A cloud platform-based intelligent park resource management method and system

Abstract

This application relates to the field of cloud computing and smart park operation and maintenance management technology, and discloses a smart park resource management method and system based on a cloud platform. The method includes: acquiring data such as hazard sources and environmental monitoring time series within the target area to construct a heterogeneous time series graph; determining the risk parameters of computing nodes and the business survival budget vector based on the heterogeneous time series graph to form a risk state tensor; acquiring operation and maintenance constraints and generating prohibited migration edges and risk weight matrices based on hazard arrival times to form a propagation constraint attention mask; filtering and generating a safe candidate computing node sub-pool based on the mask and redundancy conditions; determining risk penalties based on the tensor and constructing an objective function based on interruption losses, and solving to generate a scheduling action strategy set; executing cross-domain migration of business instances based on the strategy set and triggering physical domain collaborative actions. This application realizes adaptive resource scheduling combining virtual and physical environments under sudden disasters, improving the continuity of critical business operations in the park.

Description

Technical Field

[0001] This application relates to the fields of cloud computing and smart park operation and maintenance management technology, and in particular to a smart park resource management method and system based on a cloud platform. Background Technology

[0002] With the deep integration of cloud computing, the Internet of Things, and artificial intelligence technologies, smart parks have become the standard form of modern urban industry and digital management. Their core business systems are highly dependent on the overall planning and scheduling of computing resources on the underlying cloud platform. Currently, under normal operating conditions, the cloud infrastructure of smart parks typically adopts a static scheduling strategy based on resource utilization or network load balancing to achieve cost reduction and efficiency improvement in computing power.

[0003] However, a smart park is not merely an information network node, but a high-density physical space containing a large amount of electromechanical equipment, chemical raw materials, and complex personnel activities. When facing sudden physical disasters such as fires, toxic gas leaks, or power outages, ensuring the continuity of core business operations within the park is paramount. This necessitates that cloud platform resource management strategies possess both timely disaster awareness and accurate business redirection capabilities. However, existing cloud platform resource scheduling mechanisms exhibit significant limitations when responding to sudden physical disasters.

[0004] Specifically, traditional virtualization scheduling systems have long been disconnected from underlying data center infrastructure such as environmental monitoring, power distribution, and fire protection systems, lacking a unified spatiotemporal mapping model. This prevents the system from perceiving the spatial spread path and evolution speed of physical disasters in real time. When danger occurs, the system often only passively responds to network or power outage alarms. Triggering cross-fault domain migration of services at this point usually results in data loss and service crashes due to missing the optimal escape window. Furthermore, most existing disaster recovery scheduling algorithms simply pursue resource idle rates, ignoring the constraints of cascading failures of electromechanical equipment caused by disasters and the mutual exclusion constraints of high availability for services themselves. This means that the calculated migration target node may be located in a dangerous area about to be affected by the disaster or facing a power bottleneck, leading to blind scheduling and serious consequences of secondary downtime. Even when some systems introduce basic risk warning mechanisms, they often rely on unidirectional open-loop static assessments, failing to fully consider the chaotic and unpredictable nature of disaster evolution in the physical world. Once the wind direction changes abruptly or the fire spreads rapidly in the real environment, deviating from the initial prediction model, the existing scheduling strategies become completely ineffective. Meanwhile, under extremely severe conditions where the scale of a disaster far exceeds expectations, and conventional cluster resources are completely destroyed, existing cloud platform architectures generally lack ultimate computing power fallback measures with multiple physical and electrical isolations, as well as reverse-order anti-shake rollback mechanisms. This disconnect between the virtual and physical worlds, and the lack of closed-loop verification and bottom-line defense, severely restricts the disaster resilience and self-healing capabilities of smart park cloud platforms.

[0005] Therefore, there is an urgent need for a dynamic resource management method that can break down cross-domain barriers, integrate disaster simulation, and possess adaptive correction and extreme fallback capabilities. Summary of the Invention

[0006] To address the aforementioned technical issues, this application provides a cloud-based smart park resource management method and system. This system is designed to resolve the problem of blind resource scheduling and business failures caused by the disconnect between physical space evolution and information domain scheduling when smart parks face sudden physical disasters. It achieves adaptive disaster recovery resource scheduling that combines virtual and physical elements and provides a safety net for business continuity.

[0007] Firstly, this application provides a smart park resource management method based on a cloud platform, the method comprising: Step S1: Obtain data on hazard sources, environmental monitoring time series, target business dependencies, and computing node deployment within the target area, and construct a heterogeneous time series diagram; Step S2: Based on the heterogeneous timing diagram, determine the danger arrival time interval, failure probability and isolation status of each computing node, and generate a business survival budget vector according to the recovery time limit and priority level of the target business to form a risk state tensor; Step S3: Obtain the power supply fault domain, network reachability of the isolation zone, and service mutual exclusion co-location rules, and combine them with the danger arrival time interval to generate a set of prohibited migration edges pointing to the computing node and a migration risk weight matrix between each computing node, forming a propagation constraint attention mask. Step S4: Based on the propagation constraint attention mask, the remaining capacity of each computing node, the cross-fault domain redundancy condition, and the service survival budget vector, perform node screening to form a sub-pool of security candidate computing nodes corresponding to each target service. Step S5: Determine the danger co-domain exposure penalty and propagation front migration penalty corresponding to each candidate computing node based on the risk state tensor, construct an objective function in combination with the service interruption time loss, and solve the objective function under the condition of satisfying the preset minimum replica number constraint to generate a scheduling action strategy set; Step S6: Based on the scheduling action policy set, perform cross-fault domain migration, replica startup, and network routing switching of the service instance, and issue linkage instructions to the execution nodes in the target area to trigger physical domain collaborative actions.

[0008] Secondly, this application provides a cloud-based smart park resource management system, the system comprising: The mapping unit is used to acquire data on hazard sources, environmental monitoring time series, target business dependencies, and computing node deployment within the target area, and to construct a heterogeneous time series diagram. The risk unit is used to determine the danger arrival time interval, failure probability and isolation status of each computing node based on the heterogeneous time sequence diagram, and generate a business survival budget vector according to the recovery time limit and priority level of the target business, forming a risk state tensor. The masking unit is used to obtain the power supply fault domain, the network reachability of the isolation zone, and the service mutual exclusion co-location rules. Combined with the danger arrival time interval, it generates a set of prohibited migration edges pointing to the computing node and a migration risk weight matrix between each computing node, forming a propagation constraint attention mask. The filtering unit is used to filter nodes based on the propagation constraint attention mask, the remaining capacity of each computing node, the cross-fault domain redundancy condition and the service survival budget vector, to form a sub-pool of security candidate computing nodes corresponding to each target service. The strategy unit is used to determine the danger co-domain exposure penalty and propagation front migration penalty corresponding to each candidate computing node based on the risk state tensor, construct an objective function in combination with the service interruption time loss, and solve the objective function under the condition of satisfying the preset minimum replica number constraint to generate a set of scheduling action strategies. The linkage unit is used to perform cross-fault domain migration, replica startup, and network routing switching of service instances based on the scheduling action policy set, and to issue linkage instructions to the execution nodes in the target area to trigger physical domain collaborative actions.

[0009] Compared with the prior art, the beneficial effects of the present invention are at least as follows: 1. First, this application breaks down the data silo between physical space and information space. By acquiring heterogeneous data to construct a time series diagram and deriving the arrival time interval of danger and failure probability to generate a risk state tensor, it successfully aligns the dynamic disaster risk characteristics of the underlying physical space with the disaster recovery baseline requirements of the upper-layer application into a unified high-dimensional coordinate system. This not only gives the cloud base the ability to accurately predict the spatiotemporal evolution trajectory of physical disasters, but also eliminates the scheduling blind spot caused by the single dimension of risk assessment, thus gaining a valuable pre-disaster business escape time window for the system.

[0010] 2. Secondly, this application significantly improves the security and efficiency of resource scheduling under complex disaster conditions. In the node selection stage, the introduction of propagation constraint attention masking enables one-click hard shielding of potential power outage and network disconnection areas. Combined with multi-dimensional capacity and cross-domain compliance pruning, it ensures the dual reliability of candidate nodes in terms of physical and computing power, fundamentally preventing secondary resource squeeze caused by business cluster migration. Subsequently, by constructing a multi-objective function that integrates time interruption cost and spatiotemporal distance penalty for heuristic optimization, it effectively avoids the trap of migration followed by disaster caused by traditional local greedy scheduling, and achieves global cost minimization.

