Intelligent exhaust dynamic scheduling method and system based on federated learning and edge computing

Abstract

The application provides an intelligent exhaust dynamic scheduling method and system based on federated learning and edge computing, which comprises the following steps: taking a current edge node as a target node, collecting data in units of the target node, and obtaining a predicted exhaust intensity value through a deep neural network; based on the predicted exhaust intensity value, combining the local state of the current target node, constructing a scheduling strategy model based on reinforcement learning, and outputting the parameters of the scheduling strategy model and the local interaction trajectory data of the target node; calculating the structure disturbance perception weight according to the local interaction trajectory data of each edge node to quantify the strategy contribution degree of each node; according to the structure disturbance perception weight, the federated server calculates the aggregation weight normalization result of each node, then performs fusion calculation on the strategy parameters to obtain global guiding strategy parameters; and applying the global guiding strategy parameters to the physical exhaust equipment for intelligent scheduling of the physical exhaust equipment.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of federated learning and edge computing, and particularly relates to an intelligent exhaust dynamic scheduling method and system based on federated learning and edge computing. BACKGROUND

[0002] In modern complex spatial environments such as urban infrastructure, underground tunnels, large industrial parks, and intelligent buildings, the exhaust system is a key support for environmental regulation and pollution management. Its intelligent scheduling capability directly affects air quality, energy consumption, and personnel safety. With the acceleration of urbanization and the improvement of environmental standards, the exhaust system is evolving from traditional static, rule-driven control mode to data-driven, autonomous decision-making system. However, the current widely deployed exhaust control system generally relies on preset timing programs, threshold triggering mechanisms, or centralized computing and instruction issuing based on a central server. This architecture has obvious bottlenecks: on the one hand, the exhaust demand of each region has obvious spatial and temporal differences, influenced by pollution source intensity, ventilation path, spatial structure, and human flow dynamics, and centralized control is difficult to accurately respond to local fine-grained environmental changes; on the other hand, centralized systems rely heavily on the network, and when the number of nodes is large or the sensors are densely distributed, the computing and communication overheads increase rapidly, the system response time delay becomes larger, and even may face the risk of bottleneck failure or communication interruption. In addition, many exhaust scenarios (such as industrial parks, hospitals, underground buildings, etc.) involve sensitive data or high-frequency sensor information, and direct uploading to the central platform will face a higher risk of data privacy leakage. Therefore, in the multi-region, high-dynamic, and multi-constrained exhaust environment, a new exhaust scheduling architecture is needed that has distributed autonomous decision-making capability, real-time dynamic response capability, and at the same time takes into account privacy protection and system collaboration efficiency.

[0003] Although existing research attempts to introduce some intelligent means to alleviate the above problems, such as using machine learning to predict exhaust demand, using local feedback to adjust the intensity of the fan, or introducing Internet of Things technology to improve information collection density, it still cannot fundamentally solve the core challenges of the exhaust system, such as lack of autonomy, insufficient dynamic optimization, weak system coordination, and lack of data privacy protection. The single-point prediction and control method cannot cope with the mutual influence between devices, such as the feedback regulation of upstream exhaust behavior on downstream pressure; while the central server unified optimization has a global perspective, but it cannot reflect the rapid changes of the edge state in real time, especially in the context of increasing number of nodes, the scheduling delay and precision decline problem is more prominent. More importantly, there is currently a lack of an exhaust scheduling system that can fully utilize the local knowledge of each edge node and achieve resource coordination and strategy consistency on a global scale. Therefore, the traditional control system and some intelligent attempts still have obvious deficiencies in multi-node collaborative scheduling, data security protection, and dynamic decision strategy optimization, which restricts the landing and promotion of intelligent exhaust systems in large-scale and complex environments. SUMMARY

[0004] The purpose of the present application is to propose an intelligent exhaust dynamic scheduling method and system based on federated learning and edge computing, which combines the autonomy of edge computing and the federated collaborative mechanism to achieve distributed learning and global collaborative optimization of exhaust strategies while ensuring data privacy.

