Time-sharing zoning AI temperature compensation and air conditioner energy-saving control method and equipment considering airport passenger comfort preference self-learning

Abstract

The invention discloses a time-sharing zoning AI temperature compensation and air conditioner energy-saving control method and device considering airport passenger comfort preference self-learning, and is applied to the technical field of data processing.The method focuses on airport air conditioner energy saving and passenger comfort, firstly, passenger behavior tracks, real-time environment parameters and comfort feedback data are collected; behavior, environment and preference features are extracted through deep learning, and a multi-dimensional comfort preference data set is generated through cross-modal fusion. And constructing a target directed multi-view dynamic graph containing multiple types of nodes according to spatial partitioning and time granularity layering, and generating node embedding information through a self-adaptive multi-view dynamic graph neural network. And finally, the AI temperature compensation parameters are injected into reinforcement learning double networks, time-sharing and partitioned AI temperature compensation parameters and air conditioner energy-saving control instructions are generated through scene migration reasoning in combination with parameter efficient fine adjustment and double-target antagonism optimization, and comfort and energy conservation are both considered.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a time-sharing and zone-based AI temperature compensation and air conditioning energy-saving control method and equipment that takes into account airport passenger comfort preferences through self-learning. Background Technology

[0002] Airport terminals, as large-space, high-traffic public buildings, require air conditioning systems that simultaneously meet passenger comfort needs and energy-saving operational goals. This is a recognized challenge in the industry, and current technologies mainly suffer from the following problems:

[0003] Current airport air conditioning control relies heavily on fixed parameters or manual adjustments based on experience, making it difficult to adapt to the dynamic spatial and temporal changes of airport terminals. Spatially, passenger density varies significantly across areas such as security checkpoints, boarding gates, and waiting areas, leading to substantial fluctuations in localized cooling demand. Fixed control methods can result in excessively high (or low) temperatures in high-density areas and wasted energy in low-density areas. Temporally, peak / off-peak flight times and day / night passenger flow, along with environmental factors (such as sunlight intensity), fluctuate dramatically. Air conditioning systems exhibit significant temperature adjustment lags (e.g., it takes a long time to reach the target temperature after full operation), failing to accurately match real-time demand and resulting in high per capita air conditioning energy consumption. For example, under traditional manual control, per capita air conditioning energy consumption in some airport terminals can reach 1.004 kWh / six months, with substantial fluctuations in passenger comfort satisfaction.

[0004] Current air conditioning control systems mostly use fixed temperature thresholds (such as a uniform setting of 26°C) as the control target, without considering the differences in individual passenger comfort preferences. On the one hand, existing technologies do not effectively integrate passenger behavior trajectories (such as areas of stay, activity intensity, and stay duration) with subjective feedback (such as comfort level evaluation) data, making it impossible to quantify the weight of passenger comfort needs under different scenarios (such as peak passenger flow and extreme weather). On the other hand, the lack of a dynamic self-learning mechanism makes it difficult to iteratively optimize control strategies based on passenger feedback and environmental changes, resulting in poor adaptability of control schemes. For example, the temperature preference differences between areas where elderly passengers congregate and areas where young passengers congregate cannot be identified, further reducing the overall comfort experience.

[0005] Existing technologies mostly focus on localized control of individual air conditioning units (such as chillers and fan coil units), failing to establish a multi-dimensional collaborative optimization system encompassing the environment, passengers, and air conditioning. Firstly, data such as environmental parameters (temperature, humidity, wind speed), passenger behavior characteristics, and air conditioning operating status are not deeply integrated, making it difficult to generate a globally optimal control strategy. Secondly, the control algorithms lack scenario transfer capabilities, failing to simulate temperature compensation effects and energy consumption changes under unconventional scenarios such as sudden changes in passenger flow and extreme weather, resulting in limited control accuracy (e.g., temperature prediction errors often exceed ±1℃). Thirdly, existing solutions are mostly "black box" decisions, lacking a visual explanation of feature importance and decision logic, making it difficult for maintenance personnel to trace the basis for control measures, hindering troubleshooting and strategy optimization.

[0006] Existing airport air conditioning energy-saving control technologies are mostly single-point innovations, failing to form a closed-loop system of "demand forecasting - supply regulation - effect evaluation". Although some solutions introduce simple prediction models, they do not integrate sub-algorithms such as multi-area room temperature prediction and dynamic energy consumption trend prediction, resulting in insufficient prediction accuracy. Furthermore, the lack of standard energy consumption models (such as peak and valley time division and equipment energy efficiency threshold setting) makes it impossible to quantify the energy-saving effect and reasonable energy consumption boundaries. This leads to a lack of scientific decision-making basis when the goals of "energy saving" and "comfort" conflict. Some solutions even sacrifice passenger comfort in the pursuit of energy saving, or cause energy consumption to exceed standards due to an emphasis on comfort. Summary of the Invention

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0008] A time- and zone-based AI-driven temperature compensation and air conditioning energy-saving control method considering self-learning of airport passenger comfort preferences includes: acquiring a target dataset and a training sample set, wherein the training sample set includes an airport passenger behavior trajectory dataset, a real-time environmental parameter dataset, and a passenger comfort feedback dataset; based on the passenger behavior trajectory dataset, using a deep learning model to extract features of passenger dwell area, activity intensity, and dwell time, and combining it with the real-time environmental parameter dataset, converting passenger behavior semantic features into comfort demand vectors through a cross-modal feature fusion module, quantifying the preference levels of passenger comfort feedback data into weighted features, and generating a multi-dimensional passenger comfort preference dataset for temperature compensation and energy-saving control; processing the multi-dimensional passenger comfort preference dataset, layering it according to spatial partitioning and temporal granularity, with each layer containing area nodes, environmental parameter nodes, passenger preference nodes, and air conditioning operation nodes, and aggregating real-time features from different layers through a dynamic perception window, generating... A target directed multi-view dynamic graph is generated, consisting of environmental conditions, passenger preferences, and air conditioning operation status. The target directed multi-view dynamic graph is modeled and processed based on an adaptive multi-view dynamic graph neural network. A passenger preference self-learning unit is introduced to dynamically update comfort preference weights under different scenarios, generating node embedding information. This node embedding information is injected into each layer of the value network and policy network of the reinforcement learning model. A parameter-efficient fine-tuning method is used to optimize the model parameters, and an energy consumption-comfort dual-objective adversarial optimization module is introduced. Simultaneously, a decision interpretation chain is generated through feature importance visualization, forming a semantic association information between fused environmental features, passenger preference features, and air conditioning control strategies. Based on the reinforcement learning model and the fused semantic association information, the target data set is processed. A scenario transfer inference module simulates the temperature compensation effect and energy consumption changes under different passenger flow and environmental change scenarios, generating time- and zone-specific AI temperature compensation parameters and air conditioning energy-saving control commands.

