Airport terminal air conditioner energy saving method and system based on passenger density prediction model

Abstract

The invention relates to the technical field of airport equipment energy saving, in particular to an airport terminal air conditioner energy saving method and system based on a passenger density prediction model.The method comprises the steps that a basic data set used for air conditioner energy saving adjustment is obtained, and the obtained basic data set is preprocessed; constructing a passenger density prediction model by taking the preprocessed basic data set as input; carrying out model training configuration on the constructed passenger density prediction model; carrying out air conditioner load calculation based on a predicted value of passenger density; carrying out control strategy optimization according to the air conditioner load; according to the method, the optimal operation parameters of all subsystems of the air conditioner are solved through model prediction control with the minimum total energy consumption of the system as the target and in combination with indoor comfort constraint and equipment operation constraint, the energy saving performance and the comfort are effectively balanced, and the energy consumption of the air conditioning system of the airport terminal is remarkably reduced.

Description

Technical Field

[0001] This invention relates to the field of airport equipment energy-saving technology, and in particular to a terminal air conditioning energy-saving method and system based on a passenger density prediction model. Background Technology

[0002] With the rapid development of the global air transport industry, the scale of modern airport terminals continues to expand, and their functional zoning becomes increasingly complex. They exhibit significant characteristics such as large spaces, high personnel mobility, and long operating hours. As typical high-energy-consuming public buildings, the air conditioning system of airport terminals accounts for a large proportion of their total building energy consumption. Achieving efficient and energy-saving operation of the air conditioning system while ensuring the thermal comfort needs of passengers has become a key issue that airport operators and related technical fields urgently need to address.

[0003] In existing technologies, the control strategies for airport terminal air conditioning systems are mainly divided into two categories: one is the traditional control method based on fixed schedules or simple temperature feedback, which monitors the return air temperature or real-time indoor temperature and adjusts parameters such as the opening of the air conditioning water valve and the fan speed to maintain a stable indoor temperature. This method does not consider the dynamic changes in passenger density in the terminal with flight takeoffs and landings, resulting in the air conditioning system often being in a state of over- or under-powered operation. The other category, such as CN202410391999.0, "A Terminal Air Conditioning Control System and Method Based on the Internet of Things," divides the air conditioning control units, determines the effective number of people in each unit, and classifies them into levels to achieve control based on real-time passenger flow. While density-based tiered air conditioning control can adjust cooling capacity according to real-time differences in passenger distribution in different areas, thus optimizing energy efficiency to some extent, the large size of airport terminal buildings, long air conditioning system pipelines, and high thermal inertia mean that even if the equipment operation is adjusted immediately based on real-time passenger density, it still takes a long time for the air conditioning system to produce a significant change in indoor temperature. This sensing and reactive control logic cannot adapt to the actual scenario of rapid fluctuations in passenger density within the terminal building due to flight schedules, and it is still difficult to balance the timeliness, comfort, and energy efficiency of cooling supply. At present, there is a need for an energy-saving method and system for terminal air conditioning based on a passenger density prediction model. Summary of the Invention

[0004] To address the problem in existing airport terminal air conditioning systems that cannot accurately balance cooling supply comfort and energy efficiency while taking into account dynamic changes in time and space, this invention provides an energy-saving method and system for terminal air conditioning based on a passenger density prediction model.

[0005] In a first aspect, the present invention provides an energy-saving method for terminal air conditioning based on a passenger density prediction model, which adopts the following technical solution: An energy-saving method for airport terminal air conditioning based on a passenger density prediction model includes: S1. Obtain the basic dataset for air conditioning energy-saving regulation and preprocess the obtained basic dataset, including obtaining data from the airport flight information management system, outdoor meteorological data and air conditioning system operation data; S2. Construct a passenger density prediction model using the preprocessed basic dataset as input, including model construction using a bidirectional LSTM-Transformer hybrid model with attention mechanism as the framework; S3. Configure the model training for the completed passenger density prediction model, including training the model using a multi-task learning loss function and an adaptive optimizer. S4. Calculate the air conditioning load based on the predicted value of passenger density, including combining outdoor meteorological forecast parameters and using the total load calculation formula to calculate the cooling load demand and heating load demand of air conditioning in each zone for each time period in the future. S5. Optimize the control strategy according to the air conditioning load, including using optimization algorithms to solve for the optimal setpoint of each air conditioning subsystem with the goal of minimizing the total energy consumption of the system, generating control commands and sending them to the air conditioning actuators. S6. Real-time collection of environmental data for each zone through end sensors, comparison with target parameters and fine-tuning of control commands.

[0006] Furthermore, the preprocessing of the acquired basic dataset specifically includes: S11. Linear interpolation or forward padding is used to maintain the continuity of the time series for the sensor data stream in the basic dataset. For missing fields in flight data, the missing fields are filled by backtracking the historical records of the airport flight information management system. For missing values ​​in outdoor meteorological data, weighted replacement is used with data from nearby meteorological stations. S12. Using a sliding time window, the average, maximum and minimum values ​​of the second-level high-frequency sensor data are aggregated within the window, the hour-level low-frequency external data are filled forward, an airport geographic grid coordinate system is established, meteorological data are associated with the corresponding area according to grid ID, and flight dynamic data are matched with the GPS coordinates of the parking position. S13. Using flight number as the primary key, associate the airport flight information management system's planned data with radar real-time data, establish a topological relationship based on meteorological station ID and sensor deployment location, establish a multidimensional index through spatiotemporal hashing of longitude, latitude, and time window, and resolve data conflicts according to preset priorities. S14. Finally, standardize all data in the basic dataset and verify the validity of the preprocessed data using a data quality assessment formula, which is: , in, This is a data quality assessment value. This represents the total number of missing values ​​in the preprocessed dataset. The total amount of data in the dataset. This represents the total number of outliers in the preprocessed dataset. To standardize the data consistency coefficient, , and These are different weighting coefficients.

[0007] Furthermore, the construction of the passenger density prediction model includes first extracting the time features, spatial features, event features, and environmental features of the basic dataset from the input layer to form a multi-dimensional feature vector and dividing the input sequence. Then, through the embedding layer, the comprehensive features in a unified feature space are output by embedding classification features, transforming and concatenating continuous features by linear projection. The comprehensive features are then input into a bidirectional LSTM layer, and after forward and backward calculations and concatenation, they are fed into a two-layer Transformer encoder layer. Long-term dependencies are captured by the attention formula. Next, the features are input into a spatiotemporal attention mechanism layer, and after multi-head attention calculations in the time dimension and spatial dimension, as well as tensor transpose processing, spatiotemporal fusion features are obtained. Finally, the fully connected output layer is processed by a multilayer perceptron to output the predicted passenger density values ​​for each zone in the future.

[0008] Furthermore, the construction of the passenger density prediction model also includes a model input feature selection mechanism. Based on the preprocessed basic dataset, the correlation between each feature and passenger density is calculated using mutual information values. A correlation threshold is set for feature selection to obtain the input feature set. Simultaneously, L1 regularization is used to constrain the feature weights of the embedding layer and the linear projection layer. The formula for calculating the mutual information value is: , in, For a single input feature, For passenger density labels, Let X be the joint probability distribution of Y. and Let X and Y be the marginal probability distributions, respectively.