[0005] To achieve the above purpose, in the first aspect of the present application, an intelligent exhaust dynamic scheduling method based on federated learning and edge computing is provided, which comprises the following steps:

[0006] Taking the corresponding control unit of each independent exhaust area as an edge node, setting the current edge node as the target node, collecting data from the target node, collecting multi-source sensor data of the area corresponding to the target node, and obtaining the predicted exhaust intensity value through a deep neural network;

[0007] Based on the predicted exhaust intensity value, combined with the local state of the current target node, a scheduling strategy model based on reinforcement learning is constructed, and the parameters of the scheduling strategy model and the local interaction trajectory data of the target node are output; wherein the local interaction trajectory data includes the predicted exhaust intensity value, the local state, the current scheduling behavior, the current scheduling reward, the actual exhaust intensity value, and the air disturbance influence term;

[0008] According to the local interaction trajectory data of each edge node, the structural disturbance perception weight is calculated to quantify the strategy contribution degree of each node; according to the structural disturbance perception weight, the federated server calculates the aggregation weight normalization result of each node, and then performs fusion calculation of the strategy parameters to obtain the global guidance strategy parameters;

[0009] The global guiding strategy parameter is actually applied to the physical exhaust equipment, and intelligent scheduling of the physical exhaust equipment is performed.

[0010] Further, the multi-source sensor data of the area corresponding to the target node is collected, specifically including:

[0011] The environmental state data of the node is collected at each period t, including pollutant concentration, air pressure, wind speed, temperature, and equipment operating state.

[0012] Further, the deep neural network is a pollution-driven exhaust prediction model, and the calculation is specifically as follows:

[0013] ;

[0014] wherein, is a predicted exhaust intensity value, and the value closest to 1 indicates that the current area has the most urgent exhaust demand; is a node environmental state vector at the current time; is a weight matrix for linear mapping of the basic state output; is a Sigmoid function for normalizing the final prediction value to ; is a gradient response intensity of the state to the pollution concentration, used to strengthen the identification of the rapid rising trend of the pollution; is a regular weight of the pollution trend term; is a set of exhaust nodes adjacent to ; is a pollution coupling influence degree of the node to ; is a difference degree of the current node and the adjacent node state, indicating the risk of possible pollution propagation; is a coupling interference regular weight.

[0015] Further, the parameters of the scheduling strategy model are obtained by updating a strategy gradient algorithm containing a target deviation penalty and a disturbance response term; wherein the target deviation penalty is obtained according to the error between the predicted exhaust intensity value and the actual exhaust intensity value; and the disturbance response term is an air disturbance influence term for punishing the negative air flow impact on other areas caused by the exhaust behavior.

[0016] Further, the structural disturbance perception weight is calculated according to the local interaction trajectory data of each edge node, specifically as follows:

[0017] The structural disturbance perception weight is calculated in combination with the difference absolute value between the predicted exhaust intensity value and the actual exhaust intensity value, the action intensity of the air disturbance influence term, and the structural priority weight of the node.

[0018] Further, after obtaining the global guiding strategy parameter, for the node with long-term deviation of error, the model parameter is forced to update to the global guiding strategy parameter, and for the node with stable error within the acceptable range, only the global guiding strategy parameter is used for light fusion, and the local strategy adaptive ability is retained.

[0019] Further, the global guiding strategy parameter is actually applied to the physical exhaust equipment to perform intelligent scheduling of the physical exhaust equipment, specifically including:

[0020] Each edge node uses the synchronized global guiding strategy parameter to generate a current scheduling action in combination with the current predicted exhaust intensity value and the equipment running state, wherein the scheduling action is used to determine the control behavior of the equipment in the current period.

[0021] Further, in the intelligent scheduling of the physical exhaust equipment, further includes:

[0022] Based on the node-related attribute calculation of the exhaust behavior between the current node and its adjacent nodes, the system disturbance index of the current node in the current scheduling period is determined, indicating the potential physical conflict risk.

[0023] When the system disturbance index exceeds the safety threshold set by the system, the current action may cause air flow instability to the system, and disturbance suppression is required before execution to obtain a reconciled executable action.

[0024] All reconciled executable actions are issued as actual control instructions to the device end for execution.