[0009] A time-sharing and zone-based AI temperature compensation and air conditioning energy-saving control device that takes into account airport passenger comfort preferences through self-learning, the device being configured to perform any of the time-sharing and zone-based AI temperature compensation and air conditioning energy-saving control methods that take into account airport passenger comfort preferences through self-learning.

[0010] Its beneficial effects are as follows: This invention provides a time-sharing and zone-based AI temperature compensation and air conditioning energy-saving control method that considers self-learning of airport passenger comfort preferences. This application focuses on the core needs of airport air conditioning energy saving and passenger comfort. First, it collects three types of data: passenger behavior trajectory, real-time environmental parameters, and comfort feedback. It uses deep learning to extract behavioral, environmental, and preference features and generates a multi-dimensional comfort preference dataset through cross-modal fusion. Then, it constructs a target directed multi-view dynamic graph containing multiple types of nodes by spatial partitioning and temporal granularity. It generates node embedding information through an adaptive multi-view dynamic graph neural network, injects reinforcement learning dual networks, and combines efficient parameter fine-tuning and adversarial optimization of energy consumption and comfort dual objectives to generate a decision interpretation chain. Finally, through scene transfer reasoning, it generates time-sharing and zone-based AI temperature compensation parameters and air conditioning energy-saving control commands, forming a complete closed loop of "data collection - feature fusion - model modeling - optimization decision - command generation".

[0011] This invention precisely adapts to dynamic demands, significantly improving comfort: Through a self-learning unit for passenger preferences and a spatiotemporal hierarchical strategy, it accurately identifies differences in comfort needs across different regions and time periods, achieving a temperature control error of ≤0.3℃ and increasing passenger comfort satisfaction to over 4 points. Dual-objective adversarial optimization and time- and zone-based control reduce airport air conditioning energy consumption per capita by over 15%, avoiding energy waste in low-load areas. Feature importance visualization provides a full-link explanation chain, and scenario migration reasoning supports complex scenarios such as sudden changes in passenger flow and extreme weather, improving operational efficiency by 30%. Deep fusion and closed-loop management of multi-source data enable collaborative optimization of the "environment-passengers-air conditioning" system, allowing for dynamic adjustment of control strategies without manual intervention, adapting to complex airport operational scenarios. Attached Figure Description

[0012] Figure 1 A flowchart of a time-sharing and zone-based AI temperature compensation and air conditioning energy-saving control method that takes into account airport passenger comfort preferences and self-learning is provided for an embodiment of the present invention;

[0013] Figure 2 This is a schematic diagram of a time-sharing and zone-based AI temperature compensation and air conditioning energy-saving control device that takes into account the comfort preferences of airport passengers, provided as an embodiment of the present invention. Detailed Implementation

[0014] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention. Figure 1 This application describes a time-sharing and zone-based AI temperature compensation and air conditioning energy-saving control method that takes into account airport passenger comfort preferences and is based on an exemplary embodiment of this application.

[0015] In this application embodiment, a time-sharing and zone-based AI temperature compensation and air conditioning energy-saving control method that considers airport passenger comfort preferences through self-learning is provided, such as... Figure 1 As shown:

[0016] S101, obtain the target data set and training sample set.

[0017] In one implementation, the data acquisition process focuses on the core needs of airport temperature compensation and air conditioning energy-saving control, simultaneously acquiring the target data set and the training sample set. The training sample set includes three types of core data: airport passenger behavior trajectory dataset, real-time environmental parameter dataset, and passenger comfort feedback dataset. These three types of data together provide comprehensive data support for subsequent comfort preference modeling and control strategy generation.

[0018] The system employs a collaborative data collection approach using visual sensors and positioning devices deployed at the airport, covering all functional areas including the terminal hall, boarding gates, waiting areas, and transfer corridors. Sensors capture passenger movement trajectories, stopping locations, and activity states in real time, which, combined with spatial coordinate information output from the positioning devices, forms continuous behavioral trajectory data. The data includes the start and end times of passenger stays in each area, the specific spatial range of the stay area, movement path nodes, and the amplitude and frequency of limb movements during activity (used to characterize activity intensity), ensuring a complete record of passenger behavioral characteristics within the airport. This lays the data foundation for subsequent extraction of key features such as stopping areas, activity intensity, and stay duration.

[0019] Environmental sensors are distributed throughout the airport's functional areas, continuously acquiring environmental data at a frequency of once per minute. The data collection covers four core environmental factors: temperature, humidity, wind speed, and light intensity. Temperature data is accurate to 0.1℃, humidity data is recorded as a percentage of relative humidity, wind speed data is measured in meters per second, and light intensity is measured in lux. The sensor data is uploaded to the data processing center in real time via wireless transmission modules. A data verification mechanism is also in place to automatically remove abnormal fluctuations (such as values ​​exceeding reasonable ranges due to sensor malfunctions), ensuring the real-time nature, accuracy, and continuity of environmental parameter data. This provides precise data support for analyzing the impact of the environment on passenger comfort perception.

[0020] A multi-channel collaborative data collection model was adopted, including three core collection methods: the feedback portal built into the airport's official app, self-service feedback terminals in the terminal, and feedback QR codes on the back of flight seats. The feedback content was designed in a combination of structured and semi-structured formats. Structured content included comfort level ratings (1-5 points, 1 point being extremely uncomfortable, 5 points being extremely comfortable) and satisfaction with the current temperature (satisfied / neutral / unsatisfied). Semi-structured content allowed passengers to input text descriptions (such as "the temperature is too high" or "the air conditioning is too strong"). Simultaneously, semantic information about passenger behavior captured by visual sensors (such as actions like wiping sweat, tightening clothing, and opening windows for ventilation) was used as supplementary indirect feedback data. The collected feedback data underwent semantic parsing and standardization to form a unified format passenger comfort feedback dataset, providing a data source for subsequent preference level quantification and sensitivity coefficient calculation.

[0021] The three types of datasets collected are uploaded to the data storage center in real time via an encrypted transmission protocol. A distributed storage architecture ensures the security and accessibility of data storage. Data is stored using a dual indexing system: timestamps accurate to the second, and region identifiers corresponding to specific functional zones and sensor numbers, facilitating subsequent data retrieval and correlation analysis. Simultaneously, a data backup mechanism is established, with daily full backups to prevent data loss and ensure the integrity and availability of the training sample set. This provides stable data support for subsequent steps such as generating multi-dimensional comfort preference datasets and constructing graph structures.

[0022] S102, based on passenger behavior trajectory dataset, uses a deep learning model to extract features of passenger dwelling area, activity intensity, and dwelling duration. Combined with real-time environmental parameter dataset, it transforms passenger behavior semantic features into comfort demand vectors through a cross-modal feature fusion module, quantifies the preference level of passenger comfort feedback data into weighted features, and generates a multi-dimensional passenger comfort preference dataset for temperature compensation and energy-saving control.

