A smart park resource scheduling method and system

Abstract

The present application relates to the technical field of resource scheduling, in particular to a smart park resource scheduling method and system, the method comprising: constructing a heterogeneous information network representing the current state of the park; using a graph attention convolutional network to extract features of the heterogeneous information network, generating a state embedding vector that integrates high-order neighborhood information; inputting the state embedding vector into a scheduling strategy model to predict the expected change in future park energy consumption, security index and traffic efficiency of the candidate scheduling action; determining the optimization target of the current scheduling period from the preset operation paradigm according to the state embedding vector and external management instruction, and assigning a dynamic weight; calculating the comprehensive utility score of each candidate scheduling action, and issuing the candidate scheduling action with the highest score as the final scheduling instruction for execution. The present application improves the fine level of park management and overall operating efficiency.

Description

Technical Field

[0001] The present invention relates to the technical field of resource scheduling, and in particular to a method and system for scheduling resource in a smart park. Background Art

[0002] With the development of the Internet of Things, big data, and artificial intelligence technologies, smart campuses have become complex systems integrating security, energy consumption, transportation, and office functions. Efficiently and intelligently coordinating the massive heterogeneous resources within the campus to improve overall operational efficiency, reduce operating costs, and ensure safety remains a core challenge for current campus management. Existing campus resource management systems mostly adopt a divide-and-conquer strategy, with each subsystem (such as building automation, video surveillance, and access control) operating independently. This creates severe data silos and hinders global coordination. Their scheduling logic often relies on preset static rules or simple statistical models based on historical data. This approach is slow to respond and cannot effectively address dynamic and sudden changes in campus traffic, vehicle flow, environmental factors, and other factors. This results in limited refinement and intelligence in resource allocation, often leading to resource waste and untimely security responses.

[0003] To address these issues, some research has attempted to incorporate machine learning models for prediction and decision-making. However, these approaches still face limitations in their application. For one thing, most models focus on mining correlations between data rather than deeper causal relationships. For example, a model may find a positive correlation between foot traffic and energy consumption, but it cannot accurately predict the marginal causal effect of a specific action, such as "raising the air conditioning temperature in area A by one degree," on future energy consumption and occupant comfort, thus compromising decision-making accuracy. Furthermore, existing scheduling strategies often employ rigid optimization objectives, such as the single pursuit of maximizing energy savings, lacking flexibility. In special scenarios such as sudden security incidents or major event support, the system is unable to dynamically balance and switch between conflicting objectives, such as energy conservation, safety, and efficiency, based on real-time situational awareness and management instructions. Furthermore, how to effectively integrate multimodal sensor data to construct a unified model that comprehensively represents the complex coupling relationships between physical entities, virtual resources, and dynamic events, and based on this model for high-level reasoning, remains a challenge that current technologies have yet to adequately address. Summary of the Invention

[0004] To solve the technical problem of how to construct a model of complex coupling relationships and perform reasoning, the present invention provides the following solution.

[0005] A method for resource scheduling in a smart park, S1, obtains multimodal real-time sensor data in the park and constructs a heterogeneous information network that represents the current state of the park, wherein the nodes of the heterogeneous information network include physical resource nodes, virtual resource nodes and event entity nodes, and the edges of the heterogeneous information network represent the physical adjacency, functional coupling and logical subordination between nodes; S2, uses a graph attention convolutional network to extract features from the heterogeneous information network and generates a state embedding vector that integrates high-order neighborhood information; S3, inputs the state embedding vector into a scheduling strategy model, and uses the scheduling strategy model to The decoder of the type generates a set of candidate scheduling actions and uses its causal forest module to predict the expected changes in the future campus energy consumption, security index and traffic efficiency caused by the candidate scheduling actions; S4, based on the state embedding vector and external management instructions, determines the optimization target of the current scheduling cycle from the preset operating paradigm, and assigns dynamic weights to the three dimensions of campus energy consumption, security risk and traffic efficiency; S5, based on the dynamic weights and the predicted expected changes, calculates the comprehensive utility score of each candidate scheduling action, and issues the candidate scheduling action with the highest score as the final scheduling instruction for execution.

[0006] Furthermore, a heterogeneous information network representing the current state of the park is constructed, including: defining the sensor equipment, control equipment and infrastructure within the park as physical resource nodes; defining the logical areas divided according to functions or management needs as virtual resource nodes; defining preset alarms or status changes as event entity nodes; and establishing corresponding physical adjacent edges, functional coupling edges or logical subordinate edges based on the physical location relationship, functional collaboration relationship or logical management relationship between nodes.

