A method and device for predicting and controlling air conditioning load in time and in zones based on multi-modal perception of metro passenger flow characteristics

Abstract

This invention provides a multimodal sensing air conditioning load prediction and time-sharing / zonal energy-saving control method and device adapted to the tidal characteristics of subway passenger flow, applicable to the field of rail transit energy data management. For subway air conditioning tidal energy-saving control, it collects six types of multimodal data: passenger flow, temperature and humidity, equipment, meteorology, location, and operation. After spatiotemporal regularization to unify the spatiotemporal benchmark, a standardized sensing dataset is constructed. A six-channel feature tensor is built through interpolation, normalization, and attribute mapping, dividing the time-series samples into peak, off-peak, and low-peak periods. A dual-branch spatiotemporal prediction network is constructed to extract passenger flow-load coupling features, and adversarial training under operating conditions is used to solidify the tidal load benchmark model. A two-layer precision verification quantifies the energy-saving potential, outputting time-sharing / zonal load curves. Gradient attribution and clustering are used to mine passenger flow-load coupling patterns, precipitating typical operating condition control paradigms, and finally outputting a load prediction model and air conditioning energy-saving strategy library adapted to the tidal characteristics of subways.

Description

Technical Field

[0001] This invention relates to the field of energy data management and control in rail transit, and in particular to a multimodal sensing air conditioning load prediction and time-sharing and zone-based energy-saving control method and device adapted to the tidal characteristics of subway passenger flow. Background Technology

[0002] Existing conventional methods for calculating subway air conditioning load fall into two categories. The first category is static design load calculation, which calculates the rated cooling load based solely on the station's zoned building area and fixed design passenger flow density, and cannot adapt to the dynamic changes in daily real-time tidal passenger flow. The second category is a simple time-series statistical forecasting method, which uses only passenger flow as a single variable to fit the load curve, without integrating multi-dimensional data such as temperature and humidity, equipment operating conditions, outdoor weather, and spatial location. This results in weak multi-source feature coupling and mining capabilities, large load forecasting deviations, and an inability to achieve high-precision load output by zone and minute.

[0003] Existing load forecasting models simply splice together basic data such as passenger flow and temperature, without distinguishing between dynamic temporal features and static spatial attributes. They lack standardized temporal regularization and unified processing procedures for partitioned spaces. They have not built a multi-channel feature tensor structure, making it impossible to perform differentiated normalization and temporal broadcast adaptation for six types of heterogeneous data: passenger flow, environment, equipment, meteorology, space, and operations. There is serious interference between features, making it difficult to accurately extract the passenger flow-load coupling evolution law. Furthermore, they do not perform sample windowing isolation for peak, off-peak, and low-peak tidal conditions, and the model training is affected by the imbalance of sample distribution, resulting in poor adaptability to tidal scenarios.

[0004] Existing single-branch time-series regression networks only output load prediction values ​​and do not have independent discriminant verification branches, making it impossible to quantify the degree of fit between the prediction results and the actual operating conditions in real time. Without the introduction of adversarial iteration and knowledge distillation mechanisms, the scarce peak and off-peak operating condition samples cannot share the load mapping patterns of the sufficient off-peak samples, resulting in insufficient model generalization ability and a significant increase in prediction error during operating condition switching intervals, making it difficult to form a standardized load benchmark model applicable to the entire tidal cycle.

[0005] Existing technologies only output uniform load values ​​for the entire station, without establishing a two-layer quantization mechanism for zonal time-series deviation and zonal spatial weighting; they do not utilize the intermediate feature space of the model to reverse solve the weights of load influencing factors in each dimension, lack load thermal visualization mapping methods, and cannot output refined load curves for different zones and times. Energy-saving potential is only roughly estimated based on fixed differences, without combining prediction accuracy for error calibration. The reliability of energy-saving space calculation results is low, making it difficult to support differentiated energy-saving regulation in different zones.

[0006] Conventional prediction models are black-box models, lacking the ability to analyze gradient attribution features and quantify the contribution of single-zone or single-class features to air conditioning load. They can only output load data and cannot use clustering algorithms to uncover the inherent coupling patterns between passenger flow tides and load fluctuations. They cannot distinguish between typical operating conditions such as passenger flow-driven, weather-driven, and equipment-driven conditions, making it difficult to extract time-sharing and zone-based control paradigms adapted to tidal characteristics. They lack a systematic energy-saving control strategy library, and the model output cannot be directly converted into executable control commands for air conditioning equipment.

[0007] Most existing subway air conditioning energy-saving controls adopt a unified control logic for the entire station, without establishing independent control units for station halls, platforms, and equipment areas. They also fail to achieve a closed-loop process of prediction, quantification, pattern mining, and strategy output. The load prediction results are disconnected from the control parameters of on-site air conditioning fans, valves, and surface cooling equipment. There is a lack of a complete technical link from multi-source sensing data to adaptive zonal control. Dynamic energy-saving control under tidal passenger flow changes is difficult to implement, and there are significant gaps in the refined and intelligent energy-saving technologies for rail transit station air conditioning.

[0008] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of this disclosure and does not constitute information on prior art known to those skilled in the art. Summary of the Invention

[0009] According to one aspect of this application, a multimodal sensing air conditioning load prediction and time-sharing and zonal energy-saving control method adapted to the tidal characteristics of subway passenger flow is provided, including: collecting six categories of multimodal sensing data, namely passenger flow tidal distribution, station temperature and humidity, air conditioning operating parameters, outdoor weather, spatial location, and operating time period; performing a spatiotemporal regularization process to unify the minute-level time series benchmark and the station hall zonal spatial benchmark, generating a unified spatiotemporal sensing dataset for the subway air conditioning system; constructing six types of feature channels based on time series interpolation completion, data normalization alignment, and static spatial attribute mapping; calculating various sensing indicators with a single zone and single time period as the smallest statistical unit; constructing feature tensors for six channels: passenger flow, environment, equipment, weather, space, and operation; and simultaneously completing the time series windowing isolation of peak, off-peak, and low-peak tidal operating condition samples; and using a multimodal spatiotemporal prediction dual-branch network to extract passenger flow-load coupling evolution characteristics. The system generates a multi-step air conditioning load prediction sequence through temporal convolution and verifies the prediction fitting accuracy through a spatiotemporal discriminative branch. Based on the adversarial iterative convergence of tidal operating condition samples after windowing, a benchmark model for tidal load prediction is solidified. Based on the output of the benchmark model, a mechanism for configuring zonal temporal deviation, station-wide spatial weighted double-layer prediction accuracy, and energy-saving potential quantification is configured. By modeling the load influence factor weights through the intermediate feature space of the model, combined with spatial weight allocation and load thermal mapping, the system outputs the time-sharing and zonal load prediction curves and energy-saving spatial temporal results for the entire station. Relying on the prediction curves and model feature weights, the system extracts the load contribution of a single zone and single channel through feature gradient attribution. A clustering algorithm is used to mine the coupling law between passenger flow tides and load fluctuations, and typical operating condition control paradigms are precipitated to form a time-sharing and zonal load prediction model and energy-saving control strategy library adapted to the tidal characteristics of the subway.