[0009] Furthermore, the configuration for training the constructed passenger density prediction model includes configuring a multi-task learning loss function. This multi-task learning loss function integrates point prediction loss, interval prediction loss, and uncertainty estimation loss through weight coefficients. Specifically, the point prediction loss uses Huber loss, the interval prediction loss uses quantile loss, and the uncertainty estimation loss is used to quantify the confidence of the prediction results. An adaptive optimizer and learning rate scheduling strategy are configured. An adaptive optimizer with weight decay is selected to update the model parameters, and a periodic learning rate scheduler is used to decay the learning rate from the initial value to the minimum value according to a cosine curve at a preset period before restarting.

[0010] Furthermore, the air conditioning load calculation based on passenger density predictions includes obtaining passenger density predictions for each zone in future time periods and corresponding outdoor weather prediction parameters. The load calculation engine is then invoked to calculate the air conditioning cooling and heating loads by zone and time period. Using the passenger density predictions as the core input, and combining building heat transfer losses, equipment heat dissipation, and fresh air heat exchange requirements, the total load calculation formula integrates passenger load, building load, and fresh air load to obtain the cooling and heating load values ​​for each zone at each future time step. The total load calculation formula is as follows: , in, For the cooling and heating load generated by a unit of passengers, This is a predicted value for passenger density in each zone. To accommodate the building's load in different zones, The fresh air load for each zone.

[0011] Furthermore, the calculation of air conditioning load based on passenger density predictions also includes obtaining outdoor meteorological forecast parameters in real time through a meteorological data interface and aligning them with the passenger density prediction time step. The unit passenger load is dynamically adjusted according to the outdoor meteorological forecast parameters. The fresh air load is calculated by combining outdoor temperature and humidity with indoor target temperature and humidity. Simultaneously, basic building load parameters are preset according to zoning type, and the heat transfer characteristics of the building envelope in different areas are adapted through correction coefficients to ultimately form a load demand matrix. The formula for calculating the fresh air load is: , in, air density, For the fresh air volume of each zone, The specific heat capacity of air at constant pressure. Forecast outdoor temperature, The target indoor temperature.

[0012] Furthermore, the control strategy optimization based on air conditioning load includes minimizing total system energy consumption as the core objective, while simultaneously constraining indoor temperature to remain within a preset comfort range. An optimization model is constructed using a model predictive control algorithm, taking the future cooling and heating load demands of each zone as inputs to the optimization model. The optimal operating parameters for air conditioning subsystems such as chillers, water pumps, fresh air units, and VAV terminals are then calculated to form zone-specific time-based control strategies. These optimal operating parameters are then converted into control commands conforming to the industrial bus protocol and sent to the corresponding programmable logic controllers. The objective function of the optimization model is: , in, To optimize the overall objective value, To predict the number of time steps, Let t be the total operating power of the air conditioning system in step t. Let t be the predicted value of the actual indoor temperature at step t. The target indoor temperature For temperature tolerance, , These are the power weighting coefficient and the temperature weighting coefficient, respectively.

[0013] Furthermore, the optimization of the control strategy based on the air conditioning load also includes setting equipment operation constraints and comfort constraints during the optimization algorithm solution process. The equipment operation constraints include air volume constraints and supply air temperature change rate constraints. After obtaining the optimal setpoint, corresponding control commands are generated according to the equipment type, and a command issuance verification mechanism is established. The formula for the supply air temperature change rate constraint is: , in, The setpoint for the supply air temperature at step t. The supply air temperature setpoint for step t-1. The interval between adjacent time steps.

[0014] Secondly, an energy-saving terminal air conditioning system based on a passenger density prediction model includes: The data acquisition module is configured to: acquire the basic dataset for air conditioning energy-saving adjustment, and preprocess the acquired basic dataset, including acquiring data from the airport flight information management system, outdoor meteorological data, and air conditioning system operation data; The model module is configured to build a passenger density prediction model using the preprocessed base dataset as input, including model building using a bidirectional LSTM-Transformer hybrid model with an attention mechanism as the framework. The training module is configured to: perform model training on the completed passenger density prediction model, including training the model using a multi-task learning loss function, an adaptive optimizer, and a phased training strategy; The load module is configured to calculate the air conditioning load based on the predicted value of passenger density, including combining outdoor meteorological forecast parameters and using the total load calculation formula to calculate the cooling load demand and heating load demand of air conditioning in each zone for each time period in the future. The optimization module is configured to optimize the control strategy based on the air conditioning load, including using optimization algorithms to solve for the optimal setpoints of each air conditioning subsystem with the goal of minimizing the total energy consumption of the system, generating control commands and sending them to the air conditioning actuators. The output module is configured to: collect environmental data of each zone in real time through the end sensor, compare it with the target parameters, and fine-tune the control commands.

[0015] In summary, the present invention has the following beneficial technical effects: 1. This invention solves the problems of heterogeneous data formats, spatiotemporal misalignment, and inconsistent quality of multi-source data by preprocessing data from airport flight information management systems, outdoor meteorological data, and air conditioning system operation data. This includes targeted completion of missing values, spatiotemporal alignment, priority resolution of data conflicts, and quantitative verification of data quality. It forms a standardized and highly reliable basic dataset, providing solid data support for the accurate construction of subsequent passenger density prediction models, avoiding prediction deviations caused by data problems, and ensuring the effectiveness of air conditioning energy-saving strategies.

[0016] 2. This invention constructs a bidirectional LSTM-Transformer hybrid model with an attention mechanism, supplemented by a feature selection mechanism. Mutual information value calculation filters highly correlated features, and L1 regularization constraints eliminate redundant features. It can accurately capture the long-term dependence and periodic patterns of passenger density changes over time, and can also deeply explore the correlation between different spatial areas. This improves the accuracy and generalization ability of passenger density prediction for each zone in the next 1-2 hours, breaks through the limitations of traditional post-event perception, and provides core decision-making basis for the early control of air conditioning systems, solving the problem of air conditioning lag in large-space buildings from the source.

[0017] 3. This invention integrates personnel load, building load, and fresh air load by dynamically calculating the load based on passenger density prediction values ​​and combining outdoor meteorological prediction parameters. It also dynamically adjusts the unit passenger load and adapts to the characteristics of the building envelope in each zone as needed, thereby achieving accurate quantification of the air conditioning cooling / heating load of each zone in each time period in the future. This avoids the problem of over- or under-supply of energy caused by the rough estimation of traditional load calculation, reduces energy waste, and improves energy utilization efficiency.

[0018] 4. This invention utilizes Model Predictive Control (MPC) optimization technology with the goal of minimizing total system energy consumption. By combining indoor comfort constraints and equipment operation constraints, it solves for the optimal operating parameters of each subsystem of the air conditioning system and generates control commands that conform to the industrial bus protocol. At the same time, it constructs a closed-loop control through real-time feedback from end sensors and a command fine-tuning mechanism. This ensures both the thermal comfort needs of passengers and the overall energy efficiency of the air conditioning system, effectively balancing energy saving and comfort, and significantly reducing the energy consumption of the airport terminal air conditioning system. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall process of an energy-saving method for terminal air conditioning based on a passenger density prediction model according to an embodiment of the present invention. Detailed Implementation

[0020] The present invention will be further described in detail below with reference to the accompanying drawings.