[0007] Furthermore, the preset alarms or status changes include: illegal intrusion and fire alarms.

[0008] Furthermore, the use of a graph attention convolutional network to extract features from the heterogeneous information network includes: inputting the adjacency relationship and initial node features of the heterogeneous information network into a multi-layer graph attention network; assigning attention weights to the neighboring nodes of each node through a multi-head attention mechanism, and performing weighted aggregation on the features of the neighboring nodes based on the weights to update the features of the corresponding nodes; and finally, through a graph-level pooling operation, aggregating the features of all nodes into the state embedding vector.

[0009] Furthermore, the graph attention convolutional network is a HAN model.

[0010] Furthermore, a set of candidate scheduling actions is generated using the decoder of the scheduling strategy model, including: inputting the state embedding vector into the decoder of the scheduling strategy model, sampling and decoding it from the learned potential distribution, and generating a preset number of candidate scheduling actions containing specific resource control parameters.

[0011] Furthermore, the scheduling strategy model adopts variational autoencoder VAE.

[0012] Furthermore, determining the optimization target of the current scheduling cycle and assigning dynamic weights includes: switching to the corresponding operating paradigm when a specified external management instruction is received, or when the state embedding vector reflects that the park state meets the preset switching conditions; and setting a set of corresponding weight coefficients for the three dimensions of park energy consumption, security risk and traffic efficiency according to the current operating paradigm.

[0013] Furthermore, the calculation of the comprehensive utility score of each candidate scheduling action includes: for each candidate scheduling action, its comprehensive utility score U is calculated by the following formula: U=Wenergy consumption×ΔE+Wsecurity×ΔS+Wefficiency×ΔT, wherein Wenergy consumption, Wsecurity, and Wefficiency are the dynamic weights of the park energy consumption, security risk, and traffic efficiency determined in the current scheduling cycle; ΔE, ΔS, and ΔT are the normalized values ​​of the predicted changes in park energy consumption, security risk, and traffic efficiency, respectively.

[0014] The present invention also provides a smart park resource scheduling system, including a processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement the above-mentioned smart park resource scheduling method.

[0015] Compared with the existing technology, the beneficial effects of the present invention are: by constructing a heterogeneous information network including physical resources, virtual resources and event entities, various types of information that were originally isolated in the park are uniformly modeled, achieving a comprehensive and in-depth characterization of the park's operating status, and overcoming the problems of data fragmentation and one-sided models in the existing technology. More importantly, by utilizing the causal forest module, this method can accurately predict the causal effects of specific scheduling actions on key indicators such as park energy consumption, security risks, and traffic efficiency, rather than just staying at the correlation analysis. This changes the decision-making basis from fuzzy correlation to quantitative causality, significantly improving the accuracy and reliability of resource scheduling instructions. In addition, by assigning weights to different optimization dimensions and combining causal effect prediction for comprehensive utility evaluation, the present invention can effectively balance multiple goals such as park energy consumption, security risks, and traffic efficiency, and generate a scheduling plan that best meets the current overall interests based on the management needs and operating status of specific scenarios, thereby improving the refinement level of park management and overall operational efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 is a schematic diagram schematically illustrating a heterogeneous information network according to the present invention;

[0017] Figure 2is a schematic diagram schematically illustrating feature extraction using a graph attention convolutional network according to the present invention;

[0018] Figure 3 FIG. 1 is a schematic diagram schematically illustrating dynamic weight allocation according to the present invention. DETAILED DESCRIPTION

[0019] The technical solutions in the embodiments of the present invention will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present invention.

[0020] A smart park resource scheduling method, comprising:

[0021] S1, obtain multimodal real-time sensor data within the park and build a heterogeneous information network that represents the current state of the park. The nodes of the heterogeneous information network include physical resource nodes, virtual resource nodes and event entity nodes. The edges of the heterogeneous information network represent the physical adjacency, functional coupling and logical subordination between nodes.

[0022] Specifically, the image stream of video surveillance, temperature and humidity readings of environmental sensors, card swiping records of access control systems, and personnel positioning beacon data are aggregated in real time through the MQTT protocol and HTTPAPI interface; these data constitute the multimodal real-time sensing data.

