An industrial load control method based on cloud-edge collaboration technology

Abstract

The present invention relates to the technical field of industrial load management, and discloses an industrial load control method based on cloud-edge collaboration technology, including: Step 1, by deploying load sensors, environmental monitoring devices, and equipment status monitoring devices, historical load data, environmental data, and equipment operation status data are collected. The historical load data is used to represent the load change situation of the system at multiple time points. The environmental data includes externally obtained real-time temperature and electricity price information. The equipment operation status data includes the operation status parameters of the equipment. The historical load data, environmental data, and equipment operation status data are combined for data fusion. By introducing a dynamic weighting mechanism in the data fusion stage, the weights are adaptively adjusted according to the importance of the historical load data, environmental data, and equipment status data, so as to accurately depict the contribution of different features to the fused features, and the effect of significantly reducing the load prediction error is obtained.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial load management, and specifically provides an industrial load control method based on cloud-edge collaboration technology. Background Art

[0002] The industrial field has an increasing demand for intelligence and efficiency, and industrial load control has become a key technology for achieving energy conservation, emission reduction, and optimizing energy distribution. However, current industrial load management methods have many deficiencies when facing dynamic and complex industrial environments.

[0003] On the one hand, traditional industrial load forecasting methods mainly rely on historical load data, ignoring the significant impact of environmental data and equipment operating status on load changes, resulting in poor adaptability and accuracy of the forecasting model in complex scenarios. At the same time, existing multi-modal data fusion technologies fail to effectively measure the importance of different data sources, cannot fully utilize the complementary characteristics of multi-modal data, and have limitations in characterizing load features.

[0004] On the other hand, during the load regulation process, it is usually assumed that the correlation between devices remains fixed, unable to handle the complex collaboration requirements brought about by dynamic load changes in industrial systems. If the load distribution strategy among industrial devices lacks flexible adjustment of real-time collaboration relationships, it is prone to uneven load distribution and resource waste, reducing the overall operating efficiency of the system.

[0005] In addition, although current load control methods based on cloud-edge collaboration have made certain progress in computing resource allocation and communication efficiency, in high-frequency data interaction scenarios, the communication overhead problem of cloud-edge collaboration is particularly prominent. Since the cloud needs to frequently interact with edge devices, redundant communication causes a significant increase in system energy consumption, restricting the efficient operation of the system.

[0006] In summary, the deficiencies of existing technologies in load forecasting, load regulation, and cloud-edge collaboration communication severely restrict the performance of industrial load control systems. Therefore, those skilled in the art provide an industrial load control method based on cloud-edge collaboration technology to solve the above-mentioned problems. Summary of the Invention

[0007] Aiming at the deficiencies of the existing technology, the present invention provides an industrial load control method based on cloud-edge collaboration technology to solve the problems raised in the above background art.

[0008] To achieve the above objectives, the present invention is realized through the following technical solutions: An industrial load control method based on cloud-edge collaboration technology, comprising:

[0009] Step 1: By deploying load sensors, environmental monitoring devices, and device status monitoring devices, historical load data, environmental data, and device operation status data are collected. The historical load data is used to represent the load change of the system at multiple time points. The environmental data includes externally obtained real-time temperature and electricity price information. The device operation status data includes device operation status parameters. Combine the historical load data, environmental data, and device operation status data for data fusion to generate a unified fusion feature as the input for load forecasting.

[0010] Step 2: Input the fusion feature generated in Step 1 into the load forecasting model. Process the input feature through the encoding module of the model to extract potential load correlation patterns. Then, through the decoding module, generate the load forecasting value at future time points based on the potential patterns. The forecasting model is constructed based on the correlation between the load historical change trend, environmental data, and device status.

[0011] Step 3: Combine the load forecasting value generated in Step 2 and the industrial device network structure to establish a dynamic graph model reflecting the load transfer relationship between devices. The dynamic graph model includes a set of device nodes, the connection relationships between nodes, and an edge weight matrix reflecting the load transfer intensity between devices. The dynamic graph is updated in real time through the load forecasting value and the device correlation relationship.

[0012] Step 4: Based on the dynamic graph model established in Step 3, define the local optimization objective of the device nodes. For the cooperation relationship between device nodes, construct the load regulation cost between nodes. Coordinate the load distribution of each node through the global optimization framework. The optimization result is calculated in real time by the edge device and the processing result is fed back to the cloud.

