A climate digital twin and dynamic optimization system based on adaptive edge and cloud collaboration

Abstract

This invention belongs to the field of climate big data and artificial intelligence technology, and discloses a climate digital twin and dynamic optimization system based on adaptive edge and cloud collaboration. The system includes: an edge adaptive fusion module: performing weighted Kalman fusion on grouped collected edge data; a multi-resolution spatiotemporal graph neural network: constructing a graph according to resolution levels of grid nodes, and performing cross-layer graph attention and temporal convolution in parallel; an online Bayesian error correction closed-loop module: this module calculates the system error, calculates the Kalman gain, and updates the parameters to achieve correction closed loop; a hierarchical federated learning and model pruning network: edge nodes train sub-model gradients locally, and use differential privacy and pruning techniques to process uploaded data, aggregating and pruning the information uploaded by each edge node in the cloud; and a multi-objective dynamic optimization module: achieving multi-objective dynamic optimization based on a joint objective. This invention achieves advantages such as robust fusion of multi-source observations, cross-resolution spatiotemporal modeling, and online correction closed loop.

Description

Technical Field

[0001] This invention belongs to the field of climate big data and artificial intelligence technology, specifically relating to a climate digital twin and dynamic optimization system based on adaptive edge and cloud collaboration. Background Technology

[0002] In the fields of climate big data and artificial intelligence, accurate simulation and prediction of spatiotemporal climate are of great significance for environmental monitoring, disaster early warning, and energy management. Existing spatiotemporal climate simulation and prediction technologies largely rely on centralized cloud models, constructing global climate models to simulate and predict climate elements at spatiotemporal scales. However, these technologies suffer from the following significant drawbacks in practical applications: Data silos exist at the edge: a large amount of local time-series data collected by sensor clusters cannot be integrated into the simulation in real time; Multi-resolution unification is difficult: cloud-edge networks require frequent switching between different scale grid models, resulting in high computational overhead; Online correction lags: simulation errors need to be analyzed offline, making it impossible to form a real-time closed loop; Optimization involves multiple conflicting objectives: there is a lack of unified optimization strategies for accuracy, latency, and computing resources. Summary of the Invention

[0003] The purpose of this invention is to provide a climate digital twin and dynamic optimization system based on adaptive edge and cloud collaboration, so as to solve the technical problems of existing technologies, such as isolated edge data, difficulty in unifying multiple resolutions, lagging online correction, and conflicting objectives during optimization.

[0004] The aforementioned climate digital twin and dynamic optimization system based on adaptive edge and cloud collaboration includes:

[0005] Edge adaptive fusion module: The collected edge data is grouped according to terrain and microclimate and then weighted Kalman fusion is performed to realize edge cluster adaptive sensor fusion;

[0006] Multi-resolution spatiotemporal graph neural network: Constructs a graph by grid nodes according to resolution hierarchy, and performs cross-layer graph attention and temporal convolution in parallel;

[0007] Online Bayesian error correction closed-loop module: This module calculates the error using the real observation vector fused from the edge sensors, calculates the Kalman gain and updates the parameters to achieve a correction closed loop;

[0008] Hierarchical Federated Learning and Model Pruning Network: The gradients of sub-models are trained locally at edge nodes, and the uploaded information is processed using differential privacy and pruning techniques. The information uploaded by each edge node is aggregated and pruned in the cloud.

[0009] Multi-objective dynamic optimization module: In this module, a joint objective is pre-designed, and then an online evolution strategy is adopted to dynamically adjust the weights, thereby achieving multi-objective dynamic optimization based on the joint objective.