[0023] In an optional embodiment, the construction of a heterogeneous information network representing the current state of the park includes: Figure 1 As shown in the figure, the sensing devices (such as camera C1, temperature sensor T1), control devices (such as access control D1, air conditioner ACU2) and infrastructure in the park are defined as physical resource nodes; the logical areas divided according to functions or management needs (such as office area A, underground garage, public leisure area) are defined as virtual resource nodes; the preset alarms or status changes (such as time E2: illegal intrusion) are defined as event entity nodes; and based on the physical location relationship, functional collaboration relationship or logical management relationship between nodes, corresponding physical adjacent edges, functional coupling edges or logical subordinate edges are established.

[0024] For example, when a fire alarm occurs, the system will generate an event entity node of fire alarm E1, which contains information such as the alarm level and occurrence time.

[0025] Specifically, different types of nodes are connected through edges, forming a complex network. For example, camera C1 located in Office Area A is connected to the virtual resource node Office Area A via a logical subordinate edge, indicating their jurisdictional relationship. Camera C1 is also connected to its physically adjacent access control point D1 via a physical adjacency edge. If event E2 is set to trigger high-definition recording with camera C1, a functional coupling edge is established between event E2 and camera C1. This heterogeneous information network comprehensively and profoundly depicts the physical layout, functional linkage, and logical management status of the entire campus.

[0026] S2, uses a graph attention convolutional network to extract features from the heterogeneous information network and generate a state embedding vector that integrates high-order neighborhood information.

[0027] Specifically, the graph attention convolutional network can adopt the heterogeneous graph attention network, namely the HAN model.

[0028] In an optional embodiment, the use of a graph attention convolutional network to extract features from the heterogeneous information network includes: inputting the adjacency relationship and initial node features of the heterogeneous information network into a multi-layer graph attention network; assigning attention weights to the neighboring nodes of each node through a multi-head attention mechanism, and performing weighted aggregation on the features of the neighboring nodes based on the weights to update the features of the corresponding nodes; and finally, through a graph-level pooling operation, aggregating the features of all nodes into the state embedding vector.

[0029] Specifically, combined Figure 2 Consider a camera C1 node in a heterogeneous information network. Its initial features may include data such as operating status, power consumption, and video bitrate. In the first layer of the graph attention network, the importance of all neighboring nodes of camera C1, such as the virtual nodes Office Area A, Access Control D1, and Event E2, to C1 is calculated. If Event E2 is currently occurring, the multi-head attention mechanism assigns a higher attention weight (for example, 0.8) to Event E2, while assigning a lower weight (for example, 0.1) to Access Control D1 and Office Area A, which are in normal states.

[0030] Specifically, the model weights and sums the features of neighboring nodes according to the weights, thereby updating the features of the camera C1 node so that it not only contains its own information, but also incorporates important state information of neighboring nodes. The process propagates layer by layer in the network. For example, after three-layer network calculations, the features of each node aggregate the information within its three-hop neighbors. The graph-level pooling operation aggregates the final 64-dimensional feature vectors of all nodes, such as a total of 1,000 nodes in the park, into a single, for example, 256-dimensional state embedding vector by averaging or summing. The vector is a digital expression of the current macroscopic state of the entire park.

[0031] S3: Input the state embedding vector into a scheduling strategy model, use the decoder of the scheduling strategy model to generate a set of candidate scheduling actions, and use its causal forest module to predict the expected changes in the future campus energy consumption, security index and traffic efficiency caused by the candidate scheduling actions.

[0032] Specifically, the scheduling strategy model may use a variational autoencoder (VAE). The decoder of the variational autoencoder (VAE) samples in the learned action potential space to generate multiple candidate scheduling actions.

[0033] In an optional embodiment, a decoder of the scheduling policy model is used to generate a set of candidate scheduling actions, including: inputting the state embedding vector into the decoder of the scheduling policy model, sampling and decoding it from the learned potential distribution, and generating a preset number of candidate scheduling actions containing specific resource control parameters.

[0034] The candidate scheduling actions include: for example, action one is to increase the air conditioning temperature in area A by 2 degrees, action two is to dispatch 2 additional security personnel to patrol area B, and action three is to adjust the green light duration of the signal light at exit C by 10 seconds; for each candidate scheduling action, a pre-trained causal forest model is used to predict the specific impact of each candidate scheduling action on the total energy consumption, security risk score and traffic efficiency (which can be quantified as the average vehicle travel time) of the park in the next 30 minutes. For example, it is predicted that the causal effect of action one on the energy consumption of the park is to reduce it by 0.8 kWh, and the effect on security risk and traffic efficiency is close to zero.