[0021] Example 1 Reference Figure 1This embodiment of an energy-saving method for terminal air conditioning based on a passenger density prediction model includes: S1. Obtain the basic dataset for air conditioning energy-saving regulation and preprocess the obtained basic dataset, including obtaining data from the airport flight information management system, outdoor meteorological data and air conditioning system operation data; S2. Construct a passenger density prediction model using the preprocessed basic dataset as input, including model construction using a bidirectional LSTM-Transformer hybrid model with attention mechanism as the framework; S3. Configure the model training for the completed passenger density prediction model, including training the model using a multi-task learning loss function, an adaptive optimizer, and a phased training strategy. S4. Calculate the air conditioning load based on the predicted value of passenger density, including combining outdoor meteorological forecast parameters and using the total load calculation formula to calculate the cooling load demand and heating load demand of air conditioning in each zone for each time period in the future. S5. Optimize the control strategy according to the air conditioning load, including using optimization algorithms to solve for the optimal setpoint of each air conditioning subsystem with the goal of minimizing the total energy consumption of the system, generating control commands and sending them to the air conditioning actuators. S6. Real-time collection of environmental data for each zone through end sensors, comparison with target parameters and fine-tuning of control commands.

[0022] Specifically, an energy-saving method for terminal air conditioning based on a passenger density prediction model includes the following steps: like Figure 1 As shown, S1, obtain the basic dataset for air conditioning energy-saving regulation, and preprocess the obtained basic dataset, including obtaining airport flight information management system data, outdoor meteorological data and air conditioning system operation data; First, three types of core basic data are synchronously acquired through edge gateways and standardized data interfaces deployed in the airport terminal: The first type is Airport Flight Information Management System (AODB) data, which automatically captures flight schedule data for the next 1-2 hours through the airport information integration platform API interface, including flight arrival / departure type, aircraft type, standard number of seats, historical on-time rate and corresponding parking stand number, with the data update frequency every 5 minutes; the second type is outdoor meteorological data, which collects outdoor temperature, relative humidity and solar radiation intensity in real time by connecting to the China Meteorological Administration API and airport local meteorological station sensors, with the data from the China Meteorological Administration updated hourly and the data from the airport local meteorological station updated every 10 minutes; the third type is air conditioning system operation data, which collects the current indoor temperature, air volume and air conditioning water valve opening of each functional area of ​​the terminal through the Modbus protocol interface of the air conditioning control system, and collects real-time indoor humidity data through the terminal temperature and humidity sensors deployed in each area. All of the above air conditioning system operation data are stored in the local cache of the edge gateway at a frequency of seconds.

[0023] To address the different characteristics of the three types of data, a differentiated missing value handling strategy is adopted: For sensor data streams (such as indoor temperature and air volume) in the air conditioning system operation data, if data is missing (the criterion is that no valid data is obtained for three consecutive sampling periods), the missing value is calculated using linear interpolation based on the valid data before and after the missing period. ,in, This is the most recent valid data before the missing period. for The corresponding time, This is the most recent valid data after the missing period. for For the corresponding time period, if the missing time period is located at the beginning of the data sequence, a forward imputation method is used to fill in subsequent missing values ​​with the first valid data value, ensuring the temporal continuity of the sensor data stream. For Airport Flight Information Management System (AODB) data, if fields such as aircraft type and number of seats are missing, the system's historical flight records for the past three months are traced back to find historical operation data for the same flight number, and the missing fields are filled with the aircraft type and number of seats that appear most frequently in the historical data; if it is a newly added flight with no historical records, the airport airline registration database is called to obtain the standard parameters of the corresponding aircraft type for the flight to fill in the missing data.

[0024] For outdoor meteorological data, if local meteorological station data at the airport is missing (e.g., due to sensor malfunction), weighted data from nearby meteorological stations will be used as a substitute. First, three nearby meteorological stations (denoted as Station 1, Station 2, and Station 3) within a 50km radius of the airport are identified, and the straight-line distance between each station and the airport is calculated. The weight of data from each site The missing meteorological data for the period is filled with values. ,in, The data are meteorological data from three neighboring stations during the same period, and distance weighting is used to ensure consistency between the supplementary data and the actual meteorological conditions at the airport.

[0025] In terms of temporal alignment, a 5-minute sliding time window is used as the uniform time granularity: for high-frequency sensor data collected at the second level, the average (avg), maximum (max), and minimum (min) values ​​of all valid data within each sliding time window are calculated, and the average value is used as the representative value of that time window to achieve dimensionality reduction and aggregation of high-frequency data; for low-frequency data updated at the hour level (such as meteorological data from the China Meteorological Administration and basic flight schedule data from AODB), a forward filling method is used to cover all missing data in the 5-minute sliding time window within the current data update cycle (e.g., 1 hour). In terms of spatial alignment, a 100m×100m airport geographic grid coordinate system is first established. Using the center point of the airport terminal as the origin (0,0), with the east-west direction as the X-axis and the north-south direction as the Y-axis, the T3 terminal and its surrounding areas (such as the apron and approach gates) are divided into grids. Each grid is assigned a unique 10-digit grid ID (the first 6 digits represent the X-axis coordinate range, and the last 4 digits represent the Y-axis coordinate range). Outdoor meteorological data is associated with the corresponding grid ID according to the spatial coverage of the meteorological station (e.g., if the local airport meteorological station is located in the area with grid ID 123456-7890, its collected data is associated with that grid and its three adjacent grids). The GPS coordinates (latitude and longitude) of the parking positions in the AODB flight dynamic data are converted into coordinate values ​​in the airport geographic grid coordinate system using a coordinate transformation formula, and then matched to the corresponding grid ID.

[0026] Data fusion is centered on entity association and multidimensional index construction: using flight number as the unique primary key, flight schedule data in the Airport Flight Information Management System (AODB) is associated with real-time flight data collected by the airport radar system to form a complete dynamic flight dataset. Based on the spatial topology relationship between meteorological station IDs and sensor deployment locations, outdoor meteorological data is associated with air conditioning system sensor data (such as indoor temperature and air volume) within the corresponding grid ID, establishing a mapping relationship between meteorological conditions and air conditioning operating status. Simultaneously, a multidimensional index is built using a spatiotemporal hash algorithm based on longitude, latitude, and time window. The longitude and latitude values ​​corresponding to each data point are each retained to two decimal places, and the time window is converted into a timestamp encoding with a 5-minute granularity. These three elements are concatenated to form a hash key, enabling rapid querying and association of data from different sources. When data conflicts occur, they are resolved according to preset priorities: When flight status data (such as whether a flight is delayed) conflicts, the priority order is: tower instruction data > radar real-time data > AODB planned data. That is, the latest instruction data issued by the tower is used first. If the tower data has not been updated, the radar real-time monitoring data is used, and finally the AODB initial planned data is used. When outdoor meteorological data (such as temperature) conflicts, the priority order is: data issued by the Central Meteorological Observatory > data collected by the airport's local meteorological station > data obtained from third-party APIs, to ensure the authority and accuracy of meteorological data. When multiple sensor data in the same area conflict, a majority voting mechanism is used. The readings of each sensor within the same time window are counted, and the reading with the highest frequency is selected as the valid data. If there is a tie, the average of all sensor readings is taken to ensure the reliability of sensor data.