[0010] Preferably, the algorithm for weighted Kalman fusion of edge adaptive fusion modules includes:

[0011] 1) Calculate the weights of edge data: The input is The formula for calculating the corresponding weight is:

[0012] ,

[0013] In the formula, S k For the first k The sensor index set in each fusion group is obtained by clustering based on geography and microclimate; x i For the first i The observations from sensor number 1; For the first i The noise variance observed by the sensor reflects the measurement uncertainty. To calculate the normalized weights, The smaller the noise variance, the larger the proportion;

[0014] 2) Edge Cluster Adaptive Sensor Fusion: The fusion calculation formula is:

[0015] ,

[0016] This is a local state estimate after fusion, used for real-time edge correction and reporting.

[0017] Preferably, in a multi-resolution spatiotemporal graph neural network, the resolution level is... l =1 , . . . ,L ,picture G (l) = ( V,E (l) ), V The set of all grid nodes, number of nodes | V |= N ; E (l) For the first l The edge set of a layer resolution map consists of the spatial distance or physical correlation between adjacent grid points.

[0018] Preferably, in a multi-resolution spatiotemporal graph neural network, layers l The update expression is as follows:

[0019] ,

[0020] ,

[0021] ,

[0022] ,

[0023] in, Indicates the first l Layer time node i eigenvectors; Indicates the first l Layer feature linear transformation matrix; Indicates the first l Layer attention mechanism parameter vector; LeakyReLU represents a vector concatenation operation, which concatenates the features of the source and target nodes; · ) is a ReLU activation function with a negative slope; The attention weights after normalization are used to measure the nodes. j right i Its influence; For the first l Layer and Node i A set of connected neighbors; Conv τ This is a temporal convolution operation, with a kernel size of [size missing]. t Used to capture time dependencies; s ( · ) represents a non-linear activation function. Indicates the first l Layer nodes j right i The correlation, Indicates the first l After layer update, the node i The feature representation is obtained by aggregating information from its neighboring nodes.

[0024] Preferably, the multi-resolution spatiotemporal graph neural network couples cross-layer graph attention with temporal convolution updates in parallel. The parallel coupling can be achieved on the spectrum to: (i) resist alias alignment; (ii) preserve teleconnections; and (iii) alleviate graph oversmoothing.

[0025] Preferably, the process of implementing the correction closed loop in the online Bayesian error correction closed loop module includes:

[0026] 1) The system will vectorize the initial simulation results from the cloud. T sim The data is distributed as real observation vectors fused from edge sensors. T obs Calculation error: Δ= T obs T sim Δ represents the simulation error, used for correction;

[0027] 2) Kalman gain and parameter update, the expression is as follows: K = PH ( HPH + R ) 1 , i ← i + K Δ, where P This is the parameter covariance matrix, reflecting the parameter uncertainty; H The observation matrix is ​​used to map the parameter space to the observation space; R To observe the noise covariance, which reflects the distribution of measurement errors, K Kalman gain is a correction factor used to balance prediction and observation. i This is the vector of parameters to be estimated in the simulation model.

[0028] The preferred process for implementing hierarchical federated learning and model pruning is as follows:

[0029] 1) The gradient of the sub-model is trained locally at the edge nodes, using differential privacy and pruning techniques. The corresponding expression is: , , ,in For the first i Local gradients of edge nodes; Let i represent the local loss function at the i-th edge node. i Indicates local model parameters. Let represent the clipping function, and have , That is to The corresponding threshold; for The results of pruning This represents Gaussian noise, used for differential privacy protection. Let Variance be the variance of the Gaussian noise. I It is the identity matrix; Δ i For the first i The gradient information with added noise uploaded by the edge nodes;

[0030] 2) Aggregate and prune the information uploaded by each edge node in the cloud. The corresponding expression is:

[0031] ,

[0032] in, or The learning rate; These are global model parameters in the cloud. This indicates model pruning, retaining the importance scores before... r The weight of the proportion.