[0013] Step 5: Combine the load distribution result optimized in Step 4 to construct a game model containing device nodes in the cloud. The device nodes act as participants in the game model. According to the comprehensive influence of the short-term load forecasting error, long-term trend deviation, and communication overhead, dynamically adjust the load regulation strategy to reach the global equilibrium state through the game model. Finally, send the regulation strategy parameters calculated by the cloud to the edge device for execution to complete the global coordination and dynamic optimization of the load.

[0014] Preferably, the data fusion in Step 1 is completed through the following weighting mechanism:

[0015] ,

[0016] where is the i-th type of input data, is the number of feature categories, is the i-th type of input feature vector, is calculated by the following formula:

[0017] ,

[0018] Among them, is the weight coefficient of various types of data, is the attention score of the i-th type of feature, is the number of feature categories, is the exponential function,

[0019] is the normalization term of the exponential value of the feature attention score.

[0020] Preferably, the load prediction model in step 2 generates the load prediction value at the future time point through the following process:

[0021] The encoder extracts the potential association pattern of the fusion features:

[0022] ,

[0023] Among them, is the potential dependent variable, is the feature mean, is the feature standard deviation, is the random noise;

[0024] The decoder generates the predicted load value through the potential dependent variable , where is the current time point, is the prediction time interval.

[0025] Preferably, the dynamic graph model established in step 3 satisfies the following relationship:

[0026] ,

[0027] Among them, is the time The edge weight between nodes i and j, and are the load values of nodes i and j, represents the difference degree of the node load.

[0028] Preferably, the local optimization objective function in step 4 is defined as:

[0029] ,

[0030] Among them, is the actual load of node i, is the predicted load of node i,

[0031] is the square of the error between the actual load and the predicted load of node i,

[0032] is the prediction error function, is the load value of node i.

[0033] Preferably, the load regulation cost function between the nodes is defined as:

[0034] ,

[0035] where, is the cooperation cost function between node i and node j, is the load value of node i, is the load value of node j, is the penalty coefficient for load regulation between nodes, is the square of the difference in load values between node i and node j.

[0036] Preferably, the global optimization framework is solved by the following objective function:

[0037] ,

[0038] where, is the optimization objective, is the load value of device node i, is the set of device nodes, is the set of edges between device nodes, is the prediction error function, is the cooperation cost function between node i and node j, is the load value of device node j.

[0039] Preferably, the game model in step 5 is constructed by the following utility function:

[0040] ,

[0041] where, is the utility function of device node i, is the short-term load prediction error, is the long-term load trend error, is the communication overhead, , , are weight parameters.

[0042] Preferably, the utility function is optimized by the following equilibrium conditions:

[0043] , , ,

[0044] Among them, , , are weight parameters, is the utility function of device node i,

[0045] is the partial derivative of the utility function with respect to the short-term error weight coefficient .

[0046] is the partial derivative of the utility function with respect to the short-term error weight coefficient .

[0047] is the partial derivative of the utility function with respect to the short-term error weight coefficient .

[0048] Preferably, the equilibrium strategy parameters generated by the cloud through game optimization are periodically sent to the edge devices. The edge devices adjust the load control strategy in real time according to the received strategy parameters combined with local data, and upload the load control results to the cloud for updating the next round of strategy parameters.

[0049] The present invention provides an industrial load control method based on cloud-edge collaboration technology. It has the following beneficial effects:

[0050] 1. By introducing a dynamic weighting mechanism in the data fusion stage, the present invention adaptively adjusts the weights according to the importance of historical load data, environmental data, and device status data, realizes the precise characterization of the contribution of different features to the fused features, and obtains the effect of significantly reducing the load prediction error.

[0051] 2. By constructing a dynamic graph model based on the load prediction results, the present invention updates the load transfer relationship between device nodes in real time, realizes the flexible adjustment of the load cooperation relationship between devices, and obtains the effects of improving the load control accuracy and enhancing the system dynamic response ability.

[0052] 3. By comprehensively considering the communication overhead in game optimization, the present invention dynamically adjusts the communication frequency and data transmission volume between the cloud and the edge devices, realizes the goal of reducing cloud-edge communication redundancy, and obtains the effects of reducing the system operation energy consumption and improving the collaboration efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 is the flowchart of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0054] To enable those skilled in the art to understand the solution of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Based on the embodiments of the present invention, other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the scope of protection of the present invention.