[0010] Another aspect of this application is a multimodal sensing air conditioning load prediction and time-sharing energy-saving control device adapted to the tidal characteristics of subway passenger flow, used to execute executable instructions to perform the above-mentioned multimodal sensing air conditioning load prediction and time-sharing energy-saving control method adapted to the tidal characteristics of subway passenger flow.

[0011] According to another aspect of this application, an electronic device includes: a first processor; and a memory for storing executable instructions of the first processor; wherein the first processor is configured to execute the above-described multimodal sensing air conditioning load prediction and time-sharing zone energy-saving control method adapted to the tidal characteristics of subway passenger flow by executing the executable instructions.

[0012] According to another aspect of this application, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a second processor, implements the above-described method for multimodal sensing air conditioning load prediction and time-sharing and zone-based energy-saving control adapted to the tidal characteristics of subway passenger flow.

[0013] The beneficial technical effects of this application are as follows: Addressing industry pain points such as tidal fluctuations in subway passenger flow, crude control of traditional air conditioning, and low accuracy of load forecasting, a multimodal sensing, time-sharing, and zone-based air conditioning energy-saving full-process technical solution is constructed. The system collects six types of sensing data: passenger flow, temperature and humidity, air conditioning equipment, outdoor weather, spatial location, and operation. Through temporal interpolation, normalization, and spatial mapping, spatiotemporal regularization is achieved, constructing a standardized six-channel feature tensor, and windowing the operating condition samples according to peak, off-peak, and valley periods. A cross-modal decoupled generation-discrimination dual-branch network is built, integrating adversarial iteration and operating condition knowledge distillation to train the tidal load benchmark model; based on the intermediate features of the model, a temporal and spatial dual-layer quantization system is established, outputting refined load and energy-saving thermal time-series results; furthermore, load influencing factors are analyzed through gradient attribution and clustering algorithms, precipitating standardized zone control paradigms, and encapsulating them into a prediction model and energy-saving strategy library. The entire technology achieves a complete closed loop of standardized processing of multi-source heterogeneous data, high-precision load prediction throughout the tidal cycle, quantifiable energy-saving potential of different zones, and reusable operating condition control, making it suitable for subway operation scenarios with large fluctuations in passenger flow during morning and evening.

[0014] This application aims to construct a six-channel tensor through unified regularization of multimodal data, differentiate between dynamic and static features, eliminate interference from sampling frequency, dimensions, and spatial granularity, and significantly improve the accuracy of feature coupling mining; a dual-branch adversarial network combined with cross-condition knowledge distillation solves the problem of scarce samples during peak and trough periods, and controls the full-condition load prediction error to within 5%.

[0015] By employing dual-layer spatiotemporal quantification and thermal visualization, the energy-saving space of each zone is accurately calculated, providing a quantitative basis for differentiated air conditioning control. By adopting gradient attribution and cluster analysis to resolve load-driven patterns, a control paradigm adapted to tidal conditions is automatically derived, eliminating the need for repeated manual tuning of control parameters. This forms a closed loop of prediction-quantification-strategy output, enabling time-based independent cooling for station halls, platforms, and equipment areas, significantly reducing the overall energy consumption of subway air conditioning.

[0016] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0017] Figure 1 This document shows a flowchart illustrating a multimodal sensing air conditioning load prediction and time-sharing and zone-based energy-saving control method adapted to the tidal characteristics of subway passenger flow, provided in an embodiment of this application. Figure 2 The diagram shows a schematic of the structure of a multimodal sensing air conditioning load prediction and time-sharing energy-saving control device adapted to the tidal characteristics of subway passenger flow, provided in an embodiment of this application. Detailed Implementation

[0018] 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.

[0019] In one implementation, Figure 1 The diagram illustrates a flowchart of a multimodal sensing air conditioning load prediction and time-sharing and zone-based energy-saving control method adapted to the tidal characteristics of subway passenger flow, according to an embodiment of this application.

[0020] S101 collects six categories of multimodal sensing data, including passenger flow tidal distribution, station temperature and humidity, air conditioning operating parameters, outdoor weather, spatial location, and operating hours. It then performs a spatiotemporal regularization process to unify minute-level time series benchmarks and station hall zoning spatial benchmarks, generating a unified spatiotemporal sensing dataset for the subway air conditioning system.

[0021] In one implementation, six categories of multimodal sensing data are first collected. Passenger flow tidal distribution data is acquired through the gate entry / exit counting system and video passenger flow statistics terminals in each zone of the station. The collected data includes real-time personnel density, entry / exit passenger flow rate, and cumulative passenger flow in each zone. The sampling frequency is set to once per minute to accurately capture the rapid fluctuations in passenger flow during morning and evening peak hours. Temperature and humidity data within the station are collected through temperature and humidity sensors deployed at various locations in the station hall, platform, and equipment area. At least three collection points are arranged in each air conditioning control zone. The collected data includes dry-bulb temperature and relative humidity, with a sampling interval of one minute, used to characterize the thermal environment status within the station. Air conditioning operating parameter data is obtained from the station building automation system. The collected data includes chiller supply and return water temperatures, air conditioning fan operating frequency, damper opening, real-time equipment operating power, and cumulative power consumption. The data update frequency is aligned with the sensor collection frequency. Outdoor meteorological data is collected via the local meteorological department's public data interface or the station's outdoor meteorological data acquisition terminal. The collected data includes outdoor dry-bulb temperature, relative humidity, atmospheric pressure, wind speed, and wind direction, with an initial sampling interval of 5 minutes. This data is used to characterize the impact of the external environment on the station's load. Spatial location data is static attribute data, including the station's line affiliation, number of underground floors, building area of ​​each air-conditioning control zone, orientation, heat transfer coefficient of the building envelope, and design cooling load indicators. This data is used to support spatial load difference analysis. Operating time period data is extracted from the official subway timetable. The collected data includes the daily start and end times of operation, and the time boundaries for morning peak hours, evening peak hours, off-peak hours, and low-peak hours. This data is used for subsequent tidal condition classification.

[0022] Next, the spatiotemporal normalization process is executed, firstly to unify the minute-level time series benchmark. A linear interpolation time series alignment algorithm is used, with the natural minute mark of the standard time server as the globally unified time node, aligning various data sources with different sampling frequencies and time offsets to a 1-minute granularity. For example, the original sampling interval for outdoor meteorological data is 5 minutes. Piecewise linear interpolation is performed on the four minute nodes between every two adjacent sampling points. The corresponding value of each minute node is obtained by proportionally weighting the values ​​of the preceding and following sampling points with the time distance, ensuring that all data sources have a consistent time scale. For data collected by devices with clock deviations, timestamp offset calibration is first performed. The clock offset of each device is calculated based on the standard time, and the timestamps of all data are uniformly corrected to the standard time axis, eliminating the clock asynchrony problem between devices. For a small number of minute nodes with packet loss, the moving average of adjacent 3-minute data in the same partition is used for temporary completion. The completed data is marked as pending correction, and then further verified using historical benchmark data from the same time period.