[0027] Finally, standardization operations were performed on all data in the basic dataset, including unit conversion and coordinate system transformation. For coordinate system transformation, all data containing geographic location information (such as GPS coordinates of parking positions and weather station coordinates) in the WGS84 coordinate system were converted to the airport's local geographic grid coordinate system using a coordinate transformation matrix. Coordinate system parameters provided by the airport's surveying department were used during the conversion process to ensure accurate spatial coordinate matching. For time base alignment, timestamps from all data sources were uniformly converted to the UTC time zone, and the timestamp precision was standardized to the millisecond level to avoid data misalignment caused by time zone differences or inconsistencies in precision.

[0028] After standardization, the preprocessed dataset is validated using a data quality assessment formula, which is: , in, This is a data quality assessment value. This represents the total number of missing values ​​in the preprocessed dataset. The total amount of data in the dataset. This represents the total number of outliers in the preprocessed dataset. To standardize the data consistency coefficient, , and Different weighting coefficients are used when calculating If the dataset is deemed to be qualified for preprocessing, it is otherwise re-examined for parameter settings in missing value handling, spatiotemporal alignment, or conflict resolution. The processing logic is then optimized accordingly before preprocessing is performed again.

[0029] S2. Construct a passenger density prediction model using the preprocessed basic dataset as input, including model construction using a bidirectional LSTM-Transformer hybrid model with attention mechanism as the framework; Based on the preprocessed dataset from step S1, a model input feature filtering mechanism is constructed to ensure the effectiveness and conciseness of the input features: Time features (hours, days of the week, holiday status), spatial features (regional grid ID, region type such as check-in hall / security checkpoint / boarding gate), event features (flight arrival / departure type, aircraft type, number of seats, on-time rate), environmental features (outdoor temperature, humidity, solar radiation intensity), and system features (air conditioning airflow, water valve opening) are extracted from the preprocessed dataset, totaling 56 initial features. For each initial feature X, mutual information (MI) is used to calculate its correlation with the passenger density label Y. The formula for calculating the mutual information value is: , in, For a single input feature, For passenger density labels, Let X be the joint probability distribution of Y. and The marginal probability distributions of X and Y are set, and the correlation threshold is set to 0.3. Features with mutual information values ​​≥0.3 are included in the initial input feature set, and redundant features with low correlation are removed. At the same time, during the parameter training of the subsequent embedding layer and linear projection layer, L1 regularization is introduced to constrain the feature weights. That is, by adding a penalty term proportional to the absolute value of the feature weights to the loss function, the weights corresponding to redundant features gradually approach 0, thereby further realizing the dynamic filtering of features. Finally, an input feature set containing 42 effective features is formed. Based on the filtered input feature set, the model input sequence is constructed: the preprocessed dataset is sampled at 15-minute intervals, and the feature samples of 24 consecutive time steps are used as an input sequence. Each input sequence corresponds to a tensor with dimensions (24, 42) (24 is the time step length, 42 is the feature dimension). The feature vector of each time step contains the time, space, event, environment and system features at that moment.

[0030] The core function of the embedding layer is to transform categorical and continuous features in the input feature set into dense vectors in a unified feature space: Initializing the embedding dictionary, for categorical features, the hour feature uses an embedding layer with an embedding dimension of 64, mapping each hour integer to a 64-dimensional vector; the day of the week feature (values ​​range 0-6, 7 categories) uses an embedding layer with an embedding dimension of 32, matching its characteristic of having a small number of categories; the holiday / holiday feature (values ​​0 or 1, 2 categories) uses an embedding layer region type with an embedding dimension of 16. For continuous features (such as outdoor temperature, air volume, etc.), a linear projection layer is used for dimensionality transformation and numerical normalization. The input dimension of the linear projection layer is the number of continuous features (28), and the output dimension is... The output dimension is set to 64, consistent with the embedding dimension of the hourly features. Continuous features are mapped to a 64-dimensional space through linear transformation. At the same time, Min-Max normalization is used to compress continuous feature values ​​to the [0,1] interval to avoid interference from numerical range differences in model training. In terms of feature dimension, the embedding vector of the classification feature is concatenated with the transformed continuous feature vector (64-dimensional) to obtain a comprehensive feature vector with a dimension of (64+32+16+32+64)=208. The input feature vector at each time step is processed into a 208-dimensional comprehensive feature vector through the embedding layer, so that the input sequence is transformed from (24,42) dimensions to (24,208) dimensions, providing a unified input format for subsequent feature extraction.

[0031] A bidirectional LSTM layer is used to capture the temporal dependency of passenger density changes in the input sequence: a three-layer bidirectional LSTM structure is set up, each layer containing two directions, forward LSTM and backward LSTM, with 256 hidden units in each direction; the forward LSTM processes the comprehensive feature vector sequentially from the first time step to the 24th time step of the input sequence and calculates the forward hidden state. ,in, Let be the comprehensive feature vector at time step t. The hidden state of the forward LSTM is the (t-1)th time step. The reverse LSTM processes the synthesized feature vector from the 24th time step to the 1st time step of the input sequence and calculates the reverse hidden state. ,in, To obtain the hidden state of the inverse LSTM at time step t-1, the forward and inverse hidden states of each layer are concatenated along the feature dimension to obtain the bidirectional LSTM output of that layer. The dimensions are (24, 512) (512 = 256 × 2). A Dropout layer is introduced between each bidirectional LSTM layer, with the Dropout coefficient set to 0.2, which randomly sets the output of 20% of the hidden units to 0 to avoid model overfitting. The three bidirectional LSTM layers are stacked in sequence, and the final output is a time-dependent feature tensor with dimensions (24, 512). This tensor contains historical information about each time step in the input sequence, both forward and backward, which provides support for capturing short-term and medium-term time dependencies.

[0032] The Transformer encoder layer is used to enhance the capture of long-term dependencies in the input sequence: a two-layer Transformer encoder is set, each layer containing a multi-head self-attention mechanism and a feedforward neural network; the multi-head self-attention mechanism calculates the association weights between features through eight parallel attention heads, and the attention calculation of each attention head follows the formula: , in, For query vector, For key vectors, For value vectors, To obtain a multi-head attention output with dimensions (24, 512), the outputs of the eight attention heads are concatenated along the feature dimension. The feedforward neural network contains two linear transformation layers and a ReLU activation function. The first linear transformation layer maps the 512-dimensional features to 2048-dimensional features. After ReLU activation, the second linear transformation layer maps the 2048-dimensional features back to 512-dimensional features, realizing non-linear transformation and dimension restoration of the features. Two Transformer encoder layers are stacked sequentially. The first encoder layer receives the (24, 512) tensor output from the bidirectional LSTM layer and outputs the processed (24, 512) tensor. The second encoder layer receives the output of the first layer and processes it further to finally obtain a (24, 512) feature tensor containing long-term dependency information.