[0011] 3. Finally, this application endows the system with powerful adaptive correction capabilities and resilience under extreme disasters. By issuing linkage instructions to the lower layer, it realizes physical domain coordination of computing power migration and data center temperature control and power distribution, avoiding local overload. More importantly, it uses updated measured data to perform dual-track verification of the executed schedule. When the disaster deviates from the prediction, the system can dynamically correct the propagation parameters to reconstruct the candidate pool and seamlessly roll back by relying on the reverse order cancellation mechanism of building first and then dismantling. In the desperate situation where conventional defenses fail completely, it can forcibly switch the core business to the edge node of the pre-set isolation zone, building an indestructible ultimate digital security cabin for the critical lifeline business of the smart park. Attached Figure Description

[0012] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a flowchart illustrating the steps of a cloud-based smart park resource management method in an embodiment of this application. Figure 2 This is a schematic diagram of a heterogeneous timing diagram in an embodiment of this application; Figure 3 This is a structural diagram of a cloud-based smart park resource management system according to an embodiment of this application. Detailed Implementation

[0014] This application provides a cloud-based smart park resource management method and system. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0015] Example 1: For ease of understanding, the specific process of the embodiments of this application is described below, such as... Figure 1 The cloud-based smart park resource management method shown in this embodiment includes: Step S1: Obtain data on hazard sources, environmental monitoring time series, target business dependencies, and computing node deployment within the target area, and construct a heterogeneous time series diagram.

[0016] Step S1 includes: determining hazard source nodes based on hazard source data, determining sensing nodes based on environmental monitoring time-series data, determining business nodes based on target business dependency data, and determining computing nodes and execution nodes based on computing node deployment data; establishing physical propagation edges based on the spatial distribution of hazard source nodes and sensing nodes, establishing data flow edges based on the data calling rules between sensing nodes and business nodes, establishing deployment mapping edges based on the operational bearing association between business nodes and computing nodes, establishing network topology edges based on multiple computing nodes, and establishing linkage control edges based on the physical spatial binding and control relationship between computing nodes and execution nodes; integrating nodes and edges to generate a heterogeneous time-series graph, such as... Figure 2 As shown.

[0017] Specifically, when facing sudden physical disasters, traditional cloud platform resource scheduling systems in smart parks only focus on computing power utilization and network latency, failing to perceive the direction of accident spread in the physical space and the linkage relationship of underlying facilities, leading to highly susceptible passive failure of services. To address the technical problem of long-term data fragmentation between the physical domain accident propagation state and the information domain service scheduling process, this technical solution aims to break down cross-domain barriers and construct an underlying data structure mapping that can uniformly express the laws of physical disaster propagation, the deployment relationship of information systems, and the physical collaborative control relationship, as detailed below.

[0018] First, acquire hazard source data, environmental monitoring time-series data, target business dependency data, and computing node deployment data within the target area. The target area refers to the physical space of the smart park requiring resource management and accident risk control. Hazard source data refers to the attribute information of equipment or media within the target area that may cause accidents and have a diffuse impact on the surrounding area. Environmental monitoring time-series data refers to the time-varying sequence of physical quantities continuously collected by various sensors at a preset sampling frequency. Target business dependency data describes the various information services carried within the park and their inter-service relationships, including operational constraints such as service request volume, response time limits, and minimum available replica counts. Computing node deployment data refers to information such as the deployment location, associated computer room, associated power distribution circuit, and available resource capacity of equipment carrying computing tasks and performing physical actions. Based on this, by connecting with the park's existing safety management system, fire protection IoT platform, and business orchestration platform, the above multi-source heterogeneous data is aggregated to the cloud platform according to a unified data access protocol.

[0019] Next, hazard source nodes are determined based on hazard source data, and sensing nodes are determined based on environmental monitoring time-series data. Specifically, each hazard source with an independent spatial location is abstracted as a hazard source node, and its three-dimensional coordinates, potential release rate, and influence radius benchmark values ​​are extracted as node attributes. Each independently deployed sensing device is abstracted as a sensing node, and its sampling type, spatial coordinates, and sampling frequency are extracted as node attributes. Business nodes are determined based on target business dependency data, and each type of information business that needs to ensure continuity is abstracted as a business node, and its business identifier, minimum number of copies, and priority level are extracted as node attributes. Computing nodes and execution nodes are determined based on computing node deployment data, and each server hardware that carries out business operations is abstracted as a computing node, and various underlying devices such as automated switches, electric valves, and air conditioning units that need to perform physical actions according to scheduling results are abstracted as execution nodes.

[0020] After identifying five types of abstract nodes, connectivity edges reflecting the spatiotemporal and logical relationships between nodes are further established. Physical propagation edges are established based on the spatial distribution of hazard source nodes and sensing nodes. This process involves extracting the three-dimensional coordinates of both hazard source nodes and sensing nodes, calculating the Euclidean distance between them, and establishing a directed physical propagation edge between them when the Euclidean distance is less than the baseline value of the influence radius in the corresponding attribute of the hazard source node. A baseline propagation delay and attenuation coefficient are then assigned to this edge. Data flow edges are established based on the data calling rules between sensing nodes and business nodes. By parsing the data subscription relationship in the network layer protocol, the data output address of the sensing device is matched and associated with the data receiving address of the business service, thereby generating data flow edges. Finally, a partial connection is established based on the operational bearing association between business nodes and computing nodes. The deployment mapping edge is established by calling the container scheduling registry of the cloud platform to extract the host hardware identifier currently occupied by the business process and establish a deployment mapping edge from the business node to the corresponding computing node. A network topology edge is established based on multiple computing nodes. The topology edge reflecting the interconnection status of computing power network and link bandwidth attributes is generated by calling the routing table of the campus switch. A linkage control edge is established based on the physical space binding and control relationship between computing nodes and execution nodes. The specific calculation and matching process is to obtain the rack identifier and data center space code of the computing node, and compare the string consistency of the space code with the coverage control area code of each execution node. When it is determined that the two belong to the same physical fault domain space, a linkage control edge is generated between the corresponding computing node and execution node to identify the physical defense equipment that can be triggered when the computing node is damaged.

[0021] Finally, by integrating the nodes and edges generated above, a heterogeneous time series graph is generated. The heterogeneous time series graph is essentially a composite spatiotemporal graph data structure model. It maps the accident propagation path and sensor data flow in the physical world with the business deployment location, network topology, and physical control links of underlying devices in the information world in the same high-dimensional topological coordinate system. It also adds attribute slices that change over time to each edge of the graph model, so that the connection state in the graph can be dynamically updated as the time series window progresses, thereby providing data input support for cross-domain linkage and risk simulation.

[0022] The above content constructs a heterogeneous time sequence diagram integrating hazard sources, perception, business, computing power, and execution equipment, opening up the mapping channel from physical disasters to information domain scheduling. It provides a multi-dimensional perception data foundation for disaster prevention and coordinated risk avoidance based on spatiotemporal constraints. Through automated spatial distance calculation and registry configuration comparison, it accurately establishes the propagation flow and control binding association between various heterogeneous components. The introduced linkage control edge clarifies the physical spatial coupling relationship between computing nodes and underlying fire protection and environmental control equipment, effectively eliminating the decision-making blind spot that traditional information-side scheduling systems cannot perceive the underlying physical execution mechanism. It significantly enhances the spatiotemporal granularity of the system's characterization of sudden physical disaster boundaries and its ability to coordinate and respond in a coordinated manner.

[0023] Step S2: Based on the heterogeneous time sequence diagram, determine the danger arrival time interval, failure probability and isolation status of each computing node, and generate a business survival budget vector according to the recovery time limit and priority level of the target business to form a risk state tensor.

[0024] Step S2 includes: deriving the danger arrival time interval of each computing node through the physical propagation edge in the heterogeneous time sequence graph, and determining the failure probability and isolation state of each computing node based on the danger arrival time interval; and fusing the danger arrival time interval, failure probability, isolation state and business survival budget vector to form a multi-dimensional risk state tensor, wherein the risk state tensor includes the spatial coordinate dimension of the computing node, the time sequence dimension and the business requirement dimension corresponding to the business survival budget vector.

[0025] Specifically, after completing the construction of the heterogeneous time sequence diagram, this application possesses a static mapping association architecture between physical space and information space. However, traditional cloud platforms lack quantitative assessment methods for the dynamic evolution sequence of sudden accidents and the differentiated risk resistance capabilities of different businesses when scheduling resources. It is difficult to accurately measure the survival threat posed by disasters to specific computing nodes in the future time scale. To solve this technical problem, this technical solution aims to perform dynamic disaster spread simulation based on the heterogeneous time sequence diagram, aligning physical spread risk and business disaster recovery requirements into the same multi-dimensional spatiotemporal measurement system, as detailed below.