[0033] Preferably, in the multi-objective dynamic optimization module, the specific method for implementing multi-objective dynamic optimization is as follows: 1) Pre-design a joint objective. The joint objective is designed according to requirements. One expression for a joint objective is as follows:

[0034] ,

[0035] Where Δ is the simulation error vector. T delay For the round-trip delay between the end and the cloud, Indicates the sparsity of the model. l i This represents the target weight, i=1,2,3. The target weight is dynamically updated online based on system load and error. k Indicates the first k Sensors in a fusion group;

[0036] 2) An online evolution strategy is adopted to dynamically adjust the weights; the formula for Hedge updates is as follows:

[0037] ,

[0038] in, Represents the time t. i The target weight of the item or t Let be the learning rate at time t. For time t, the first i The pruning results of each edge node; using a constrained lightweight NSGA-II with decision variable Θ=( i , k Returns an approximate Pareto set.

[0039] Preferably, this system employs a visualization-user-driven scenario feedback technology, allowing users to define a set of climate control parameters. The system then triggers a simulation-correction-optimization closed loop to achieve control of the set of parameters. The update expression is: ,in U This is a multi-objective utility function used to integrate error and resource cost, where γ is the step size (learning rate). t grid The grid trigger threshold is used for scene selection. For regulating parameter set The gradient.

[0040] The technical advantages of this invention are as follows: Under edge-cloud collaborative conditions, this invention achieves robust fusion of multi-source observations through weighted Kalman fusion grouped by terrain and microclimate; it completes cross-resolution spatiotemporal modeling and operation through parallel coupling of cross-layer graph attention and temporal convolution updates, avoiding switching between grid models of multiple scales; it achieves a closed-loop correction of cloud simulation and edge observation through online Bayesian correction; it achieves training of cloud-edge models through differential privacy and structured sparsity; it completes multi-objective online optimization for extreme events based on online evolution strategies and joint design objectives, and ensures existence, stability, and convergence with rigorous theoretical proof when experimental gains are lacking. Attached Figure Description

[0041] Figure 1 This is a flowchart illustrating the edge sensor data fusion process in the edge adaptive fusion module of this invention.

[0042] Figure 2 This is a flowchart of the Multi-Resolution Spatiotemporal Graph Neural Network (MR-STGNN) in this invention.

[0043] Figure 3 This is a flowchart of the online Bayesian correction closed loop in this invention.

[0044] Figure 4 This is a flowchart of the hierarchical federated learning and pruning network in this invention.

[0045] Figure 5 This is a flowchart illustrating the optimization process of the multi-objective dynamic optimization module in this invention.

[0046] Figure 6 This is a flowchart of the visualization-user-driven contextual feedback technology used in this invention. Detailed Implementation

[0047] The following detailed description of the embodiments, with reference to the accompanying drawings, will further illustrate the specific implementation of the present invention, in order to help those skilled in the art to have a more complete, accurate, and in-depth understanding of the inventive concept and technical solution of the present invention.

[0048] like Figure 1-Figure 6 As shown, this invention provides a climate digital twin and dynamic optimization system based on adaptive edge and cloud collaboration, employing the SE-DO (Digital Twin for Spatiotemporal Encoding with Dynamic Optimization) framework, and including the following modules:

[0049] Edge Adaptive Fusion Module: This module performs weighted Kalman fusion on the acquired edge data, grouping it according to terrain and microclimate, to achieve adaptive sensor fusion for edge clusters. The algorithm for weighted Kalman fusion based on grouping includes:

[0050] 1) Calculate the weights of edge data: The input is The formula for calculating the corresponding weight is:

[0051] ,

[0052] In the formula, S k For the first k The sensor index set in each fusion group is obtained by clustering based on geography and microclimate; x i For the first i The observed values ​​of the sensor (such as temperature, humidity, etc.); For the first i The noise variance observed by the sensor reflects the measurement uncertainty. To calculate the normalized weights, The smaller the noise variance, the larger the proportion.

[0053] 2) Edge Cluster Adaptive Sensor Fusion: The fusion calculation formula is:

[0054] ,

[0055] This module provides the fused local state estimate for real-time edge correction and reporting. It also enables adaptive filtering of local noise and spatiotemporal completion.