[0025] In a second aspect of the application, an intelligent exhaust dynamic scheduling system based on federated learning and edge computing is provided, which includes:

[0026] A demand prediction module is configured to use each independent exhaust area corresponding to a control unit as an edge node, set a current edge node as a target node, collect multi-source sensor data of the area corresponding to the target node through data collection in units of the target node, and obtain a predicted exhaust intensity value through a deep neural network.

[0027] An edge individualized learning module is configured to construct a scheduling strategy model based on reinforcement learning based on the predicted exhaust intensity value and the local state of the current target node, output parameters of the scheduling strategy model and local interaction trajectory data of the target node, and the local interaction trajectory data includes the predicted exhaust intensity value, the local state, the current scheduling behavior, the current scheduling reward, the actual exhaust intensity value and the air disturbance influence term.

[0028] A federal collaborative strategy aggregation module is configured to calculate a structure disturbance perception weight according to local interaction trajectory data of each edge node to quantify the strategy contribution of each node, and the federal server calculates an aggregation weight normalization result of each node according to the structure disturbance perception weight, and then performs fusion calculation of the strategy parameter to obtain a global guiding strategy parameter.

[0029] A conflict detection and strategy reconciliation execution module is configured to apply the global guiding strategy parameter to the physical exhaust equipment to perform intelligent scheduling of the physical exhaust equipment.

[0030] The present application has at least the following beneficial technical effects:

[0031] The present application proposes an intelligent dynamic scheduling method and system architecture for a multi-region exhaust system, which is characterized by combining the autonomous ability of edge computing and the federal collaborative mechanism to realize distributed learning and global collaborative optimization of exhaust strategy under the premise of protecting data privacy. The present application constructs an edge intelligent agent network with local intelligence, autonomous perception and sustainable learning ability. Each edge node can respond quickly to local exhaust demand according to real-time sensor information, and at the same time, through the collaborative mechanism, the scheduling experience of other nodes is fused, thereby breaking through the limitations of single-point prediction under the condition of multi-region interference. In addition, the system constructs a structured modeling mechanism based on the relationship between air flow channels, exhaust equipment and regional load, supports accurate modeling and behavior prediction of the exhaust system state in a dynamic environment, and provides semantic support for the generation of multi-objective scheduling strategy.

[0032] Compared with the existing centralized control system, the present application not only significantly reduces the computing and communication burden of the center server, but also improves the response efficiency, adaptability and robustness of the system.

[0033] At the same time, by establishing a collaborative learning mechanism that exchanges model parameters instead of raw data between edge nodes, global strategy fusion is realized without exposing sensitive data, effectively addressing the technical bottlenecks of current intelligent exhaust systems in data privacy, safe scheduling and complex environment adaptation. BRIEF DESCRIPTION OF DRAWINGS

[0034] The present application is further illustrated by the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present application. For ordinary skilled in the art, other drawings can be obtained without creative labor on the premise of the following drawings.

[0035] Figure 1 The present application is further illustrated by the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present application. For ordinary skilled in the art, other drawings can be obtained without creative labor on the premise of the following drawings. DETAILED DESCRIPTION

[0036] Embodiments of the present application are described below in detail, examples of which are shown in the accompanying drawings, wherein the same or similar notations represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by reference to the drawings are exemplary only, for the purpose of explanation, and are not to be understood as limiting the present application.

[0037] As Figure 1 shown, the intelligent exhaust dynamic scheduling method based on federated learning and edge computing provided by the embodiments of the present application comprises:

[0038] S1, taking the corresponding control unit of each independent exhaust area as an edge node, taking the current edge node as a target node, collecting data in units of the target node, collecting multi-source sensor data of the area corresponding to the target node, and obtaining a predicted exhaust intensity value through a deep neural network.

[0039] Specifically, in the intelligent exhaust system, each edge node corresponds to a control unit of an independent exhaust area, such as a certain ventilation fan control module or a controllable air pipe valve unit. Each is deployed locally with edge computing capability, capable of collecting multi-source sensor data of the area and performing local prediction. The system collects the environmental state data of the node in each cycle , including the concentration of pollutants (such as PM2.5 or VOC), air pressure , wind speed , temperature , and device operating state (such as fan opening power, damper angle, etc. Control parameters). In addition, in order to reflect the pollution diffusion and airflow coupling between areas, the node also needs to maintain a neighborhood interference quantity , indicating the airflow pressure or pollution backflow risk assessment from the adjacent exhaust node to the current node.