[0023] In one implementation, key behavioral elements are analyzed from the passenger behavior trajectory dataset to extract behavioral features including spatial coordinates of stopping areas, activity intensity levels, and stay duration intervals. A temporal behavior analysis algorithm is used to mine key elements from the passenger behavior trajectory dataset, extracting three core types of features. Spatial coordinates of stopping areas are determined through cluster analysis of location points in the trajectory data. Using the airport plane coordinate system as a reference, the boundary coordinate range and center coordinate values ​​of each stopping area are output, accurately pinpointing the spatial locations of passengers' high-frequency stops. Activity intensity levels are quantitatively assessed using temporal data of the speed change rate and limb movement amplitude of the behavior trajectory. A fuzzy comprehensive evaluation method is used to divide activity intensity into three levels: low, medium, and high. Low intensity corresponds to a seated waiting state, medium intensity corresponds to a slow moving state, and high intensity corresponds to a fast walking or baggage handling state. Stay duration intervals are obtained by calculating the difference between the start and end times of passengers in each stopping area, divided into three intervals according to time span: short duration (0-30 minutes), medium duration (30 minutes-2 hours), and long duration (over 2 hours), comprehensively representing the stay characteristics of passengers in different areas.

[0024] Environmental factors were decomposed from the real-time environmental parameter dataset to obtain environmental features including temperature, humidity, wind speed, and light intensity. A structured decomposition algorithm for environmental factors was used to process the real-time environmental parameter dataset, extracting four core environmental features. Temperature features were directly obtained from raw data collected by sensors, retaining one decimal place of precision, covering the real-time temperature status of various areas of the airport. Humidity features, using relative humidity as the characterization index, were directly obtained from sensor data, reflecting the impact of air moisture content on passenger comfort perception. Wind speed features were extracted from airflow velocity data collected by sensors, measured in meters per second, accurately capturing airflow changes caused by air conditioning and natural ventilation. Light intensity features were obtained from data collected by light sensors, measured in lux, covering the regional lighting conditions under the combined effects of natural and artificial lighting, providing data support for analyzing the correlation between light intensity and temperature perception.

[0025] A preference semantic transformation was performed on the passenger comfort feedback dataset, mapping subjective evaluations to comfort preference levels and sensitivity coefficients to generate preference features. The dataset was processed using semantic parsing and quantification mapping algorithms to generate two types of preference features. The comfort preference levels adopted a 5-level quantification standard, mapping passengers' subjective evaluations of "extremely uncomfortable," "uncomfortable," "neutral," "comfortable," and "extremely comfortable" to quantitative scores of 1-5, achieving standardized transformation of subjective feelings. The sensitivity coefficient was calculated through correlation analysis between feedback data and corresponding environmental parameters. The Pearson correlation coefficient algorithm was used to quantify the sensitivity of passenger comfort evaluations to changes in environmental factors such as temperature and humidity. The coefficient ranged from 0 to 1; the closer the coefficient was to 1, the more sensitive the passenger was to environmental changes, and the closer the coefficient was to 0, the lower the sensitivity, comprehensively representing the differences in comfort preferences among different passengers.

[0026] A cross-modal feature fusion module is used to align and correlate behavioral, environmental, and preference features, constructing a multi-dimensional feature association matrix. The module employs temporal alignment and spatial association algorithms to achieve accurate fusion of the three types of features. First, using timestamps as a benchmark, behavioral, environmental, and preference features are temporally synchronized to ensure corresponding matching of the three types of features within the same time dimension. Then, based on spatial partitioning identifiers, the three types of features from different regions are spatially associated, establishing a three-dimensional "region-time-feature" association relationship. Building upon this, a matrix construction algorithm is used to organize the associated feature data into a multi-dimensional feature association matrix. The row dimension of the matrix corresponds to different time-region combination units, and the column dimension corresponds to various feature indicators, achieving structured integration of multi-source heterogeneous features and laying the foundation for subsequent weight allocation and vector generation.

[0027] A dynamic weight allocation mechanism is introduced to prioritize and weight fused features, generating weighted features. This mechanism employs a weighting strategy combining the Analytic Hierarchy Process (AHP) and entropy weighting to quantify the priority of fused features. First, an AHP-based three-layer evaluation system is constructed: a target layer (temperature compensation and energy-saving control adaptability), a criterion layer (comfort perception impact, energy-saving correlation, and real-time response), and an indicator layer (various feature indicators), determining the subjective weights of each feature. Then, the entropy weighting method is used to calculate the information entropy and objective weights of each feature, reflecting the dispersion and effective information content of the feature data. Finally, a weighted summation method is used to fuse the subjective and objective weights, obtaining the comprehensive weights of each feature. The weights range from 0 to 1; higher weight values ​​indicate a greater impact of the feature on comfort needs and energy-saving control. Based on the comprehensive weights, the features in the feature correlation matrix are weighted to generate a weighted feature matrix, highlighting the role of high-impact features.

[0028] Based on the weighted feature matrix, a standardized comfort demand vector is generated. This vector is then integrated with quantified weighted features to construct a multi-dimensional passenger comfort preference dataset for temperature compensation and energy-saving control. A vector normalization algorithm is used to process the weighted feature matrix, generating a standardized comfort demand vector. Each feature index in the weighted feature matrix is ​​normalized, mapping the feature values ​​to the [0,1] interval to eliminate the influence of differences in feature dimensions. Subsequently, by time-region combination unit, the normalized feature values ​​are integrated into a one-dimensional vector, i.e., the standardized comfort demand vector. Each dimension of the vector corresponds to a standardized feature index, comprehensively representing the passenger comfort demand characteristics under that time-region unit. Finally, the standardized comfort demand vectors and weighted feature matrices of all time-region units are integrated to form a multi-dimensional passenger comfort preference dataset for temperature compensation and energy-saving control. This dataset covers complete information in three dimensions: space, time, and features, providing standardized data support for subsequent graph structure construction and model modeling.

[0029] S103 processes the multi-dimensional passenger comfort preference dataset, dividing it into spatial partitions and temporal granularities. Each layer contains regional nodes, environmental parameter nodes, passenger preference nodes, and air conditioning operation nodes. Through a dynamic perception window, it aggregates real-time features from different layers to generate a target directed multi-view dynamic graph composed of environmental status, passenger preferences, and air conditioning operation status.

