A backup scheduling and bandwidth allocation method for multi-active disaster tolerance network

Abstract

This invention relates to the field of network resource scheduling technology, and provides a backup scheduling and bandwidth allocation method for multi-active disaster recovery networks, solving the problems of insufficient forward-looking perception, limited dynamic adjustment capabilities, and weak collaborative optimization capabilities of traditional methods. The method includes: reading historical traffic data, topological associations, and time period information from multiple nodes, performing preprocessing and spatiotemporal feature embedding; performing spatial association modeling, temporal evolution modeling, and feature fusion on the input representation to predict future time-slot node traffic changes and obtain a forward-looking load characterization; constructing a node state space by combining current traffic observations, prediction results, queue status, remaining backup job volume, and node resource information; generating backup job receiving actions and weighted fair queue weight configuration actions based on the state space, interacting with the environment to obtain reward information; and performing value evaluation and policy updates based on the state, actions, and reward information to achieve collaborative optimization of backup scheduling and bandwidth allocation.

Description

Technical Field

[0001] This invention relates to the field of computer network technology, specifically providing a backup scheduling and bandwidth allocation method for multi-active disaster recovery networks. Background Technology

[0002] With the continuous development of cloud computing, distributed storage, and multi-datacenter collaborative technologies, multi-active disaster recovery architectures are widely used in scenarios with high requirements for business continuity and data reliability. In such networks, multiple nodes need to perform local backup jobs periodically or triggered by events, while also ensuring the concurrent transmission quality of online services and backup traffic. Therefore, the job scheduling and egress bandwidth allocation on the node side directly affect the end-to-end backup timeliness, resource utilization efficiency, and system operational stability.

[0003] However, in multi-node parallel backup scenarios, multiple sites often trigger backup processes simultaneously within similar time windows. This, coupled with the temporal fluctuations in online business traffic, causes significant time-varying and highly competitive time-dependent computing, storage, and egress queue resources on the node side. Short-term congestion on a single node not only causes queue backlog and backup delays but can also induce cross-node load shifts through business flow rerouting and backup window overlap, ultimately evolving into a global, spatiotemporally coupled load jitter. This becomes a crucial factor restricting the quality of multi-active disaster recovery services.

[0004] Currently, most methods for node-side resource scheduling and bandwidth allocation still employ threshold-based heuristic strategies, such as static quota allocation based on priority queues, greedy scheduling based on remaining resources, or concentrated backup execution during off-peak periods to avoid peak contention. While these methods are simple to implement, they typically rely on the assumption that the load is approximately stable or that local states fully reflect the global state. Faced with complex scenarios such as non-stable business traffic, sudden backup requests, and parallel triggering by multiple nodes, existing methods suffer from the following main shortcomings:

[0005] 1. Node resource usage and queue backlog have significant time correlation and cross-node correlation. Relying solely on instantaneous indicators is insufficient to characterize load evolution trends, resulting in a lack of foresight in scheduling decisions.

[0006] 2. Rule-driven task rearrangement and weight adjustment lack continuous dynamic correction capabilities, making it difficult to achieve stable and efficient adaptive allocation based on changes in resource status;

[0007] 3. Although some existing adaptive methods have introduced reinforcement learning for scheduling optimization, most of them still encode the state from the perspective of a single node or a single agent, lacking the ability to jointly model the temporal evolution pattern and spatial correlation structure. Therefore, it is difficult to achieve stable collaborative optimization in scenarios with large-scale node concurrency and strong coupling load.

[0008] In summary, the limitations of existing technologies restrict the effectiveness of network node-side resource scheduling and bandwidth allocation in multi-active disaster recovery networks. It is necessary to propose a backup scheduling and bandwidth allocation method for multi-active disaster recovery networks to improve the ability to proactively perceive changes in node load and the adaptive optimization ability for multi-node collaborative scheduling, thereby reducing the risk of resource contention and queue congestion, and improving system throughput, resource utilization efficiency, and operational stability. Summary of the Invention

[0009] The purpose of this invention is to solve the problems of inefficient resource allocation, low resource utilization, increased queue congestion risk, and degraded system service performance in existing technologies for large-scale parallel backup of nodes in multi-active disaster recovery networks. These problems arise because existing technologies struggle to accurately depict the spatiotemporal dynamic changes of node load and lack a joint adaptive decision-making mechanism for backup job scheduling and bandwidth allocation.

[0010] To achieve the above objectives, the present invention adopts the following technical solution:

[0011] This invention proposes a backup scheduling and bandwidth allocation method for multi-active disaster recovery networks, comprising:

[0012] S1: Read historical traffic data, topology association information and time period information of multiple nodes, preprocess the data and embed spatiotemporal features to obtain a unified spatiotemporal input representation;

[0013] S2: Perform spatial correlation modeling and temporal evolution modeling on the spatiotemporal input representation, extract spatiotemporal dynamic features, fuse the spatiotemporal dynamic features, predict the node traffic changes in several future time slots, and obtain a forward-looking network traffic representation.

[0014] S3: Based on the current traffic observation of the node, the forward-looking network traffic representation, the egress queue status, the remaining amount of backup jobs, and the node's storage resources, computing resources, and bandwidth resources, construct the node's local state, and then splice the local states of each node to form the global state.

[0015] S4: Based on the local and global states of the node, extract context information related to the current scheduling decision from the forward-looking network traffic representation, and generate backup job receiving action and weighted fair queue weight configuration action according to the context information, thereby forming a joint action;

[0016] S5: Interact with the network environment to obtain immediate feedback and the next time slot observation results; perform value assessment and policy update based on the global state, joint actions and immediate feedback to achieve coordinated optimization of backup scheduling and bandwidth allocation.

[0017] In the above method, the feature extraction in S1 is described in detail below:

[0018] S11: Read historical traffic data from multiple nodes and construct a historical input traffic tensor according to the historical time window length, number of nodes, and traffic feature dimensions, denoted as... ,in, Indicates the length of the history window. Indicates the number of nodes. This represents the traffic characteristic dimension of each node;

[0019] Simultaneously read network topology information and construct a topology graph. and its adjacency matrix ,in, Represents a set of network nodes. Represents a set of links;

[0020] S12: Embed the original traffic features into the historical input traffic tensor, and project the original traffic data onto a unified latent space through linear mapping to obtain the original traffic feature representation:

[0021]

[0022] in, For linear embedding operations, Indicates the embedded spatial dimension;

[0023] S13: Based on the network topology information, spatial feature embedding is performed to aggregate neighboring node information while preserving the node's own information during feature propagation, and to optimize the adjacency matrix. Introduce node self-connections and normalize the adjacency matrix after introducing self-connections:

[0024] ,

[0025] in, express 3D identity matrix Indicates and The corresponding degree matrix. Then, using the initial features of the nodes. As input, topological relationship features between nodes are extracted through graph convolution operations to obtain a spatial embedding representation:

[0026]

[0027] in, Indicates the number of nodes. and This represents the learnable weight parameters. and Indicates the bias parameter. Represents spatial embedding features, Indicates the activation function;

[0028] S14: Based on the time period information, perform time feature embedding, and encode the weekday index and intraday time slot index corresponding to each time step using one-hot encoding, and concatenate them to obtain the time feature vector:

[0029]

[0030] in, Indicates a time step. Indicates time step The corresponding weekday feature vector, Indicates time step The corresponding intraday time slot feature vector, This represents a vector concatenation operation, which stacks the temporal feature vectors of all time steps to form a concatenation vector. The temporal embedding representation is obtained through two fully connected layers:

[0031]

[0032] in, and For learnable weight parameters, and For bias parameters, This represents the activation function. For temporal embedding features;

[0033] S15: Expand and stack the spatial embedding features along the time dimension to obtain: And by aligning the temporal embedding features to all nodes, we get: ,in, For the expanded spatial embedding features, These are the aligned temporal embedding features.