[0040] The above data is continuously collected for a period of time to form a multi-time input sequence , which is used to perceive and predict whether the area needs to be exhausted and how strong the exhaust is. For this problem, the present application proposes a pollution-driven exhaust prediction model (PDEP, Pollution-Driven Exhaust Prediction) for generating a target exhaust intensity prediction value of the current cycle as the core target input for subsequent scheduling strategy learning.

[0041] The model first extracts the environmental evolution trend of the last time step through the time sequence structure to obtain a state vector , which represents the comprehensive pollution and air flow situation of the region at the current time. Then the prediction output is made through the following innovative structure:

[0042] ;

[0043] wherein: represents the predicted exhaust intensity value, the closer to 1 indicating the more urgent the current regional exhaust demand; represents the node environmental state vector at the current time; represents the weight matrix, used for linear mapping of the basis state output; represents the Sigmoid function, which normalizes the final prediction value to ; represents the gradient response intensity of the state to the pollution concentration, used to strengthen the identification of the rapid rising trend of pollution; represents the regularization weight of the pollution trend term; represents the exhaust node set adjacent to ; represents the pollution coupling influence degree of node to ; represents the difference degree of the current node and the adjacent node state, indicating the risk of pollution propagation that may occur; represents the coupling interference regularization weight.

[0044] It can be understood that the two innovative items of the model design have key invention. The first item of pollution change rate response item can identify the precursor of pollution outbreak in advance, so that the system can be scheduled in advance instead of responding after the pollution has occurred. The second item of spatial coupling item highlights the systematic risk assessment mechanism under the air dynamics characteristics, which can prevent pollution backflow or cross-regional diffusion caused by scheduling misjudgment.

[0045] For example, in a certain underground parking lot, corresponds to the west air outlet, corresponds to the north main ventilation pipe section. If the PM2.5 concentration is detected to be continuously rising, and the wind direction is from north to south, then increases, causing the system to predict to rise even if the local pollution concentration has not exceeded the standard, so as to schedule the west air fan to exhaust in advance to avoid the spread of pollution from north to south.

[0046] Finally, the output will be used as the input of the strategy learning model in the second step to drive the exhaust behavior learning.

[0047] S2, based on the predicted exhaust intensity value, combined with the local state of the current target node, a scheduling strategy model based on reinforcement learning is constructed, and the parameters of the scheduling strategy model and the local interaction trajectory data of the target node are output; wherein the local interaction trajectory data includes the predicted exhaust intensity value, the local state, the current scheduling behavior, the current scheduling reward, the actual exhaust intensity value and the air disturbance influence term.

[0048] Specifically, the present step aims to construct a personalized scheduling strategy model for each exhaust region edge node . , which can intelligently select the exhaust behavior (e.g. set the fan power to 80% and continue ventilation for 12 minutes) according to the exhaust intensity prediction target obtained in the previous step and the current local state (such as the running status of the device motor, the cumulative working time, the current valve angle, etc.). The strategy model is independently trained by each edge node, aiming to meet the regional exhaust demand while taking into account the actual operation constraints such as minimum energy consumption, system disturbance reduction and scheduling stability. Unlike traditional reinforcement learning, the present invention first proposes a target-behavior coupling calibration mechanism, which not only optimizes the immediate reward of the current behavior during the scheduling strategy learning process, but also introduces a penalty for the deviation of the predicted exhaust target , thereby establishing a strong closed loop of predicted target → actual effect → strategy update. This design makes the agent no longer train only based on environmental rewards, but dynamically adjusts behavior around the target exhaust intensity , with obvious scene adaptability and scheduling precision improvement ability.