[0055] The present invention will be described in detail below with reference to the accompanying drawings: Embodiment

[0056] Please refer to the attached Figure 1 , the embodiment of the present invention provides an industrial load control method based on cloud-edge collaboration technology, including:

[0057] Step 1: By deploying load sensors, environmental monitoring devices, and device status monitoring devices, historical load data, environmental data, and device operation status data are collected. The historical load data is used to represent the load change of the system at multiple time points. The environmental data includes externally obtained real-time temperature and electricity price information. The device operation status data includes the operation status parameters of the device. The historical load data, environmental data, and device operation status data are combined for data fusion to generate a unified fusion feature as the input for load prediction.

[0058] Step 2: The fusion feature generated in Step 1 is input into the load prediction model. The input feature is processed by the encoding module of the model to extract potential load correlation patterns. Then, the decoding module generates the load prediction value for future time points based on the potential patterns. The prediction model is constructed based on the relevance of the load historical change trend, environmental data, and device status.

[0059] Step 3: Combining the load prediction value generated in Step 2 and the industrial equipment network structure, a dynamic graph model reflecting the load transmission relationship between devices is established. The dynamic graph model includes a set of device nodes, the connection relationship between nodes, and an edge weight matrix reflecting the load transmission intensity between devices. The dynamic graph is updated in real time through the load prediction value and the correlation between devices.

[0060] Step 4: Based on the dynamic graph model established in Step 3, the local optimization objective of the device nodes is defined. For the cooperation relationship between device nodes, the load regulation cost between nodes is constructed. The global optimization framework coordinates the load distribution of each node, and the optimization result is calculated in real time by the edge device and the processing result is fed back to the cloud.

[0061] Step 5: Combine the optimized load distribution results in Step 4 to build a game model containing device nodes in the cloud. The device nodes, as participants in the game model, dynamically adjust the load control strategy according to the comprehensive influence of short-term load prediction errors, long-term trend deviations, and communication overheads, and reach a global equilibrium state through the game model. Finally, the control strategy parameters calculated in the cloud are sent to the edge devices for execution to complete the global coordination and dynamic optimization of the load.

[0062] Benefits of Step 1: By deploying load sensors, environmental monitoring devices, and device status monitoring devices, comprehensively collect historical load, environmental data, and device operation status data, providing rich data support for load prediction. Combine multimodal data for fusion processing, dynamically adjust the weights of different data sources, effectively improve the data expression ability and the characterization accuracy of load characteristics, and solve the problem of the limitation of a single data source in traditional load prediction.

[0063] Benefits of Step 2: By introducing a load prediction model, using the encoding module to extract potential load correlation patterns, and generating load prediction values at future time points through the decoding module, it can effectively capture the non-linear change trend of the load. The model is constructed based on the correlation between historical load, environmental data, and device status, significantly improving the accuracy of load prediction.

[0064] Benefits of Step 3: Combine the load prediction value and the device network structure to establish a dynamically changing device node relationship graph. By real-time updating the connection relationship and load transmission intensity of device nodes, the dynamic graph can flexibly adapt to the load transmission requirements and cooperation changes between devices, solve the problem that the traditional static topology model cannot reflect dynamic load changes, and improve the real-time performance and flexibility of load control.

[0065] Benefits of Step 4: By defining the local optimization objectives of device nodes and the load control costs between nodes, use the global optimization framework to coordinate the load distribution of each node, and achieve distributed management of load control. The optimization results are calculated by the edge devices in real time and fed back to the cloud, reducing the over-reliance on cloud computing, enhancing the adaptive ability of the system, and solving the problem of high computational latency in centralized load control.

[0066] Benefits of Step 5: By constructing a game model, comprehensively consider the influence of short-term load prediction errors, long-term trend deviations, and communication overheads, dynamically adjust the load control strategy, and reach a global equilibrium state. The cloud sends the equilibrium strategy parameters to the edge devices for execution, ensuring the multi-objective balance of the control objectives, reducing communication redundancy at the same time, lowering the system operation energy consumption, and solving the problem that the load control strategy in traditional methods cannot take into account both global coordination and communication efficiency.

[0067] The data fusion in Step 1 is completed through the following weighted mechanism:

[0068] ,

[0069] Among them, is the i-th type of input data, is the number of feature categories, is the i-th type of input feature vector, which is calculated by the following formula:

[0070] ,

[0071] Among them, is the weight coefficient of each type of data, is the attention score of the i-th type of feature, is the number of feature categories, is the exponential function,

[0072] is the normalization term of the exponential value of the feature attention score.