[0034] The extended spatial embedding feature Aligned temporal embedding features Compared with the original flow characteristics representation By performing element-wise addition, a unified spatiotemporal input representation is obtained:

[0035]

[0036] in, To unify the spatiotemporal input representation.

[0037] In multi-active disaster recovery networks, service traffic and backup requests are constantly and dynamically changing, and node-side computing, storage, and bandwidth resources exhibit significant time-varying and spatial coupling characteristics. Existing scheduling methods based on static rules or instantaneous states struggle to map the evolution trend of network load in real time and are prone to perception lag in scenarios involving sudden traffic spikes, multi-node parallel backups, and resource contention, leading to a mismatch between subsequent scheduling decisions and the actual network state. To address this, step S2 introduces STAF-Former, combining spatiotemporal attention and an adaptive spatiotemporal fusion mechanism to perform spatiotemporal joint modeling of node historical traffic information, obtaining a fused spatiotemporal feature representation, and further outputting node traffic prediction results for several future time slots, forming a forward-looking network traffic characterization. This step provides advanced state information for subsequent backup scheduling and bandwidth allocation, improving the accuracy and stability of resource scheduling in complex dynamic environments.

[0038] The above method, in S2, provides a detailed description of the forward-looking traffic prediction:

[0039] S21: To model the dynamic spatial correlation between nodes, in each time slot Next, taking nodes as the modeling object, the input features of that time slot are extracted. And through linear mapping, query vector, key vector, and value vector are generated to obtain spatial attention representation:

[0040]

[0041]

[0042] in, This represents the node input features under time slot t. , and These represent the query vector, key vector, and value vector in spatial attention, respectively. , and Indicates learnable parameters, This indicates spatial attention output. Indicates the dimension of attention projection;

[0043] S22: To model the temporal dependency of nodes across time slots, using the time dimension as the modeling object, a linear mapping is performed on the representation sequence of each node within the historical window to generate query vectors, key vectors, and value vectors, resulting in a temporal attention representation.

[0044]

[0045]

[0046] in, Represents a node Feature sequences within a historical time window, , and These represent the query vector, key vector, and value vector in temporal attention, respectively. , and Indicates learnable parameters, Indicates time-based attention output. Indicates the attention projection dimension;

[0047] S23: Perform feature enhancement processing on the spatial attention output and temporal attention output, and map the spatial branch and temporal branch through residual connection, layer normalization and feedforward network respectively to obtain spatiotemporal feature representation;

[0048] The spatial branching process for generating its features is as follows:

[0049]

[0050]

[0051] in, Indicates input features, This indicates spatial attention output. This represents the intermediate features in the space. This represents the enhanced spatial feature representation. This represents a feedforward neural network;

[0052] The feature generation process of time branch is as follows:

[0053]

[0054]

[0055] in, Indicates time-based attention output. This represents the intermediate features over time. This represents the enhanced temporal feature representation;

[0056] S24: In multi-node traffic scenarios, different nodes are not equally sensitive to spatial correlation and temporal dependence: the state changes of some nodes are more affected by neighborhood interactions, while some nodes exhibit strong autoregressive and periodic driving characteristics.

[0057] Therefore, the spatial and temporal feature representations are adaptively fused to achieve node-differentiated feature selection, thereby addressing the spatiotemporal heterogeneity problem of nodes. First, the graph generated by graph convolution operations is embedded... As a topological representation of nodes, it is also used as a spatial feature obtained from S23. With time characteristics Based on the above Generate a node query vector, and generate spatial key and time key vectors based on the spatial and temporal features respectively:

[0058] Q

[0059] ,

[0060] in , , These are the learnable parameters. and These are the spatial key and temporal key vectors, respectively. Then, based on the scaled dot product results between the node query vector and the spatial key vector and temporal key vector, the spatial branch score and temporal branch score are calculated respectively:

[0061] ,

[0062] in, This indicates the degree of matching between the current node and the spatial features. This indicates the degree to which the current node matches the time features.

[0063] The time and space matching scores are further normalized to obtain the corresponding fusion weights, which are calculated using the following formula:

[0064]

[0065] in, These represent the corresponding fusion weights. Based on the fusion weights corresponding to the spatial and temporal branches, the features of the two branches are weighted and aggregated to obtain the fused spatiotemporal feature representation, calculated using the following formula:

[0066]

[0067] in, This represents the spatiotemporal features after fusion.

[0068] S25: Input the fused spatiotemporal feature representation into the prediction layer, and output the node traffic prediction results for several future time slots to obtain a forward-looking network traffic representation. The calculation formula is as follows:

[0069]

[0070] in, This represents the forward-looking forecast of node traffic over several future time slots. This represents the linear mapping operation of the prediction layer.

[0071] In the above method, the state space construction in S3 is described in detail below:

[0072] S31: Obtain the current traffic observation value of the node and the traffic prediction results for several future time slots, where the node... In the time slot The current flow observation value is denoted as The corresponding future The flow prediction sequence for each time slot is represented as follows:

[0073]

[0074] in, Represents a node In the time slot The corresponding forward-looking flow forecast sequence, Represents a node In the future The predicted flow rate for each time slot, Indicates the prediction step size;

[0075] S32: Obtain the transmission queue status on the node's egress side and construct a queue status vector, represented as:

[0076]

[0077] in, Represents a node In the time slot The queue state vector, Represents a node In the time slot The The amount of data waiting to be sent in each logical queue Indicates the total number of logical queues;

[0078] S33: Obtain the total remaining data of all backup jobs currently on the node, to characterize the node's potential future load intensity, expressed as:

[0079]

[0080] in, Represents a node In the time slot The remaining amount of backup jobs, This indicates that node v is in the time slot. A collection of backup jobs, Indicate homework The corresponding remaining data size;

[0081] S34: Obtain the status of node storage resources, computing resources, and bandwidth resources, and update the current traffic observation values. The forward-looking flow forecast sequence obtained from S31 The queue state vector obtained by S32 The remaining amount of backup jobs obtained by S33 The node resource information is then concatenated to construct the local state of the node, which is represented as follows:

[0082]

[0083] in, , and Representing nodes respectively In the time slot The remaining storage resources, available computing resources, and available bandwidth resources;

[0084] S35: Concatenate the local states of each node according to the node dimension to construct the global state, represented as:

[0085]

[0086] in, Indicates time slot The global state under, This represents the local state of node v. This represents the total number of nodes in the network. The global state is used for subsequent value assessment and joint scheduling decisions.