[0049] We define the parameters of the strategy model as , and its update process is expressed by the following improved policy gradient expression which includes target deviation penalty and disturbance response term. In the present invention, the optimization of the strategy not only depends on the immediate reward , but also combines the target exhaust error and system disturbance, making the scheduling strategy more robust. Specifically, the objective function we propose is:

[0050] ;

[0051] wherein, denotes the learning rate; denotes the scheduling action probability of the given input state under the current strategy model; denotes the local immediate reward (such as the weighted combination of the decrease in pollution concentration and the decrease in energy consumption) brought by the current scheduling behavior ; represents the actual exhaust effect measured after the dispatching behavior is executed; represents the target exhaust error, indicating whether the behavior achieves the predicted target; represents the target deviation penalty factor (e.g., set to 10 to represent a penalty of 0.1 for each deviation of 0.1); represents the air disturbance influence term, used to penalize the negative air flow impact generated by the exhaust behavior on other areas; represents the disturbance penalty factor, dynamically set according to the current load of the system and the air duct flow safety threshold.

[0052] In particular, is not a gradient regularization term in the general sense, but a physical disturbance perception term specially designed for the exhaust system, defined as:

[0053] Node Execute action , along its ventilation path to other nodes, the weighted sum of the wind speed disturbance transmitted.

[0054] In specific implementation, the system estimates the action applied to the system ventilation network and whether the pressure fluctuation caused by the action exceeds the threshold. If it exceeds, the value is accumulated into . This can significantly reduce the system-level risk of abnormal air pressure in area B caused by excessive exhaust in area A.

[0055] To improve the stability and diversity of policy learning, the system also introduces a target entropy regularization term to avoid excessive convergence of the policy to a single high-frequency action and loss of scheduling adaptability. The final loss function is defined as follows:

[0056] ;

[0057] where, represents the scheduling policy learning loss of node ; represents the entropy of the policy output, used to encourage exploration diversity; represents the entropy regularization coefficient, preventing the policy from degenerating into a fixed behavior template.

[0058] For example, if a node of an underground garage exhausts in the last period , the current device state is , and the policy decides to take to turn on 85% wind for 10 minutes, the actual execution measures , and causes the adjacent node If the air pressure fluctuation exceeds the limit, the behavior will be appropriately weakened in the next round of training according to the loss function, prompting the strategy to learn to control the disturbance to the system while meeting the exhaust target.

[0059] The final output of this step is the personalized strategy model of each node and its training parameters , as well as the complete local interaction trajectory data . These contents not only provide input basis for the next step of federated learning module, but also serve as the core decision logic for the independent operation of local controller.

[0060] S3, calculate the structure disturbance perception weight according to the local interaction trajectory data of each edge node, to quantify the strategy contribution of each node; according to the structure disturbance perception weight, the federated server calculates the aggregation weight normalization result of each node, and then performs fusion calculation of the strategy parameter to obtain the global guiding strategy parameter.

[0061] Specifically, the goal of this step is to unify the cross-node collaborative optimization of the personalized scheduling strategy model parameters learned in the previous step and its corresponding local interaction trajectory data , while keeping the original data unshared, so as to form a system-level coordinated exhaust scheduling behavior. This step is in a key position connecting individual strategy to global coordination in the present invention, and determines whether the intelligent exhaust system can move from edge autonomy to structural consistency. Unlike existing general federated learning methods, the present invention takes into account four difficulties in exhaust systems, including physical space coupling, device type heterogeneity, uneven ventilation path structure, and asymmetric scheduling risk, and therefore designs a federated strategy aggregation mechanism specially applicable to exhaust scheduling systems, called structure and disturbance-aware federated aggregation (SD-FedAgg, Structure&Disturbance-Aware Federated Aggregation).

[0062] In actual exhaust systems, main air duct nodes (such as building central exhaust outlets) and end nodes (such as office local exhaust valves) play completely different roles in exhaust action and air flow control. If the model parameters are combined by the conventional average method, the following three problems are likely to occur:

[0063] The main node strategy is diluted by the edge node, causing abnormal scheduling of the main channel;

[0064] The end node strategy is forcibly overwritten, and its regional adaptability cannot be preserved;

[0065] The overall system disturbance is intensified, and the air pressure fluctuation is difficult to control.

[0066] To this end, the policy contribution of each node needs to be re-modeled before aggregation, introducing three joint factors: target alignment, disturbance robustness, and structure position weight, to define the structure disturbance-aware weight of each node As follows:

[0067] ;

[0068] Wherein, represents the average deviation between the policy output and the target exhaust intensity, which is used to measure the policy achievement ability, the smaller the deviation, the higher the score; represents a constant to prevent division by zero; represents the disturbance index defined in step two, which represents the strength of the current policy behavior to the system air pressure or flow rate disturbance, the greater the disturbance, the smaller the contribution; represents the disturbance penalty weight; represents the structure priority weight of the node , which is preset by the system topology. The main ventilation node (such as the air supply well of the subway station) is set to , the edge branch is set to , and the ordinary node is 1; represents the structure guiding coefficient, which is used to adjust the weight of the structure influence in aggregation.