[0073] By introducing a dynamic weighting mechanism, input data of different categories can adaptively adjust their weights according to their importance in specific scenarios. The dynamic weighting method solves the deficiency of assigning fixed weights to data in traditional feature fusion, enhances the flexibility and accuracy of feature expression, and improves the effect of load prediction.

[0074] The attention mechanism dynamically allocates weights according to the attention scores of various types of data, ensuring that the key characteristics of multi-modal data are retained in different industrial operating environments. For example, when the environment changes significantly, the weight of environmental data automatically increases, and when the device state changes drastically, the weight of device state data will increase, effectively adapting to the dynamic changes of load characteristics.

[0075] By calculating the weight coefficient through the exponential function and normalization processing, the weighted values of different feature categories can be obtained quickly, avoiding the high computational overhead brought by complex feature fusion methods. This method realizes the efficient fusion of multi-modal data at a low computational cost and is suitable for the real-time prediction requirements of large-scale industrial load systems.

[0076] The load prediction model in step 2 generates the load prediction value for future time points through the following process:

[0077] The encoder extracts the potential correlation patterns of the fused features:

[0078] ,

[0079] Among them, is the potential dependent variable, is the feature mean, is the feature standard deviation, is the random noise;

[0080] The decoder generates predicted load values through latent dependent variables , where is the current time point, is the prediction time interval.

[0081] By introducing the encoder to extract the latent correlation patterns of the fusion features, the complex non-linear relationships in industrial load data can be effectively captured, especially the implicit correlations between multi-modal data. The generation process of latent latent variables integrates the global and local characteristics of multi-modal data, solves the problem that traditional methods cannot predict complex load changes, and significantly improves the accuracy of load prediction.

[0082] The generation process of latent latent variables ensures the adaptability of the model to the random fluctuations and dynamic changes of load data through the dynamic adjustment of feature means and feature standard deviations, and the introduction of random noise. By generating future load prediction values from latent variables through the decoder, the sudden changes in industrial load scenarios can be handled, and the robustness and reliability of the model can be improved.

[0083] Through the structural design of extracting the latent pattern of features by the encoder and decoding to generate load prediction values, the model can adapt to different industrial scenarios and various load change patterns. The generation of latent latent variables provides a unified feature representation for the model, avoids overfitting problems, and enhances the generality and generalization ability of the model under different data sets and operating scenarios.

[0084] The process of generating predicted load values from latent latent variables by the decoder is efficient and stable, avoiding direct complex calculations on multi-modal data and reducing the consumption of computing resources. In the industrial real-time load prediction scenario, it can meet the efficiency requirements of high-frequency and large-scale predictions.

[0085] The dynamic graph model established in step 3 satisfies the following relationship:

[0086] ,

[0087] where is the time the edge weight between node i and node j, and are the load values of node i and node j, represents the difference degree of node loads.

[0088] By defining the edge weight as a function of the load difference between node i and node j, the dynamic graph model can reflect the load transfer intensity between nodes in real time. The dynamic adjustment mechanism solves the problem that traditional static models cannot adapt to the load changes of industrial equipment, making the cooperation relationship between equipment flexible.

[0089] The edge weight The calculation is closely related to the difference degree of node loads, enabling the model to dynamically capture the imbalance of load distribution among devices and timely optimize the load regulation strategy. By updating the relationships between nodes in real time, the load regulation is ensured to be accurate, meeting the requirements of complex industrial scenarios.

[0090] By dynamically adjusting the edge weights, reducing the connection weights of load differences between nodes, and reducing unnecessary load transfer collaborations. This mechanism can optimize the load regulation cost between nodes and improve the overall operation efficiency of the industrial load regulation system.

[0091] The dynamic graph model is updated in real time based on the load prediction values and device association relationships, and can quickly adapt to scenarios where the device operating states and load characteristics change, improving the self - adaptability and robustness of the model, especially in industrial environments with load fluctuations.

[0092] The local optimization objective function in Step 4 is defined as:

[0093] ,

[0094] where, is the actual load of node i, is the predicted load of node i,

[0095] is the square of the error between the actual load and the predicted load of node i,

[0096] is the prediction error function, is the load value of node i.

[0097] The load regulation cost function between nodes is defined as:

[0098] ,

[0099] where, is the collaboration cost function between node i and node j, is the load value of node i, is the load value of node j, is the penalty coefficient for load regulation between nodes, is the square of the difference in load values between node i and node j.