[0023] Subsequently, a unified spatial benchmark for the station concourse was established. Based on the physical coverage of the air conditioning terminal control units, the station space was divided into three main areas: the concourse public area, the platform public area, and the equipment management area. Each area was further subdivided into 8 to 12 independent sub-zones according to physical partitions and airflow organization boundaries. Each sub-zone corresponds to an independent set of air conditioning terminal control equipment. A spatial mapping rule was established, assigning all sensors and data acquisition terminals to their corresponding sub-zones based on their physical installation location. For example, the three temperature and humidity sensors installed on the north side of the concourse were uniformly assigned to the North Concourse Zone, and the video passenger flow terminal covering the south side of the concourse was assigned to the South Concourse Zone, ensuring that all types of data within each zone have spatial correspondences. Static spatial location data were assigned values ​​one by one according to the zone dimension. Each sub-zone corresponds to an independent building area, maximum personnel density, design cooling load, and building envelope parameters, forming a unique spatial index for each zone and eliminating the problem of inconsistent spatial granularity between different data sources.

[0024] The final result is a unified spatiotemporal perception dataset for the subway air conditioning system. The dataset uses a single zone and a single minute as the smallest statistical unit. Each unit fully contains corresponding fields for six major data categories: passenger flow, temperature and humidity, air conditioning equipment, outdoor weather, spatial attributes, and operating hours. For example, a data record for the south zone of the station hall at a given moment includes the zone's average passenger flow density, average temperature and relative humidity, fan frequency and real-time power of the corresponding air conditioning terminal, current outdoor dry-bulb temperature and humidity, zone building area and design cooling load, and the label of whether it belongs to peak or off-peak operating hours. The dataset is arranged continuously in chronological order, while retaining the index identifiers for the zone dimensions. All data dimensions and numerical ranges have undergone preliminary unified verification, providing a standardized input data source for subsequent feature channel construction and load prediction model training.

[0025] S102 constructs six feature channels based on temporal interpolation completion, data normalization alignment, and static spatial attribute mapping. It calculates various perception indicators with a single partition and a single time period as the smallest statistical unit, and constructs feature tensors for six channels: passenger flow, environment, equipment, meteorology, space, and operation. It also completes temporal window isolation of peak, off-peak, and low-peak tidal conditions samples simultaneously.

[0026] In one implementation, combining the requirements for constructing multimodal features of subway air conditioning load and dividing tidal condition samples, three types of feature processing methods are introduced: temporal interpolation completion, data normalization alignment, and static spatial attribute mapping, to build a six-channel standardized feature tensor construction system. To adapt to the input format requirements of subsequent multimodal spatiotemporal prediction networks and solve the problems of inconsistent sampling frequencies, inconsistent dimensions, and misaligned spatiotemporal attributes of multi-source data, three types of feature processing methods are introduced: temporal interpolation completion, data normalization alignment, and static spatial attribute mapping, to build a six-channel standardized feature tensor construction system.

[0027] The temporal interpolation completion algorithm comprises two execution stages: multi-sampling rate temporal alignment interpolation and outlier removal and missing value completion. Multi-sampling rate temporal alignment interpolation uses a global standard time axis of 1 minute increments as a reference, performing piecewise linear interpolation on data sources with sampling intervals greater than 1 minute. The calculation logic involves taking the values ​​of two adjacent valid sampling moments and linearly distributing the values ​​of intermediate moments according to their time distance proportions. The value of any intermediate moment is equal to the preceding sample value plus the product of the difference between the preceding and following sample values ​​and the time proportion. For example, if the original sampling interval for outdoor meteorological data is 5 minutes, and the outdoor dry-bulb temperature measured at 7:00 is 28 degrees Celsius, and the measured value at 7:05 is 28.5 degrees Celsius, then the temperature value at 7:01 is equal to 28 plus 0.5 multiplied by one-fifth, resulting in a final calculated value of 28.1 degrees Celsius. This process is repeated to complete the value completion for all intermediate minute nodes, unifying the time granularity of all data sources to 1 minute. Outlier removal employs the 3σ criterion. First, the mean and standard deviation of valid data for a single indicator in a single partition over the past 30 minutes are calculated. Values ​​exceeding the mean plus or minus three times the standard deviation are considered outliers and directly replaced with the arithmetic mean of the two preceding and following valid values. Missing value imputation is handled in stages based on the duration of the missing value. When the duration of consecutive missing values ​​is no more than 3 minutes, the arithmetic mean of seven valid data points (three minutes before and after the missing value) is used to impute the missing value. When the duration of consecutive missing values ​​exceeds 3 minutes, the historical mean of the same period and working conditions within the past seven days is used to impute the missing value. Simultaneously, the imputed data is marked with a confidence level indicator to prevent outlier imputations from interfering with model training.

[0028] The data normalization and alignment algorithm employs differentiated normalization and encoding rules for features of different attributes, eliminating differences in units and numerical magnitudes. Dynamic time-series features utilize min-max normalization, separately calculating the historical maximum and minimum values ​​within the entire training set for each feature type, linearly mapping the original values ​​to the range of 0 to 1. The normalized value equals the original value minus the feature's minimum value, then divided by the difference between the feature's maximum and minimum values. For example, if the historical minimum for passenger flow density is 0 people per square meter and the historical maximum is 6 people per square meter, and the original passenger flow density at a certain moment is 4.2 people per square meter, the normalized value is 4.2 minus 0 divided by 6, resulting in 0.7. Static spatial attribute features employ the same min-max normalization rules, with the statistical scope covering all partition data of all stations along the entire line, ensuring that the spatial attributes of different stations have a uniform and comparable scale. The operational period classification labels are converted using a one-hot encoding method. Peak, off-peak, and low-peak operating conditions each correspond to three independent coding bits. Each bit is represented by 0 or 1 to indicate whether it belongs to the corresponding operating condition, avoiding the introduction of false weights due to differences in classification values. For example, the coding bits for peak operating conditions are 1 for the first bit, 0 for the second bit, and 0 for the third bit. These three coding bits are included as independent features in the operational channel.

[0029] The core of the static spatial attribute mapping algorithm is spatial index matching and temporal broadcast expansion, achieving dimensional alignment between static attributes and dynamic time-series data. The first step establishes a spatial index mapping relationship between zones. Each air conditioning control zone is assigned a unique spatial identifier, corresponding to a fixed set of static attribute parameters, including building area, building envelope heat transfer coefficient, design cooling load index, and orientation code. The second step performs a temporal broadcast expansion operation. For all time-step samples of a single zone, the static attribute values ​​of that zone are copied to the feature vector of each time step, ensuring that the static attributes and dynamic time-series data maintain consistency in feature length. For example, if a zone has 1440 minute-level time steps per day, each time step originally contained 4 dynamic features, after broadcast expansion, 4 new static features are added, and the feature length of each time step is uniformly 8, ensuring complete alignment of feature dimensions across all time steps. Static attributes can participate in full-time-series feature calculations.