[0087] To effectively inject the forward-looking prediction information and spatiotemporal dynamic representation output by STAF-Former in S2 into the node resource scheduling decision-making process and improve the collaborative scheduling capability in multi-node parallel backup scenarios, this invention constructs a multi-agent joint action generation mechanism in step S4. Each node is treated as a collaborative decision-making agent, and based on the local state obtained in S3, it retrieves contextual information related to the current decision from the spatiotemporal features output by STAF-Former using a cross-attention mechanism to generate backup job receiving actions and weighted fair queue bandwidth allocation actions. The node agents further aggregate to form global joint actions, achieving integrated collaborative control of node-side backup scheduling and bandwidth allocation.

[0088] In S4, the steps for constructing the combined backup scheduling and bandwidth allocation action described above are as follows:

[0089] S41: Generate decision query vectors based on the local state of nodes, and then... In the time slot Local state Input the mapping network to obtain the decision query vector:

[0090]

[0091] in, , Representing nodes respectively In the time slot The decision query vector and local state, Indicates a query mapping network;

[0092] S42: Considering the frequent traffic changes and complex node interactions in network task scheduling scenarios, relying solely on the current local observations of the executor can easily lead to short-sighted decision-making and suboptimal scheduling. Therefore, we utilize the feature retrieval and aggregation capabilities of cross-attention mechanisms to drive the executor to adaptively retrieve highly relevant spatiotemporal contexts, extract key environmental information, and make execution decisions.

[0093] Therefore, based on the decision query vector, cross-attention interaction is performed using spatiotemporal situational features to extract contextual information related to the current decision, resulting in a spatiotemporal context representation:

[0094]

[0095] in, This represents the spatiotemporal context of node v in time slot t. Represents nodes In the time slot The corresponding spatiotemporal feature sequence, , and Indicates learnable parameters, Represents the dimension of the latent space. For activation functions;

[0096] S43: Jointly represent the spatiotemporal context representation with the local state of the node, output the scoring result of the candidate action set, and obtain the action probability distribution through the activation function:

[0097]

[0098]

[0099] in, Represents a node In the time slot Action scoring vector, This represents the concatenation result of the node's local state and its spatiotemporal context representation. Represents a policy mapping network. Indicates candidate actions The probability of choosing, Indicates the total number of candidate actions;

[0100] S44: Based on the action probability distribution described in S44, determine the backup job receiving action and the weighted fair queue bandwidth weight configuration action for the node in the current time slot, respectively, as follows:

[0101]

[0102]

[0103] in, Represents a node In the time slot Homework The decision to receive, This represents the set of jobs for node v in time slot t. Represents a node In the time slot The queue weight configuration vector, Indicates the first The weight values ​​corresponding to each logical queue Indicates the total number of logical queues;

[0104] S45: To achieve coordinated control of pending queue orchestration and egress bandwidth share adjustment, the backup job receiving action and the weighted fair queue bandwidth weight configuration action are combined to obtain a joint node action, represented as:

[0105]

[0106] in, This indicates the action of node v in time slot t. This indicates that the backup job has been received. This indicates the action of configuring the weights of the weighted fair queue.

[0107] S46: Combine the actions of each node according to the node dimension to form a joint action, which is used for subsequent environmental interaction and collaborative scheduling control.

[0108]

[0109] in, Indicates time slot The joint action under, This represents the total number of nodes in the network.

[0110] To address the problem of lacking a globally consistent evaluation signal and difficulty in stably optimizing multi-agent strategies after joint actions are executed in multi-node parallel backup scenarios, this invention constructs a centralized state evaluation and network update mechanism based on global state, joint actions, and reward information in step S5. It aggregates local node rewards to form a global real-time reward and updates the policy network by combining multi-step reward objectives and advantage information, so as to achieve global collaborative optimization of backup scheduling and bandwidth allocation strategies.

[0111] In S5, the above method performs value assessment based on global state, joint actions, and reward information, and updates the policy network through multi-step rewards. The specific process is as follows:

[0112] S51: To enable agents to jointly optimize task timeliness and transmission quality, a local real-time report for each node is constructed based on the node's task progress, resource pressure status, and egress queue service quality in the current time slot, represented as:

[0113]

[0114] in, This represents the local instantaneous report of node v in time slot t. Indicates the benefits of work progress. This represents the service quality cost of the exit queue. and This represents the corresponding weighting coefficient;

[0115] S52: Aggregate the local real-time reports of each node to obtain the global real-time report for the current time slot, represented as:

[0116]

[0117] S53: Based on the aforementioned global instantaneous reward and subsequent state information, considering the distribution offset introduced by policy sampling and historical data reuse, to mitigate the inherent error of single-step temporal differencing and improve learning stability, a multi-step reward objective is constructed and a global state value assessment is performed, expressed as:

[0118]

[0119] in, Indicates time slot t Step-by-step return goals Indicates the discount factor. Indicates the first Global state under time slot, This represents the estimation result of the value assessment network for the global state. Indicates the network parameters for value assessment;

[0120] S54: Construct a value assessment loss based on the multi-step return objective and update the value assessment network, expressed as:

[0121]

[0122] in, Indicates a loss in value assessment. Indicates the value assessment network for time slots Next global state The estimation results;

[0123] S55: Calculate advantage information based on the value assessment results, and update the policy network accordingly to optimize backup scheduling and bandwidth allocation strategies, as shown below:

[0124]

[0125]

[0126] in, Represents a node In the time slot The advantage value below, Indicates the strategy loss. This indicates that node v is in a local state. Select action The probability, Represents a node Corresponding policy network parameters;

[0127] S56: Minimize via gradient descent Complete the policy network parameter update:

[0128]

[0129] in is the learning rate of the policy network.

[0130] Because the present invention employs the above-described technical means, it possesses the following beneficial effects:

[0131] 1. Unified Spatiotemporal Feature Embedding: By jointly encoding historical traffic data, topological association information, and time period information from multiple nodes, the original traffic features, spatial structure features, and temporal semantic features are embedded into a unified representation space, enabling a more comprehensive characterization of the node operating status in a multi-active disaster recovery network. This method effectively overcomes the shortcomings of traditional scheduling methods, which rely solely on local instantaneous information and fail to adequately utilize network spatiotemporal correlations, providing a more complete input foundation for subsequent traffic prediction and scheduling decisions.

[0132] 2. Efficient Spatiotemporal Dependency Joint Modeling: By separately modeling the spatial relationships between nodes and the temporal evolution relationships across time slots, and then adaptively fusing them, the model can more accurately reflect the load propagation and resource contention patterns in multi-node parallel backup scenarios, thereby improving its adaptability to complex dynamic network environments.

[0133] 3. High-precision forward-looking traffic prediction: Based on the spatiotemporal feature modeling results, the node traffic changes in several future time slots are predicted to obtain a forward-looking traffic characterization. Compared with scheduling based solely on the current state, this method can detect potential congestion and load change trends in advance, providing a predictive basis for backup job scheduling and bandwidth allocation, thereby enhancing the foresight and proactivity of scheduling decisions.

[0134] 4. Enhanced Multi-Node Collaborative Optimization Capability: By constructing a value assessment mechanism oriented towards the global state, a unified evaluation and collaborative optimization of the scheduling behavior of each node is achieved in multi-node parallel backup scenarios. This method can effectively alleviate the local optimum problem caused by independent decision-making by a single node, reduce the risk of resource contention and queue congestion, and enhance the stability and adaptability of the system in strongly coupled, highly dynamic network environments.