[0033] The spatiotemporal attention mechanism layer is used to dynamically focus on key information in the temporal and spatial dimensions of the input sequence: Two multi-head attention layers with 8 attention heads are initialized, namely the temporal attention layer and the spatial attention layer, with each attention head having an input / output dimension of 512. First, the feature tensor output by the Transformer encoder layer, with dimensions (batch_size, 24, num_areas, 512), where batch_size is the batch size and num_areas is the number of terminal partitions, is directly input into the temporal attention layer. Using the features of 24 time steps within each partition as Q, K, and V, the attention weights of the temporal dimension are calculated, allowing the model to focus on historical time steps that significantly influence the prediction of current passenger density (such as the time step 1 hour before concentrated flight arrivals). The output dimension remains (batch_size, 24, num_areas, 512). A temporal attention feature tensor of (batch_size, num_areas, 24, 512) is generated. Then, the temporal attention feature tensor is transposed to adjust its dimensions to (batch_size, num_areas, 24, 512), making the spatial dimension (num_areas) the core dimension for attention calculation. This is input into the spatial attention layer, using the features of all partitions within each time step as Q, K, and V to calculate the attention weights of the spatial dimension. This allows the model to focus on partitions with passenger flow associations, outputting a spatial attention feature tensor of (batch_size, num_areas, 24, 512). Finally, the spatial attention feature tensor is transposed again to restore its dimensions to (batch_size, 24, num_areas, 512), resulting in a spatiotemporal fusion feature tensor that integrates key temporal and spatial information.

[0034] The fully connected output layer is used to convert spatiotemporal fusion features into passenger density predictions. A three-layer Multilayer Perceptron (MLP) structure is employed. The first layer has an input dimension of 512 and an output dimension of 256, using ReLU activation. The second layer has an input dimension of 256 and an output dimension of 128, also using ReLU activation. The third layer has an input dimension of 128 and an output dimension of 1, using Linear (no activation). The spatiotemporal fusion feature tensor (batch_size, 24, num_areas, 512) is expanded over time steps. The features from the last time step are used as the input to the MLP. Through three layers of linear transformation and nonlinear activation, the predicted passenger density for each sample corresponds to the next eight time steps (one time step every 15 minutes, for a total of two hours) and each partition. The final output tensor has dimensions of (batch_size, 8, num_areas), where 8 represents the number of future time steps and num_areas represents the number of terminal partitions, achieving accurate prediction of passenger density for each partition in the next 1-2 hours.

[0035] S3. Configure the model training for the completed passenger density prediction model, including training the model using a multi-task learning loss function, an adaptive optimizer, and a phased training strategy. To simultaneously optimize the model's point prediction accuracy, interval prediction reliability, and uncertainty estimation capability, a multi-task learning loss function is designed. This function integrates the point prediction loss (Huber loss), interval prediction loss (quantile loss), and uncertainty estimation loss through weighted coefficients. The specific configuration is as follows: First, the general expression of the multi-task learning loss function is determined as follows: , in, , and These represent the corresponding weight coefficients. For point prediction loss, For interval prediction loss, Estimate the loss for uncertainty, and satisfy the following conditions: .

[0036] Among them, point prediction loss Huber loss is employed, which manifests as mean squared error (MSE) when the prediction error is small and as mean absolute error (MAE) when the error is large. This effectively reduces the interference of outliers on model training. Its calculation formula is: when the actual passenger density y differs from the predicted passenger density... absolute value of error hour, ,when hour, ,in, The error distribution statistics are based on historical passenger density data from Terminal 3 of the airport, serving as the Huber loss threshold. Interval predicted loss. Quantile loss is used to ensure the coverage of the actual passenger density by the prediction interval. Three quantiles, 0.1, 0.5, and 0.9 (corresponding to the lower boundary, median, and upper boundary of the prediction interval, respectively), are selected. The calculation formula is as follows: , Where N is the number of training samples. Let i be the actual passenger density of the i-th sample. For the i-th sample at the quantile The predicted value, the uncertainty estimate of the loss KL divergence loss is used to quantify the difference between the model's predicted distribution and the actual distribution, thereby assessing the confidence level of the prediction results. First, it is assumed that the passenger density predicted by the model follows a normal distribution. ,in, To predict the mean, To predict variance, assuming the actual passenger density follows an empirical distribution P, the KL divergence loss is calculated using the following formula: , in, The loss function is the normal probability density function predicted by the model. By training the model with this loss function, the model can output a predicted distribution with a smaller deviation from the actual distribution, providing a confidence reference for subsequent air conditioning load calculation.

[0037] To balance model training speed and parameter optimization effectiveness, the AdamW optimizer with weight decay was selected to update model parameters, along with a periodic cosine annealing learning rate scheduler. The specific configuration is as follows: the initial learning rate of the AdamW optimizer was set to 1e-3, and the weight decay coefficient was set to 1e-4. This was achieved by adding [a specific parameter to the loss function]. The penalty items, is the weight decay coefficient, w is the model parameter, which suppresses overfitting caused by excessively large model parameters; the momentum coefficient is set to (0.9, 0.999), where 0.9 is the exponential decay rate of the first-order momentum (gradient mean) and 0.999 is the exponential decay rate of the second-order momentum (gradient variance), ensuring the stability of gradient estimation.

[0038] The periodic learning rate scheduler uses a cosine annealing strategy and sets a scheduling period. In each training epoch, the initial and minimum learning rates are adapted to the initial learning rate of the AdamW optimizer and the model convergence requirements, respectively. The initial learning rate is set to 1e-3 and the minimum to 1e-6. Its working mechanism is as follows: within each scheduling cycle, the learning rate decreases from the initial value according to a cosine function... Where t is the current training round, Let be the learning rate in round t. Regular decay, when the training rounds reach When the learning rate drops to its minimum value of 1e-6, the scheduling cycle is restarted to improve the final convergence performance.

[0039] S4. Calculate the air conditioning load based on the predicted value of passenger density, including combining outdoor meteorological forecast parameters and using the total load calculation formula to calculate the cooling load demand and heating load demand of air conditioning in each zone for each time period in the future. Two types of core basic data are needed to obtain air conditioning load calculation: The first type is passenger density prediction data, which is the passenger density prediction value of each zone for the next 1-2 hours output by the model trained in step S3. This prediction value is divided into 8 time steps (corresponding to 2 hours) in 15-minute time steps. Each time step contains the passenger density prediction value of all functional zones in Terminal 3 (such as check-in hall A, security check area B, boarding gates C1-C10, etc., a total of 32 zones). (Unit: person), along with a confidence score (0.00-1.00) for each predicted value; only predicted values ​​with a confidence score ≥0.85 are used in load calculation. The second category is outdoor meteorological forecast parameters, which obtain meteorological data for the next 1-2 hours that are perfectly aligned with the passenger density forecast time step by connecting to the API of the Central Meteorological Observatory and the real-time data interface of the airport's local meteorological station, including the predicted outdoor temperature. outdoor relative humidity Solar radiation intensity (Unit: W / ㎡, accuracy ±10W / ㎡), the time step of meteorological data is also set to 15 minutes to ensure that it corresponds one-to-one with the time step of passenger density prediction and avoid load calculation deviation due to time misalignment.