[0056] Multi-resolution Spatiotemporal Graph Neural Network (MR-STGNN): In this network, grid nodes are arranged according to resolution levels. l =1 , . . . ,L Construction Graph G (l) = ( V,E (l) It performs cross-layer graph attention and temporal convolution in parallel. V The set of all grid nodes, number of nodes | V |= N ; E (l) For the first l The edge set of a layer resolution map consists of the spatial distance or physical correlation between adjacent grid points.

[0057] In this network, layers l The update expression is as follows:

[0058] ,

[0059] ,

[0060] ,

[0061] ,

[0062] in, Indicates the first l Layer time node i eigenvectors; Indicates the first l Layer feature linear transformation matrix; Indicates the first l Layer attention mechanism parameter vector; LeakyReLU represents a vector concatenation operation, which concatenates the features of the source and target nodes; · ) is a ReLU activation function with a negative slope; The attention weights after normalization are used to measure the nodes. j right i Its influence; For the first l Layer and Node i A set of connected neighbors; Conv τ This is a temporal convolution operation, with a kernel size of [size missing]. t Used to capture time dependencies; s ( · ) represents a nonlinear activation function (such as ReLU or GELU). Indicates the first l Layer nodes j right i The correlation, Indicates the first l After layer update, the node i The characteristic representation is obtained by aggregating information from its neighboring nodes. The information flow process in this network is as follows: ,E (l) → → → → .

[0063] Multi-resolution spatiotemporal graph neural networks couple cross-layer graph attention in parallel with temporal convolution updates. This parallel coupling can be achieved on the spectrum for: (i) anti-aliasing alignment (temporal phase correction); (ii) teleconnection preservation (long-distance dependence across scales); and (iii) mitigating graph oversmoothing (expanding spectral bandwidth).

[0064] Online Bayesian error correction closed-loop module: The process of implementing the correction closed loop in this module includes:

[0065] 1) The system will vectorize the initial simulation results from the cloud. T sim The data is distributed as real observation vectors fused from edge sensors. T obs Calculation error: Δ= T obs T sim Δ represents the simulation error, used for correction.

[0066] 2) Kalman gain and parameter update, the expression is as follows: K = PH ( HPH + R ) 1 , i ← i + K Δ, where P This is the parameter covariance matrix, reflecting the parameter uncertainty; H The observation matrix is ​​used to map the parameter space to the observation space; R To observe the noise covariance, which reflects the distribution of measurement errors, K Kalman gain is a correction factor used to balance prediction and observation. i This is the vector of parameters to be estimated in the simulation model.

[0067] Hierarchical Federated Learning and Model Pruning Module: This module implements hierarchical federated learning and model pruning, as detailed below:

[0068] 1) The gradient of the sub-model is trained locally at the edge nodes, using differential privacy and pruning techniques. The corresponding expression is: , , ,in No. i Local gradients of edge nodes; Let i represent the local loss function at the i-th edge node. i Indicates the current global model parameters. Let represent the clipping function, and have , That is to The corresponding threshold; This represents Gaussian noise, used for differential privacy protection. Let Variance be the variance of the Gaussian noise. IIt is the identity matrix; or The learning rate; i , r () indicates model pruning, retaining the components before the importance score. r The weight of the proportion; Δ i For the first i The gradient information with added noise uploaded by the edge node.

[0069] 2) Aggregate and prune the information uploaded by each edge node in the cloud. The corresponding expression is:

[0070] ,

[0071] in, or The learning rate; These are global model parameters in the cloud. This indicates model pruning, retaining the importance scores before... r The weights of the proportions are determined. Finally, the global model parameters obtained from the pruning results are sent to each edge node for peer processing. The information flow process is as follows:

[0072] Upward convergence Downlink synchronization.