[0135] 5. Significantly Improved Overall Performance: Through the above technical solutions, this invention achieves proactive perception of node load changes in multi-active disaster recovery networks, dynamic response to resource status changes, and collaborative optimization of the multi-node concurrent backup process. Compared with traditional methods, this method can improve bandwidth allocation stability while ensuring backup scheduling timeliness, reduce performance losses caused by queue backlog and resource contention, thereby improving system throughput, resource utilization efficiency, and operational stability.

[0136] 6. Network spatiotemporal prediction (steps S1-S2) focuses on traffic load change estimation, while backup scheduling and bandwidth allocation decision-making based on multi-agent reinforcement learning (S3-S5) focuses on policy optimization based on current network observations. There is a lack of a technical mechanism to effectively transform forward-looking prediction information into joint scheduling actions and further use it for global policy updates. This invention does not simply use prediction models and reinforcement learning models side-by-side, but rather injects forward-looking prediction information and spatiotemporal dynamic representations output by STAF-Former into the multi-agent joint action generation process (S4) through a cross-attention mechanism. That is, each agent can "efficiently retrieve" spatiotemporal information with high correlation to corresponding nodes through cross-attention, reducing the interference of irrelevant information on decision-making, achieving high-quality transformation of prediction information into reference information guiding agent decisions, and further combining global state, joint actions, and multi-step rewards for centralized value evaluation and policy updates (S5).

[0137] 7. Unlike other existing technologies that limit the use of prediction metrics as optimization targets or merely inject them into the state space for concatenation, this invention enables the prediction module to directly guide action generation, allowing the reinforcement learning agent to focus on the highly relevant spatiotemporal context of the network and efficiently utilize future load trends to conduct cross-node collaborative optimization.

[0138] In summary, this invention achieves collaborative optimization of backup scheduling and bandwidth allocation for multi-active disaster recovery networks by introducing unified embedding of spatiotemporal features, joint modeling of spatiotemporal dependencies, forward-looking load prediction, and a joint scheduling mechanism driven by reinforcement learning. It can effectively improve resource management capabilities and system performance in complex dynamic environments, and has strong advanced value and engineering application value. Attached Figure Description

[0139] Figure 1 A flowchart illustrating a backup scheduling and bandwidth allocation method for multi-active disaster recovery networks, provided for the implementation of this invention;

[0140] Figure 2 This is an overview diagram of backup scheduling and bandwidth allocation in an example of the present invention;

[0141] Figure 3 This is a structural diagram of the spatiotemporal prediction model established in an example of the present invention;

[0142] Figure 4 The reinforcement learning model structure diagram is established in the example of this invention. Detailed Implementation

[0143] The embodiments of the present invention will be described in detail below. Although the present invention will be described and illustrated in conjunction with some specific embodiments, it should be noted that the present invention is not limited to these embodiments. On the contrary, any modifications or equivalent substitutions made to the present invention should be covered within the scope of the claims of the present invention.

[0144] Furthermore, to better illustrate the present invention, numerous specific details are set forth in the following detailed embodiments. Those skilled in the art will understand that the present invention can be practiced without these specific details.

[0145] This invention proposes a backup scheduling and bandwidth allocation method for multi-active disaster recovery networks, comprising:

[0146] S1: Read historical traffic data, topology association information and time period information of multiple nodes, preprocess the data and embed spatiotemporal features to obtain a unified spatiotemporal input representation;

[0147] S2: Perform spatial correlation modeling and temporal evolution modeling on the spatiotemporal input representation, extract spatiotemporal dynamic features, fuse the spatiotemporal dynamic features, predict the node traffic changes in several future time slots, and obtain a forward-looking load characterization.

[0148] S3: Based on the current traffic observation of the node, the forward load characterization, the egress queue status, the remaining amount of backup jobs, and the node's storage resources, computing resources, and bandwidth resources, construct the node's local state, and then combine the local states of each node to form the global state.

[0149] S4: Based on the local and global states of the node, extract context information related to the current scheduling decision from the forward-looking load representation, and generate backup job receiving action and weighted fair queue bandwidth weight configuration action according to the context information, thereby forming a joint action;

[0150] S5: Interact with the network environment to obtain immediate feedback and the next time slot observation results; perform value assessment and policy update based on the global state, joint actions and immediate feedback to achieve coordinated optimization of backup scheduling and bandwidth allocation.

[0151] In the above method, the feature extraction in S1 is described in detail below:

[0152] S11: Read historical traffic data from multiple nodes and construct a historical input traffic tensor according to the historical time window length, number of nodes, and traffic feature dimensions, denoted as... ,in, Indicates the length of the history window. Indicates the number of nodes. This represents the traffic characteristic dimension of each node;

[0153] Simultaneously read network topology information and construct a topology graph. and its adjacency matrix ,in, Represents a set of network nodes. Represents a set of links;

[0154] S12: Embed the original traffic features into the historical input traffic tensor, and project the original traffic data onto a unified latent space through linear mapping to obtain the original traffic feature representation:

[0155]

[0156] in, For linear embedding operations, Indicates the embedded spatial dimension;

[0157] S13: Based on the network topology information, spatial feature embedding is performed to aggregate neighboring node information while preserving the node's own information during feature propagation, and to optimize the adjacency matrix. Introduce node self-connections and normalize the adjacency matrix after introducing self-connections:

[0158] ,

[0159] in, express 3D identity matrix Indicates and The corresponding degree matrix. Then, using the initial features of the nodes. As input, topological relationship features between nodes are extracted through graph convolution operations to obtain a spatial embedding representation:

[0160]

[0161] in, Indicates the number of nodes. and This represents the learnable weight parameters. and Indicates the bias parameter. Represents spatial embedding features, Indicates the activation function;

[0162] S14: Based on the time period information, perform time feature embedding, and encode the weekday index and intraday time slot index corresponding to each time step using one-hot encoding, and concatenate them to obtain the time feature vector:

[0163]

[0164] in, Indicates a time step. Indicates time step The corresponding weekday feature vector, Indicates time step The corresponding intraday time slot feature vector, This represents a vector concatenation operation, which stacks the temporal feature vectors of all time steps to form a concatenation vector. The temporal embedding representation is obtained through two fully connected layers:

[0165] ,

[0166] in, and For learnable weight parameters, and For bias parameters, This represents the activation function. Temporal embedding features;

[0167] S15: Expand and stack the spatial embedding features along the time dimension to obtain: And by aligning the temporal embedding features to all nodes, we get: ,in, For the expanded spatial embedding features, These are the aligned temporal embedding features.

[0168] The extended spatial embedding feature Aligned temporal embedding features Compared with the original flow characteristics representation By performing element-wise addition, a unified spatiotemporal input representation is obtained:

[0169]

[0170] in, To unify the spatiotemporal input representation.

[0171] The above method, in S2, provides a detailed description of the forward-looking traffic prediction:

[0172] S21: To model the dynamic spatial correlation between nodes, in each time slot Next, taking nodes as the modeling object, the input features of that time slot are extracted. And through linear mapping, query vector, key vector, and value vector are generated to obtain spatial attention representation:

[0173]

[0174]

[0175] in, This represents the node input features under time slot t. , and These represent the query vector, key vector, and value vector in spatial attention, respectively. , and Indicates learnable parameters, This indicates spatial attention output. Indicates the attention projection dimension;

[0176] S22: To model the time dependency of nodes across time slots, using the time dimension as the modeling object, a linear mapping is performed on the representation sequence of each node within the historical window to generate query vector, key vector, and value vector, resulting in a time attention representation.