[0030] Following the single-zone, single-time-period minimum statistical unit accounting standard, the multimodal sensing data across the entire domain is used as the basic input source to connect with feature mapping rules, completing the unified accounting and feature tensor construction of six types of sensing indicators: passenger flow, environment, equipment, meteorology, space, and operation. Also following the single-zone, single-time-period minimum statistical unit accounting standard, the unified spatiotemporal sensing dataset is used as the basic input source to connect with feature mapping rules. Through three steps—multi-point weighted accounting, multi-channel matrix construction, and multi-channel tensor concatenation—the unified accounting and feature tensor construction of six types of sensing indicators—passenger flow, environment, equipment, meteorology, space, and operation—are completed.

[0031] The multi-point zoning unified calculation uses a distance-weighted average algorithm to solve the problem of fusing data from multiple similar sensors within a single zone. Using the geometric center of the zone as the reference point, the weight of each sensor is inversely proportional to its distance from the reference point, and the sum of the weights of all sensors is normalized to 1. The closer a sensor is to the center, the higher its weight; the farther away, the lower its weight, thus eliminating errors in the representative value of the zone caused by sensor installation position deviations. For example, if three temperature and humidity sensors are installed in a zone, with distances of 5 meters, 10 meters, and 15 meters from the zone center, corresponding to weights of 0.55, 0.3, and 0.15 respectively, and the temperature readings from the three sensors are 24.1 degrees Celsius, 24.3 degrees Celsius, and 24.2 degrees Celsius, the average temperature of the zone is equal to the sum of each reading multiplied by its corresponding weight, resulting in a final calculated value of 24.18 degrees Celsius. The same weighted calculation logic is used for multi-point indicators such as passenger flow and equipment to obtain a unified representative value for each zone per minute.

[0032] The feature matrix construction for each of the six channels is performed separately, with each channel corresponding to multiple subdivided feature dimensions. The features of all time steps within a single zone are combined to form a two-dimensional feature matrix, with rows corresponding to time steps and columns corresponding to feature dimensions. Specifically, the passenger flow channel includes three features: average passenger flow density per zone, minute-by-minute arrival rate, and minute-by-minute departure rate; the environmental channel includes two features: average dry-bulb temperature and average relative humidity per zone; the equipment channel includes four features: fan operating frequency, valve opening, real-time operating power, and supply / return water temperature difference; the meteorological channel includes three features: outdoor dry-bulb temperature, outdoor relative humidity, and atmospheric pressure; the spatial channel includes four features: normalized building area, building envelope heat transfer coefficient, design cooling load index, and orientation code; and the operational channel includes two features: three-digit unique heating condition code and the duration from peak hours. After each channel's feature matrix is ​​constructed separately, the time step length and data integrity are uniformly verified.

[0033] The six-channel feature tensor concatenation is performed along the feature dimension, horizontally concatenating the feature matrices of the six channels to form a complete temporal feature matrix for a single partition. Then, the feature matrices of all partitions across the entire site are vertically stacked along the spatial dimension, ultimately generating a three-dimensional six-channel feature tensor. The three dimensions of the tensor correspond to the number of partitions, the number of consecutive time steps, and the total number of feature dimensions, respectively. The total number of feature dimensions is the sum of the feature dimensions of the six channels. For example, if a single site has 10 partitions, 1440 time steps per day, and a total feature dimension of 18, then the final tensor dimension will be 10 x 1440 x 18, perfectly matching the input format requirements of subsequent spatiotemporal prediction networks.

[0034] Based on the core requirement of accurate isolation of passenger flow tidal condition samples, the constructed time-series feature tensor is processed by independent windowing for three types of conditions: peak, off-peak, and valley. The sample set is split according to the time window to achieve standardized isolation of tidal condition samples. This is achieved through three steps: condition boundary calibration, sliding window segmentation, and sample balancing verification.

[0035] The working condition time boundary calibration algorithm generates a daily working condition time interval array based on the subway operating timetable. Each working condition corresponds to a precise start and end timestamp. The morning peak working condition time interval is 7:00 to 9:00, and the evening peak working condition time interval is 17:00 to 19:00, with these two periods together forming the peak working condition sample set. The off-peak working condition time intervals are 9:00 to 17:00 and 19:00 to 21:00; and the low-peak working condition time intervals are 6:00 to 7:00 and 21:00 to 23:00. For each sample window to be generated, the start and end timestamps of the window are extracted, and inclusion judgment is performed with the time intervals of the three working conditions. Only when the start and end times of the window fall completely within the time interval of the same working condition is it determined to be a valid sample window. If the window crosses the boundary of two working conditions, it is directly determined to be invalid and discarded to avoid the mixing of features of different working conditions affecting the working condition adaptability of model training.

[0036] The fixed-step sliding window segmentation algorithm uses minute-level time steps as the smallest unit, with a window length of 30 minutes, corresponding to 30 consecutive time steps. The sliding step size is set to 5 minutes, meaning the window moves forward 5 time steps each time to generate the next sample. The segmentation operation starts from the start time of the first work condition of the day and slides sequentially until the end of the last work condition of the day. For example, the morning peak period lasts 120 minutes. Using a 30-minute window and a 5-minute step size, 19 consecutive peak work condition sample windows can be generated. Each sample window corresponds to a continuous 3D feature tensor fragment, retaining complete partition, time, and channel dimension information, and also includes the corresponding work condition classification label.

[0037] Sample balance verification and optimization address the issue of significant differences in sample sizes among the three operating conditions by performing equalization processing. For off-peak operating conditions with a large sample size, random undersampling is employed, randomly selecting a subset with a sample size ratio matching that of peak conditions for training. For peak operating conditions with a small sample size, window offset enhancement is used, generating derivative samples by shifting the window start time by 1 to 2 time steps before or after the peak condition, thus expanding the sample size for peak conditions. Ultimately, the sample size ratio of peak, off-peak, and low-peak operating conditions is controlled to be within 1:1.2:1 to avoid model bias towards the operating condition category with the larger sample size. After processing, the three types of samples are stored in independent operating condition sample sets, with all samples having completely identical tensor dimensions and uniform label formats, allowing them to be directly input into subsequent prediction networks for training.

[0038] S103 employs a multimodal spatiotemporal prediction dual-branch network to extract passenger flow-load coupling evolution features. It generates multi-step air conditioning load prediction sequences through temporal convolution and verifies the prediction fitting accuracy through spatiotemporal discrimination branches. Based on the adversarial iterative convergence of tidal operating condition samples after windowing, it solidifies and forms a benchmark model for tidal scenario load prediction.

[0039] In one implementation, to address the need for high-precision adaptation and prediction of air conditioning load under scenarios of strong tidal fluctuations in subway passenger flow, three core mechanisms are introduced: cross-modal coupling feature decoupling, generation-discrimination dual-track collaborative computation, and tidal condition adversarial knowledge distillation. This establishes a spatiotemporal dual-domain multimodal load prediction modeling system. To address the issues of insufficient load prediction accuracy and poor adaptability to operating condition switching under scenarios of strong tidal fluctuations in subway passenger flow, three core mechanisms are introduced: cross-modal coupling feature decoupling, generation-discrimination dual-track collaborative computation, and tidal condition adversarial knowledge distillation.