[0040] Based on the functional types of each zone in Terminal 3 (check-in hall, security checkpoint, boarding gates, commercial area, baggage claim area), the basic parameters for the building's load are preset. Simultaneously, the unit passenger load is dynamically adjusted based on outdoor weather forecast parameters, as follows: For the basic parameters for the building's load, the basic heat transfer coefficient is preset according to the characteristics of the building envelope for different zone types (e.g., multiple glass curtain walls in the check-in hall, multiple solid walls in the security checkpoint, and differences in boarding gate floor heights). (Unit: W / (㎡)) ℃), solar radiation absorption coefficient For example, in check-in hall area A, where glass curtain walls account for 60%, the pre-designed... Security checkpoint area B, where solid walls account for 80% of the area, is pre-designed... Simultaneously, a correction factor is introduced. (Value range 0.8-1.2), updated quarterly based on the actual maintenance status of each zone (such as the cleanliness of the glass curtain wall and the integrity of the wall insulation layer), ultimately determining the building's load. The parameters are calculated using basic parameters and correction factors to ensure compatibility with heat transfer characteristics in different regions.

[0041] For unit passenger load Unit: W / person, including sensible and latent heat, dynamically adjusted based on outdoor weather forecast parameters: when the outdoor predicted temperature... And relative humidity In high-temperature and high-humidity environments, the human body primarily dissipates heat through latent heat. (Of which sensible heat is 40W / person and latent heat is 80W / person); when and Under comfortable weather conditions, the ratio of sensible heat to latent heat is balanced, and the setting is appropriate. (Sensible heat 50W / person, latent heat 50W / person); when In a low-temperature, dry environment, the human body primarily dissipates heat through sensible heat. (Sensible heat 60W / person, latent heat 30W / person), dynamic adjustment ensures that the load per passenger unit matches the heat dissipation characteristics of the human body in the actual environment.

[0042] Passenger load is the sum of the cooling / heating loads generated by all passengers within a zone, expressed as passenger density forecasts. With dynamically adjusted unit passenger load The core calculation parameter is calculated using the following formula: , in, For the total workload of personnel in the zone, This is the dynamically adjusted unit passenger load. For the predicted passenger density, the building load is the sum of heat loss from the zoned envelope (walls, roof, windows) and heat gain from solar radiation, calculated using the following formula: ， in, For correction factor, The total area of ​​the zoned enclosure structure was obtained from the airport architectural drawings. The area of ​​the windows / glass curtain wall in the zone.

[0043] Fresh air load is the cooling / heating load required to treat outdoor fresh air to the target indoor temperature and humidity, using the formula: ,in, air density, For the fresh air volume of each zone, The specific heat capacity of air at constant pressure. Forecast outdoor temperature, The target indoor temperature.

[0044] The above-mentioned personnel load, building load, and fresh air load are calculated using the total load formula: , The total air conditioning load for each zone at each future time step is obtained.

[0045] S5. Optimize the control strategy according to the air conditioning load, including using optimization algorithms to solve for the optimal setpoint of each air conditioning subsystem with the goal of minimizing the total energy consumption of the system, generating control commands and sending them to the air conditioning actuators. With the minimum total system energy consumption as the core objective, and combining indoor comfort constraints and equipment operation constraints, a Model Predictive Control (MPC) optimization model is constructed: the core objective is quantified by an objective function, the formula of which is: , in, To optimize the overall objective value, To predict the number of time steps, Let t be the total operating power of the air conditioning system in step t. Let t be the predicted value of the actual indoor temperature at step t. The target indoor temperature For temperature tolerance, , These are power weighting coefficients and temperature weighting coefficients, respectively. The constraints are divided into two categories: the first category is comfort constraints, which explicitly require the indoor temperature to be maintained within a preset comfort range, i.e. (The lower limit ensures winter heating needs, while the upper limit restricts summer cooling energy consumption); the second category is equipment operation constraints, including air volume constraints and supply air temperature change rate constraints. The air volume constraint requires that the supply air volume of each air conditioning subsystem not be less than 30% of the rated air volume (to avoid excessively low air volume leading to deterioration of indoor air quality). ,in, Let t be the air volume supplied in step t. The supply air temperature change rate is constrained by the formula for the rated air volume of the equipment: , in, The setpoint for the supply air temperature at step t. The supply air temperature setpoint for step t-1. The interval between adjacent time steps, here Therefore, the constraint is simplified to the supply air temperature change not exceeding 2℃ between adjacent time steps, to avoid the temperature change being too rapid and affecting the passenger comfort experience, while protecting the air conditioning equipment from damage caused by frequent and large adjustments.

[0046] The load demand matrix generated in step S4 is used as input to the optimization model to clarify the cooling / heating load demand of each zone at each future time step. Based on the performance curves of each subsystem of the air conditioning system (such as the COP curve of the chiller unit, the flow-power curve of the pump, and the air volume-energy consumption curve of the fresh air unit, obtained by fitting the equipment's factory parameters with actual operating data), a mapping relationship between load demand and equipment operating parameters is established. That is, by adjusting parameters such as the chiller unit's supply water temperature, the pump speed, the fresh air unit's supply air temperature, and the opening degree of the VAV terminal damper, the system's cooling / heating capacity can be changed accordingly to meet the load demand. The gradient descent method is used to solve the optimization model, guided by minimizing the overall optimization objective value J, and the operating parameters of each air conditioning subsystem are iteratively adjusted: the initial parameters are set to 70% of the equipment's rated operating parameters, and then the corresponding parameters are calculated in each iteration. and Substituting the values ​​into the objective function, we obtain the J value. By calculating the gradient of J with respect to each parameter, we adjust the parameter values ​​(increase the parameter when the gradient is negative and decrease the parameter when the gradient is positive). We iterate until the J value converges (the difference between the J values ​​of two adjacent iterations is less than 1%) or reaches the maximum number of iterations. Finally, we obtain the optimal operating parameters of the air conditioning subsystem corresponding to each time step and each zone, including the chiller supply water temperature, water pump speed, fresh air unit supply air temperature, and VAV terminal air valve opening, forming a zone-by-zone and time-by-time control strategy matrix.

[0047] Based on the optimal operating parameters obtained from the solution, corresponding control instructions are generated according to the equipment type: for equipment driven by frequency converters, frequency control instructions are generated; for electric valve equipment, opening control instructions are generated; for chiller units, water supply temperature setting instructions are generated. All control instructions are encapsulated according to industrial bus protocols (BACnet or Modbus protocol) to ensure communication compatibility with field programmable logic controllers (PLCs) or direct digital controllers (DDCs).