[0030] In one implementation, based on the feature distribution characteristics of a multi-dimensional passenger comfort preference dataset and airport spatial-temporal association rules, a hierarchical aggregation strategy is determined through negotiation between the data processing module and the graph structure generation engine. The node hierarchy and association dimensions are determined through spatial partition attribute parsing and temporal granularity partitioning. Relying on the feature distribution characteristics of the multi-dimensional passenger comfort preference dataset and airport spatial-temporal association rules, a bidirectional interaction mechanism is constructed between the data processing module and the graph structure generation engine to collaboratively negotiate the hierarchical aggregation strategy. Spatial partition attribute parsing employs a regional functional clustering algorithm, dividing the airport terminal into partitions based on its functional attributes (such as departure hall, security checkpoint, boarding gate cluster, waiting area, transfer channel, and arrival hall), clearly defining the spatial scope and attribute identifiers of each partition. Temporal granularity partitioning employs a time-series segmentation algorithm, dividing time into three core intervals—peak hours, off-peak hours, and nighttime hours—based on airport passenger flow patterns, while further subdividing each interval into hourly sub-granularities. The above analysis determines that the node hierarchy is a two-level structure of "core partition - sub-partition" and "main time period - sub-time period". The association dimensions are set as three categories: spatial association, temporal association and feature association, to ensure that the hierarchical aggregation strategy is accurately adapted to the actual operation scenario of the airport.

[0031] A dual-dimensional partitioning mechanism (partition-time division) was employed to split the multi-dimensional passenger comfort preference dataset. Corresponding feature data was extracted according to terminal functional zones and time intervals, and core elements such as area, environment, preferences, and air conditioning operation were associated to generate preliminary hierarchical data blocks. This mechanism systematically split the multi-dimensional passenger comfort preference dataset. Feature data was extracted for each functional zone to ensure complete correspondence between behavioral, environmental, and preference features. Feature data for each time period was extracted according to defined time intervals to achieve accurate feature matching over time. During the splitting process, a feature association algorithm was used to deeply associate the four core elements—area, environment, preferences, and air conditioning operation—establishing a three-dimensional data association relationship of "area-time-element." Based on this association, the split feature data was integrated into multiple preliminary hierarchical data blocks. Each data block corresponds to a "functional zone + time interval" combination unit, containing complete feature data of the four core elements within that unit, providing a basic data unit for subsequent graph structure construction.

[0032] The initial layered data blocks undergo validity verification. For data with ambiguous spatial boundaries, a regional calibration service is initiated. For data with inconsistent timestamps, a time-series synchronization mechanism is triggered. For samples with missing node features, a feature completion process is performed, generating an optimized layered data scheme that includes a regional calibration plan, a time-series synchronization strategy, and a feature completion method. A multi-dimensional validity verification process is initiated for the initial layered data blocks to ensure data quality meets the requirements for graph structure construction. For data with ambiguous spatial boundaries, a regional calibration service is initiated, using a spatial coordinate correction algorithm based on the actual airport spatial layout map to correct the ambiguous spatial coordinates of the stopping areas, clarifying the precise spatial boundaries corresponding to each data block. For data with inconsistent timestamps, a time-series synchronization mechanism is triggered, using a timestamp alignment algorithm based on the time of a standard time server to uniformly calibrate the timestamps of data from different sources, achieving time synchronization of all feature data within the same data block. For samples with missing node features, a feature completion process is performed, using a similarity-based feature interpolation algorithm to accurately complete the missing feature values ​​by mining the feature data patterns of similar "region-time" units. Through the above verification and processing, an optimized hierarchical data scheme is generated, which includes regional calibration scheme, timing synchronization strategy and feature completion method, to ensure the integrity, accuracy and consistency of each hierarchical data block.

[0033] The optimized hierarchical data scheme is integrated and implemented, introducing a dynamic sensing window mechanism with a window sliding cycle of 5 minutes to capture the dynamic changes of feature data at each level in real time. During the integration process, node association weights are generated synchronously through a weight calculation algorithm. The weight values ​​are determined based on a combination of feature influence and data reliability, ranging from 0 to 1. Higher weights indicate stronger associations between nodes. Simultaneously, a feature aggregation progress feedback mechanism is established to monitor the feature aggregation status of each "functional zone + time interval" unit in real time, ensuring the aggregation process is efficient and controllable. Based on the integrated feature data and node association weights, a hierarchical network is constructed, containing four types of core nodes: area nodes (corresponding to each functional zone of the terminal), environmental parameter nodes (corresponding to environmental factors such as temperature, humidity, wind speed, and light intensity), passenger preference nodes (corresponding to preference features such as comfort preference level and sensitivity coefficient), and air conditioning operation nodes (corresponding to data related to air conditioning operation status). Based on this hierarchical network and combined with the multi-view construction algorithm, three views are constructed according to comfort demand constraints, energy consumption optimization constraints, and real-time response constraints. Each view highlights the node relationships under the corresponding constraints, and finally forms a target directed multi-view dynamic graph consisting of environmental status, passenger preferences, and air conditioning operation status. This graph can dynamically reflect the relationship evolution of various elements under different spatiotemporal scenarios.

[0034] S104 models the target directed multi-view dynamic graph based on an adaptive multi-view dynamic graph neural network, introduces a passenger preference self-learning unit, dynamically updates the comfort preference weights under different scenarios, and generates node embedding information.

[0035] In one implementation, the target directed multi-view dynamic graph is structurally analyzed, layered according to spatial partition type and temporal granularity. For each layer, dynamic features such as the environmental sensitivity of regional nodes, the demand intensity of preference nodes, and the control response speed of air conditioning nodes are extracted, generating feature subsets for each layer. A graph structure layered analysis algorithm is used to structurally process the target directed multi-view dynamic graph, achieving a two-layer split according to spatial partition type and temporal granularity. The spatial partition type corresponds to the terminal building's functional partitions, and the temporal granularity corresponds to peak, off-peak, nighttime, and hourly sub-periods, forming a two-dimensional "space-time" layered structure. For each structural layer, three core dynamic features are extracted: Regional node environmental sensitivity is quantified by assessing the impact of environmental parameter changes on regional comfort. The feature value is calculated using a gradient descent algorithm, ranging from 0 to 1; a higher value indicates a greater impact of environmental changes on passenger comfort perception in that area. Preference node demand intensity is calculated based on a weighted average of passenger comfort preference levels and sensitivity coefficients, generated using a linear weighting algorithm, directly reflecting the urgency of temperature regulation needs at that level. Air conditioning node control response speed is determined through statistical analysis of the time difference between "control command issuance" and "temperature reaching target" in historical air conditioning operation data, using a time-series statistical algorithm, characterizing the air conditioning's response efficiency to temperature regulation needs at that level. Based on these feature extraction results, a subset of node features corresponding to each level is generated, ensuring accurate matching of feature data for each level with the spatiotemporal scenario.