[0040] The cross-modal coupling feature decoupling algorithm splits feature paths according to data attribute differences and achieves adaptive coupling of dynamic and static features through a gated fusion mechanism, solving the feature interference problem of direct fusion of multi-source data. The algorithm's first step is feature set splitting, dividing the 18-dimensional input features into two groups along the feature dimensions: a dynamic temporal feature set containing 12 time-varying features across four categories (passenger flow, environment, equipment, and weather), and a static attribute feature set containing 6 time-invariant attribute features across two categories (space and operation). The second step performs dual-path independent feature extraction: the dynamic temporal feature set follows a temporal convolution extraction path, extracting short-, medium-, and long-term temporal evolution patterns through multiple dilated convolutions; the static attribute feature set follows a fully connected mapping path, using two fully connected layers to perform feature dimensionality upscaling, mapping to the same high-dimensional space as the dynamic features. The third step performs gated adaptive fusion, setting a learnable fusion gate unit. The two feature paths are concatenated and input into a fully connected layer, outputting two weight values. After normalization, the sum of the weights is 1. The weights are then applied to the two feature paths element-wise and summed to obtain the coupled features. The gating weights will be automatically adjusted according to the current operating conditions. During peak periods with drastic fluctuations in passenger flow, the dynamic feature weight can reach 0.75, and the static feature weight is 0.25. During off-peak periods with stable passenger flow, the dynamic feature weight is adjusted to 0.6, and the static feature weight is increased to 0.4, thus achieving adaptive allocation of feature contribution under different operating conditions.

[0041] The dual-track collaborative computation mechanism for generation and discrimination adopts a game-theoretic dual-branch architecture. This architecture uses adversarial training logic to simultaneously improve prediction accuracy and verification reliability, matching the dual business requirements of subway air conditioning energy-saving control for prediction accuracy and stability. The optimization goal of the generation branch is to output a prediction sequence that closely approximates the actual load distribution, reducing the discriminative power of the discrimination branch. The optimization goal of the discrimination branch is to accurately identify whether the input sequence belongs to real historical data or generated prediction data; the two form a collaborative optimization relationship through mutual competition. The computation process employs a dual-loss linkage rule. The total loss of the generation branch consists of two weighted components: a 0.7x regression loss to measure the deviation between the predicted and actual values, and a 0.3x adversarial loss to measure the discrimination branch's judgment of the generated sequence. The loss of the discrimination branch is composed of both the real sequence classification loss and the generated sequence classification loss. The training adopts an alternating update mode. Each batch of training is executed in two steps. The first step fixes the weight of the discrimination branch and updates the generation branch and the shared feature layer to minimize the prediction error while maximizing the misclassification rate of the discrimination branch. The second step fixes the weight of the generation branch and updates the discrimination branch to maximize the accuracy of distinguishing between real and generated sequences. The performance of the two branches is improved simultaneously through bidirectional game theory.

[0042] The tidal load condition adversarial knowledge distillation mechanism addresses the business pain points of insufficient sample size and model generalization under peak and off-peak load conditions. It achieves feature transfer across load conditions through soft-label knowledge transfer, improving prediction adaptability throughout the entire tidal cycle. The core algorithm is to transfer general load patterns learned from load conditions with sufficient samples to load conditions with scarce samples, compensating for fitting bias caused by uneven data distribution. During computation, a teacher model and student models are first selected. In each round, a training model from one type of load condition is randomly selected as the teacher model, and the models from the other two types of load conditions serve as student models. The three types of load conditions take turns acting as teachers to ensure bidirectional knowledge transfer. Soft labels are then generated, and the training samples corresponding to the student model are input into the teacher model with frozen weights. The predicted sequence output by the teacher model is divided by a temperature coefficient of 3 to soften the distribution. The temperature coefficient is calibrated through a grid search, traversing the range of 1 to 5 with a step size of 0.5, and selecting the value with the lowest error in the peak load test set as the final parameter. The temperature coefficient of 3 corresponds to the optimal balance between transfer generalization and load specificity, ensuring the cross-load transfer of basic load patterns while preserving the differences in load distribution between different load conditions, retaining more similarity information between load conditions. The higher the temperature coefficient, the smoother the soft label distribution and the richer the transfer generalization information. Finally, distillation training is performed. The total loss of the student model is composed of a weighted average of 0.7 times the true hard label loss and 0.3 times the teacher soft label loss. Backpropagation only updates the weights of the student model. For example, if the peak load sample size is only 30% of the off-peak sample size, using the off-peak load model as the teacher can transfer the basic mapping pattern of temperature, humidity and load in the off-peak scenario to the peak model, reducing the risk of overfitting caused by insufficient peak samples.

[0043] Following the standard for in-depth mining of passenger flow-load coupling evolution characteristics, the six-channel spatiotemporal feature tensor after windowing isolation is used as the basic input source and connected to the dual-branch convolution operation paradigm to complete the bidirectional linkage adaptation of multi-step load prediction sequence generation and fitting accuracy closed-loop verification. The overall network architecture is divided into three levels: a shared cross-modal decoupling feature extraction module, a temporal convolution generation branch, and a spatiotemporal discrimination branch. Input data is split into two branches after the coupling features are extracted by the shared layer, and bidirectional linkage adaptation is achieved through loss linkage.

[0044] The shared cross-modal decoupled feature extraction module is a pre-feature extraction unit shared by both branches, responsible for converting the original multimodal features into high-dimensional coupled feature maps. The module input is a three-dimensional feature tensor, with dimensions equal to the batch size multiplied by the time step size multiplied by the feature dimension. During training, sample batches are organized independently by partition, with samples within each batch originating from the same partition. Different partitions independently complete model training and parameter iteration, ensuring the model learns the specific load mapping rules for each partition. During inference, all partitions across the entire site call the same set of baseline model weights in parallel, combining the specific spatial attribute features of each partition to generate partition-specific prediction results, balancing model generality and partition adaptability. A single batch is set to 16 samples by default, with each sample containing 30 consecutive time steps and 18-dimensional fused features. Internally, the module first sets a feature splitting layer, splitting along the feature dimension into two parallel paths: a 12-dimensional dynamic feature and a 6-dimensional static feature. The dynamic feature path uses two dilated temporal convolutional layers. The first layer has a kernel size of 3, a dilation coefficient of 1, and 32 output channels, extracting short-term temporal fluctuation features. The second layer has a kernel size of 3, a dilation coefficient of 2, and 64 output channels, expanding the temporal receptive field to 7 time steps to capture medium-term passenger flow trends. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function to avoid gradient vanishing and accelerate training convergence. The static feature path uses two fully connected layers. The first layer has an output dimension of 32, and the second layer has an output dimension of 64, perfectly aligned with the number of output channels of the dynamic features, mapping low-dimensional static attributes to a high-dimensional feature space. A gated fusion layer is set at the end of the module. After calculating the adaptive fusion weights, the two feature paths are weighted and summed, outputting a coupled feature map with a dimension of batch size multiplied by time step size multiplied by 64, which is simultaneously sent to the two downstream branches. For example, with 16 samples in a single batch and 30 time steps per sample, after splitting, extraction, and fusion, 16 sets of 30x64 coupled feature matrices are output.