[0048] Establish a command issuance verification mechanism to verify the safety and rationality of the generated control commands. The verification includes whether the command exceeds the safe operating range of the equipment. If the command exceeds the range, it is automatically corrected to the equipment safety threshold. At the same time, the logical consistency of the command is verified. If there is a logical conflict, it is backtracked to the parameter solving stage for iterative optimization.

[0049] After successful verification, control commands are sent to the corresponding PLC or DDC controller via industrial bus (BACnet or Modbus). Upon receiving the commands, the controller drives the actuators (frequency converters, electric valves, etc.) to adjust the equipment operating status (such as adjusting the water pump speed according to the frequency command and adjusting the air valve opening according to the opening command), so that the air conditioning system can run in advance according to the optimal operating parameters to prepare for future load changes. For example, for the case of a cooling load demand of 464.928kW in area A of the check-in hall at a certain time step, the control commands sent are: chiller water supply temperature 7℃, water pump frequency 45Hz, fresh air unit supply air temperature 18℃, and VAV terminal air valve opening 85%, to ensure that the system meets the cooling load demand of the area with the lowest energy consumption.

[0050] S6. Real-time collection of environmental data for each zone through end sensors, comparison with target parameters and fine-tuning of control commands.

[0051] High-precision temperature and humidity sensors have been deployed in key locations and high-traffic areas of all functional zones in Terminal 3 (32 zones including check-in hall, security checkpoint, and boarding gates) to collect real-time indoor temperatures in each zone. Actual humidity The collected environmental data is transmitted to the edge gateway via the LoRa wireless communication module. The edge gateway performs real-time caching and format conversion (converting sensor protocol data to JSON format) on the data before transmitting it to the central controller via industrial Ethernet, ensuring that the data transmission delay is controlled within 5 seconds and guaranteeing timely feedback.

[0052] After receiving real-time environmental data from each zone, the central controller compares it with the indoor target parameters and target humidity determined in step S5, calculating the environmental parameter deviation values, namely temperature deviation and humidity deviation. When the deviation is within the set error range, the central controller maintains the control command issued in step S5 unchanged. When the deviation is greater than the set error, the control command fine-tuning mechanism is triggered. During command fine-tuning, the fine-tuning amplitude is determined according to the direction and magnitude of the deviation: when the actual temperature is too high, if the current control command is in cooling mode, the chiller unit supply water temperature is fine-tuned (reduced by 0.2℃-0.5℃, not lower than the 7℃ safety threshold) or the VAV terminal air valve opening is increased (increased by 5%-10%, not exceeding the 100% upper limit); when the actual temperature is too low, if it is in heating mode, the hot water valve opening is fine-tuned (increased by 5%-10%, not exceeding the 100% upper limit) or the fresh air ratio of the fresh air handling unit is reduced (reduced by 5%-8%, not lower than 30%). (Lower limit); when the actual humidity is too high, increase the operating power of the dehumidification module; when the actual humidity is too low, start the humidification module (adjusting gradually in 5% humidity increments). The fine-tuned control commands also need to pass the command verification mechanism in step S5. After verification, the central controller reissues the commands to the corresponding PLC / DDC controllers of the partitions to drive the actuators to adjust their operating status. At the same time, the central controller synchronously stores the real-time environmental data of each partition, the control commands before and after fine-tuning, and the equipment operating parameters (such as the adjusted air valve opening and water supply temperature) to the time-series database. On the one hand, this is used to trace the cause of environmental deviations (such as deviations caused by a sudden increase in passenger flow or equipment failure). On the other hand, this historical feedback data is used as incremental samples for the online learning of the passenger density prediction model in step S3. It is used periodically (such as every morning) to update the model parameters, so that the prediction model can better adapt to the actual environmental change patterns and further improve the prediction accuracy.

[0053] Example 2 The difference between this embodiment and Embodiment 1 is that this embodiment provides an energy-saving terminal air conditioning system based on a passenger density prediction model, including: The data acquisition module is configured to: acquire the basic dataset for air conditioning energy-saving adjustment, and preprocess the acquired basic dataset, including acquiring data from the airport flight information management system, outdoor meteorological data, and air conditioning system operation data; The model module is configured to build a passenger density prediction model using the preprocessed base dataset as input, including model building using a bidirectional LSTM-Transformer hybrid model with an attention mechanism as the framework. The training module is configured to: perform model training on the completed passenger density prediction model, including training the model using a multi-task learning loss function, an adaptive optimizer, and a phased training strategy; The load module is configured to calculate the air conditioning load based on the predicted value of passenger density, including combining outdoor meteorological forecast parameters and using the total load calculation formula to calculate the cooling load demand and heating load demand of air conditioning in each zone for each time period in the future. The optimization module is configured to optimize the control strategy based on the air conditioning load, including using optimization algorithms to solve for the optimal setpoints of each air conditioning subsystem with the goal of minimizing the total energy consumption of the system, generating control commands and sending them to the air conditioning actuators. The output module is configured to: collect environmental data of each zone in real time through the end sensor, compare it with the target parameters, and fine-tune the control commands.

[0054] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A terminal air conditioning energy-saving method based on a passenger density prediction model, characterized in that, include: S1. Obtain the basic dataset for air conditioning energy-saving regulation and preprocess the obtained basic dataset, including obtaining data from the airport flight information management system, outdoor meteorological data and air conditioning system operation data; S2. Construct a passenger density prediction model using the preprocessed basic dataset as input, including model construction using a bidirectional LSTM-Transformer hybrid model with attention mechanism as the framework; S3. Configure the model training for the completed passenger density prediction model, including training the model using a multi-task learning loss function and an adaptive optimizer. S4. Calculate the air conditioning load based on the predicted value of passenger density, including combining outdoor meteorological forecast parameters and using the total load calculation formula to calculate the cooling load demand and heating load demand of air conditioning in each zone for each time period in the future. S5. Optimize the control strategy according to the air conditioning load, including using optimization algorithms to solve for the optimal setpoint of each air conditioning subsystem with the goal of minimizing the total energy consumption of the system, generating control commands and sending them to the air conditioning actuators. S6. Real-time collection of environmental data for each zone through end sensors, comparison with target parameters and fine-tuning of control commands.

2. The terminal air conditioning energy-saving method based on a passenger density prediction model according to claim 1, characterized in that, The preprocessing of the acquired basic dataset specifically includes: S11. Linear interpolation or forward padding is used to maintain the continuity of the time series for the sensor data stream in the basic dataset. For missing fields in flight data, they are filled by backtracking the historical records of the airport flight information management system. For missing values ​​in outdoor meteorological data, weighted replacement is used with data from nearby meteorological stations. S12. Using a sliding time window, the average, maximum and minimum values ​​of the second-level high-frequency sensor data are aggregated within the window, and the hour-level low-frequency external data are forward-filled. An airport geographic grid coordinate system is established, and meteorological data is associated with the corresponding area according to the grid ID. Flight dynamic data is matched with the GPS coordinates of the parking position. S13. Using flight number as the primary key, associate the airport flight information management system's planned data with radar real-time data, establish a topological relationship based on meteorological station ID and sensor deployment location, create a multidimensional index through spatiotemporal hashing of longitude, latitude, and time window, and resolve data conflicts according to preset priorities. S14. Finally, standardize all data in the basic dataset and verify the validity of the preprocessed data using a data quality assessment formula, which is: ， in, This is a data quality assessment value. This represents the total number of missing values ​​in the preprocessed dataset. The total amount of data in the dataset. This represents the total number of outliers in the preprocessed dataset. To standardize the data consistency coefficient, , and These are different weighting coefficients.