[0036] A node-linked architecture with regional nodes as the core hub is established, clarifying the functional positioning of environmental parameter nodes, passenger preference nodes, and air conditioning operation nodes: environmental parameter nodes, as influencing factor nodes, provide external environmental input; passenger preference nodes, as demand-oriented nodes, clarify temperature regulation targets; and air conditioning operation nodes, as execution feedback nodes, provide feedback on the regulation effect. A spatial topology mapping algorithm is used to spatially link regional nodes with other nodes, combined with a preference type identification algorithm to distinguish different levels of preference feature types, generating dynamic attribute information including node spatial location, functional type, and feature values. Edge weight association information is generated through two types of algorithms: parameter influence is calculated using the mutual information entropy algorithm to quantify the degree of influence of environmental parameter nodes and passenger preference nodes on regional nodes; the preference-regulation response time-series relationship is analyzed using a time-series association algorithm to characterize the response delay and association strength between passenger preference nodes and air conditioning operation nodes. Edge weights range from 0 to 1; higher weight values ​​indicate a tighter association and more significant interaction between nodes.

[0037] An adaptive directed graph is constructed, comprising all regional nodes, multi-dimensional environmental nodes, differentiated preference nodes, and intelligent control nodes. It is divided into three views based on comfort demand constraints, energy consumption optimization constraints, and real-time response constraints. A preference-control collaborative attention mechanism optimizes the cross-view edge association strength, and a passenger preference self-learning unit is incorporated to dynamically update scenario-based preference weights. The graph is constructed by integrating all regional nodes, multi-dimensional environmental nodes, differentiated preference nodes, and intelligent control nodes. Based on constraint type, the graph is divided into three views: a comfort demand constraint view focuses on passenger comfort preference satisfaction; an energy consumption optimization constraint view emphasizes air conditioning energy consumption control targets; and a real-time response constraint view highlights the timeliness of control command execution. A preference-control collaborative attention mechanism is introduced to optimize the cross-view edge association strength. This mechanism uses an attention weight allocation algorithm to adjust the edge weight ratio of each view according to the core objective of the current scenario (e.g., prioritizing comfort needs in high-density passenger flow scenarios), achieving collaborative fusion of cross-view features. The system incorporates a passenger preference self-learning unit, which uses an online learning algorithm to dynamically update the comfort preference weights under different spatiotemporal scenarios, taking real-time collected passenger comfort feedback data as input. The update cycle is consistent with the dynamic perception window (5 minutes / time), ensuring that the preference weights can adapt to scenario changes in real time and improving the directed graph's ability to dynamically represent passenger comfort needs.

[0038] An adaptive multi-view dynamic graph neural network architecture was constructed, comprising four layers: an input layer, a feature fusion layer, a spatiotemporal attention layer, and an output layer. The input layer receives optimized multi-constraint directed graph data. The feature fusion layer uses a weighted summation algorithm to fuse features from the three views, with the fusion weights determined by a preference-regulation collaborative attention mechanism. The spatiotemporal attention layer captures the dynamic interaction patterns between nodes through a spatiotemporal attention module. Temporal convolutional networks are used to extract feature evolution trends in the temporal dimension, while graph attention networks are used in the spatial dimension to strengthen the inter-node association features. The output layer uses a fully connected layer to output node embedding vectors. During model training, minimizing node feature reconstruction error is the objective. The Adam optimizer is used to adjust parameters, with a learning rate of 0.001 and 100 iterations. After each iteration, the model performance is verified using a validation set to ensure convergence. This neural network performs hierarchical feature fusion and dynamic evolution modeling, ultimately outputting node embedding information containing environment-preference-regulation semantic associations. The embedding vector dimension is set to 256 dimensions, comprehensively representing the feature associations and semantic information of each node in the spatiotemporal scene.

[0039] S105 injects node embedding information into each layer of the value network and policy network of the reinforcement learning model, optimizes the model parameters using an efficient parameter fine-tuning method, introduces an energy consumption-comfort dual-objective adversarial optimization module, and generates a decision interpretation chain through feature importance visualization to form a semantic association information between fused environmental features, passenger preference features and air conditioning control strategies.

[0040] In one implementation, based on the requirements of dual-objective optimization and semantic fusion, a fusion architecture is constructed, comprising a node embedding cross-layer injection module and a reinforcement learning value-policy dual network, to deeply inject and semantically model environmental features, passenger preference features, and air conditioning control features. Based on the core requirements of energy consumption-comfort dual-objective optimization and multi-feature semantic fusion, a deep fusion architecture of "node embedding cross-layer injection module + reinforcement learning value-policy dual network" is constructed. The node embedding cross-layer injection module adopts a fully connected mapping method, converting 256-dimensional node embedding information into feature dimensions adapted to the reinforcement learning network, achieving smooth feature transition and deep injection. Both the reinforcement learning value network and the policy network adopt a 3-layer fully connected layer structure. The value network is responsible for evaluating the comprehensive benefits of temperature compensation and energy-saving control under the current state, and the policy network is responsible for outputting the specific control policy distribution. The architecture establishes semantic modeling of environmental features, passenger preference features, and air conditioning control features through a two-way information interaction mechanism. Environmental features provide external constraint input, passenger preference features clarify the core demand orientation, and air conditioning control features provide feedback on execution effects. The three are fused and semantically associated at each layer through the cross-layer injection module, ensuring that the architecture can simultaneously respond to the dual-objective optimization requirements.

[0041] Based on the energy consumption-comfort collaborative optimization strategy, the feature injection logic is designed, clarifying the injection dimensions and feature fusion ratios of node embedding in the value assessment layer of the value network and the decision output layer of the strategy network, and generating network-feature adaptation rules. Based on the energy consumption-comfort collaborative optimization strategy, the injection logic and fusion ratios of node embedding in the reinforcement learning dual network are clarified. The injection dimensions of the value assessment layer of the value network are set to 128 dimensions, with a feature fusion ratio of 40% for node embedding features, 30% for environmental features, 20% for passenger preference features, and 10% for air conditioning control features, focusing on adapting to the comprehensive needs of benefit assessment. The injection dimensions of the decision output layer of the strategy network are set to 64 dimensions, with a feature fusion ratio of 30% for node embedding features, 20% for environmental features, 30% for passenger preference features, and 20% for air conditioning control features, highlighting the strong correlation between decision-making, demand, and execution. Through the clarification of the above dimensions and ratios, network-feature adaptation rules are generated. The rules define the feature input types, dimension matching standards, and fusion weight allocation basis for different network levels, ensuring that node embedding and various features achieve accurate adaptation and efficient fusion in the dual network, avoiding feature conflicts or information redundancy.