[0069] Based on the above structure disturbance-aware weight , the federal server calculates the aggregation weight normalization result of each node, and then performs fusion calculation of the policy parameters to obtain the global guiding policy parameters:

[0070] ;

[0071] Wherein, represents the set of edge nodes participating in this round of aggregation; represents the policy parameters obtained by local training on the node ; represents the fused global guiding policy parameters, which are used for subsequent synchronization and fine-tuning.

[0072] Wherein, the aggregation mechanism has the following advantages:

[0073] Coupling system disturbance factors: taking as the risk quantification index of the exhaust scheduling strategy for the whole system, for the first time, the influence of aerodynamics is considered in the policy fusion;

[0074] Introducing structure topology priority: through , the importance difference of different nodes to the global system stability is made clear, so that the main air duct strategy has a higher weight, avoiding decision conflict;

[0075] Dynamic weight update based on target deviation: The aggregation weight is not only determined by the number of node contributions, but also depends on the actual control effect of the strategy, with the characteristics of precision-oriented reinforcement learning.

[0076] After obtaining , the system does not force all edge nodes to replace their own strategy parameters immediately, but performs individualized synchronization according to the deviation. For nodes with long-term deviation of error (such as for two cycles), the system will force its model parameters to be ; while for nodes with error stable in an acceptable range (such as less than 0.05), only is used for lightweight fusion, retaining the local strategy adaptation ability. This mechanism is called federated soft synchronization strategy (FSS, Federated Soft Synchronization). For example, if node is an air outlet controller of an underground air exchange main pipeline, the exhaust prediction for the last cycle is , the actual exhaust effect is , and the disturbance index is , then its will be much higher than that of other nodes, having a significant influence in aggregation. While edge node (such as the exhaust at the end of the bathroom) has a strategy deviation and a disturbance , it will be guided to update by the system to eliminate the risk of exhaust scheduling imbalance caused by regional strategy inconsistency.

[0077] S4, apply the global guiding strategy parameters to the physical exhaust equipment for intelligent scheduling of the physical exhaust equipment.

[0078] Specifically, the role of this step is to apply the synchronized strategy parameters output in step three to the physical exhaust equipment, and through the system-level conflict detection mechanism to ensure that the execution behavior does not cause regional interference in air flow field, air pressure, wind speed, etc. physical indicators, is the only physical action execution closed loop step in the intelligent exhaust system of the present application. Each edge node uses its synchronized strategy , combined with the current exhaust prediction and equipment state , to generate the current scheduling action . This action determines the control behavior of the equipment (such as fans, dampers) in the current cycle, such as wind level, running time, etc.

[0079] To avoid air pressure conflicts or wind volume mutual exclusion when multiple nodes execute the strategy at the same time, the system introduces a disturbance index , calculate the potential physical conflict risk of exhaust behavior between the current node and its adjacent nodes:

[0080] ;

[0081] wherein, represents the node System disturbance index under the current scheduling period; represents the set of adjacent nodes connected with There is a wind channel connection relationship; , represents the scheduling action output by the current node; represents the air power coupling strength, reflecting the influence of the wind channel structure.

[0082] If exceeds the safety threshold set by the system , the current action may cause airflow instability to the system, and disturbance suppression needs to be performed before execution to obtain the reconciled executable action :

[0083] ;

[0084] wherein, represents the control action reconciled for actual execution; represents the suppression proportion coefficient, generally .

[0085] All actions will be issued as actual control instructions to the device end for execution, such as adjusting the output of the fan, opening the valve, etc. After the control is completed, the system obtains the actual execution result through the sensor and feeds it back for the next round of learning period, realizing the complete control closed loop of perception-prediction-strategy-execution.