[0026] First, by using the physical propagation edges in the heterogeneous time series graph, the arrival time interval of hazards for each computing node is deduced. This arrival time interval refers to the set of the earliest and latest possible times required for the impact frontier of an accident to propagate from the hazard source and reach the physical location of a computing node, assuming an accident occurs at a certain hazard source. Specifically, during the deduction and calculation process, for each hazard source node, a topological traversal is performed along the physical propagation edges in the heterogeneous time series graph to its downstream sensing nodes. The spatial coordinates of the geometric regions covered by the sensing nodes and the three-dimensional spatial coordinates of the computing nodes are extracted. A spatial bounding box intersection algorithm is executed to determine collisions. If spatial region intersection is determined, the inherent baseline propagation delay and attenuation coefficient on the physical propagation edge are obtained. Wind speed, wind direction, and temperature from the current environmental monitoring time series data are read as dynamic environmental correction factors. These dynamic environmental correction factors are used to compensate for the baseline propagation delay. The specific process for time delay compensation calculation is as follows: First, extract the spatial geometric vector pointing from the hazard source node to the calculation node, and calculate the relative angle between the current wind direction and the spatial geometric vector; then, use the cosine value of the relative angle to perform vector projection calculation on the current wind speed to obtain the effective wind speed component with or against the wind, and combine it with the temperature difference between the actual temperature and the preset standard reference temperature to generate dimensionless wind speed and wind direction compensation coefficients and temperature compensation coefficients respectively; the wind speed and wind direction compensation coefficient is less than one when with the wind to indicate acceleration, and greater than one when against the wind to indicate deceleration; the temperature compensation coefficient is less than one when at high temperatures to indicate increased thermodynamic diffusion; then, multiply the reference propagation time delay by the wind speed and wind direction compensation coefficient and the temperature compensation coefficient in sequence to obtain the dynamically corrected expected arrival time delay; finally, combine the attenuation coefficient on the physical propagation edge to calculate a time fluctuation margin, and subtract and add the time fluctuation margin to the expected arrival time delay respectively, thereby accurately outputting the lower and upper bound times of the hazard arrival time interval corresponding to the calculation node.

[0027] Next, based on the calculated arrival time interval of the hazard, the failure probability and isolation status of each computing node are determined. The failure probability is used to characterize the quantitative possibility that the target computing node will completely lose its normal computing service capability due to physical disaster within the arrival time interval of the hazard. The specific calculation and acquisition process is as follows: the median of the arrival time interval of the hazard is extracted as the time reference parameter, combined with the concentration decay model of the hazardous medium at the corresponding spatial distance to output the expected exposure concentration parameter, and matched with the hardware protection level coefficient parameter of the target computing node itself; since the time reference parameter, the expected exposure concentration parameter and the hardware protection level coefficient parameter are... Since the parameters are inconsistent, they are mapped to a dimensionless interval of zero to one using an extreme value normalization function. Then, the weight coefficients corresponding to the normalized parameters are weighted and jointly evaluated to calculate the failure probability value between zero and one. At the same time, the isolation status is used to describe whether a specific computing node is covered by the physical isolation barrier specified in the park's emergency plan. By extracting the physical location information of the computing node and indexing and comparing it in the isolation zone division table pre-set in the cloud platform, if it is confirmed that the location has been marked as a forced power outage or fire door closed area, the isolation status of the computing node is directly assigned as isolated.

[0028] Simultaneously, a business survival budget vector is generated based on the recovery time limit and priority level corresponding to the target business. The recovery time limit refers to the maximum tolerable time threshold for a business instance to resume operation after an unexpected interruption. The priority level reflects the relative importance weight of the business in the global resource game of the cloud platform. The cloud platform extracts the above attribute data of various target businesses and transforms them into a structured set of multi-dimensional constraint indicators, namely the business survival budget vector. This vector encapsulates the allowable interruption time budget and the allowable replica loss budget, which are used to clearly define the disaster recovery bottom line of different businesses in disaster scenarios.

[0029] Finally, a multi-dimensional risk state tensor is formed by integrating the danger arrival time interval, failure probability, isolation status, and business survival budget vector. This risk state tensor is specifically represented as a multi-dimensional mathematical data structure model. Its inputs are the underlying multi-source disaster state data and business budget data obtained through dynamic deduction and analysis, and its output is a unified, integrated high-dimensional tensor matrix. During tensor construction, various state indicators are aligned and mapped to specified tensor dimensions. Specifically, the risk state tensor includes the spatial coordinate dimension of the computing node, the time series dimension, and the business demand dimension corresponding to the business survival budget vector. The spatial coordinate dimension records the three-dimensional physical distribution matrix of the computing node; the time series dimension is sliced ​​according to a preset time step, recording the dynamic evolution numerical sequence of the aforementioned failure probability and isolation status within various future time windows; the business demand dimension maps to the specific disaster recovery budget parameters of different business instances. The final risk state tensor is generated through the Cartesian product concatenation operation of the multi-dimensional array. Each element in this tensor model uniquely and accurately represents the comprehensive physical disaster risk score of a computing node at a specific future time from the perspective of specific business continuity.

[0030] The above-described structure constructs a risk state tensor that integrates the dynamic evolution of disasters with the disaster recovery baseline of business operations. This effectively solves the technical challenge of traditional scheduling systems being unable to perform cross-domain joint quantitative assessment of the propagation sequence of physical accidents and the continuity requirements of multiple services. By introducing a spatial bounding box intersection algorithm and an environmental correction factor, it accurately extrapolates the spatiotemporal impact boundary of disasters, transforming the previously difficult-to-measure physical accident evolution into a calculable time interval and failure probability. Combined with a multi-dimensional risk state tensor data structure model constructed using business survival budget vectors, it successfully aligns the dynamic risk characteristics of the underlying physical space with the disaster recovery baseline requirements of upper-layer applications into a unified high-dimensional mathematical coordinate system. This not only eliminates the data semantic gap between the physical disaster domain and the information scheduling domain, effectively avoiding scheduling decision delays and misjudgments caused by a single or fragmented risk assessment dimension, but also provides a high-precision, fine-grained quantitative decision-making data foundation for subsequent forward-looking computing power risk-avoidance scheduling in complex accident scenarios.

[0031] Step S3: Obtain the power supply fault domain, network reachability of the isolation zone, and service mutual exclusion co-location rules. Combined with the danger arrival time interval, generate a set of prohibited migration edges pointing to computing nodes and a migration risk weight matrix between each computing node to form a propagation constraint attention mask.

[0032] Step S3 includes: determining whether each computing node is within a power failure domain; if the target computing node is within a power failure domain, assigning the network topology edge pointing to the target computing node to the prohibited migration edge set; determining the initial migration risk weight between each computing node based on the danger arrival time interval and the network reachability of the isolation zone, and adjusting the initial migration risk weight according to the service mutual exclusion co-location rule to obtain the migration risk weight matrix; performing numerical masking on the target network topology edge through the prohibited migration edge set, and combining the adjusted migration risk weight matrix to form the final propagation constraint attention mask.

[0033] Specifically, after completing the node-level quantitative assessment based on the risk state tensor, this application can grasp the disaster sequence of each independent computing node. However, when the cloud platform actually performs business redirection and resource scheduling, it is necessary not only to consider the security of a single node, but also to take into account the network connectivity of the migration path, the risk of coordinated power cut-off of the underlying power equipment in the data center, and the high availability and disaster recovery deployment specifications of the upper-layer business itself. Traditional methods often ignore the power outage linkage of the underlying electromechanical equipment and the mutual exclusion constraints of high availability of the business, resulting in secondary failures of the calculated migration path during physical execution. This technical solution aims to solve the problem of migration path connectivity and security assessment under multiple cross-domain operation and maintenance constraints. It performs pre-emptive mandatory risk filtering of potential migration paths by transforming physical rules into graph structure masks, as follows.

[0034] In practice, the first step is to obtain the power supply fault domain, network reachability of the isolation zone, and service mutual exclusion co-location rules. Specifically, the cloud platform reads the power supply topology of the equipment from the underlying power distribution operation and maintenance system, and divides the set of computing nodes that are under the same incoming switch, the same transformer, or the same uninterruptible power supply system and have the characteristic of "one failure causing all failures" into power supply fault domains. At the same time, the isolation zone network reachability is extracted from the contingency plan management system. This feature parameter is used to quantitatively describe whether the computing nodes inside and outside the area still have the network connectivity to maintain the underlying protocol interoperability and data transmission based on wireless backhaul links or backup links after a specific physical area is forcibly isolated or the main power supply is cut off by the emergency plan. Then, the service mutual exclusion co-location rules are read from the business orchestration strategy library. The service mutual exclusion co-location rules refer to the mandatory distributed deployment strategy constraints that strictly limit the simultaneous deployment of multiple primary and backup replica instances of the same business in the same physical rack, the same data center, or the same power supply circuit based on the security compliance requirements for preventing single point failures.

[0035] Based on the aforementioned basic constraint data, each computing node is individually assessed to determine whether it is within a power supply failure domain. The core logic is to pre-screen those hidden high-risk nodes that have not yet been physically affected by the disaster front, but are destined to lose computing service capabilities because their associated power distribution zones are about to be forcibly shut down by the emergency plan. If, after table comparison, it is determined that the target computing node is indeed within the aforementioned power supply failure domain, then in the global network topology structure, all unidirectional network topology edges pointing to the target computing node are forcibly assigned to the prohibited migration edge set. The prohibited migration edge set is a graph edge index set specifically used to explicitly block and shield the channels for sending service instance migration instructions to the soon-to-be-failed nodes in subsequent scheduling simulations.