[0042] To adapt to varying airport passenger flow densities, environmental fluctuations, and energy-saving target thresholds, a dynamic weight balancing mechanism is established. In high-density passenger flow scenarios, the weight of comfort features is increased; in low-load operation scenarios, the proportion of energy consumption optimization features is strengthened; and in extreme environmental scenarios, the weight allocation of both objectives is balanced. Based on the adaptation needs of actual airport operation scenarios, a scenario-aware dynamic weight balancing mechanism is constructed to achieve dynamic adaptation between energy consumption and comfort objectives. In high-density passenger flow scenarios, the criterion is ≥3 passengers per square meter. In this case, the weight of comfort features is increased to 0.6, and the weight of energy consumption features is decreased to 0.4, prioritizing the comfort experience of large numbers of passengers. In low-load operation scenarios, the criterion is <1 passenger per square meter and air conditioning energy consumption exceeding a preset threshold by 15%. In this case, the weight of energy consumption optimization features is increased to 0.7, and the weight of comfort features is decreased to 0.3, focusing on achieving energy-saving targets. In extreme environmental scenarios, the criterion is temperature >35℃ or <5℃ and humidity >80% or <30%. The weight of both objectives is set to 0.5, balancing comfort experience and energy consumption control, and avoiding operational risks caused by excessive emphasis on a single objective. The weight adjustment is triggered in real time through the scene recognition algorithm, and the adjustment cycle is consistent with the dynamic perception window to ensure the real-time responsiveness of the mechanism.

[0043] An adversarial optimization module with dual objectives of energy consumption and comfort is introduced. This module iteratively optimizes network parameters using the LoRA parameter efficient fine-tuning method, while simultaneously launching a feature importance visualization tool to generate a complete explanatory chain including feature-decision-effect. The module employs an adversarial training framework, constructing a dual-objective loss function with the two adversarial objectives of maximizing comfort satisfaction and minimizing energy consumption. The LoRA parameter efficient fine-tuning method iteratively optimizes the dual network parameters of the reinforcement learning system. The LoRA adapter dimension is set to 64, the rank to 8, the learning rate to 0.0001, and the number of iterations to 200 rounds. After each iteration, the dual-objective loss value is calculated. Optimization stops when the loss value converges to a preset threshold (≤0.01), ensuring efficient parameter updates without reconstructing the network structure. A feature importance visualization tool is launched simultaneously, employing a gradient-weighted activation mapping algorithm to quantify the contribution of each feature to the decision result. This generates a full-link explanation chain of "feature input - weight allocation - decision output - effect feedback," which is presented in the form of a visual chart. This clearly shows the role path and decision logic of core features in different scenarios, improving the interpretability and credibility of the decision.

[0044] The system integrates and executes the fusion architecture, injection rules, dynamic weighting mechanism, and adversarial optimization module to output time-zone temperature compensation and energy-saving control decision-making foundational data containing dynamic semantic associations of environment, preference, and regulation. The system integrates and executes the fusion architecture, feature injection rules, dynamic weighting mechanism, and dual-objective adversarial optimization module to construct an end-to-end processing flow. First, the fusion architecture completes deep feature injection and semantic modeling. Then, based on the injection rules, it achieves accurate adaptation between features and the network. Next, the dynamic weighting mechanism adjusts the priority of the two objectives. Finally, the adversarial optimization module completes parameter optimization and interpretation chain generation. During the integration and execution process, a data interaction protocol and synchronization mechanism are established between modules to ensure efficient and timely collaboration. The final output is time-zone temperature compensation and energy-saving control decision-making foundational data containing dynamic semantic associations of environment, preference, and regulation. The data covers core information such as comfort demand priorities, energy consumption control thresholds, temperature compensation ranges, and air conditioning operating parameter suggestions for each functional zone and time period, providing comprehensive and accurate decision support for subsequent control command generation.

[0045] S106 processes the target data set based on a reinforcement learning model combined with fused semantic association information. Through the scene transfer reasoning module, it simulates the temperature compensation effect and energy consumption changes under different passenger flow and environmental change scenarios, and generates time-division and zone-division AI temperature compensation parameters and air conditioning energy-saving control instructions.

[0046] In one implementation, the reinforcement learning model uses fused semantic association information as its core input, combining real-time environmental parameters, passenger behavior trajectories, comfort feedback, and other data from the target dataset to construct a "state-action-reward" reinforcement learning framework. The state space is defined as a high-dimensional vector containing environmental features, passenger preference features, and air conditioning operating status features. The action space is set as a combination of air conditioning control parameters such as temperature adjustment range, air supply speed, and operating mode. The reward function is designed as a weighted sum of an energy consumption penalty term and a comfort satisfaction reward term, where comfort satisfaction is calculated based on passenger feedback data and preference features, and the energy consumption penalty term is quantified based on the product of air conditioning operating power and time. The model updates the parameters of the value network and policy network through a temporal difference learning algorithm, utilizing dynamic association rules from the fused semantic association information to accurately capture the mapping relationship between the environment, preferences, and control strategies, ensuring that the model's decisions are highly adapted to the needs of the actual scenario.

[0047] A scenario migration reasoning module was built, employing an architecture combining case-based reasoning and generative adversarial networks (GANs) to accurately simulate various scenarios involving changes in passenger flow and environment. The module first constructs a scenario feature library, including core feature dimensions such as passenger flow density (low / medium / high), environmental parameters (normal / extreme ranges for temperature / humidity / wind speed / light intensity), and time period type (peak / off-peak / nighttime). Each feature dimension corresponds to multiple scenario values. A case retrieval algorithm matches similar scenarios from historical data as base cases. Then, a GAN is used to mutate and recombine the features of these base cases, generating new scenario instances covering various complex scenarios such as sudden changes in passenger flow, extreme weather, and time period transitions. During the simulation, the module outputs predicted values ​​for temperature compensation effects (e.g., time to reach target area temperature, improvement in comfort and satisfaction) and energy consumption changes (e.g., increase / decrease in energy consumption per unit time, and achievement of total energy consumption control threshold) for each scenario, providing data support for the generation of control commands.

[0048] Based on the simulation results from the scene transfer inference module, and combined with spatial zoning and temporal granularity division rules, targeted AI temperature compensation parameters are generated. At the spatial zoning level, compensation parameters are calculated separately for each functional zone of the terminal (departure hall, boarding gate, waiting area, etc.). The target temperature range and temperature adjustment rate for each zone are determined based on differences in passenger density, environmental sensitivity, and preference intensity. At the temporal granularity level, the parameters are divided into hourly time periods, and the update frequency and adjustment range are dynamically adjusted based on passenger flow patterns and environmental change trends during each time period. The generation of temperature compensation parameters employs a multi-objective optimization algorithm, with the optimization objectives of maximizing comfort and minimizing energy consumption. Constraints are set as follows: temperature adjustment range (18℃-26℃), upper limit of adjustment rate (0.5℃ / minute), and energy consumption threshold (set based on historical best energy consumption data). The optimal solution is obtained through particle swarm optimization, outputting core parameters such as target temperature, temperature compensation difference, and compensation activation timing for each zone and time period.