[0086] The embodiment of the application also provides an intelligent exhaust dynamic scheduling system based on federated learning and edge computing, which comprises:

[0087] A demand prediction module is configured to take each independent exhaust area corresponding to a control unit as an edge node, set a current edge node as a target node, collect data in units of the target node, collect multi-source sensor data of the area corresponding to the target node, and obtain a predicted exhaust intensity value through a deep neural network.

[0088] An edge personalized learning module is configured to construct a scheduling strategy model based on reinforcement learning based on the predicted exhaust intensity value and the local state of the current target node, and output parameters of the scheduling strategy model and local interaction trajectory data of the target node; wherein the local interaction trajectory data comprises the predicted exhaust intensity value, the local state, the current scheduling behavior, the current scheduling reward, the actual exhaust intensity value and the air disturbance influence term.

[0089] A federal collaborative strategy aggregation module is configured to calculate a structure disturbance perception weight according to the local interaction trajectory data of each edge node to quantify the strategy contribution degree of each node, and calculate an aggregation weight normalization result of each node according to the structure disturbance perception weight, and then perform fusion calculation of the strategy parameters to obtain global guiding strategy parameters.

[0090] A conflict detection and strategy reconciliation execution module is configured to apply the global guiding strategy parameters to the physical exhaust equipment to perform intelligent scheduling of the physical exhaust equipment.

[0091] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the system, device and unit described above can refer to the corresponding process in the foregoing method embodiments, which will not be repeated here.

[0092] In several embodiments provided in the present application, it should be understood that the disclosed system, device and method can be implemented in other ways. For example, the device embodiments described above are only schematic, for example, the division of the units is only a logical function division, and actual implementation can have another division manner, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the units shown or discussed can be indirect coupling or communication connection through some interface, device or unit, and can be electrical, mechanical or other forms.

[0093] If the functions are implemented in the form of software function units and sold or used as independent products, they can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application essentially or the parts that contribute to the prior art or parts of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0094] Although the embodiments of the present application have been shown and described, those skilled in the art can understand that various changes, modifications, replacements and deformations can be made to the embodiments without departing from the principles and purposes of the present application, and the scope of the present application is defined by the claims and their equivalents.

Claims

1. An intelligent exhaust dynamic scheduling method based on federated learning and edge computing, characterized in that, The method comprises the following steps: With the corresponding control unit of each independent exhaust area as an edge node, taking the current edge node as the target node, collecting data in units of the target node, collecting multi-source sensor data of the area corresponding to the target node, and obtaining a predicted exhaust intensity value through a deep neural network; Based on the predicted exhaust intensity value, combining the local state of the current target node, constructing a scheduling strategy model based on reinforcement learning, outputting the parameters of the scheduling strategy model and the local interaction trajectory data of the target node; wherein the local interaction trajectory data includes the predicted exhaust intensity value, the local state, the current scheduling behavior, the current scheduling reward, the actual exhaust intensity value, and the air disturbance influence term; According to the local interaction trajectory data of each edge node, the structural disturbance perception weight is calculated to quantify the strategy contribution degree of each node; according to the structural disturbance perception weight, the federal server calculates the aggregation weight normalization result of each node, and then performs fusion calculation of the strategy parameters to obtain the global guidance strategy parameters; The global guidance strategy parameters are actually applied to the physical exhaust equipment for intelligent scheduling of the physical exhaust equipment; The deep neural network is a pollution-driven exhaust prediction model, and the specific calculation is as follows: ； wherein, 1 indicates the most urgent exhaust demand of the current area; is a node environment state vector at the current time; is a weight matrix for linear mapping of the basis state output; is a Sigmoid function for normalizing the final prediction value to ; is a gradient response intensity of the state to the pollution concentration, used to strengthen the identification of the rapid rising trend of the pollution; is a regularization weight of the pollution trend term; is a set of exhaust nodes adjacent to ; is a pollution coupling influence degree of the node to ; is a difference degree of the current node and the adjacent node state, indicating the risk of pollution propagation that may occur; is a coupling interference regularization weight; The objective function of the scheduling strategy model is: ； wherein, denotes a parameter of the policy model, denotes a learning rate; denotes a scheduling action probability of a given input state under the current policy model; denotes a local immediate reward brought by the scheduling behavior in this round; denotes an actual measured exhaust effect after the scheduling behavior is executed; denotes a target exhaust error, indicating whether the behavior achieves the predicted target; denotes a target deviation penalty factor; denotes an air disturbance influence item, used to punish the negative air flow impact on other areas caused by the exhaust behavior; denotes a disturbance penalty factor, dynamically set according to the current load of the system and the air duct flow safety threshold. wherein, defined as a node performing an action after, along its ventilation path weighted sum of wind speed disturbances passed on to other nodes.