[0035] Specifically, taking the previously generated 256-dimensional state embedding vector as input, the variational autoencoder's decoder samples from its internally learned space of efficient scheduling strategies, generating, for example, 50 candidate scheduling actions. One candidate action might be to raise the air conditioning temperature in Office Area A by 2 degrees Celsius and set the two charging stations at the parking lot entrance to fast charging mode. Another candidate action might be to turn off 50% of the landscape lighting in the leisure area and deploy an additional security drone to patrol the north fence.

[0036] Specifically, for each generated candidate action, the causal forest module is called to predict its effect. For example, for the actions of raising the air conditioning temperature and turning on fast charging, the causal forest model, based on historical data analysis, predicts that these actions will increase the total energy consumption of the park by 30 kWh within the next hour, keep the security risk unchanged, and increase traffic efficiency by 0.5 units. For the actions of turning off the lights and deploying additional drones, the model may predict a reduction in energy consumption by 15 kWh, a reduction in security risk by 2.3 units, and unchanged traffic efficiency. In this way, each abstract scheduling action is given a clear, quantifiable expectation of its impact on future states.

[0037] S4, based on the state embedding vector and external management instructions, determine the optimization target of the current scheduling cycle from the preset operation paradigm, and assign dynamic weights to the three dimensions of the park energy consumption, security risk and traffic efficiency.

[0038] Specifically, the operating paradigms include energy conservation first, safety first, and efficiency first paradigms.

[0039] Combine Figure 3 , using a multi-layer perceptron classifier (such as Figure 3 The MLP classifier in the example above takes the state embedding vector and a vector representing external management instructions as input. If there are no external instructions and the state vector indicates that the campus load is stable, the MLP classifier selects the energy-saving priority paradigm and outputs weights of 0.7 for energy consumption, 0.1 for security index, and 0.2 for traffic efficiency. If an external management instruction related to a fire alarm is received, the MLP classifier immediately switches to the safety priority paradigm and outputs weights of 0.05 for energy consumption, 0.9 for security index, and 0.05 for traffic efficiency.

[0040] In an optional embodiment, determining the optimization target of the current scheduling cycle and assigning dynamic weights includes: switching to the corresponding operating paradigm when a specified external management instruction is received, or when the state embedding vector reflects that the park state meets the preset switching conditions; and setting a set of corresponding weight coefficients for the three dimensions of park energy consumption, security risk and traffic efficiency according to the current operating paradigm.

[0041] Specifically, combined Figure 3 During weekday nights, the system automatically or upon administrator instruction switches to an energy-saving priority mode. Under this energy-saving priority mode, the optimization objective prioritizes energy conservation, with weighting coefficients set to 0.7 for campus energy consumption, 0.2 for security risk, and 0.1 for traffic efficiency. When the analysis of the state embedding vector indicates a sharp increase in the flow of people and vehicles on campus, meeting the conditions for entering peak hours in the morning and evening, the system automatically switches to a peak traffic mode (i.e., an efficiency-first mode). To prioritize the smooth passage of people and vehicles, the weights are dynamically adjusted to 0.2 for campus energy consumption, 0.2 for security risk, and 0.6 for traffic efficiency. If an abnormal intrusion alarm is detected in one or more areas, the system immediately switches to a safety-first mode, increasing the security weight to 0.9 and reducing the other two weights accordingly, ensuring that all scheduling decisions prioritize maximizing campus safety.

[0042] S5, based on the dynamic weight and the predicted expected change, calculating the comprehensive utility score of each candidate scheduling action, and issuing the candidate scheduling action with the highest score as the final scheduling instruction for execution.

[0043] In an optional embodiment, the calculation of the comprehensive utility score of each candidate scheduling action includes: for each candidate scheduling action, the comprehensive utility score U is calculated by the following formula: U=W 能耗 ×ΔE+W 安防 ×ΔS+W 效率 ×ΔT where W 能耗 、W 安防 、W 效率 The dynamic weights of the park energy consumption, security risk, and traffic efficiency determined in the current scheduling cycle; ΔE, ΔS, and ΔT are the normalized values ​​of the predicted changes in park energy consumption, security risk, and traffic efficiency, respectively. Specifically, assuming that the current system is in the energy-saving priority paradigm, the weight coefficient is set to W 能耗 Equal to 0.7, W 安防 Equal to 0.2, W 效率 Equal to 0.1. There are two candidate actions, A and B. By comparison, if the score of action A is higher than that of action B, then candidate action A is selected because it best fits the current operating paradigm with energy conservation as the primary goal.