[0100] The global optimization framework is solved through the following objective function:

[0101] ,

[0102] where, is the optimization objective, is the load value of device node i, is a set of device nodes, is a set of edges between device nodes, is a prediction error function, is a cooperation cost function between node i and node j, is the load value of device node j.

[0103] By defining the local optimization objective function , taking the error between the actual load and the predicted load of each device node as the optimization objective can ensure the accuracy of the load prediction result, solve the problem that the single-node error in traditional load prediction cannot be effectively controlled, and improve the accuracy of load distribution.

[0104] The cooperation cost function between nodes By punishing the square of the difference in node load distribution, it can dynamically balance the load transfer intensity between nodes, reduce the system instability caused by load differences, effectively optimize the load cooperation cost between nodes, and improve the overall operation efficiency of the system.

[0105] The global optimization framework is solved through the objective function . While ensuring the minimization of the local load error of each node, it can optimize the global load distribution through the cooperation relationship between nodes, achieve the global coordination of load prediction and regulation, and solve the conflict problem between the single-node and the overall system objectives in the industrial load system.

[0106] The result of local optimization is calculated in real time by the edge device, reducing the computing pressure and data transmission delay of the cloud. The edge device and the cloud cooperate with each other, and can adjust the load regulation strategy in a timely manner according to the dynamic load change, enhancing the real-time and adaptive capabilities of the industrial load management system.

[0107] The game model in step 5 is constructed through the following utility function:

[0108] ,

[0109] where, is the utility function of device node i, is the short-term load prediction error, is the long-term load trend error, is the communication overhead, 、 、 are weight parameters.

[0110] The utility function is optimized through the following equilibrium conditions:

[0111] , , ,

[0112] where, , , is a weight parameter, is the utility function of device node i,

[0113] is the partial derivative of the utility function with respect to the short-term error weight coefficient .

[0114] is the partial derivative of the utility function with respect to the short-term error weight coefficient .

[0115] is the partial derivative of the utility function with respect to the short-term error weight coefficient .

[0116] The equilibrium strategy parameters generated by the cloud through game optimization are periodically sent to the edge devices. The edge devices adjust the load control strategy in real time according to the received strategy parameters combined with local data, and upload the load control results to the cloud for updating the next round of strategy parameters.

[0117] By constructing the utility function in the game model, the impacts of short-term load prediction errors, long-term trend errors, and communication overheads are comprehensively considered. By using the dynamic adjustment of the weight parameters , and , the conflicts between multiple objectives can be effectively balanced, and the problem that traditional single-objective optimization methods cannot simultaneously meet the global objective requirements can be solved.

[0118] By optimizing the equilibrium condition of the utility function, and , the equilibrium strategies of each device node are solved, so that the benefits of each node are maximized, and the overall system reaches the global optimal state, significantly improving the equilibrium of the load control strategy and ensuring the stability of system operation.

[0119] In game optimization, the communication overhead is added to the utility function. By optimizing and adjusting the communication weight , the data transmission frequency and communication volume between the cloud and the edge devices can be dynamically controlled, effectively reducing redundant communication, lowering system energy consumption, and ensuring the real-time nature of the control strategy.

[0120] The equilibrium strategy parameters generated by the cloud are periodically sent to the edge devices. The edge devices adjust the load control strategy in real time according to the received strategy parameters combined with local data, and feedback the control results to the cloud for updating. This method ensures that the load control strategy can be continuously optimized according to the dynamic changes of the industrial system, enhancing the adaptive ability of the system.

[0121] By comprehensively considering short-term errors, long-term trends, and communication overhead, the game model can optimize the resource allocation of the entire system, significantly reduce the operating costs of industrial systems, and improve economic efficiency.