[0177]

[0178]

[0179] in, Represents a node Feature sequences within a historical time window, , and These represent the query vector, key vector, and value vector in temporal attention, respectively. , and Indicates learnable parameters, Indicates time-based attention output. Indicates the attention projection dimension;

[0180] S23: Perform feature enhancement processing on the spatial attention output and temporal attention output, and map the spatial branch and temporal branch through residual connection, layer normalization and feedforward network respectively to obtain spatiotemporal feature representation;

[0181] The spatial branching process for generating its features is as follows:

[0182]

[0183]

[0184] in, Indicates input features, This indicates spatial attention output. This represents the intermediate features in the space. This represents the enhanced spatial feature representation. This represents a feedforward neural network;

[0185] The feature generation process of time branch is as follows:

[0186]

[0187]

[0188] in, Indicates time-based attention output. This represents the intermediate features over time. This represents the enhanced temporal feature representation;

[0189] S24: In multi-node traffic scenarios, different nodes are not equally sensitive to spatial correlation and temporal dependence: the state changes of some nodes are more affected by neighborhood interactions, while some nodes exhibit strong autoregressive and periodic driving characteristics.

[0190] Therefore, the spatial and temporal feature representations are adaptively fused to achieve node-differentiated feature selection, thereby addressing the spatiotemporal heterogeneity problem of nodes. First, the graph generated by graph convolution operations is embedded... As a topological representation of nodes, it is also used as a spatial feature obtained from S23. With time characteristics Based on the above Generate a node query vector, and generate spatial key and time key vectors based on the spatial and temporal features respectively:

[0191] Q

[0192] ,

[0193] in , , These are the learnable parameters. and These are the spatial key and temporal key vectors, respectively. Then, based on the scaled dot product results between the node query vector and the spatial key vector and temporal key vector, the spatial branch score and temporal branch score are calculated respectively:

[0194] ,

[0195] in, This indicates the degree of matching between the current node and the spatial features. This indicates the degree to which the current node matches the time features.

[0196] The time and space matching scores are further normalized to obtain the corresponding fusion weights, which are calculated using the following formula:

[0197]

[0198] in, These represent the corresponding fusion weights. Based on the fusion weights corresponding to the spatial and temporal branches, the features of the two branches are weighted and aggregated to obtain the fused spatiotemporal feature representation, calculated using the following formula:

[0199]

[0200] in, This represents the spatiotemporal features after fusion.

[0201] S25: Input the fused spatiotemporal feature representation into the prediction layer, and output the node traffic prediction results for several future time slots to obtain the forward load characterization, the calculation formula of which is:

[0202]

[0203] in, This represents the forward-looking forecast of node traffic over several future time slots. This represents the linear mapping operation of the prediction layer.

[0204] In the above method, the state space construction in S3 is described in detail below:

[0205] S31: Obtain the current traffic observation value of the node and the traffic prediction results for several future time slots, where the node... In the time slot The current flow observation value is denoted as The corresponding future The flow prediction sequence for each time slot is represented as follows:

[0206]

[0207] in, Represents a node In the time slot The corresponding forward-looking flow forecast sequence, Represents a node In the future The predicted flow rate for each time slot, Indicates the prediction step size;

[0208] S32: Obtain the transmission queue status on the node's egress side and construct a queue status vector, represented as:

[0209]

[0210] in, Represents a node In the time slot The queue state vector, Represents a node In the time slot The The amount of data waiting to be sent in each logical queue Indicates the total number of logical queues;

[0211] S33: Obtain the total remaining data of all backup jobs currently on the node, to characterize the node's potential future load intensity, expressed as:

[0212]

[0213] in, Represents a node In the time slot The remaining amount of backup jobs, This indicates that node v is in the time slot. A collection of backup jobs, Indicate homework The corresponding remaining data size;

[0214] S34: Obtain the status of node storage resources, computing resources, and bandwidth resources, and update the current traffic observation values. The forward-looking flow forecast sequence obtained from S31 The queue state vector obtained by S32 The remaining amount of backup jobs obtained by S33 The node resource information is then concatenated to construct the local state of the node, which is represented as follows:

[0215]

[0216] in, , and Representing nodes respectively In the time slot The remaining storage resources, available computing resources, and available bandwidth resources;

[0217] S35: Concatenate the local states of each node according to the node dimension to construct the global state, represented as:

[0218]

[0219] in, Indicates time slot The global state under, This represents the local state of node v. This represents the total number of nodes in the network. The global state is used for subsequent value assessment and joint scheduling decisions.

[0220] In S4, the steps for constructing the combined backup scheduling and bandwidth allocation action described above are as follows:

[0221] S41: Generate decision query vectors based on the local state of nodes, and then... Input the mapping network to obtain the decision query vector:

[0222]

[0223] in, , Representing nodes respectively In the time slot The decision query vector and local state, Indicates a query mapping network;

[0224] S42: Considering the frequent traffic changes and complex node interactions in network task scheduling scenarios, relying solely on the current local observations of the executor can easily lead to short-sighted decision-making and suboptimal scheduling. Therefore, the feature retrieval and aggregation capabilities of cross-attention mechanisms are utilized to drive the executor to adaptively retrieve highly relevant spatiotemporal contexts, extract key environmental information, and make execution decisions.

[0225] Therefore, based on the decision query vector, cross-attention interaction is performed using spatiotemporal situational features to extract contextual information related to the current decision, resulting in a spatiotemporal context representation:

[0226]

[0227] in, This represents the spatiotemporal context of node v in time slot t. Represents nodes In the time slot The corresponding spatiotemporal feature sequence, , and Indicates learnable parameters, Represents the dimension of the latent space. For activation functions;

[0228] S43: Jointly represent the spatiotemporal context representation with the local state of the node, output the scoring result of the candidate action set, and obtain the action probability distribution through the activation function:

[0229]

[0230]

[0231] in, Represents a node In the time slot Action scoring vector, This represents the concatenation result of the node's local state and its spatiotemporal context representation. Represents a policy mapping network. Indicates candidate actions The probability of choosing, Indicates the total number of candidate actions;

[0232] S44: Based on the action probability distribution described in S44, determine the backup job receiving action and the weighted fair queue bandwidth weight configuration action for the node in the current time slot, respectively, as follows:

[0233]

[0234]

[0235] in, Represents a node In the time slot Homework The decision to receive, Represents a node In the time slot The set of assignments, Represents a node In the time slot The queue weight configuration vector, Indicates the first The weight values ​​corresponding to each logical queue Indicates the total number of logical queues;

[0236] S45: To achieve coordinated control of pending queue orchestration and egress bandwidth share adjustment, the backup job receiving action and the weighted fair queue bandwidth weight configuration action are combined to obtain a joint node action, represented as:

[0237]

[0238] in, This indicates the action of node v in time slot t. This indicates that the backup job has been received. This indicates the action of configuring the weights of the weighted fair queue.

[0239] S46: Combine the actions of each node according to the node dimension to form a joint action, which is used for subsequent environmental interaction and collaborative scheduling control.