[0073] Multi-objective dynamic optimization module: In this module, a joint objective is pre-designed, and then an online evolution strategy is used to dynamically adjust the weights to achieve multi-objective dynamic optimization based on the joint objective. Details are as follows:

[0074] 1) Pre-design the joint objective. The joint objective is designed according to the requirements. One expression for the joint objective is as follows:

[0075] ,

[0076] Where Δ is the simulation error vector. T delay For the round-trip delay between the end and the cloud, This indicates the sparsity of the model (i.e., the number of non-zero parameters). l i This represents the target weight, i=1,2,3. The target weight is dynamically updated online based on system load and error. k Indicates the first k Sensors in a fusion group;

[0077] 2) Employ an online evolution strategy to dynamically adjust weights; the online evolution strategy includes:

[0078] The outer layer: The Hedge update formula is as follows:

[0079] ,

[0080] in, Represents the time t. i The target weight of the item or t Let be the learning rate at time t. For time t, the first i The pruning results of each edge node.

[0081] Inner layer: Employs a constrained lightweight NSGA-II with decision variable Θ=( i , k The function returns an approximate Pareto set. The theoretical framework is as follows:

[0082]

[0083] The corresponding information flow process is as follows:

[0084] Target calculation evolutionary strategy .

[0085] Visualization-User-Driven Contextual Feedback: This system employs visualization-user-driven contextual feedback technology, as detailed below:

[0086] User-defined climate control parameter set The system then triggers a simulation-correction-optimization closed loop to achieve control of the set of parameters. The update expression is: ,in U This is a multi-objective utility function used to integrate error and resource cost, where γ is the step size (learning rate). t grid The grid trigger threshold is used for scene selection. For regulating parameter set The gradient. The corresponding information flow process is as follows:

[0087] Simulation—Correction—Optimization .

[0088] The present invention has been described above by way of example with reference to the accompanying drawings. Obviously, the specific implementation of the present invention is not limited to the above-described manner. Any non-substantial improvements made using the inventive concept and technical solution of the present invention, or the direct application of the inventive concept and technical solution of the present invention to other occasions without modification, are all within the protection scope of the present invention.