[0122] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art will understand that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An industrial load control method based on cloud-edge collaborative technology, characterized in that, Including: Step 1: By deploying load sensors, environmental monitoring devices, and device status monitoring devices, historical load data, environmental data, and device operation status data are collected. The historical load data is used to represent the load change of the system at multiple time points. The environmental data includes externally acquired real-time temperature and electricity price information. The device operation status data includes device operation status parameters. The historical load data, environmental data, and device operation status data are combined for data fusion to generate a unified fusion feature as the input for load forecasting. Step 2: The fusion feature generated in Step 1 is input into a load forecasting model. The input feature is processed by the encoding module of the model to extract potential load correlation patterns. Then, the decoding module generates the load forecasting value for future time points based on the potential patterns. The forecasting model is constructed based on the correlation of the load historical change trend, environmental data, and device status. Step 3: Combining the load forecasting value generated in Step 2 and the industrial device network structure, a dynamic graph model reflecting the load transmission relationship between devices is established. The dynamic graph model includes a set of device nodes, the connection relationships between nodes, and an edge weight matrix reflecting the load transmission intensity between devices. The dynamic graph is updated in real time based on the load forecasting value and the association relationship between devices. Step 4: Based on the dynamic graph model established in Step 3, a local optimization objective for device nodes is defined. For the cooperation relationship between device nodes, the load regulation cost between nodes is constructed. The global optimization framework coordinates the load distribution of each node. The optimization result is calculated in real time by the edge device and the processing result is fed back to the cloud. Step 5: Combining the load distribution result optimized in Step 4, a game model including device nodes is constructed in the cloud. The device nodes are the participants in the game model. According to the comprehensive influence of the short-term load forecasting error, long-term trend deviation, and communication overhead, the load regulation strategy is dynamically adjusted. The global equilibrium state is achieved through the game model. Finally, the regulation strategy parameters calculated by the cloud are sent to the edge device for execution to complete the global coordination and dynamic optimization of the load.

2. The industrial load control method based on cloud-edge collaboration technology according to claim 1, wherein, The data fusion in Step 1 is completed through the following weighting mechanism: ， Among them, is the i-th type of input data, is the number of feature categories, is the i-th type of input feature vector, which is calculated by the following formula: ， Among them, is the weight coefficient of various types of data, is the attention score of the i-th type of feature, is the number of feature categories, is the exponential function, It is the normalization term of the exponential value of the feature attention score.

3. The industrial load control method based on cloud-edge collaboration technology according to claim 1, characterized in that, The load forecasting model in Step 2 generates the load forecasting value for future time points through the following process: The encoder extracts the potential correlation patterns of the fusion feature: ， wherein, is the potential dependent variable, is the feature mean, is the feature standard deviation, is the random noise; The decoder generates a predicted load value through a latent dependent variable , where is the current time point, is the prediction time interval.

4. The industrial load control method based on cloud-edge collaboration technology according to claim 1, wherein The dynamic graph model established in step 3 satisfies the following relationship: ， Among them, is the time is the edge weight between node i and node j, and are the load values of node i and node j, represents the difference degree of node loads.

5. The industrial load control method based on cloud-edge collaboration technology according to claim 4, wherein, The local optimization objective function in Step 4 is defined as: ， Among them, is the actual load of node i, is the predicted load of node i, is the square of the error between the actual load and the predicted load of node i, is the prediction error function, is the load value of node i.

6. The industrial load control method based on cloud-edge collaboration technology according to claim 5, characterized in that, The load regulation cost function between nodes is defined as: ， Among them, is the collaboration cost function between node i and node j, is the load value of node i, is the load value of node j, is the penalty coefficient for load regulation between nodes, is the square of the difference in load values between node i and node j.

7. The industrial load control method based on cloud-edge collaboration technology according to claim 6, wherein The global optimization framework is solved through the following objective function: ， Among them, is the optimization objective, is the load value of device node i, is the set of device nodes, is the edge set between device nodes, is the prediction error function, is the collaboration cost function between node i and node j, is the load value of device node j.

8. The industrial load control method based on cloud-edge collaboration technology according to claim 1, wherein, The game model in Step 5 is constructed through the following utility function: ， Among them, is the utility function of device node i, is the short-term load forecasting error, is the long-term load trend error, is the communication overhead, , , are the weight parameters.

9. The industrial load control method based on cloud-edge collaboration technology according to claim 8, wherein, The utility function is optimized through the following equilibrium condition: ， ， ， Among them, , , are weight parameters, is the utility function of device node i, is the partial derivative of the utility function with respect to the short-term error weight coefficient , is the partial derivative of the utility function with respect to the short-term error weight coefficient , is the partial derivative of the utility function with respect to the short-term error weight coefficient .

10. The industrial load control method based on cloud-edge collaboration technology according to claim 9, wherein The equilibrium strategy parameters generated by game optimization in the cloud are periodically sent to the edge device. The edge device adjusts the load regulation strategy in real time according to the received strategy parameters combined with local data, and uploads the load regulation result to the cloud for updating the next round of strategy parameters.

Images