[0036] Based on this, and combining the danger arrival time intervals output in the previous step, the initial migration risk weights between each computing node are determined according to the danger arrival time intervals and the network reachability of the isolation zone. In the specific calculation process, the upper and lower bound parameters of the danger arrival time intervals for the migration source node and each candidate migration target node are extracted, and the ratio of their overlap time on the time axis to the total length of the interval is calculated as a time-dimensional risk indicator. A higher overlap ratio means a greater probability that the two nodes will be simultaneously destroyed by the same wave of disaster. Simultaneously, based on the aforementioned isolation zone network reachability status, the communication paths between nodes are Boolean-wise evaluated. Conditional Judgment: If the determination result is unreachable, i.e., physical isolation causes the underlying protocol to be interrupted, then the initial risk weight for this migration path is directly assigned a preset maximum penalty value to completely block migration to this path at the algorithm level; if the determination result is reachable, i.e., communication can still be maintained through backup or wireless links, then on this basis, the actual remaining available bandwidth and message transmission latency of the communication path between nodes are extracted through the network flow measurement component, and normalized as a network dimension risk indicator; finally, the time dimension risk indicator and the network dimension risk indicator are weighted and fused to obtain a comprehensive initial migration risk weight.

[0037] Furthermore, the initial migration risk weights are penalized and adjusted according to the service mutual exclusion co-location rule to obtain the migration risk weight matrix. All candidate migration target nodes are traversed, and it is checked whether the host machine where the target node is located and other existing replica hosts of the service have triggered the service mutual exclusion co-location rule. If it is determined that the candidate target node and the existing replica are in the same power failure domain or the same physical sub-region of the data center, a large positive mutual exclusion penalty increment is directly superimposed on the initial migration risk weight value associated with the corresponding network topology edge. By increasing the risk weight value, the probability of the path being selected in the subsequent optimal path solution is greatly reduced. After traversing and probing the entire network topology and superimposing and correcting the penalty parameters, the final output is a two-dimensional matrix containing the global node migration risk assessment results, namely the migration risk weight matrix.

[0038] Finally, numerical masking is applied to the target network topology edges using the prohibited edge set, and combined with the adjusted migration risk weight matrix to form the final propagation-constrained attention mask. This propagation-constrained attention mask is essentially a graph-structured filtering data model. It does not require training on historical data using backpropagation; instead, it takes the underlying global network topology graph, the prohibited edge set, and the migration risk weight matrix as input data, directly mapping the output to a structured graph mask tensor. During mask construction, each network topology edge in the graph data structure is traversed. If a match indicates that the edge exists in the prohibited edge set, the corresponding value is set to a minimum value approaching negative infinity using the masking mechanism. This achieves hard numerical masking and completely cuts off the possibility of resource scheduling in that direction. For other regular network topology edges that do not fall into the prohibited edge set, their corresponding mask values ​​are directly assigned to the corresponding continuous numerical components in the migration risk weight matrix, thus generating the final propagation-constrained attention mask that combines rigid masking constraints with flexible risk guidance capabilities.

[0039] The above content establishes a dynamic weight adjustment mechanism that integrates power topology disconnection logic and high-availability deployment strategies for services. This effectively solves the problem of secondary failures in resource migration paths caused by underlying electromechanical linkages or upper-layer rule conflicts under complex disaster environments. By combining a power supply fault domain identification mechanism with a prohibited migration edge set, it can proactively identify computing nodes that are about to "passively fail" due to emergency power outage plans before the actual arrival of a physical disaster, and completely block migration paths to these nodes from the network topology, avoiding "ineffective disaster relief" of service resources and waste of scheduling computing power. A cross-dimensional initial weight based on an extreme value normalization function is introduced. By recomputing and combining the incremental mutual exclusion penalties generated by the mutual exclusion co-location rules of services, the risk of overlapping disasters in the time dimension, the risk of communication disruption in the network dimension, and the risk of deployment aggregation in the compliance dimension are unified and quantified into the same evaluation system. The resulting propagation constraint attention mask graph model does not require computing power for model training and can directly provide a strong decision boundary input for subsequent graph neural networks or heuristic search algorithms. This ensures that every business scheduling instruction output by the cloud platform perfectly matches the dual red lines of underlying physical availability and business availability, significantly improving the business continuity assurance capability of smart parks in extreme disaster scenarios.

[0040] Step S4: Based on the propagation constraint attention mask, the remaining capacity of each computing node, the cross-fault domain redundancy conditions, and the service survival budget vector, the nodes are screened to form a sub-pool of security candidate computing nodes corresponding to each target service.

[0041] Step S4 includes: removing computation nodes associated with prohibited migration edges by propagating a constraint attention mask to obtain a preliminary set of selected nodes; determining whether the remaining capacity of each node in the preliminary set of selected nodes meets the resource operation conditions required by the business survival budget vector, selecting nodes that meet the resource operation conditions, and grouping and further filtering the nodes that meet the resource operation conditions according to the cross-fault domain redundancy conditions to ensure that the business replica distribution has isolation diversity, and obtaining a safe candidate computation node sub-pool corresponding to each target business.

[0042] Specifically, after generating the propagation constraint attention mask containing the global node scheduling red line, the challenge is how to quickly identify a set of host nodes that are both safe and available and meet business resource requirements from a massive number of nodes. If resource scheduling is directly solved in the global topology, not only will the computational overhead explode exponentially, but it is also highly likely that business migration will fail due to insufficient actual available resources of the target nodes or disaster recovery compliance conflicts. This technical solution aims to solve the problem of excessively large business scheduling solution space and complex multi-dimensional constraints in disaster scenarios. By introducing multi-level hard pruning and flexible filtering algorithms, a set of candidate bearer nodes with extremely high physical security and resource feasibility is generated in advance for each business, as detailed below.

[0043] In specific implementation, based on the propagation constraint attention mask output from the aforementioned steps, and combined with the remaining capacity of each computing node acquired in real time, the preset cross-fault domain redundancy conditions, and the business survival budget vector of the corresponding target business, multi-level node screening is performed. First, the computing nodes associated with the prohibited migration edge set are removed by the propagation constraint attention mask to obtain a preliminary set of screened nodes. Specifically, the global network topology graph at the bottom layer of the cloud platform is traversed, and the mask value of each network topology edge pointing to the candidate computing node is checked. If the mask value is read as a minimum value set for complete shielding, such as tending to negative infinity, it is logically determined that the node is located in an absolutely high-risk area that is about to be powered off or has been damaged by a disaster. Then, the node is hard-removed from the global set of available nodes using set difference operation. Conversely, if the mask value pointing to the node is a normal risk weight value, it is retained and included in the preliminary set of screened nodes. Through this graph structure mask filtering algorithm, efficient one-click pruning of high-risk nodes at the physical space level is achieved.

[0044] Subsequently, the real-time resource status of each computing node in the initial screening node set is obtained. It is then determined whether the remaining capacity of each node in the initial screening node set meets the resource operating conditions required by the business survival budget vector, thereby filtering out nodes that meet the resource operating conditions. Here, remaining capacity refers to the multi-dimensional physical resource vector of a computing node after deducting the currently allocated business load at the physical host level, including the number of available CPU cores, available memory capacity, storage read / write bandwidth, and network throughput. The business survival budget vector refers to the multi-dimensional feature vector of the minimum resources required to maintain the basic core functions of the target business after disaster recovery and migration. In the specific mathematical algorithm implementation, the multi-dimensional data structures of the above two are extracted, and vector element comparison operations are performed on each resource dimension. Only when the value of the remaining capacity vector of a candidate computing node is greater than or equal to the value of the corresponding dimension of the business survival budget vector in all specific resource dimensions is it determined to meet the resource operating conditions and retained. This step can accurately eliminate any node with a bottleneck in a single key resource dimension, preventing secondary downtime caused by severe resource exhaustion after business migration.