Claims

1. A climate digital twin and dynamic optimization system based on adaptive edge and cloud collaboration, characterized in that: include: Edge adaptive fusion module: The collected edge data is grouped according to terrain and microclimate and then weighted Kalman fusion is performed to realize edge cluster adaptive sensor fusion; Multi-resolution spatiotemporal graph neural network: Constructs a graph by dividing grid nodes according to resolution levels, and performs cross-layer graph attention and temporal convolution in parallel; the resolution levels are... l =1 , . . . ,L ,picture G (l) = ( V,E (l) ), V The set of all grid nodes, number of nodes | V |= N ; E (l) For the first l The edge set of a layer resolution map consists of the spatial distance or physical correlation between adjacent grid points; layer l The update expression is as follows: ， ， ， ， in, Indicates the first l Layer time node i eigenvectors; Indicates the first l Layer feature linear transformation matrix; Indicates the first l Layer attention mechanism parameter vector; LeakyReLU represents a vector concatenation operation, which concatenates the features of the source and target nodes; · ) is a ReLU activation function with a negative slope; The attention weights after normalization are used to measure the nodes. j right i Its influence; For the first l Layer and Node i A set of connected neighbors; Conv τ This is a temporal convolution operation, with a kernel size of [size missing]. τ Used to capture time dependencies; σ ( · ) represents a non-linear activation function. Indicates the first l Layer nodes j right i The correlation, Indicates the first l After layer update, the node i The feature representation is obtained by aggregating information from its neighboring nodes; Online Bayesian error correction closed-loop module: This module calculates the error using the real observation vector fused from the edge sensors, calculates the Kalman gain and updates the parameters to achieve a correction closed loop; The online Bayesian error correction closed-loop module implements the correction closed-loop process as follows: 1) The system will vectorize the initial simulation results from the cloud. T sim The data is distributed as real observation vectors fused from edge sensors. T obs Calculation error: Δ= T obs T sim Δ represents the simulation error, used for correction; 2) Kalman gain and parameter update, the expression is as follows: K = PH ( HPH + R ) 1 , θ ← θ + K Δ, where P This is the parameter covariance matrix, reflecting the parameter uncertainty; H The observation matrix maps the parameter space to the observation space. R To observe the noise covariance, which reflects the distribution of measurement errors, K Kalman gain is a correction factor used to balance prediction and observation. θ The vector of parameters to be estimated in the simulation model; Hierarchical Federated Learning and Model Pruning Network: The gradients of sub-models are trained locally at edge nodes, and the uploaded information is processed using differential privacy and pruning techniques. The information uploaded by each edge node is aggregated and pruned in the cloud. The specific process for implementing hierarchical federated learning and model pruning is as follows: 1) The gradient of the sub-model is trained locally at the edge nodes, using differential privacy and pruning techniques. The corresponding expression is: , , ,in For the first i Local gradients of edge nodes; Let i represent the local loss function at the i-th edge node. θ The vector of parameters to be estimated in the simulation model. Let represent the clipping function, and have , That is to The corresponding threshold; for The results of pruning This represents Gaussian noise, used for differential privacy protection. Let Variance be the variance of the Gaussian noise. I It is the identity matrix; Δ i For the first i The gradient information with added noise uploaded by the edge nodes; 2) Aggregate and prune the information uploaded by each edge node in the cloud. The corresponding expression is: , in, η The learning rate; These are global model parameters in the cloud. This indicates model pruning, retaining the scores before importance scores. ρ The weight of the proportion; Multi-objective dynamic optimization module: In this module, a joint objective is pre-designed, and then an online evolution strategy is adopted to dynamically adjust the weights, thereby achieving multi-objective dynamic optimization based on the joint objective; In the multi-objective dynamic optimization module, the specific methods for implementing multi-objective dynamic optimization are as follows: 1) Pre-design the joint objective. The joint objective is designed according to requirements. One expression for the joint objective is as follows: ， Where Δ is the simulation error vector. T delay For the round-trip delay between the end and the cloud, Indicates the sparsity of the model. λ i This represents the target weight, i=1,2,3. The target weight is dynamically updated online based on system load and error. k Indicates the first k Sensors in a fusion group; 2) An online evolution strategy is adopted to dynamically adjust the weights; the online evolution strategy includes: outer layer: the formula for updating the Hedge is as follows: ， in, Represents the time t. i The target weight of the item η t Let be the learning rate at time t. For time t, the first i The pruning results of each edge node; Inner layer: Employs a constrained lightweight NSGA-II with decision variable Θ=( θ , k Returns an approximate Pareto set.

2. The climate digital twin and dynamic optimization system based on adaptive edge and cloud collaboration according to claim 1, characterized in that: The algorithm for weighted Kalman fusion by grouping edge adaptive fusion modules includes: 1) Calculate the weights of edge data: The input is The formula for calculating the corresponding weight is: ， In the formula, S k For the first k The sensor index set in each fusion group is obtained by clustering based on geography and microclimate; x i For the first i The observations from sensor number 1; For the first i The noise variance observed by the sensor reflects the measurement uncertainty. To calculate the normalized weights, The smaller the noise variance, the larger the proportion; 2) Edge Cluster Adaptive Sensor Fusion: The fusion calculation formula is: ， This is a local state estimate after fusion, used for real-time edge correction and reporting.

3. The climate digital twin and dynamic optimization system based on adaptive edge and cloud collaboration according to claim 1, characterized in that: This system employs a visualization-driven, user-defined scenario feedback technology, allowing users to define their own climate control parameter sets. The system then triggers a simulation-correction-optimization closed loop to achieve control of the set of parameters. The update expression is: ,in U This is a multi-objective utility function used to integrate error and resource cost, where γ is the step size. τ grid The grid trigger threshold is used for scene selection. For regulating parameter set The gradient.

Images