[0240]

[0241] in, Indicates time slot The joint action under, This represents the total number of nodes in the network.

[0242] In S5, the above method performs value assessment based on global state, joint actions, and reward information, and updates the policy network through multi-step rewards. The specific process is as follows:

[0243] S51: To enable the agent to jointly optimize task timeliness and transmission quality, a local real-time report of the node is constructed based on the node's task progress in the current time slot and the service quality of the exit queue, as shown below:

[0244]

[0245] in, This represents the local instantaneous report of node v in time slot t. Indicates the benefits of work progress. This represents the service quality cost of the exit queue. and This represents the corresponding weighting coefficient;

[0246] S52: Aggregate the local real-time reports of each node to obtain the global real-time report for the current time slot, represented as:

[0247]

[0248] S53: Based on the aforementioned global instantaneous reward and subsequent state information, considering the distribution offset introduced by policy sampling and historical data reuse, to mitigate the inherent error of single-step temporal differencing and improve learning stability, a multi-step reward objective is constructed and a global state value assessment is performed, expressed as:

[0249]

[0250] in, Indicates time slot t Step-by-step return goals Indicates the discount factor. Indicates the first Global state under time slot, This represents the estimation result of the value assessment network for the global state. Indicates the network parameters for value assessment;

[0251] S54: Construct a value assessment loss based on the multi-step return objective and update the value assessment network, expressed as:

[0252]

[0253] in, Indicates a loss in value assessment. Indicates the value assessment network for time slots Next global state The estimation results;

[0254] S55: Calculate advantage information based on the value assessment results, and update the policy network accordingly to optimize backup scheduling and bandwidth allocation strategies, as shown below:

[0255]

[0256]

[0257] in, Represents a node In the time slot The advantage value below, Indicates the strategy loss. This indicates that node v is in a local state. Select action The probability, Represents a node Corresponding policy network parameters;

[0258] S56: Minimize via gradient descent Complete the policy network parameter update:

[0259]

[0260] in is the learning rate of the policy network.

[0261] Example

[0262] 1. Experimental Dataset Processing Instructions

[0263] This paper selects the GEANT network topology and its corresponding public traffic matrix dataset GEANT-TM as experimental data. The GEANT topology contains 23 nodes and 37 links, which can well reflect the cross-node communication relationships in complex backbone networks. GEANT-TM records the traffic changes between node pairs in the form of an end-to-end traffic matrix sequence, with a sampling interval of 15 minutes and a time span of 4 months, containing a total of 10769 records. Based on this, a multi-service dynamic competition environment is constructed to simulate the different transmission ratios of multiple services at the same time, so as to evaluate the adaptability of the scheduling method under complex service ratio changes, sudden traffic fluctuations, and resource-constrained conditions under simulation.

[0264] 2. Simulation Environment Setup Instructions

[0265] The experiment consisted of two phases: predictive model training and scheduling policy evaluation. Predictive model training was conducted on an Ubuntu 22.04 platform using Python 3.10, CUDA 12.1, and PyTorch 2.1.0, with an 18-core AMD EPYC9754 processor and an NVIDIA RTX 4090 GPU. The scheduling evaluation phase was carried out on an SDN simulation platform. The control plane used the Ryu controller, and the data plane used Mininet 2.3.0 and OpenvSwitch, communicating via the OpenFlow 1.3 protocol. Service traffic was injected via iperf3, and the experimental platform's GPU was a GeForce RTX 3090.

[0266] 3. Evaluation Indicators Explanation

[0267] To comprehensively evaluate the effectiveness of the backup scheduling and bandwidth allocation method for multi-active disaster recovery networks proposed in this invention, the experiment sets evaluation indicators from two aspects: spatiotemporal prediction performance and scheduling optimization performance. In the spatiotemporal prediction task, root mean square error (RMSE) and mean absolute error (MAE) are used as evaluation indicators; in the scheduling optimization task, average round reward (Reward), job throughput (Thr), resource pressure cost (Press), and weighted fair queue cost (Cost) are used as evaluation indicators. The calculation formulas for each indicator are as follows:

[0268]

[0269]

[0270] in, Represents the true value. Indicates the predicted value. This represents the total number of samples. The smaller the values ​​of RMSE and MAE, the smaller the error between the predicted result and the true value, and the higher the spatiotemporal prediction accuracy.

[0271] In scheduling optimization tasks, job throughput is defined as:

[0272]

[0273] in, This represents the set of backup jobs for node v in time slot t. Indicates the dynamic priority of the task. This indicates whether the job is in the execution state during time slot t.

[0274] The cost of a weighted fair queue (WFQ) is defined as: in, , and Representing nodes respectively In the time slot The Queue backlog, average queuing latency, and packet loss. , and These are the corresponding weighting coefficients. The smaller the value, the better the outbound bandwidth allocation strategy is in reducing queue congestion, queuing delay, and packet loss.

[0275] The average convergence reward is defined as:

[0276] in, and Weighting coefficients The larger the value, the more effectively the proposed method can reduce the cost of the weighted fair queue while improving the efficiency of backup job execution, thus achieving a better overall scheduling effect.

[0277] 4. Test Instructions

[0278] This invention discloses a backup scheduling and bandwidth allocation method for multi-active disaster recovery networks. The core of this invention lies in combining spatiotemporal feature embedding, spatiotemporal dependency modeling, and forward-looking load prediction with a reinforcement learning-driven joint decision-making mechanism. This integrates current traffic observations, future traffic predictions, queue status, remaining backup job capacity, and node resource information into a unified state space. This allows the scheduling agent to dynamically generate backup job receiving actions and weighted fair queue weight configuration actions, achieving collaborative optimization of resource allocation in multi-node parallel backup scenarios. Compared to traditional methods, this invention can more accurately characterize the spatiotemporal evolution of node load, improve the forward-looking and adaptive capabilities of scheduling decisions, and enhance the synergy between backup job scheduling and bandwidth allocation, thereby effectively alleviating resource contention and queue congestion problems. This verification test is mainly used to simulate the actual application scenario of multi-node parallel backup and online service concurrent transmission in a multi-active disaster recovery network. It verifies whether the present invention can still effectively perceive future traffic changes and efficiently coordinate and schedule node resources under conditions of sudden load fluctuations, dynamic changes in node resources, and multi-service competition, thereby achieving better overall performance than traditional methods, namely higher prediction accuracy, stronger backup job advancement capability, lower resource pressure and queue service cost, as well as better system throughput and operational stability.

[0279] I. Performance Testing and Result Analysis

[0280] In the experiment, the spatiotemporal prediction part used 12 historical time slots as input windows and set a prediction step size of 4 time slots; the reinforcement learning part was trained for 1000 rounds in a static topology simulation environment and repeated experiments with 10 independent random seeds to evaluate the average performance and stability of the method.

[0281] To verify the effectiveness of the method of this invention, the proposed spatiotemporal prediction model STAF-Former is compared with typical prediction models such as ARIMA, LSTM, TCN, ATGCN, STGCN, and PDFormer in spatiotemporal prediction tasks; and the proposed reinforcement learning method STA2C is compared with classic multi-agent reinforcement learning methods such as IQL, VDN, and MAPPO in node-side scheduling tasks. Through the above comparisons, the superiority of the method of this invention can be comprehensively evaluated from the aspects of prediction accuracy and scheduling performance.