[0045] Furthermore, nodes meeting the resource operation conditions are grouped and secondary-screened based on cross-fault domain redundancy conditions to ensure the isolation diversity of service replica distribution, ultimately obtaining a sub-pool of secure candidate computing nodes corresponding to each target service. The cross-fault domain redundancy condition is a mandatory compliance constraint rule based on a high-availability architecture design, requiring multiple running replicas of the same target service to be deployed in non-overlapping power supply or network fault isolation zones. Isolation diversity is used to quantitatively characterize the degree of decoupling of candidate node groups in terms of underlying physical infrastructure dependencies. In implementing the secondary screening algorithm, the physical location and power distribution association tags of nodes are first read, and nodes meeting the resource conditions are clustered according to their respective power supply fault domains and data center physical sub-regions. Subsequently, the fault domain identifier set of the currently surviving replicas of the target service is extracted, and the overlap relationship between each candidate node group and this surviving identifier set is calculated through set intersection operations. First, nodes in groups whose intersection is not empty (i.e., those in the same fault domain) are prioritized for elimination. Simultaneously, to balance security and resource utilization in the selection process, a comprehensive redundancy suitability score for candidate nodes can be calculated. Specifically, the physical topological distance between the fault domain of the node and the fault domain of the surviving replica, as well as the node's resource slack rate, are used as evaluation indicators. Since the distance and resource rate indicators have inconsistent dimensions, a maximum-minimum-maximum value normalization function is used to uniformly map them to a dimensionless interval of zero to one. Then, a linear weighted sum is performed, assigning preset weight coefficients to both. Nodes meeting the redundancy requirements are retained from high to low based on the weighted score, ultimately generating a dedicated safe candidate computing node sub-pool for this target business. The safe candidate computing node sub-pool is a dynamic set. When updates to external environmental monitoring time-series data lead to a mitigation of the disaster propagation situation, the previously masked node resources are released again, triggering the above multi-level selection process cyclically to update the sub-pool capacity in real time.

[0046] The above content proposes a multi-level pruning and filtering mechanism that integrates graph mask hard isolation, multi-dimensional resource vector comparison, and cross-domain disaster recovery grouping. This effectively solves the technical challenge of a large cloud platform scheduling search space under complex constraints, which is prone to causing secondary physical downtime. Through a cascaded dimensionality reduction filtering algorithm, it first utilizes propagation constraint attention masks to achieve efficient "one-vote veto" of disaster frontier and power outage affected nodes, significantly reducing the ineffective computational overhead in subsequent strategy solving stages. Subsequently, it performs multi-dimensional strict alignment calculations between the remaining capacity of the underlying physical nodes and the survival budget vector of the upper-layer business, ensuring that selected nodes are not only physically isolated but also have a high degree of resource utilization. In terms of location, it is "safe" and in terms of computing power supply, it is more "available". It fundamentally avoids the performance degradation or service crash caused by the cluster migration of business in disaster emergency. Finally, it introduces a secondary screening based on cross-fault domain redundancy conditions and is supplemented by a normalized weighted scoring mechanism. This forces the isolation diversity of the multi-replica distribution of business to be guaranteed. This makes the selected safe candidate computing node sub-pool have system-level high availability resilience to resist the chain reaction of potential secondary disasters while undertaking the current business redirection. This lays a high-quality decision boundary data foundation for the accurate scheduling of the smart park cloud platform in extreme disaster environments.

[0047] Step S5: Determine the danger co-domain exposure penalty and propagation front migration penalty for each candidate computing node based on the risk state tensor, construct the objective function in combination with the service interruption time loss, and solve the objective function under the condition of satisfying the preset minimum replica number constraint to generate a set of scheduling action strategies.

[0048] Step S5 includes: calculating the service interruption time loss corresponding to each candidate computing node in the security candidate computing node sub-pool, configuring the weight coefficients corresponding to the service interruption time loss, dangerous common domain exposure penalty and propagation front migration penalty, and constructing an objective function with the goal of minimizing the global scheduling cost by weighted summation; under the condition of satisfying the minimum replica number constraint, performing a heuristic search to solve the objective function to obtain a scheduling action strategy set, which includes at least the instance migration path between computing nodes, the expansion number of the target service, and the network switching routing table.

[0049] Specifically, after identifying a secure candidate computing node sub-pool with both physical security and resource feasibility, the challenge lies in making a globally optimal deployment decision among these legitimate candidate nodes. If only random allocation or a scheduling method solely focused on resource utilization is adopted, it is highly likely that services will experience long downtimes during migration to candidate nodes, or be reassigned to dangerous edge areas too close to the forefront of accident propagation, thus facing the risk of being affected by disasters again in the short term and triggering repeated migrations. This technical solution aims to solve the scheduling optimization problem in an environment where multidimensional spatiotemporal risks conflict with business continuity requirements. By constructing a multi-objective evaluation system that takes into account both time costs and spatial penalties, a deterministic action instruction that minimizes the global scheduling cost is calculated, as follows.

[0050] In practice, the first step is to calculate the service interruption time loss for each candidate computing node in the security candidate computing node sub-pool. This service interruption time loss refers to the total unavailability caused by underlying hardware and software resource allocation during the process of scheduling the target service to that candidate computing node. Specifically, the calculation involves extracting the instance memory snapshot transmission and state replay time, the replica warm-up time on the new computing node, and the link jitter time caused by network routing switching, respectively. These three time data items are then summed scalarly to obtain the total time loss for a specific service paired with a specific node. When calculating the service interruption time loss, if the data flow is blocked due to shielding by the underlying network topology edge, a data interruption penalty term is added. Subsequently, based on the risk state tensor output in the preceding steps, the danger co-domain exposure penalty and propagation front migration for each candidate computing node are determined. The penalty, or hazard co-location exposure penalty, refers to the cost of a candidate computing node being physically close to the area affected by an accident. The spatial coordinate boundary of the current accident-affected area is extracted from the risk state tensor. The shortest Euclidean distance between the physical location of the candidate computing node's data center and this spatial coordinate boundary is calculated. If the distance is greater than zero, the reciprocal of this distance is taken as the initial value of the hazard co-location exposure penalty. The propagation front migration penalty refers to the time urgency cost imposed to prevent business instances from being immediately affected by a disaster immediately after migration. The expected migration completion time of the target business is obtained, and the lower bound of the hazard arrival time interval corresponding to the area where the candidate computing node is located is extracted from the risk state tensor. The difference between this lower bound and the expected migration completion time is calculated. If the difference is greater than zero, the reciprocal of this time difference is taken as the initial value of the propagation front migration penalty.

[0051] Because the calculated service interruption time loss, hazard co-location exposure penalty, and propagation front migration penalty correspond to the absolute value of time, the reciprocal of spatial distance, and the reciprocal of time difference, respectively, in a physical sense, there is a significant inconsistency in dimensions, making direct superposition and calculation impossible within the same mathematical space. Therefore, this method introduces a maximum-minimum-extreme value normalization function to independently map the three indicators of each candidate node to a standard dimensionless interval of zero to one. Next, corresponding weight coefficients are assigned to the normalized service interruption time loss, hazard co-location exposure penalty, and propagation front migration penalty. The value of these weight coefficients is dynamically set according to the current macro-level emergency response parameters of the smart park. For example, higher weight coefficients are assigned to the two risk penalty items under high-level emergency conditions, while higher weight coefficients are assigned to time loss under low-level operational conditions. By multiplying the three normalized indicators by their corresponding weight coefficients and performing a weighted sum, an objective function is constructed with the goal of minimizing the global scheduling cost. The smaller the output value of this objective function, the lower the overall cost of scheduling services to the candidate node.

[0052] Furthermore, under the condition of satisfying the preset minimum replica number constraint, the objective function is solved, and a set of scheduling action strategies is finally generated. The minimum replica number constraint refers to the lower limit threshold of the number of service replicas that are online and providing normal service at any time, as specified in the service orchestration strategy library, to ensure the basic availability of various services. For the solution process of this objective function, a heuristic search algorithm is adopted, using the calculation result of the objective function as heuristic information. In the specific calculation process, all candidate nodes in the candidate computing node sub-pool are sorted in ascending order according to their corresponding objective function calculation values. Then, the target services to be scheduled are traversed from high to low according to the importance of the services. For the target service currently traversed, resource allocation is attempted sequentially according to the ascending order of the nodes. During each trial allocation, the number of service replicas that are alive and unaffected in the network at the current time is strictly checked, plus the number of replicas in the current trial allocation. The system checks whether the number of replicas to be deployed is greater than or equal to the minimum replica count constraint, and whether the remaining capacity of the current preferred node is sufficient to support the service's survival budget vector. If all conditions are met, the allocation action is confirmed, an instance migration path record from the disaster source node to the target node is generated, and the corresponding resource capacity of the target node is immediately deducted from the system cache to prevent resource overselling. If the minimum replica count constraint cannot be met through pure migration operations, an instruction to add replicas is triggered on other nodes with higher priority in the sorting sequence, and the expansion quantity of the target service is recorded. All confirmed instance allocation actions and the updated access endpoint address data are summarized together to generate a scheduling action strategy set containing the specific instance migration path between compute nodes, the expansion quantity of the target service, and the network switching routing table, for the underlying cloud platform to directly execute the scheduling.