[0045] The temporal convolution generation branch is responsible for outputting multi-step air conditioning load prediction sequences. Its core employs a causal dilated convolution structure to ensure the causal rationality of the predictions; that is, the prediction at each time step relies only on historical data and does not reveal future information. The branch hierarchy consists of three layers of causal dilated temporal convolution, a global average pooling layer, and an output fully connected layer. The causal convolution operation rule involves masking the convolution results in the temporal direction, resetting the convolution weights for future positions to zero, and retaining only the calculation results of the current and historical time steps, ensuring that the prediction logic conforms to the business temporal patterns. The first causal convolution kernel has a size of 3, a dilation coefficient of 1, 64 output channels, and a receptive field covering 3 time steps, extracting short-term load change trends. The second causal convolution kernel has a size of 3, a dilation coefficient of 2, 32 output channels, and a receptive field covering 7 time steps, capturing medium-term load evolution patterns. The third causal convolution kernel has a size of 3, a dilation coefficient of 4, 16 output channels, and a receptive field covering 15 time steps, matching the time span of the prediction step. The global average pooling layer performs an averaging operation along the channel dimension, compressing the channel dimension to 1 to obtain a single-channel time-series feature vector. The output fully connected layer has an input dimension of 30 and an output dimension of 15, corresponding to the minute-by-minute air conditioning load forecast values ​​for the next 15 minutes. It outputs independently by zone to form a multi-step load forecast sequence. For example, if the input is the coupling features of the south zone of the station hall for 30 consecutive minutes, the branch outputs 15 load forecast values ​​for the 31st to 45th minutes of that zone, which are 12.3 kW, 12.7 kW and up to 14.1 kW, forming a continuous forecast sequence.

[0046] The spatiotemporal discriminant branch is responsible for evaluating the fit between the predicted sequence and the real load distribution, outputting a fit confidence score in the range of 0 to 1. Simultaneously, it drives the generation branch to optimize accuracy through error backpropagation. The branch's input is the concatenation result of the coupled feature map and the load sequence to be discriminated. Before concatenation, a 1×1 convolution layer compresses the 64-channel coupled feature map to 8 channels. Then, the single-channel load sequence is copied and expanded along the channel dimension to 8 channels. The two are concatenated along the channel dimension to form a 16-channel spatiotemporal feature map, maintaining a size of time step × 16 channels, which serves as the unified input for subsequent spatiotemporal convolutional layers. The sequence to be discriminated includes the real historical load sequence and the predicted sequence output by the generation branch, serving as positive and negative sample inputs, respectively. The branch hierarchy consists of two spatiotemporal convolutional layers, a global average pooling layer, and a sigmoid output layer. The first spatiotemporal convolutional layer has a 3x3 kernel size and performs convolution operations along both the temporal and spatial dimensions to extract the temporal fluctuation features and spatial correlation features of the sequence. It outputs 32 channels, followed by a batch normalization layer and the LeakyReLU activation function. The second spatiotemporal convolutional layer also has a 3x3 kernel size and 16 output channels, further compressing the feature dimensions and extracting higher-order discriminative features. A global average pooling layer compresses the three-dimensional features of time, space, and channels into a one-dimensional global feature vector. The sigmoid output layer outputs a single confidence score; the closer the score is to 1, the higher the match between the input sequence and the true load distribution. The bidirectional linkage logic is as follows: the generation branch outputs a predicted sequence, which is then input into the discriminant branch to obtain a confidence score. The lower the score, the greater the difference between the predicted sequence and the true distribution, corresponding to a higher adversarial loss. When backpropagating to the generation branch, the weight update magnitude is larger, pushing the generation branch to generate prediction results that are closer to the true distribution. The discriminant branch simultaneously learns the feature differences between the true sequence and the generated sequence, continuously improving its discriminative ability, forming a closed-loop verification and optimization linkage mechanism.

[0047] Based on the core requirement of generalization and solidification of the tidal load forecasting benchmark model, multi-condition cross-adversarial iteration and parameter distillation constraints are applied to the network weights. Parameter solidification is completed only when the fitting accuracy across the entire tidal cycle meets the standard and the loss function converges stably, thus completing the construction of a globally adaptable tidal load forecasting benchmark model. The training process, based on the core requirement of generalization and solidification of the tidal load forecasting benchmark model, is performed in four stages: off-peak pre-training initialization, multi-condition cross-adversarial iteration, tidal condition knowledge distillation optimization, and dual-standard convergence solidification. Multi-condition adversarial and parameter distillation constraints are applied to the network weights, and parameter solidification is completed only when the fitting accuracy across the entire tidal cycle meets the standard and the loss function converges stably.

[0048] The first phase involves pre-training and initializing for off-peak operating conditions. The most abundant off-peak operating condition samples are selected as the pre-training dataset to allow the model to learn the basic mapping relationship between passenger flow and air conditioning load, providing stable initial parameters for subsequent multi-condition training. Training parameters are set as follows: initial learning rate 0.001, Adam optimizer, batch size 16, and a preset total of 30 training epochs. During training, only the temporal convolutional generation branch and shared feature layer are optimized; the discriminant branch is not updated initially. The loss function uses mean squared error, calculating the average of the squared pointwise errors between the predicted load sequence and the actual load sequence. After each batch of samples is input into the network, the predicted sequence is output through the shared feature layer and the generation branch. After calculating the loss, backpropagation is used to update the weights of the generation branch and the shared layer sequentially. Pre-training terminates when 30 epochs of training are completed, or when the average absolute percentage error of off-peak operating condition predictions is below 6%. Pre-training stops when either condition is met, and the initial model weights are obtained.

[0049] The second phase involves multi-condition cross-adversarial iterative training. Peak, off-peak, and low-peak condition sample sets are input alternately in batches, with the batch ratio of the three conditions in each training round being 1:1.2:1 to ensure balanced training. The training parameters are adjusted to a learning rate of 0.0005, which decays to 0.8 times its original value every 20 rounds, gradually reducing the weight adjustment range. The batch size remains at 16, and the total training rounds are preset to 100. Single-batch training strictly follows a two-step alternating update rule. The first step updates the generation branch and shared feature layer, fixes the discriminant branch weights, calculates the total loss, and then backpropagates. The total loss is composed of a weighted average of 0.7 times the mean squared error regression loss and 0.3 times the adversarial loss. The formula for the total loss of the generation branch is: ,in, To generate the predicted sequence of the branch output, To determine the confidence score of the output of the discrimination branch, this step only updates the weights of the shared feature layer and the generating branch through backpropagation; the weights of the discrimination branch remain frozen. The formula for the discrimination branch loss is: The first step only updates the weights of the discriminant branch, while keeping the weights of the shared feature layer and the generator branch frozen. This avoids gradient backpropagation interfering with generator training. The adversarial loss is the square of the difference between the discriminant branch's confidence score for the generated sequence and 1, aiming to drive the generator branch to produce a high-precision prediction sequence sufficient to fool the discriminant branch. The second step updates the discriminant branch, fixing the weights of the generator branch and the shared layer. The actual load sequence and the generated prediction sequence are input respectively, labeled 1 and 0. After calculating the binary cross-entropy loss, backpropagation updates the discriminant branch weights. During training, the sample load conditions are cyclically switched in the order of peak, off-peak, and trough to ensure the model learns the load patterns of the three types of load conditions simultaneously, avoiding overfitting to a single load condition.