[0049] The generated time-sharing and zone-specific AI temperature compensation parameters are converted into executable air conditioning energy-saving control commands. These commands cover key control items such as air conditioning operation mode (cooling / heating / ventilation), set temperature, airflow speed, and operating time. During command generation, a control strategy mapping algorithm is used to adapt the temperature compensation parameters to the control protocol of the air conditioning equipment, ensuring that the command format meets the equipment's execution requirements. Simultaneously, a real-time feedback verification mechanism is introduced. Combining real-time temperature data collected by environmental sensors with passenger comfort feedback data, the execution effect of the control commands is dynamically evaluated. If the actual temperature deviates from the target temperature by more than 0.3℃ or the comfort satisfaction is lower than a preset threshold (4 points), a command adjustment mechanism is triggered. The compensation parameters are recalculated using a reinforcement learning model, and new control commands are generated. Finally, the control commands are transmitted in real-time to the air conditioning control terminals in each zone via a wireless communication module, achieving precise time-sharing and zone-specific temperature compensation and energy-saving control.

[0050] like Figure 2 As shown, a time-sharing and zone-based AI temperature compensation and air conditioning energy-saving control device that considers airport passenger comfort preferences through self-learning includes:

[0051] The data acquisition module 201 is used to acquire the target data set and the training sample set, wherein the training sample set includes the airport passenger behavior trajectory dataset, the real-time environmental parameter dataset, and the passenger comfort feedback dataset, and to complete the collection and aggregation of multi-source data.

[0052] The multi-dimensional preference dataset generation module 202 is used to extract features of passenger stay area, activity intensity and stay duration based on passenger behavior trajectory dataset through deep learning model, combine real-time environmental parameter dataset, convert behavioral semantic features into comfort demand vector through cross-modal feature fusion module, quantify the preference level of passenger comfort feedback data into weighted features, and generate multi-dimensional passenger comfort preference dataset for temperature compensation and energy-saving control.

[0053] The directed multi-view dynamic graph construction module 203 is used to process the multi-dimensional passenger comfort preference dataset, and to layer it according to spatial partitioning and temporal granularity. Each layer contains regional nodes, environmental parameter nodes, passenger preference nodes, and air conditioning operation nodes. Through the dynamic perception window, real-time features of different layers are aggregated to generate a target directed multi-view dynamic graph composed of environmental status, passenger preferences, and air conditioning operation status.

[0054] The node embedding information generation module 204 is used to model the target directed multi-view dynamic graph based on the adaptive multi-view dynamic graph neural network, introduce a passenger preference self-learning unit to dynamically update the comfort preference weights under different scenarios, and output node embedding information containing the semantic association between environment, preference and regulation.

[0055] The dual-objective fusion optimization module 205 is used to inject node embedding information into each layer of the value network and policy network of the reinforcement learning model, optimize the model parameters using an efficient parameter fine-tuning method, introduce an energy consumption-comfort dual-objective adversarial optimization module, and generate a decision interpretation chain through feature importance visualization to form the semantic association information of fused environmental features, passenger preference features and air conditioning control strategy.

[0056] The temperature compensation and energy-saving control instruction generation module 206 is used to process the target data set based on the reinforcement learning model and the fusion of semantic association information. Through the scene transfer reasoning module, it simulates the temperature compensation effect and energy consumption changes under different passenger flow and environmental change scenarios, and generates time-division and zone-division AI temperature compensation parameters and air conditioning energy-saving control instructions.

[0057] A computing device includes a memory for storing computer program instructions and a processor for executing the computer program instructions, wherein when the computer program instructions are executed by the processor, the device is triggered to execute any of the time-sharing and zone-based AI temperature compensation and air conditioning energy-saving control methods that take into account airport passenger comfort preferences through self-learning.

[0058] The methods and / or embodiments in this application can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowchart. When the computer program is executed by a processing unit, it performs the functions defined in the methods of this application.

[0059] It should be noted that the computer-readable medium described in this application can be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this application, a computer-readable medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0060] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0061] It will be apparent to those skilled in the art that this application is not limited to the details of the exemplary embodiments described above, and that this application can be implemented in other specific forms without departing from the spirit or essential characteristics of this application. Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects, and the scope of this application is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be embraced within this application.

Claims

1. A time-sharing and zone-based AI temperature compensation and air conditioning energy-saving control method that considers airport passenger comfort preferences through self-learning, characterized in that... include: Obtain the target dataset and training sample set, where the training sample set includes the airport passenger behavior trajectory dataset, the real-time environmental parameter dataset, and the passenger comfort feedback dataset; Based on the passenger behavior trajectory dataset, a deep learning model is used to extract features of passenger dwelling area, activity intensity and dwell time. Combined with the real-time environmental parameter dataset, the passenger behavior semantic features are converted into comfort demand vectors through a cross-modal feature fusion module. The preference level of passenger comfort feedback data is quantified into weighted features to generate a multi-dimensional passenger comfort preference dataset for temperature compensation and energy-saving control. The multi-dimensional passenger comfort preference dataset is processed and layered according to spatial partitioning and temporal granularity. Each layer contains regional nodes, environmental parameter nodes, passenger preference nodes, and air conditioning operation nodes. Real-time features of different layers are aggregated through a dynamic perception window to generate a target directed multi-view dynamic graph composed of environmental status, passenger preferences, and air conditioning operation status. The target directed multi-view dynamic graph is modeled and processed based on an adaptive multi-view dynamic graph neural network. A passenger preference self-learning unit is introduced to dynamically update the comfort preference weights under different scenarios and generate node embedding information. Node embedding information is injected into each layer of the value network and policy network of the reinforcement learning model. The model parameters are optimized by using an efficient parameter fine-tuning method. An adversarial optimization module with dual objectives of energy consumption and comfort is introduced. At the same time, a decision interpretation chain is generated by visualizing the importance of features, forming a semantic association information between the fused environmental features, passenger preference features and air conditioning control strategies. Based on the reinforcement learning model and combined with the fusion of semantic association information, the target data set is processed. The scene transfer reasoning module simulates the temperature compensation effect and energy consumption changes under different passenger flow and environmental change scenarios, and generates time-division and zone-division AI temperature compensation parameters and air conditioning energy-saving control instructions.

2. The time-sharing and zone-based AI temperature compensation and air conditioning energy-saving control method considering airport passenger comfort preferences as described in claim 1, is characterized in that... Based on passenger behavior trajectory datasets, deep learning models are used to extract features from passenger dwell areas, activity intensity, and dwell time. Combined with real-time environmental parameter datasets, a cross-modal feature fusion module transforms passenger behavioral semantic features into comfort demand vectors. Passenger comfort feedback data preference levels are quantified into weighted features, generating a multi-dimensional passenger comfort preference dataset for temperature compensation and energy-saving control, including: Key behavioral elements of the passenger behavior trajectory dataset are analyzed to extract behavioral features including spatial coordinates of the stay area, activity intensity level, and stay duration range. The real-time environmental parameter dataset is decomposed into environmental factors to obtain environmental features including temperature, humidity, wind speed, and light intensity. We perform preference semantic transformation on the passenger comfort feedback dataset, mapping subjective evaluations to comfort preference levels and sensitivity coefficients to generate preference features; The cross-modal feature fusion module is used to align and correlate behavioral features, environmental features, and preference features to construct a multi-dimensional feature association matrix. A dynamic weight allocation mechanism is introduced to prioritize and weight the fused features, generating weighted features. Based on the weighted feature matrix, a standardized comfort demand vector is generated, and the quantified weighted features are integrated to construct a multi-dimensional passenger comfort preference dataset for temperature compensation and energy-saving control.