[0282] Table 1 shows that the proposed method exhibits superior performance in spatiotemporal prediction on the GEANT dataset. At four time-slot prediction steps, its MAE and RMSE are 5.42 and 12.08, respectively. Compared to the suboptimal PDFormer method (6.02 and 12.97), STAF-Former reduces MAE by approximately 10.0% and RMSE by approximately 6.9%; compared to STGCN (8.10 and 16.43), its MAE and RMSE are reduced by approximately 33.1% and 26.5%, respectively. This indicates that the proposed STAF-Former can more accurately extract and fuse the spatiotemporal features of network traffic, thereby significantly improving the prediction accuracy at a 60-minute step size.

[0283] Table 1 Comparison of evaluation metrics for different models in predicting step size at 60 min

[0284] Model MAERMSE ARIMA 13.54 24.06 LSTM 11.66 22.27 TCN 9.11 21.83 ATGCN 8.46 16.98 STGCN 8.10 16.43 PDFormer 6.02 12.97 STAF-Former (This invention) 5.42 12.08

[0285] Table 2 shows that STA2C achieved the highest overall reward value (46.9) in node-side scheduling tasks, surpassing MAPPO's 45.0, VDN's 39.9, and IQL's 38.6, indicating superior overall scheduling performance. Further analysis of throughput and WFQ cost reveals that STA2C has the highest throughput (58.516) and the lowest WFQ cost (19.360), demonstrating its ability to effectively reduce queue service costs while improving backup job execution efficiency. In contrast, while MAPPO boasts higher throughput, its WFQ cost is also higher; IQL and VDN show significant differences in both throughput and overall reward. In summary, STA2C achieves a better balance between job execution benefits and queue costs, validating its effectiveness and superiority in node-side scheduling tasks.

[0286] Table 2 Comparison of Criticism Indicators for Node-Side Backup Scheduling and WFQ Bandwidth Allocation in Different Models

[0287] method Reward Value Task throughput WFQ Cost IQL 38.6 51.58 21.64 VDN 39.9 52.73 21.4 MAPPO 45.0 59.38 22.47 STA2C (This invention) 46.9 58.51 19.36

[0288] III. Test Summary

[0289] This section presents performance tests of the invention. Comparative analysis of spatiotemporal prediction experiments and node-side scheduling experiments verifies the rationality and effectiveness of the proposed backup scheduling and bandwidth allocation method for multi-active disaster recovery networks. Experimental results show that the invention achieves superior performance in both network traffic prediction accuracy and node collaborative scheduling performance. It effectively reduces queue service costs while improving backup job execution efficiency, and effectively addresses the problems of insufficient perception of spatiotemporal traffic changes, weak scheduling coordination, and inflexible bandwidth allocation in traditional methods under complex network environments.

Claims

1. A backup scheduling and bandwidth allocation method for multi-active disaster recovery networks, characterized in that, Includes the following steps: S1: Read historical traffic data, topology association information and time period information of multiple nodes, preprocess the data and embed spatiotemporal features to obtain a unified spatiotemporal input representation; S2: Perform spatial correlation modeling and temporal evolution modeling on the spatiotemporal input representation, extract spatiotemporal dynamic features, fuse the spatiotemporal dynamic features, predict the node traffic changes in several future time slots, and obtain a forward-looking load characterization. S3: Based on the current traffic observation of the node, the forward load characterization, the egress queue status, the remaining amount of backup jobs, and the node's storage resources, computing resources, and bandwidth resources, construct the node's local state, and then combine the local states of each node to form the global state. S4: Based on the local and global states of the node, extract context information related to the current scheduling decision from the forward-looking load representation, and generate backup job receiving action and weighted fair queue bandwidth weight configuration action according to the context information, thereby forming a joint action; S5: Interact with the network environment to obtain immediate feedback and the next time slot observation results; perform value assessment and policy update based on the global state, joint actions and immediate feedback to achieve coordinated optimization of backup scheduling and bandwidth allocation.

2. The backup scheduling and bandwidth allocation method for multi-active disaster recovery networks according to claim 1, characterized in that, In S1, the feature extraction is described in detail as follows: S11: Read historical traffic data from multiple nodes and construct a historical input traffic tensor according to the historical time window length, number of nodes, and traffic feature dimensions, denoted as... ,in, Indicates the length of the history window. Indicates the number of nodes. This represents the traffic characteristic dimension of each node; Simultaneously read network topology information and construct a topology graph. and its adjacency matrix ,in, Represents a set of network nodes. Represents a set of links; S12: Embed the original traffic features into the historical input traffic tensor, and project the original traffic data onto a unified latent space through linear mapping to obtain the original traffic feature representation: in, For linear embedding operations, Indicates the embedded spatial dimension; S13: Based on the network topology information, spatial feature embedding is performed to aggregate neighboring node information while preserving the node's own information during feature propagation, and to optimize the adjacency matrix. Introduce node self-connections and normalize the adjacency matrix after introducing self-connections: , in, express 3D identity matrix Indicates and The corresponding degree matrix, and then the initial features of the nodes. As input, topological association features between nodes are extracted through graph convolution operations to obtain a spatial embedding representation: in, Indicates the number of nodes. Indicates the initial feature dimension of the node. and This represents the learnable weight parameters. and Indicates the bias parameter. Represents spatial embedding features, Indicates the activation function; S14: Time feature embedding is performed based on time period information. The weekday index and intraday time slot index corresponding to each time step are one-hot encoded and concatenated to obtain the time feature vector: in, Indicates a time step. Indicates time step The corresponding weekday feature vector, Indicates time step The corresponding intraday time slot feature vector, This represents a vector concatenation operation, which stacks the temporal feature vectors of all time steps to form a concatenation vector. The temporal embedding representation is obtained through two fully connected layers: in, and For learnable weight parameters, and For bias parameters, This represents the activation function. For temporal embedding features; S15: Expand and stack the spatial embedding features along the time dimension to obtain the expanded spatial embedding features. Align the temporal embedding features to all nodes to obtain the aligned temporal embedding features. ; The extended spatial embedding feature Aligned temporal embedding features Compared with the original flow characteristics representation By performing element-wise addition, a unified spatiotemporal input representation is obtained: in, To unify the spatiotemporal input representation.