2. The intelligent exhaust dynamic scheduling method based on federated learning and edge computing according to claim 1, characterized in that, The multi-source sensor data of the area corresponding to the target node is collected, specifically including: The environmental state data of the node is collected at each period t, including pollutant concentration, air pressure, wind speed, temperature, and equipment operating state.

3. The intelligent exhaust dynamic scheduling method based on federated learning and edge computing according to claim 1, characterized in that, The parameters of the scheduling strategy model are obtained by updating the strategy gradient algorithm containing the target deviation penalty and the disturbance response term; wherein the target deviation penalty is obtained according to the error between the predicted exhaust intensity value and the actual exhaust intensity value.

4. The intelligent exhaust dynamic scheduling method based on federated learning and edge computing according to claim 1, characterized in that, The structural disturbance perception weight is calculated according to the difference absolute value between the predicted exhaust intensity value and the actual exhaust intensity value, the action strength of the air disturbance influence term, and the structural priority weight of the node. After obtaining the global guidance strategy parameters, for the nodes with long-term deviation error, the model parameters are forced to update to the global guidance strategy parameters; and for the nodes with stable error within the acceptable range, only the global guidance strategy parameters are used for lightweight fusion to retain the local strategy adaptive ability.

5. The intelligent exhaust dynamic scheduling method based on federated learning and edge computing according to claim 1, characterized in that, The global guidance strategy parameters are actually applied to the physical exhaust equipment for intelligent scheduling of the physical exhaust equipment, specifically including:

6. The intelligent exhaust dynamic scheduling method based on federated learning and edge computing according to claim 1, characterized in that, Each edge node uses its synchronized global guidance strategy parameters to generate the current scheduling action in combination with the current predicted exhaust intensity value and the equipment operating state; wherein the scheduling action is used to determine the control behavior of the equipment in the current period. In the intelligent scheduling of the physical exhaust equipment, it also includes:

7. The intelligent exhaust dynamic scheduling method based on federated learning and edge computing according to claim 6, characterized in that, Based on the node-related attribute calculation of the exhaust behavior between the current node and its adjacent nodes, the system disturbance index of the current node in the current scheduling period is determined, indicating the potential physical conflict risk; ​ When the system disturbance index exceeds the safety threshold set by the system, the current action may cause airflow instability to the system, and disturbance suppression needs to be performed before execution to obtain a harmonized executable action; All the harmonized executable actions will be issued as actual control instructions to the device end for execution.

8. The system for performing the intelligent exhaust dynamic scheduling method based on federated learning and edge computing according to claim 1, wherein, The system comprises: A demand prediction module, configured to take each independent exhaust area corresponding control unit as an edge node, take the current edge node as a target node, collect data in the target node unit, collect multi-source sensor data of the area corresponding to the target node, and obtain a predicted exhaust intensity value through a deep neural network; An edge individualized learning module, configured to construct a scheduling strategy model based on reinforcement learning based on the predicted exhaust intensity value and the local state of the current target node, and output parameters of the scheduling strategy model and local interaction trajectory data of the target node; wherein the local interaction trajectory data comprises the predicted exhaust intensity value, the local state, the current scheduling behavior, the current scheduling reward, the actual exhaust intensity value, and an air disturbance influence item; A federal collaborative strategy aggregation module, configured to calculate a structure disturbance perception weight according to the local interaction trajectory data of each edge node to quantify the strategy contribution degree of each node, calculate an aggregation weight normalization result of each node according to the structure disturbance perception weight, and then perform fusion calculation of the strategy parameters to obtain global guidance strategy parameters; A conflict detection and strategy harmonization execution module, configured to actually apply the global guidance strategy parameters to the physical exhaust equipment to perform intelligent scheduling of the physical exhaust equipment.

Images