[0053] The above content establishes a normalized multi-objective evaluation system that integrates time interruption costs and spatiotemporal risk penalties. Combined with minimum disaster recovery constraints and heuristic search, it achieves a deterministic solution for cloud-edge resource scheduling schemes under disaster degradation conditions. By transforming abstract business interruption risks and specific physical space spillover risks into independently calculable mathematical penalty terms, and supplementing this with extreme value normalization, the dimensional barriers between scheduling considerations of different dimensions are completely eliminated. This allows the system to accurately quantify the comprehensive security and business continuity costs of each candidate node in rapidly changing disaster environments. Combined with a heuristic search algorithm that aims to minimize global costs, it effectively avoids the pitfalls of traditional local greedy scheduling. This significantly reduces the probability of cascading failures due to secondary disasters, avoiding the short-sighted trap of "migrating only to be hit by a disaster again." Simultaneously, the mandatory integration of minimum replica count constraints into the optimization computation ensures that the online service capabilities of the park's core businesses remain above a safe level during complex instance migrations and topology refactoring. Ultimately, a structured scheduling action strategy set containing migration paths, expansion parameters, and routing rules is generated, which can be directly parsed and executed by the underlying cloud platform infrastructure. This enables the smart park to output optimal resource self-healing instructions with extremely low decision latency when facing sudden physical disasters, greatly improving the disaster recovery resilience and emergency response efficiency of the entire park's cloud computing architecture.

[0054] Step S6: Based on the scheduling action policy set, perform cross-fault domain migration, replica startup, and network routing switching of the service instance, and issue linkage instructions to the execution nodes in the target area to trigger physical domain collaborative actions.

[0055] Step S6 includes: parsing the scheduling action policy set, obtaining the instance migration path, expansion quantity, and network switching routing table; performing cross-fault domain migration of the target service instance using state synchronization according to the instance migration path; performing replica startup and service registration of the target service instance on the target node according to the expansion quantity; performing network routing switching through the software-defined network controller according to the network switching routing table; generating linkage instructions bound to the currently deployed node of the target service instance; and sending the linkage instructions to the execution nodes in the target area to control physical devices to perform physical domain collaborative actions.

[0056] Specifically, after calculating the set of scheduling action strategies with the minimum global scheduling cost through a multi-objective optimization engine, the ultimate challenge of this method is how to seamlessly and securely transform this static planning at the digital level into actual execution actions at the cloud platform's underlying layer. If the old instance is shut down and the new instance is started directly and rudely, it is easy to cause serious network black holes and service gaps during the migration. At the same time, if the migration of computing power ignores the coordinated adaptation of the physical infrastructure supporting the computing power, it may cause local temperature rise or power overload in the target data center due to a sudden increase in load. This technical solution aims to solve the above-mentioned technical problems of smooth transition from virtual strategy to virtual-physical combined execution and collaborative management of information domain and physical domain.

[0057] In practice, the scheduling action policy set is first parsed to obtain the instance migration path, expansion quantity, and network switching routing table. During this process, the policy set file generated in the previous steps is structured, unpacked, and grouped to extract the source node identifier, target node identifier, and corresponding container resource configuration parameters used to guide the underlying resource scheduling. Subsequently, the cross-fault domain migration of the target service instance is performed according to the instance migration path using a state synchronization method. The state synchronization method refers to the process of copying runtime memory data pages to the target node without interrupting the read / write operations of the source node, using a combination of underlying memory copying and file system snapshots for services with runtime state caching. Computing resources are pre-allocated and built on the target node. A low-level data transmission channel is established to iteratively synchronize dirty pages from the source node's memory to the target node in an incremental streaming manner. Once the data on both ends is completely consistent, the source node process is instantly frozen, and the process is immediately resumed on the target node. For situations where additional load needs to ensure redundancy, replicas of the target business instances are launched and services are registered on the target node according to the expansion quantity. The target compute node calls the underlying container engine interface to download the required business image and starts a new running replica. Service registration refers to the newly launched replica actively reporting its assigned internal virtual network address and port number to the unified service registration center of the smart park after passing the internal health check, so that all callers on the network can dynamically discover and legally access the newly added replica.

[0058] Once all instance replicas of the target node are ready, network routing is switched via a software-defined network controller (SDB) according to the network switching routing table. The SDB is a core network element component that decouples the control plane and data plane of network devices, dynamically managing network traffic flow through a centralized software platform that sends flow tables downwards. Updated routing weights and service access virtual address rules are sent to the controller, which then converts them into flow table instructions recognizable by the underlying edge switches and distributes them. This action smoothly diverts external access traffic to the newly deployed cross-domain replicas and gradually clears old routing connections pointing to the original high-risk disaster area, achieving a smooth network cutover. To achieve unified management of the cloud and physical environments, a linkage instruction bound to the currently deployed node of the target service instance is generated and sent to execution nodes within the target area to control physical devices to perform physical domain collaborative actions. Physical domain collaborative actions refer to the adaptive physical adjustment operations of the data center's supporting infrastructure triggered to ensure the stability of the underlying environment after a sudden migration of large-scale computing resources. Based on the estimated power consumption increment of the newly deployed nodes, the system generates... Command data packets containing specific air conditioning temperature control parameters or backup power switching logic are distributed to the environmental control unit of the target data center via the IoT device control bus in the park. The power consumption increment estimate refers to the expected additional power consumption and physical heat generation of the computing nodes in the target data center after accepting new service instances migrated due to scheduling, compared to their current steady-state operation. This estimate is obtained by: extracting the steady-state resource utilization rate and historical actual power consumption baseline of the service instances to be migrated on the source node before the disaster from the cloud platform's historical monitoring logs; parsing the scheduling action policy set to read the number of CPU cores and memory capacity allocated to the target computing nodes for the newly migrated instances; and finally, combining the inherent unit computing power power consumption coefficient of the target computing node's underlying hardware with a linear mapping and conversion using the historical actual power consumption baseline and the allocated new resource scale to accurately calculate the additional power load value that the batch of services will bring to the target data center after cross-domain migration, thus obtaining the power consumption increment estimate. Upon receiving the command, physical devices automatically adjust the local air supply volume or preheat the backup power supply in the data center, thereby providing a reliable computing power carrying environment for the cloud platform at the spatial physical level.

[0059] The above-described structure establishes an integrated execution mechanism from abstract scheduling strategies to cloud platform infrastructure. This mechanism enables cross-domain collaborative closed-loop scheduling for container-level service migration, seamless SDN network traffic redirection, and data center physical environment control. Through state synchronization and service registration following health checks, the cross-fault domain instance migration process remains highly transparent to upper-layer users, effectively avoiding network black holes and prolonged service downtime caused by abrupt hard starts and stops. The introduction of a software-defined network controller completely decouples network topology from physical location, enabling the gradual transition of business traffic and the secure clearing of high-risk old routes. Simultaneously, a linkage command mechanism... This breaks down the barriers between cloud platform digital business management and park physical facility operation and maintenance, enabling infrastructure such as the air conditioning and power distribution systems of the target data center to adjust their physical domain actions in advance based on the spatial transfer of computing load. This virtual-physical linkage execution mechanism not only ensures the safe escape of business systems before the impact of disasters arrives, but also guarantees the temperature control and power stability of computing power migrating into the data center. It fundamentally prevents secondary disasters caused by local physical overload due to resource scheduling, and greatly improves the business continuity and overall robustness of the smart park base in complex physical disaster scenarios.

[0060] Step S7: Based on the updated environmental monitoring time series data and cross-fault domain deployment rules, perform physical propagation and business continuity verification on the executed scheduling action strategy set and physical domain collaborative actions, and update the propagation parameters and security candidate computing node sub-pool according to the verification results. For scheduling actions that fail verification, cancel them and switch to the pre-set isolation zone edge node for carrying.

[0061] Step S7 includes: determining the current physical propagation status of the target area based on the updated environmental monitoring time-series data, and determining the current business continuity status of the target business instance based on the updated cross-fault domain deployment rules; comparing the current physical propagation status with the executed scheduling status to obtain a first comparison result, and comparing the current business continuity status with the executed scheduling status to obtain a second comparison result; fusing the first comparison result and the second comparison result to generate a verification result; when the verification result is a failure, correcting the propagation parameters based on the updated environmental monitoring time-series data, and regenerating the security candidate computing node sub-pool based on the updated propagation parameters; canceling the failed scheduling action according to the time sequence opposite to the scheduling execution, and redirecting the bearer location of the target business instance to the pre-set isolation zone edge node.

[0062] Specifically, after the dispatch instructions are issued and the physical equipment is executed in conjunction, the initial assessment and dispatch of a one-way open loop is prone to safety deviations due to sudden environmental changes in the actual disaster spread patterns in real-world environments, which are highly uncertain and dynamically variable. This technical solution aims to solve the technical problem of the failure of the original dispatch decision caused by sudden changes in disaster spread. By constructing a closed-loop verification mechanism and a multi-layered isolation fallback mechanism based on actual environmental measurement data, the solution ensures the ultimate continuity of core business in smart parks under extreme and harsh conditions.