[0044] The present invention also relates to a smart park resource scheduling system, including a processor and a memory, the memory storing a computer program, the processor can interact with the memory, can call the computer program (for example, through a bus), and then the processor executes the computer program. When the computer program is executed by the processor, a smart park resource scheduling method of the above embodiment is implemented.

[0045] In the present invention, the aforementioned memory may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. For example, the computer-readable storage medium may be any suitable magnetic storage medium or magneto-optical storage medium. This specification and the accompanying drawings are merely illustrative of the present invention as defined by the appended claims. Obviously, various modifications and variations may be made by those skilled in the art without departing from the scope of the present invention.

Claims

1. A smart park resource scheduling method, characterized in that: include: S1, acquire multimodal real-time sensor data within the park and construct a heterogeneous information network representing the current state of the park. The nodes of the heterogeneous information network include physical resource nodes, virtual resource nodes, and event entity nodes. The edges of the heterogeneous information network represent the physical adjacency, functional coupling, and logical subordination between nodes; Define logical areas divided according to functions or management requirements as virtual resource nodes; define preset alarms or status changes as event entity nodes; S2, using a graph attention convolutional network to extract features from the heterogeneous information network and generate a state embedding vector that incorporates high-order neighborhood information; S3: Input the state embedding vector into a scheduling strategy model, use the decoder of the scheduling strategy model to generate a set of candidate scheduling actions, and use its causal forest module to predict the expected changes in the future campus energy consumption, security index, and traffic efficiency caused by the candidate scheduling actions; S4, determining the optimization target of the current scheduling cycle from a preset operating paradigm based on the state embedding vector and external management instructions, and assigning dynamic weights to the three dimensions of park energy consumption, security risk, and traffic efficiency; S5, based on the dynamic weight and the predicted expected change, calculating the comprehensive utility score of each candidate scheduling action, and issuing the candidate scheduling action with the highest score as the final scheduling instruction for execution.

2. The method according to claim 1, characterized in that Build a heterogeneous information network that represents the current state of the park, including: The sensing equipment, control equipment and infrastructure within the park are defined as physical resource nodes; and based on the physical location relationship, functional collaboration relationship or logical management relationship between the nodes, corresponding physical adjacent edges, functional coupling edges or logical subordinate edges are established.

3. The method according to claim 2, characterized in that Preset alarms or status changes include: illegal intrusion and fire alarms.

4. The method according to claim 1, wherein The method of using a graph attention convolutional network to extract features from the heterogeneous information network includes: The adjacency relationship and initial node features of the heterogeneous information network are input into a multi-layer graph attention network; attention weights are assigned to the neighboring nodes of each node through a multi-head attention mechanism, and the features of the neighboring nodes are weightedly aggregated based on the weights to update the features of the corresponding nodes; finally, through a graph-level pooling operation, the features of all nodes are aggregated into the state embedding vector.

5. The method according to claim 1, wherein The graph attention convolutional network is a HAN model.

6. The method according to claim 1, characterized in that A decoder of the scheduling policy model is used to generate a set of candidate scheduling actions, including: The state embedding vector is input into the decoder of the scheduling policy model, sampled and decoded from the learned potential distribution, and a set of candidate scheduling actions with a preset number of specific resource control parameters is generated.

7. The method according to claim 6, characterized in that The scheduling strategy model adopts variational autoencoder VAE.

8. The method according to claim 1, characterized in that Determining the optimization target of the current scheduling period and assigning a dynamic weight includes: When a specified external management instruction is received, or when the state embedding vector reflects that the park state meets the preset switching conditions, it switches to the corresponding operating paradigm; and according to the current operating paradigm, a set of corresponding weight coefficients are set for the three dimensions of park energy consumption, security risk and traffic efficiency.

9. The method according to claim 1, characterized in that Calculating the comprehensive utility score of each candidate scheduling action includes: For each candidate scheduling action, its comprehensive utility score U is calculated by the following formula: U=W 能耗 ×ΔE+W 安防 ×ΔS+W 效率 ×ΔT where W 能耗 、W 安防 、W 效率 The dynamic weights of the park's energy consumption, security risk, and traffic efficiency determined in the current scheduling cycle; ΔE, ΔS, and ΔT are the normalized values ​​of the predicted changes in the park's energy consumption, security risk, and traffic efficiency, respectively.

10. A smart park resource scheduling system, characterized in that: It includes a processor and a memory, the memory stores a computer program, and the processor executes the computer program to implement the smart park resource scheduling method described in any one of claims 1 to 9.

Images