[0050] The third stage involves adversarial knowledge distillation optimization based on tidal operating conditions. After every five rounds of adversarial training, one round of knowledge distillation is performed to achieve feature knowledge transfer across operating conditions. During the operation, all weights of the teacher model are first frozen to ensure the stability of the knowledge output. In each round, one type of operating condition model is randomly selected as the teacher, and the other two types are selected as students, with the three types of operating conditions taking turns as teachers to achieve bidirectional knowledge transfer. Then, the operating condition samples corresponding to the student models are input into the teacher models. The predicted sequences output by the teacher models are divided by a temperature coefficient of 3 to perform softening processing, resulting in soft labels. Finally, the total loss of the student models is calculated, which equals 0.7 times the true hard label loss plus 0.3 times the distillation loss. The distillation loss is the mean squared error between the student prediction results and the teacher soft labels. Backpropagation only updates the weights of the student models. This stage effectively compensates for the lack of data in operating conditions with scarce samples. For example, the sample size for off-peak operating conditions is small. Through knowledge distillation of the off-peak teacher models, the basic mapping pattern between outdoor temperature and humidity changes and station load can be learned, improving the prediction generalization during off-peak periods.

[0051] The fourth stage involves convergence determination and model parameter solidification. A dual convergence criterion is established, with both criteria being met simultaneously to determine model convergence. The first criterion is achieving full-condition accuracy. The mean absolute percentage error (MASE) is calculated on independent test sets for three different conditions, requiring that the error not exceed 5% for peak conditions, 4% for off-peak conditions, and 5% for low-peak conditions, with all tidal cycle fitting accuracy meeting the criteria. The MASSE is calculated by dividing the absolute error of each prediction point by the corresponding true value and taking the average of all test samples. The second criterion is stable convergence of the loss function. The total loss value is calculated over 10 consecutive training rounds, and the coefficient of variation (i.e., standard deviation) is divided by the mean. A MASE less than 0.001 indicates stable convergence of the loss function, meaning the model weights no longer fluctuate significantly. Once both criteria are met, training is immediately stopped, and all weight parameters of the shared feature layer, temporal convolutional generation branch, and spatiotemporal discrimination branch are frozen and saved, forming a globally adaptable tidal load prediction benchmark model. If the convergence condition is not met after 100 training rounds, the model with the highest accuracy on the test set is selected as the final benchmark model. For example, at the 82nd training round, the peak test set average error is 4.3%, the off-peak is 3.7%, and the trough is 4.6%, all of which meet the accuracy requirements; the coefficient of variation of the total loss over 10 consecutive rounds is 0.0008, which is lower than the threshold of 0.001, indicating that the model has converged, and all weight parameters are fixed to output the final benchmark model.

[0052] S104, based on the output of the load prediction benchmark model for tidal scenarios, configures a mechanism for zonal time-series deviation, full-station spatial weighted double-layer prediction accuracy and energy-saving potential quantification. By modeling the load influence factor weights in the intermediate feature space of the model, and combining spatial weight allocation and load thermal mapping, it outputs the time-series and zonal load prediction curves and energy-saving spatial time-series results for the entire station.

[0053] In one implementation, combining the characteristics of the multi-step prediction sequence output by the tidal scenario load prediction benchmark model with the subway air conditioning two-layer quantitative evaluation specification, a partitioned time-series deviation calculation rule and a spatial dimension weight mapping logic are introduced. The benchmark model output is decomposed and analyzed in both time-series and spatial dimensions to generate spatiotemporal dual-domain load quantization adaptation information. Using the full-partition multi-step load prediction sequence output by the benchmark model and the intermediate convolutional features of the shared feature layer as dual input sources, a layer-by-layer decomposition and analysis is performed according to two dimensions: time-series accuracy verification and spatial weight distribution, to generate spatiotemporal dual-domain load quantization adaptation information.

[0054] The partitioned time-series deviation calculation algorithm uses historical load data from the same period and under the same operating conditions as a benchmark to complete point-by-point deviation calculation and stability assessment of the predicted sequence. This is performed in three steps. The first step constructs a historical benchmark load sequence, selecting measured load data from the same week, time period, and operating conditions within the past 14 days. Abnormal day samples corresponding to equipment failures or extreme weather are removed, and the arithmetic mean of the remaining valid samples is calculated minute by minute to generate the partitioned benchmark load sequence. For example, to calculate the benchmark load from 7:00 AM to 9:00 AM, valid data from the past 14 days' weekday morning peak are selected, resulting in 10 valid samples. The average is then calculated minute by minute to obtain 120 minute-level benchmark values. The second step performs point-by-point time-series deviation calculation, sequentially calculating three core deviation indicators: absolute deviation (the difference between the predicted load value and the benchmark load value), relative deviation (the percentage of the absolute deviation divided by the benchmark load value), and cumulative deviation (the sum of all absolute deviations within the prediction window). The third step calculates the deviation volatility, taking all relative deviation values ​​within the complete prediction window and calculating the standard deviation of the deviation and the average of the absolute deviation values. The ratio of these two values ​​is the deviation volatility, used to characterize the time-series stability of the prediction results. Simultaneously, the 3σ criterion is used to identify abnormal deviation points. Minute points within the window whose relative deviation exceeds three times the standard deviation of the mean are marked as abnormal deviation points and subsequently calibrated separately. The load forecast values ​​of abnormal deviation points are calibrated using a neighborhood weighted correction method. The forecast values ​​of the three normal deviation points before and after the abnormal point are taken, and the correction value is calculated by inversely weighting them according to the distance. The closer to the abnormal point, the higher the weight. The corrected value replaces the original forecast value. The corresponding energy-saving potential value is recalculated synchronously according to the corrected load value to ensure the continuous smoothness of the time series curve and eliminate the interference of single-point anomalies on the overall energy-saving calculation. For example, in the 15-minute forecast window of the south section of the station hall, the average relative deviation is 4.2%, the standard deviation of the deviation is 0.8%, and the corresponding deviation volatility is 19%. Among them, the relative deviation of the 12th minute reaches 8.1%, which exceeds the threshold of the mean plus three times the standard deviation, and is marked as an abnormal deviation point. The minute-by-minute deviation data, volatility data, and abnormal point markings of all sections together constitute the basic quantitative information of the time series dimension.

[0055] The spatial dimension weighting mapping algorithm combines initial attribute weighting with inverse mapping of intermediate features to generate a spatial weighting system that takes into account both physical attributes and model features. This is implemented in two steps. The first step calculates the initial spatial weights. Three core spatial parameters are selected: zone building area, upper limit of designed personnel density, and designed cooling load index. These parameters are then normalized to the minimum and maximum values ​​for all zones across the entire site, mapping their values ​​to the range of 0 to 1. Subsequently, they are weighted and summed at weight ratios of 0.3, 0.4, and 0.3 to obtain the initial weight for each zone. Finally, a global normalization is performed on all zone weights to ensure that the sum of all zone weights is 1. For example, the normalized value for the building area of ​​the North Zone of the station hall is 0.8, the normalized value for designed personnel density is 0.9, and the normalized value for designed cooling load is 0.85. After weighting, the initial weight is 0.8 x 0.3 + 0.9 x 0.4 + 0.85 x 0.3, resulting in 0.855. After global normalization, the final initial spatial weight for this zone is 0.12.