[0063] In practice, the current physical propagation state of the target area is determined based on the updated environmental monitoring time-series data, and the current business continuity state of the target business instance is determined based on the updated cross-fault domain deployment rules. The latest time-series data, such as wind speed, wind direction, or hazardous gas concentration, within the sampling window are obtained by subscribing to the park's environmental monitoring IoT platform. The current physical propagation state refers to the actual evolution boundary and velocity of the disaster in the current real physical space. A Gaussian plume diffusion model is used, with the latest acquired environmental monitoring time-series data as input parameters, to calculate and extrapolate the current actual coordinates and current diffusion velocity of the accident impact frontier, which are then used as the output. The Gaussian plume diffusion model is... The Gaussian plume diffusion model is a classic mathematical and physical model used in fluid mechanics and environmental engineering to simulate the transport of substances released from continuous point sources in three-dimensional physical space. Its derivation logic assumes that as hazardous substances such as leaked toxic gases or fire smoke drift downstream along the X-axis, their spatial concentration distribution along both the Y-axis and Z-axis follows a Gaussian normal distribution. In this scenario, real-time data such as wind speed, wind deflection angle, hazard release rate, and atmospheric stability collected by sensors are substituted into the Gaussian plume diffusion model to calculate the real-time concentration scalar field at each coordinate point within the park space. Subsequently, through... By extracting the three-dimensional envelope surface of equal concentration equal to the preset dangerous concentration threshold in the scalar field, the outermost boundary of the physical spread of the disaster can be accurately located, that is, the current actual coordinates of the frontier of the accident impact can be obtained. Then, the current spread rate can be solved by comparing the coordinate displacement rate of adjacent time slices. At the same time, the online status of migrated services can be measured in real time through the service orchestration platform. The current service continuity status refers to the actual operating indicators of the service after scheduling. The current number of online replicas, the distribution of replicas across fault domains, and the end-to-end response latency of each target service are statistically analyzed to form a set of measured service continuity values.

[0064] Subsequently, the first comparison result is obtained by comparing the current physical propagation state with the executed scheduling state, and the second comparison result is obtained by comparing the current business continuity state with the executed scheduling state. Specifically, the actual arrival time of the incident impact frontier to the target fault domain is calculated and compared with the predicted arrival time of the danger in the previous scheduling cycle. If the actual arrival time is earlier than the predicted time and the time difference exceeds the safety redundancy threshold, it is determined that the physical security has deviated, and the generated first comparison result is that the physical state has failed. Similarly, the measured end-to-end response latency and the actual number of online replicas are compared with the minimum service constraint indicators specified in the latest cross-fault domain deployment rules. If the measured response latency is higher than the minimum constraint requirement or the number of online replicas is less than the constraint number, it is determined that the business continuity has degraded, and the generated second comparison result is that the business state has failed. Then, the first comparison result and the second comparison result are merged to generate a verification result. In this process, the verification mechanism follows a strict dual security criterion. The verification result is successful only if both the first comparison result and the second comparison result are passed. If either comparison result shows a failure, the current verification result is determined to be a failure.

[0065] When the verification result fails, the propagation parameters are corrected based on the updated environmental monitoring time series data, and a new safety candidate computation node sub-pool is generated based on the updated propagation parameters. The latest wind direction deflection angle is extracted, and the actual propagation speed is obtained by calculating the ratio of the actual accident propagation distance to the evolution time. The original baseline propagation speed and wind direction correction coefficient are directly overwritten or weighted and corrected using the above real environmental feedback values. Here, the evolution time refers to the time from the initial moment when the hazard source first triggers the accident alarm or the baseline moment of the last effective environmental monitoring sampling in the system to the moment when the latest environmental monitoring time series data is obtained. The system extracts the actual time span up to a given moment and combines it with the absolute distance the disaster front actually travels in physical space within the corresponding time period to accurately calculate the actual propagation speed representing the current severe physical conditions. Subsequently, the updated propagation parameters are re-introduced into the disaster evolution spatiotemporal simulation process. By recalculating the danger arrival time interval, a risk state tensor matching the current severe conditions is generated. Based on this, the mask generation and candidate node screening process is executed again to eliminate nodes affected by the accelerated actual propagation speed, resulting in a new pool of safe candidate computation nodes.

[0066] Based on this, the scheduling actions that failed the verification are revoked according to the reverse timing sequence of the scheduling execution, and the bearer location of the target service instance is redirected to the pre-set isolation zone edge node. The reverse timing sequence refers to the "build first, then tear down" timing strategy adopted to prevent secondary service interruptions caused by the rollback process. First, a safe alternative node is selected from the new safe candidate computing node sub-pool. On the alternative node, a new service replica is launched and service registration and traffic redirection are completed. After the traffic transition is smooth, the old replica in the original threatened fault domain is shut down and the underlying computing resources are released. In extreme cases where, after multiple rounds of recalculation, the disaster impact range is too large and no alternative node that meets the cross-domain redundancy constraint can still be found in the new candidate sub-pool, the target bearer location is forcibly switched directly to the pre-set isolation zone edge node. The pre-set isolation zone edge node refers to the physical isolation zone edge node that is pre-set during the planning and design phase of the park. The location far exceeds the maximum extreme value of the impact radius of conventional disasters. The method of obtaining this value is as follows: during the initial configuration phase of the park system, based on the physical medium properties and maximum theoretical reserves of registered hazardous sources in the target area, the system calls the pre-set engineering disaster limit assessment model or directly associates with the corresponding national fire safety distance specifications to calculate the farthest absolute physical distance that disasters such as explosions, heat radiation, or toxic gas diffusion may affect under the most extreme environmental conditions. This absolute physical distance is recorded as a static threshold parameter and stored in the cloud platform configuration library. This serves as a hard geometric reference baseline for defining the spatial boundary of the isolation zone. It also has a backup calculation unit with an independent backup power supply circuit and a dedicated network exit. By using an emergency dedicated line and uninterruptible power supply combination independent of the backbone network, critical business that fails verification and has nowhere to be placed is forcibly diverted to such nodes, thereby maintaining the survival baseline of critical business when conventional defenses fail completely.

[0067] The above content covers a post-disaster dynamic resource scheduling defense line encompassing time-series feedback, dual-dimensional verification, reverse rollback, and isolation fallback. It addresses the crisis of open-loop scheduling failure and total loss of computing power caused by disaster evolution deviating from the expected model. By introducing measured data from the physical and information domains for dual-track verification, it breaks the limitations of traditional cloud-native scheduling open-loop execution, giving the system the adaptive ability to dynamically correct deviations and adjust parameters based on the actual disaster evolution. The reverse time-series strategy is used to execute smooth rollback, effectively avoiding secondary business interruptions caused by blind cancellation. At the same time, it cleverly sets up pre-positioned isolation zone edge nodes with multiple guarantees of physical, electrical, and network isolation. In the desperate situation where a major disaster causes the complete paralysis of the conventional cluster, it provides an indestructible digital safety cabin for the core lifeline business of the smart park, raising the resilience and bottom-line survivability of the entire system under extreme conditions to a new level.

[0068] Through the coordination of the above steps, this application significantly improves the resilience, adaptive correction capabilities, and ultimate survival rate of the underlying computing platform of smart parks in the face of extreme physical emergencies.

[0069] Example 2: The above describes a cloud-based smart park resource management method according to an embodiment of this application. The following describes a cloud-based smart park resource management system according to an embodiment of this application, such as... Figure 3 As shown in the figure, a cloud-based smart park resource management system in this application embodiment includes: The mapping unit is used to acquire data on hazard sources, environmental monitoring time series, target business dependencies, and computing node deployment within the target area, and to construct a heterogeneous time series diagram.

[0070] The risk unit is used to determine the danger arrival time interval, failure probability and isolation status of each computing node based on the heterogeneous time sequence diagram, and to generate a business survival budget vector based on the recovery time limit and priority level of the target business, thus forming a risk state tensor.

[0071] The masking unit is used to obtain the power supply fault domain, the network reachability of the isolation zone, and the service mutual exclusion co-location rules. Combined with the danger arrival time interval, it generates a set of prohibited migration edges pointing to computing nodes and a migration risk weight matrix between each computing node, forming a propagation constraint attention mask.

[0072] The filtering unit is used to filter nodes based on the propagation constraint attention mask, the remaining capacity of each computing node, cross-fault domain redundancy conditions, and business survival budget vector, forming a sub-pool of security candidate computing nodes corresponding to each target business.

[0073] The strategy unit is used to determine the danger co-domain exposure penalty and propagation front migration penalty for each candidate computing node based on the risk state tensor, construct the objective function in combination with the service interruption time loss, and solve the objective function under the condition of satisfying the preset minimum replica number constraint to generate a set of scheduling action strategies.

[0074] The linkage unit is used to perform cross-fault domain migration, replica startup, and network routing switching of business instances based on the scheduling action policy set, and to issue linkage instructions to the execution nodes in the target area to trigger physical domain collaborative actions.

[0075] Through the collaborative efforts of the aforementioned components, this application further enhances the resilience, adaptive correction capabilities, and ultimate survival rate of the underlying computing infrastructure of smart parks in the face of extreme physical emergencies.