3. The terminal air conditioning energy-saving method based on a passenger density prediction model according to claim 1, characterized in that, The construction of the passenger density prediction model includes first extracting time features, spatial features, event features, and environmental features from the basic dataset into a multi-dimensional feature vector and dividing the input sequence. Then, through the embedding layer, the model outputs comprehensive features in a unified feature space by embedding categorical features, transforming and concatenating continuous features by linear projection. The comprehensive features are then input into a bidirectional LSTM layer, and after forward and backward calculations and concatenation, they are fed into a two-layer Transformer encoder layer. Long-term dependencies are captured through an attention formula. The model is then input into a spatiotemporal attention mechanism layer, where it undergoes multi-head attention calculations in the time and spatial dimensions, as well as tensor transpose processing, to obtain spatiotemporal fusion features. Finally, the fully connected output layer processes the data through a multilayer perceptron to output the predicted passenger density values ​​for each zone in the future.

4. The terminal air conditioning energy-saving method based on a passenger density prediction model according to claim 3, characterized in that, The construction of the passenger density prediction model also includes a model input feature selection mechanism. Based on the preprocessed basic dataset, the correlation between each feature and passenger density is calculated using mutual information values. A correlation threshold is set for feature selection to obtain the input feature set. Simultaneously, L1 regularization is used to constrain the feature weights of the embedding layer and the linear projection layer. The formula for calculating the mutual information value is: ， in, For a single input feature, For passenger density labels, Let X be the joint probability distribution of Y. and Let X and Y be the marginal probability distributions, respectively.

5. The terminal air conditioning energy-saving method based on a passenger density prediction model according to claim 1, characterized in that, The process of configuring the constructed passenger density prediction model for training includes configuring a multi-task learning loss function. This multi-task learning loss function integrates point prediction loss, interval prediction loss, and uncertainty estimation loss through weight coefficients. Specifically, the point prediction loss uses Huber loss, the interval prediction loss uses quantile loss, and the uncertainty estimation loss is used to quantify the confidence level of the prediction results. An adaptive optimizer and learning rate scheduling strategy are also configured. An adaptive optimizer with weight decay is selected to update the model parameters, while a periodic learning rate scheduler is used to decay the learning rate from its initial value to its minimum value according to a cosine curve at a preset period before restarting.

6. The terminal air conditioning energy-saving method based on a passenger density prediction model according to claim 1, characterized in that, The air conditioning load calculation based on passenger density predictions includes obtaining predicted passenger density values ​​for each zone in future time periods and corresponding outdoor weather forecast parameters. It then calls a load calculation engine to calculate the air conditioning cooling and heating loads by zone and time period. Using the predicted passenger density as the core input, and combining it with building heat transfer losses, equipment heat dissipation, and fresh air heat exchange requirements, the total load calculation formula integrates passenger load, building load, and fresh air load to obtain the cooling and heating load values ​​for each zone at each future time step. The total load calculation formula is as follows: ， in, For the cooling and heating load generated per passenger, This is a predicted value for passenger density in each zone. To accommodate the building's load in different zones, The fresh air load for each zone.

7. The terminal air conditioning energy-saving method based on a passenger density prediction model according to claim 6, characterized in that, The air conditioning load calculation based on passenger density predictions also includes obtaining outdoor meteorological forecast parameters in real time through a meteorological data interface and aligning them with the passenger density prediction time step. The unit passenger load is dynamically adjusted according to the outdoor meteorological forecast parameters. The fresh air load is calculated by combining outdoor temperature and humidity with indoor target temperature and humidity. Simultaneously, basic building load parameters are preset according to zoning type, and correction coefficients are used to adapt to the heat transfer characteristics of the building envelope in different areas, ultimately forming a load demand matrix. The fresh air load calculation formula is as follows: ， in, air density, For the fresh air volume of each zone, The specific heat capacity of air at constant pressure. Forecast outdoor temperature, The target indoor temperature.

8. The terminal air conditioning energy-saving method based on a passenger density prediction model according to claim 1, characterized in that, The control strategy optimization based on air conditioning load includes minimizing total system energy consumption as the core objective, while simultaneously maintaining indoor temperature within a preset comfort range. An optimization model is constructed using a model predictive control algorithm. The future cooling and heating load demands for each zone are used as inputs to the optimization model. The optimal operating parameters for air conditioning subsystems such as chillers, water pumps, fresh air units, and VAV terminals are calculated to form zone-specific time-based control strategies. These optimal operating parameters are then converted into control commands conforming to the industrial bus protocol and sent to the corresponding programmable logic controllers. The objective function of the optimization model is: ， in, To optimize the overall objective value, To predict the number of time steps, Let t be the total operating power of the air conditioning system in step t. Let t be the predicted value of the actual indoor temperature at step t. The target indoor temperature For temperature tolerance, , These are the power weighting coefficient and the temperature weighting coefficient, respectively.

9. The terminal air conditioning energy-saving method based on a passenger density prediction model according to claim 1, characterized in that, The control strategy optimization based on air conditioning load also includes setting equipment operation constraints and comfort constraints during the optimization algorithm solution process. The equipment operation constraints include airflow constraints and supply air temperature change rate constraints. After obtaining the optimal setpoints, corresponding control commands are generated according to the equipment type, and a command issuance verification mechanism is established. The formula for the supply air temperature change rate constraint is: ， in, The setpoint for the supply air temperature at step t. The supply air temperature setpoint for step t-1. The interval between adjacent time steps.

10. An energy-saving air conditioning system for airport terminals based on a passenger density prediction model, comprising the method described in claim 1, characterized in that, include: The data acquisition module is configured to: acquire the basic dataset for air conditioning energy-saving adjustment, and preprocess the acquired basic dataset, including acquiring data from the airport flight information management system, outdoor meteorological data, and air conditioning system operation data; The model module is configured to build a passenger density prediction model using the preprocessed base dataset as input, including model building using a bidirectional LSTM-Transformer hybrid model with an attention mechanism as the framework. The training module is configured to: perform model training on the completed passenger density prediction model, including training the model using a multi-task learning loss function, an adaptive optimizer, and a phased training strategy; The load module is configured to calculate the air conditioning load based on the predicted value of passenger density, including combining outdoor meteorological forecast parameters and using the total load calculation formula to calculate the cooling load demand and heating load demand of air conditioning in each zone for each time period in the future. The optimization module is configured to optimize the control strategy based on the air conditioning load, including using optimization algorithms to solve for the optimal setpoints of each air conditioning subsystem with the goal of minimizing the total energy consumption of the system, generating control commands and sending them to the air conditioning actuators. The output module is configured to: collect environmental data of each zone in real time through the end sensor, compare it with the target parameters, and fine-tune the control commands.

Images