3. The backup scheduling and bandwidth allocation method for multi-active disaster recovery networks according to claim 1, characterized in that, In S2, the detailed description of forward-looking traffic forecasting is as follows: S21: To model the dynamic spatial correlation between nodes, in each time slot Next, taking nodes as the modeling object, the input features of that time slot are extracted. And through linear mapping, query vector, key vector, and value vector are generated to obtain spatial attention representation: in, This represents the node input features in time slot t. , and These represent the query vector, key vector, and value vector in spatial attention, respectively. , and Indicates learnable parameters, This indicates spatial attention output. Indicates the attention projection dimension; S22: To model the temporal dependency of nodes across time slots, using the time dimension as the modeling object, a linear mapping is performed on the representation sequence of each node within the historical window to generate query vectors, key vectors, and value vectors, resulting in a temporal attention representation. in, Represents a node Feature sequences within a historical time window, , and These represent the query vector, key vector, and value vector in temporal attention, respectively. , and Indicates learnable parameters, Indicates time-based attention output. Indicates the attention projection dimension; S23: Perform feature enhancement processing on the spatial attention output and temporal attention output, and map the spatial branch and temporal branch through residual connection, layer normalization and feedforward network respectively to obtain spatiotemporal feature representation; The spatial branching process for generating its features is as follows: in, Indicates input features, This indicates spatial attention output. This represents the intermediate features in the space. This represents the enhanced spatial feature representation. This represents a feedforward neural network; The feature generation process of time branch is as follows: in, Indicates time-based attention output. This represents the intermediate features over time. This represents the enhanced temporal feature representation; S24: Adaptively fuse the spatial feature representation and the temporal feature representation to achieve node-differentiated feature selection: First, embed the graph generated by the graph convolution operation. As a topological representation of nodes, it is also used as a spatial feature obtained from S23. With time characteristics ; Based on the above Generate a node query vector Q, and generate spatial key and time key vectors based on the spatial and temporal features respectively: Q , in , , These are the learnable parameters. and These are the spatial key and time key vectors, respectively. Then, based on the scaled dot product results between the node query vector and the spatial key vector and time key vector, the spatial branch score and time branch score are calculated respectively. ， in, This indicates the degree of matching between the current node and the spatial features. This indicates the degree to which the current node matches the time features; The time and space matching scores are further normalized to obtain the corresponding fusion weights, which are calculated using the following formula: in, and These represent the corresponding fusion weights; Based on the fusion weights corresponding to the spatial and temporal branches, the features of the two branches are weighted and aggregated to obtain the fused spatiotemporal feature representation, the calculation formula of which is: in, This represents the spatiotemporal features after fusion. S25: Input the fused spatiotemporal feature representation into the prediction layer, and output the node traffic prediction results for several future time slots to obtain the forward load characterization, the calculation formula of which is: in, This represents the forward-looking forecast of node traffic over several future time slots. This represents the linear mapping operation of the prediction layer.

4. The backup scheduling and bandwidth allocation method for multi-active disaster recovery networks according to claim 1, characterized in that, In S3, the state space construction is described in detail below: S31: Obtain the current traffic observation value of the node and the traffic prediction results for several future time slots, where the node... In the time slot The current flow observation value is denoted as The corresponding future The flow prediction sequence for each time slot is represented as follows: in, Represents a node In the time slot The corresponding forward-looking flow forecast sequence, Represents a node In the future The predicted flow rate for each time slot, Indicates the prediction step size; S32: Obtain the transmission queue status on the node's egress side and construct a queue status vector, represented as: in, Represents a node In the time slot The queue state vector, Represents a node In the time slot The The amount of data waiting to be sent in each logical queue Indicates the total number of logical queues; S33: Obtain the total remaining data of all backup jobs currently on the node, to characterize the node's potential future load intensity, expressed as: in, Represents a node In the time slot The remaining amount of backup jobs, This indicates that node v is in the time slot. A collection of backup jobs, Indicate homework The corresponding remaining data size; S34: Obtain the status of node storage resources, computing resources, and bandwidth resources, and update the current traffic observation values. The forward-looking flow forecast sequence obtained from S31 The queue state vector obtained by S32 The remaining amount of backup jobs obtained by S33 The node resource information is then concatenated to construct the local state of the node, represented as follows: in, , and Representing nodes respectively In the time slot The remaining storage resources, available computing resources, and available bandwidth resources; S35: Concatenate the local states of each node according to the node dimension to construct the global state, represented as: in, Indicates time slot The global state under, This represents the local state of node v. This represents the total number of nodes in the network. The global state is used for subsequent value assessment and joint scheduling decisions.

5. The backup scheduling and bandwidth allocation method for multi-active disaster recovery networks according to claim 1, characterized in that, In S4, the steps for constructing the joint action of backup scheduling and bandwidth allocation are as follows: S41: Generate decision query vectors based on node local states, and then... In the time slot Local state Input the mapping network to obtain the decision query vector: in, , Representing nodes respectively In the time slot The decision query vector and local state, Indicates a query mapping network; S42: Based on the decision query vector, perform cross-attention interaction with spatiotemporal situation features to extract contextual information related to the current decision, and obtain a spatiotemporal context representation: in, This represents the spatiotemporal context of node v in time slot t. Represents nodes In the time slot The corresponding spatiotemporal feature sequence, , and Indicates learnable parameters, Represents the dimension of the latent space. For activation functions; S43: Jointly represent the spatiotemporal context representation with the local state of the node, output the scoring result of the candidate action set, and obtain the action probability distribution through the activation function: in, Represents a node In the time slot Action scoring vector, This represents the concatenation result of the node's local state and its spatiotemporal context representation. Represents a policy mapping network. Indicates candidate actions The probability of choosing, Indicates the total number of candidate actions; S44: Based on the action probability distribution described in S44, determine the backup job receiving action and the weighted fair queue bandwidth weight configuration action for the node in the current time slot, respectively, as follows: in, Represents a node In the time slot Homework The decision to receive, This represents the set of jobs for node v in time slot t. Represents a node In the time slot The queue weight configuration vector, Indicates the first The weight values ​​corresponding to each logical queue Indicates the total number of logical queues; S45: To achieve coordinated control of pending queue orchestration and egress bandwidth share adjustment, the backup job receiving action and the weighted fair queue bandwidth weight configuration action are combined to obtain a joint node action, represented as: in, This indicates the action of node v in time slot t. This indicates that the backup job has been received. This indicates the action of configuring the weights of the weighted fair queue. S46: Combine the actions of each node according to the node dimension to form a joint action, which is used for subsequent environmental interaction and collaborative scheduling control. in, Indicates time slot The joint action under, This represents the total number of nodes in the network.

6. The backup scheduling and bandwidth allocation method for multi-active disaster recovery networks according to claim 1, characterized in that, In S5, the specific process of value assessment based on global state, joint actions, and reward information, and the multi-step reward update strategy, is as follows: S51: Based on the node's job progress and exit queue service quality in the current time slot, construct a local real-time report for the node, represented as follows: in, This represents the local instantaneous report of node v in time slot t. Indicates the benefits of work progress. This represents the service quality cost of the exit queue. and This represents the corresponding weighting coefficient; S52: Aggregate the local real-time reports of each node to obtain the global real-time report for the current time slot, represented as: S53: Based on the aforementioned global instantaneous reward and subsequent state information, construct a multi-step reward objective and perform a global state value assessment, expressed as: in, Indicates time slot t Step-by-step return goals Indicates the discount factor. Indicates the first Global state under time slot, This represents the estimation result of the value assessment network for the global state. Indicates the network parameters for value assessment; S54: Construct a value assessment loss based on the multi-step return objective and update the value assessment network, expressed as: in, Indicates a loss in value assessment. Indicates the value assessment network for time slots Next global state The estimation results; S55: Calculate advantage information based on the value assessment results, and update the policy network accordingly to optimize backup scheduling and bandwidth allocation strategies, as shown below: in, Represents a node In the time slot The advantage value below, Indicates strategy loss. This indicates that node v is in a local state. Select action The probability, Represents a node Corresponding policy network parameters; S56: Minimize via gradient descent Complete the policy network parameter update: in For the policy network learning rate, express Node policy network parameters The gradient.

Images