[0076] In summary, the cloud-based smart park resource management method and system proposed in this application solves the long-standing industry problem of the disconnect between physical disasters in data centers and cloud computing power scheduling. By constructing heterogeneous time series graphs and multi-dimensional risk state tensors, it gains the ability to predict the dynamic spread of disasters in the physical world. By generating attention masks and multi-level hard pruning, it compresses the massive scheduling optimization space to a very small scale and uses a multi-objective heuristic evaluation model to achieve optimal decision-making that takes into account both spatiotemporal security. More importantly, this solution not only realizes cross-domain virtual-physical linkage execution of data network flow and underlying electromechanical infrastructure, but also introduces a closed-loop verification and correction engine based on environmental measurement data and a reverse-order fallback mechanism based on independent power supply edge nodes. This complete and rigorous technical defense line enables smart parks to calmly achieve lossless business escape and bottom-line protection of digital lifeline in the face of any kind of severe disasters, regardless of their evolution and sudden changes. It has extremely high engineering implementation value and broad industrial application prospects.

[0077] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system and unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0078] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0079] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A cloud platform-based smart park resource management method, characterized in that, The method includes: Step S1: Obtain data on hazard sources, environmental monitoring time series, target business dependencies, and computing node deployment within the target area, and construct a heterogeneous time series diagram; Step S2: Based on the heterogeneous timing diagram, determine the danger arrival time interval, failure probability and isolation status of each computing node, and generate a business survival budget vector according to the recovery time limit and priority level of the target business to form a risk state tensor; Step S3: Obtain the power supply fault domain, network reachability of the isolation zone, and service mutual exclusion co-location rules, and combine them with the danger arrival time interval to generate a set of prohibited migration edges pointing to the computing node and a migration risk weight matrix between each computing node, forming a propagation constraint attention mask. Step S4: Based on the propagation constraint attention mask, the remaining capacity of each computing node, the cross-fault domain redundancy condition, and the service survival budget vector, perform node screening to form a sub-pool of security candidate computing nodes corresponding to each target service. Step S5: Determine the danger co-domain exposure penalty and propagation front migration penalty corresponding to each candidate computing node based on the risk state tensor, construct an objective function in combination with the service interruption time loss, and solve the objective function under the condition of satisfying the preset minimum replica number constraint to generate a scheduling action strategy set; Step S6: Based on the scheduling action policy set, perform cross-fault domain migration, replica startup, and network routing switching of the service instance, and issue linkage instructions to the execution nodes in the target area to trigger physical domain collaborative actions.

2. The smart park resource management method based on a cloud platform according to claim 1, characterized in that, Following step S6, the following is also included: Step S7: Based on the updated environmental monitoring time series data and cross-fault domain deployment rules, perform physical propagation and service continuity verification on the executed scheduling action strategy set and the physical domain collaborative action, update the propagation parameters and the security candidate computing node sub-pool according to the verification results, and cancel the scheduling actions that fail the verification and switch to the pre-set isolation zone edge node for carrying.

3. The smart park resource management method based on a cloud platform according to claim 1, characterized in that, Step S1 includes: Hazard source nodes are determined based on hazard source data, sensing nodes are determined based on environmental monitoring time series data, business nodes are determined based on target business dependency data, and computing nodes and execution nodes are determined based on computing node deployment data. A physical propagation edge is established based on the spatial distribution of the hazard source node and the sensing node; a data flow edge is established based on the data calling rules between the sensing node and the business node; a deployment mapping edge is established based on the operational bearing association between the business node and the computing node; a network topology edge is established based on the multiple computing nodes; and a linkage control edge is established based on the physical spatial binding and control relationship between the computing node and the execution node. Integrate nodes and edges to generate a heterogeneous time sequence graph.

4. The smart park resource management method based on a cloud platform according to claim 1, characterized in that, Step S2 includes: By using the physical propagation edges in the heterogeneous timing graph, the danger arrival time interval of each computing node is deduced, and the failure probability and isolation status of each computing node are determined based on the danger arrival time interval. The arrival time interval of the danger, the failure probability, the isolation status, and the business survival budget vector are integrated to form a multi-dimensional risk state tensor, wherein the risk state tensor includes the spatial coordinate dimension of the computing node, the time series dimension, and the business requirement dimension corresponding to the business survival budget vector.

5. The smart park resource management method based on a cloud platform according to claim 1, characterized in that, Step S3 includes: Determine whether each computing node is within the power supply fault domain. If the target computing node is within the power supply fault domain, then the network topology edge pointing to the target computing node is assigned to the prohibited migration edge set. The initial migration risk weights between each computing node are determined based on the danger arrival time interval and the network reachability of the isolation zone, and the initial migration risk weights are adjusted by penalty according to the service mutual exclusion co-location rule to obtain the migration risk weight matrix. The target network topology edges are numerically masked by the set of prohibited migration edges, and the final propagation constraint attention mask is formed by combining it with the adjusted migration risk weight matrix.

6. The smart park resource management method based on a cloud platform according to claim 1, characterized in that, Step S4 includes: By using the propagation constraint attention mask to remove computational nodes associated with the prohibited inbound edge set, a preliminary set of filtered nodes is obtained. Determine whether the remaining capacity of each node in the initial screening node set meets the resource operation conditions required by the business survival budget vector, screen out the nodes that meet the resource operation conditions, and group and screen the nodes that meet the resource operation conditions according to the cross-fault domain redundancy conditions to ensure that the business replica distribution has isolation diversity, and obtain the security candidate computing node sub-pool corresponding to each target business.

7. The cloud-based smart park resource management method according to claim 1, characterized in that, Step S5 includes: Calculate the service interruption time loss for each candidate computing node in the security candidate computing node sub-pool, and configure the weight coefficients corresponding to the service interruption time loss, the dangerous common domain exposure penalty and the propagation front migration penalty respectively. Construct an objective function with the goal of minimizing the global scheduling cost by weighted summation. Under the condition of satisfying the minimum number of replicas constraint, the objective function is solved by heuristic search to obtain a set of scheduling action strategies. The set of scheduling action strategies includes at least the instance migration path between computing nodes, the expansion quantity of the target service, and the network switching routing table.

8. The smart park resource management method based on a cloud platform according to claim 7, characterized in that, Step S6 includes: The scheduling action policy set is parsed to obtain the instance migration path, the expansion quantity, and the network switching routing table. The cross-fault domain migration of the target service instance is performed in a state synchronization manner according to the instance migration path. The replica of the target service instance is launched and service is registered on the target node according to the expansion quantity. The network routing is switched through the software-defined network controller according to the network switching routing table. A linkage instruction bound to the current deployment node of the target service instance is generated and the linkage instruction is sent to the execution node in the target area to control the physical device to perform the physical domain collaborative action.

9. A smart park resource management method based on a cloud platform according to claim 2, characterized in that, Step S7 includes: The current physical propagation status of the target area is determined based on the updated environmental monitoring time-series data, and the current service continuity status of the target service instance is determined based on the updated cross-fault domain deployment rules. A first comparison result is obtained by comparing the current physical propagation state with the executed scheduling state, and a second comparison result is obtained by comparing the current service continuity state with the executed scheduling state. The first comparison result and the second comparison result are merged to generate a verification result. When the verification result fails, the propagation parameters are corrected according to the updated environmental monitoring time series data, and a new security candidate computing node sub-pool is generated based on the updated propagation parameters. The scheduling action that failed the verification was revoked according to the reverse timing of the scheduling execution, and the bearer location of the target service instance was redirected and switched to the edge node of the preset isolation zone.

10. A cloud-based smart park resource management system, used to implement the cloud-based smart park resource management method as described in any one of claims 1-9, characterized in that, The system includes: The mapping unit is used to acquire data on hazard sources, environmental monitoring time series, target business dependencies, and computing node deployment within the target area, and to construct a heterogeneous time series diagram. The risk unit is used to determine the danger arrival time interval, failure probability and isolation status of each computing node based on the heterogeneous time sequence diagram, and generate a business survival budget vector according to the recovery time limit and priority level of the target business, forming a risk state tensor. The masking unit is used to obtain the power supply fault domain, the network reachability of the isolation zone, and the service mutual exclusion co-location rules. Combined with the danger arrival time interval, it generates a set of prohibited migration edges pointing to the computing node and a migration risk weight matrix between each computing node, forming a propagation constraint attention mask. The filtering unit is used to filter nodes based on the propagation constraint attention mask, the remaining capacity of each computing node, the cross-fault domain redundancy condition and the service survival budget vector, to form a sub-pool of security candidate computing nodes corresponding to each target service. The strategy unit is used to determine the danger co-domain exposure penalty and propagation front migration penalty corresponding to each candidate computing node based on the risk state tensor, construct an objective function in combination with the service interruption time loss, and solve the objective function under the condition of satisfying the preset minimum replica number constraint to generate a set of scheduling action strategies. The linkage unit is used to perform cross-fault domain migration, replica startup, and network routing switching of service instances based on the scheduling action policy set, and to issue linkage instructions to the execution nodes in the target area to trigger physical domain collaborative actions.

Images