3. The time-sharing and zone-based AI temperature compensation and air conditioning energy-saving control method considering airport passenger comfort preferences as described in claim 2, is characterized in that... The multi-dimensional passenger comfort preference dataset is processed, layered by spatial partitioning and temporal granularity. Each layer contains regional nodes, environmental parameter nodes, passenger preference nodes, and air conditioning operation nodes. Real-time features from different layers are aggregated through a dynamic perception window to generate a target directed multi-view dynamic graph composed of environmental status, passenger preferences, and air conditioning operation status, including: Based on the feature distribution characteristics of the multi-dimensional passenger comfort preference dataset and the airport spatial-temporal association rules, the data processing module and the graph structure generation engine negotiate to determine the hierarchical aggregation strategy, and determine the node level and association dimension through spatial partition attribute parsing and temporal granularity division. A dual-dimensional partitioning mechanism of partitioning and time-sharing was adopted to split the multi-dimensional passenger comfort preference dataset. Corresponding feature data were extracted according to the terminal functional partitions and time intervals, and the core elements of region, environment, preference and air conditioning operation were associated to generate preliminary hierarchical data blocks. The validity of the initial layered data blocks is verified, the regional calibration service is initiated for data with ambiguous spatial boundaries, the time-series synchronization mechanism is triggered for data with inconsistent timestamps, and the feature completion process is performed for samples with missing node features. An optimized layered data scheme including regional calibration scheme, time-series synchronization strategy and feature completion method is generated. The optimized hierarchical data scheme is integrated and executed. Combined with the dynamic perception window, the feature changes of each level are captured in real time. The node association weight and feature aggregation progress feedback are generated synchronously. A hierarchical network containing regional nodes, environmental parameter nodes, passenger preference nodes, and air conditioning operation nodes is constructed to form a target directed multi-view dynamic graph composed of environmental status, passenger preference, and air conditioning operation status.

4. The time-sharing and zone-based AI temperature compensation and air conditioning energy-saving control method considering airport passenger comfort preferences and self-learning, as described in claim 1, is characterized in that... The target directed multi-view dynamic graph is modeled and processed based on an adaptive multi-view dynamic graph neural network. A passenger preference self-learning unit is introduced to dynamically update the comfort preference weights under different scenarios and generate node embedding information, including: The target directed multi-view dynamic graph is structured and analyzed, and layered according to spatial partition type and time granularity level. Dynamic features of regional node environmental sensitivity, preference node demand intensity, and air conditioning node control response speed are extracted in each layer to generate node feature subsets at each level. With regional nodes as the core hub, environmental parameter nodes as influencing factor nodes, passenger preference nodes as demand-oriented nodes, and air conditioning operation nodes as execution feedback nodes, dynamic attribute information of nodes is generated by combining spatial topology mapping and preference type identification, and edge weight association information is generated according to parameter influence degree and preference-control response time sequence relationship. An adaptive directed graph containing all regional nodes, multi-dimensional environmental nodes, differentiated preference nodes, and intelligent control nodes is constructed. It is divided into three views according to comfort demand constraints, energy consumption optimization constraints, and real-time response constraints. The cross-view edge association strength is optimized through a preference-control collaborative attention mechanism, and the scenario-based preference weights are dynamically updated by incorporating a passenger preference self-learning unit. Based on the adaptive multi-view dynamic graph neural network, hierarchical feature fusion and dynamic evolution modeling are performed on the optimized multi-constraint directed association graph. The spatiotemporal attention module captures the dynamic interaction patterns between nodes and outputs node embedding information containing the semantics of environment-preference-regulation association.

5. The time-sharing and zone-based AI temperature compensation and air conditioning energy-saving control method considering airport passenger comfort preferences and self-learning as described in claim 1, characterized in that, Node embedding information is injected into each layer of the value network and policy network of the reinforcement learning model. A parameter-efficient fine-tuning method is used to optimize model parameters. An adversarial optimization module with dual objectives of energy consumption and comfort is introduced. Simultaneously, a decision interpretation chain is generated through feature importance visualization, forming a fused semantic association between environmental features, passenger preference features, and air conditioning control strategies, including: Based on the requirements of dual-objective optimization and semantic fusion, a fusion architecture including a node embedding cross-layer injection module and a reinforcement learning value-policy dual network is built to perform deep injection and associated semantic modeling of environmental features, passenger preference features, and air conditioning control features. Based on the energy consumption-comfort collaborative optimization strategy, the feature injection logic is designed, the injection dimension and feature fusion ratio of the node embedded in the value assessment layer of the value network and the decision output layer of the strategy network are clarified, and the network-feature adaptation rules are generated. Based on the adaptation requirements of airport passenger flow density, environmental fluctuation range and energy-saving target threshold, a dynamic weight balancing mechanism is set up. In high-density passenger flow scenarios, the weight of comfort features is increased; in low-load operation scenarios, the proportion of energy consumption optimization features is strengthened; and in extreme environment scenarios, the weight allocation of the two objectives is balanced. An adversarial optimization module with dual objectives of energy consumption and comfort is introduced. The network parameters are iteratively optimized through the LoRA parameter fine-tuning method. At the same time, a feature importance visualization tool is launched to generate a full-link explanation chain including features, decisions, and effects. The system integrates and executes the fusion architecture, injection rules, dynamic weighting mechanism, and adversarial optimization module to output time-division and zone-based temperature compensation and energy-saving control decision-making basis data containing dynamic semantic associations of environment, preference, and regulation.

6. A time-sharing and zone-based AI temperature compensation and air conditioning energy-saving control device that considers airport passenger comfort preferences through self-learning, characterized in that... The device is configured to perform any of the time-sharing and zone-based AI temperature compensation and air conditioning energy-saving control methods that take into account airport passenger comfort preferences.

7. An electronic device, characterized in that, include: First processor; and memory for storing executable instructions of the first processor; The first processor is configured to execute the time-sharing and zone-based AI temperature compensation and air conditioning energy-saving control method according to any one of claims 1 to 5 by executing the executable instructions.

8. A computing device comprising a memory for storing computer program instructions and a second processor for executing the computer program instructions, wherein, When the computer program instructions are executed by the second processor, the device is triggered to execute the time-sharing and zoned AI temperature compensation and air conditioning energy-saving control method that takes into account airport passenger comfort preferences as described in any one of claims 1 to 5.

Images