of this disclosure shall still fall within the scope of this disclosure. Those skilled in the art will readily conceive of other embodiments of this disclosure upon considering the specification and the disclosure of practical truth. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not described in this disclosure. The specification and embodiments are considered exemplary only, and the scope and spirit of this disclosure are defined by the claims.

Claims

1. A collaborative method for active sampling and constrained reinforcement learning at the edge, characterized in that, The method includes: Acquire sampling data and event data sent by edge computing nodes regarding carbon emissions in green industrial parks; The sampled data and event data are modeled and sparsely encoded to obtain compressed encoded data; An adaptive data stream scheduling algorithm is used to process the compressed encoded data to obtain optimized transmission data; The optimized transmission data is verified using a generative adversarial network to obtain the verification result. Based on the verification results, a reinforcement learning algorithm is used to optimize the equipment scheduling for carbon emissions in the green park, generating the optimal equipment scheduling scheme. The local models of each edge computing node are aggregated using a federated learning algorithm to obtain a global model. Differential privacy technology is used to control the privacy level of data uploads in the global model to obtain a privacy-preserving model. Based on the privacy protection model, the optimal device scheduling scheme is executed through meta-learning and online convex optimization techniques to obtain the edge collaboration strategy.

2. The active sampling and constrained reinforcement learning edge collaboration method according to claim 1, characterized in that, The step of modeling and sparsely encoding the sampled data and the event data to obtain compressed encoded data specifically includes: The sampled data and event data are modeled using mathematical models to capture the temporal characteristics of the data and generate multidimensional modeling data. The multidimensional modeling data is sparsely encoded using sparse representation techniques. By constructing a sparse dictionary and learning the dictionary based on an optimization algorithm, the compressed encoded data is obtained.

3. The active sampling and constrained reinforcement learning edge collaboration method according to claim 1, characterized in that, The process of using an adaptive data stream scheduling algorithm to process the compressed encoded data to obtain optimized transmission data specifically includes: Determine the communication parameters of the current network environment, and obtain the network conditions based on the communication parameters; Based on the timeliness and importance of different data under the network conditions, the priority of the compressed encoded data is determined, and a data stream is generated; The adaptive data stream scheduling algorithm is used to schedule the data stream in real time, automatically adjusting the data transmission frequency, batch size, and sequence, and adjusting the data transmission efficiency through a data flow control mechanism to obtain the optimized transmission data.

4. The active sampling and constrained reinforcement learning edge collaboration method according to claim 1, characterized in that, The step of performing multi-evidence verification on the optimized transmission data using a generative adversarial network to obtain the verification result specifically includes: Based on the optimized transmission data, potential hypothetical event data are generated by the generator in the generative adversarial network; The discriminator in the generative adversarial network compares the potential hypothetical event data with the optimized transmission data to obtain the verification result; If the verification result indicates a high degree of consistency between the potential hypothetical event data and the optimized transmission data, then the optimized transmission data is determined to be real data. If the verification result indicates a high degree of inconsistency between the potential hypothetical event data and the optimized transmission data, then the optimized transmission data is determined to be biased data.

5. The active sampling and constrained reinforcement learning edge collaboration method according to claim 1, characterized in that, Based on the verification results, a reinforcement learning algorithm is used to optimize the equipment scheduling for carbon emissions in the green park, generating an optimal equipment scheduling scheme, specifically including: The current status and environmental conditions of each device in the green park are obtained, and the state space of device scheduling is obtained based on the current status and environmental conditions. The current status includes device power, device health status and device operating mode, and the current environmental conditions include temperature, humidity and carbon emission level. Based on the state space, the actions corresponding to the device scheduling strategy are determined, including starting the device, adjusting the device power, and switching the device operating mode. After the control device executes the action corresponding to the device scheduling strategy, it obtains feedback and calculates a reward signal; Based on the reward signal, the equipment scheduling strategy is gradually optimized through the reinforcement learning algorithm to generate the optimal equipment scheduling scheme.

6. The active sampling and constrained reinforcement learning edge collaboration method according to claim 1, characterized in that, The process of aggregating the local models of each edge computing node using a federated learning algorithm to obtain a global model, and controlling the privacy level of data uploads in the global model using differential privacy technology to obtain a privacy-preserving model, specifically includes: Each edge computing node is controlled to train a local model based on the corresponding local environmental data and device status data. The local model includes a prediction model and a device scheduling strategy model related to carbon emission management. The parameters of the local models of each edge computing node are uploaded using the federated learning algorithm and aggregated by weighted averaging to obtain the global model. The differential privacy technique is used to inject noise into the parameters of the global model to obtain the privacy-preserving model.

7. The active sampling and constrained reinforcement learning edge collaboration method according to claim 1, characterized in that, The edge collaboration strategy, based on the privacy protection model, is obtained by executing the optimal device scheduling scheme through meta-learning and online convex optimization techniques, specifically including: Based on the privacy protection model, initialize the meta-learning model and train the meta-learning model using historical environmental change data; The equipment scheduling task is decomposed, and the meta-learning model is used to train the scheduling strategies of each equipment to obtain candidate equipment scheduling strategies. In each round of optimization, the candidate device scheduling strategy is dynamically adjusted using the online convex optimization technique to obtain the current device scheduling strategy; The current device scheduling strategy is corrected based on real-time feedback signals, and the current device scheduling strategy is optimized through gradient calculation to obtain the optimal device scheduling scheme. Based on the optimal equipment scheduling scheme, the edge collaboration strategy is generated, wherein the edge collaboration strategy is used to ensure the coordinated scheduling of equipment at each edge node in order to optimize carbon emission control and energy utilization efficiency in the green park.

8. A collaborative device for active sampling and constrained reinforcement learning at the edge, characterized in that, The apparatus is used to perform the method as described in any one of claims 1 to 7, the apparatus comprising an acquisition module and a processing module, wherein... The acquisition module is used to acquire sampling data and event data sent by edge computing nodes targeting carbon emissions in green parks; The processing module is used to model and sparsely encode the sampled data and the event data to obtain compressed encoded data; The processing module is further configured to process the compressed encoded data using an adaptive data stream scheduling algorithm to obtain optimized transmission data; The processing module is also used to perform multi-evidence verification on the optimized transmission data through a generative adversarial network to obtain a verification result; The processing module is also used to optimize the equipment scheduling for carbon emissions in the green park by using a reinforcement learning algorithm based on the verification results, and generate an optimal equipment scheduling scheme. The processing module is also used to aggregate the local models of each edge computing node through a federated learning algorithm to obtain a global model, and to control the privacy level of data upload in the global model through differential privacy technology to obtain a privacy-protected model. The processing module is further configured to execute the optimal device scheduling scheme based on the privacy protection model through meta-learning and online convex optimization techniques to obtain an edge collaboration strategy.

9. An electronic device, characterized in that, The electronic device includes a processor, a memory, a user interface, and a network interface. The memory is used to store instructions. The user interface and the network interface are both used to communicate with other devices. The processor is used to execute the instructions stored in the memory to cause the electronic device to perform the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed, perform the method as described in any one of claims 1 to 7.

Images