[0056] The second step involves reverse mapping of the intermediate feature space, using the intermediate output of the shared feature layer of the baseline model to correct the weights. The 64-channel feature map output from the second convolutional layer of the shared feature layer is extracted. The global average activation value of each channel across the entire sample is calculated. Then, the correlation between the activation value and the true payload value of each channel is calculated using the Pearson correlation coefficient algorithm; the absolute value of the correlation is the importance score of that channel. Subsequently, based on the input-output connectivity of the first convolutional layer, a weight contribution propagation method is used, employing the kernel weights of the first temporal convolutional layer of the shared feature layer as the propagation medium to calculate the contribution ratio of each original input feature to each intermediate output channel. The specific formula is: Total importance of the i-th dimension original feature... ,in, This represents the mean absolute value of the kernel weights corresponding to the j-th output channel of the i-th input feature in the first convolutional layer. For the importance score of the j-th intermediate channel, the importance scores of the 64 intermediate channels are back-assigned to the 18-dimensional original input features. Each original feature corresponds to the sum of the importance scores of all associated intermediate channels. Finally, the 18-dimensional features are merged into six major feature channels to obtain the global impact weight of each feature category on load changes. For example, after back-mapping calculation, the cumulative importance score of the passenger flow channel's 3 features accounts for 42%, the environmental channel's 2 features accounts for 18%, the equipment channel's 4 features accounts for 15%, the meteorological channel's 3 features accounts for 12%, the spatial channel's 4 features accounts for 8%, and the operational channel's 2 features accounts for 5%. After the initial spatial weights are matched and aligned with the feature impact weights, spatial dimension weight adaptation data is formed. The temporal deviation data and spatial weight data together constitute the spatiotemporal dual-domain load quantization adaptation information, serving as the unified input for subsequent hierarchical modeling.

[0057] Based on the dual-layer quantification architecture specification of time-sharing and zone-based energy-saving management requirements and prediction accuracy-energy-saving potential for subway stations, a hierarchical modeling logic for the allocation of temporal and spatial load quantification adaptation information is implemented. This includes dividing the process into a zone-based time-series deviation accuracy verification link and a station-wide spatial weighted energy-saving calculation link. A load influence factor weight mining mechanism based on the intermediate feature space of the benchmark model is configured to generate time-series-spatial layer-specific quantification feature allocation information. Furthermore, after independent computation of the two links, they are structurally integrated to generate time-series-spatial layer-specific quantification feature allocation information.

[0058] The partitioned temporal deviation accuracy verification link employs a dual-dimensional joint scoring mechanism to complete the prediction accuracy classification and correction coefficient calculation, specifically in three steps. The first step calculates the dual-dimensional accuracy score, composed of a weighted average of the relative deviation score and the fit confidence score, each with a weight of 0.5. The relative deviation score uses a reverse linear scoring rule: a score of 1 is given when the relative deviation is 0, and a score of 0 is given when the relative deviation is 10%, with intermediate values ​​calculated via linear interpolation. The fit confidence score directly uses the confidence value in the 0-1 interval output by the spatiotemporal discriminant branch. The weighted sum of the two scores yields the comprehensive accuracy score for a single partition and single window.

[0059] The second step involves classifying the accuracy into three levels. A comprehensive score of 0.8 or higher is classified as high accuracy, and the prediction results can be directly used for energy-saving calculations. A score between 0.6 and 0.8 is classified as acceptable, and the prediction results need to be calibrated with a correction coefficient before use. A score less than 0.6 is classified as low accuracy, and the corresponding results are marked as pending review and are not included in the current period's energy-saving potential calculation. The third step derives the deviation correction coefficient, which is equal to the comprehensive accuracy score multiplied by the empirical calibration coefficient of 0.98. The empirical calibration coefficient is obtained by statistically analyzing the deviation between the prediction accuracy and the measured accuracy of the same period in history, and is used to offset systematic overestimation biases. For example, the average relative deviation of the south section of the station hall is 4.2%, corresponding to a relative deviation score of 0.58; the confidence level of the discriminant branch output fit is 0.87, corresponding to a confidence score of 0.87; the weighted comprehensive score is 0.58 multiplied by 0.5 plus 0.87 multiplied by 0.5, resulting in 0.725, which is classified as acceptable; the corresponding deviation correction coefficient is 0.725 multiplied by 0.98, resulting in approximately 0.71. The accuracy level, overall score, and deviation correction factor of all partitions together constitute the output result of the timing accuracy link.

[0060] The station-wide space-weighted energy-saving calculation link uses the traditional fixed-condition control mode as a benchmark to complete the initial energy-saving potential calculation and the refined weight mining of load influence factors. This is implemented in three steps. The first step calculates the baseline energy consumption under the fixed-condition control mode, simulating the minute-by-minute energy consumption level under the traditional fixed control mode. Based on the zoned design cooling load index, it is corrected by combining the real-time outdoor temperature and the zoned personnel density benchmark value. The baseline energy consumption equals the zoned design cooling load index multiplied by the zone area, then multiplied by the outdoor temperature correction coefficient and the personnel density correction coefficient. The outdoor temperature correction coefficient is based on 26 degrees Celsius, increasing by 0.08 for every 1 degree Celsius increase. The personnel density correction coefficient is based on the design density and is linearly adjusted according to the actual density ratio. For example, the design cooling load index for the south zone of the station hall is 120 watts per square meter, with an area of ​​800 square meters. The current outdoor temperature is 28 degrees Celsius, corresponding to an outdoor correction coefficient of 1.16 and a personnel density correction coefficient of 0.7. The calculated baseline energy consumption is 120 x 800 x 1.16 x 0.7, resulting in approximately 77.95 kilowatts.

[0061] The second step calculates the initial energy-saving potential. The initial energy-saving potential per minute for a single zone equals the difference between the baseline energy consumption and the predicted load value, divided by the percentage of the baseline energy consumption, representing the energy consumption reduction potential compared to the traditional control mode. The third step deepens the mining of load influencing factor weights, generating zone-specific weights based on global feature weights. Combining the spatial attributes and operating characteristics of each zone, the weights of the six global feature categories are adjusted for zone adaptation. For public areas with high personnel density, the weight of passenger flow factors is increased; for entrance and exit zones with strong heat exchange in the building envelope, the weight of meteorological factors is increased; and for equipment areas with dense equipment, the weight of equipment operation factors is increased. After adjustment, the sum of all weights for each zone remains 1. For example, the south zone of the station hall is a public area with high passenger flow. The weight of passenger flow factors is increased from 42% to 47% globally, and the weight of meteorological factors is decreased from 12% to 9%. The weights of other categories are adjusted simultaneously to ensure a total proportion of 1. The initial energy-saving potential value of each zone and the zone-specific influencing factor weights together constitute the output result of the spatial energy-saving link.

[0062] The outputs of the time-series accuracy link and the spatial energy-saving link are matched and integrated one by one according to the partition dimension to form time-series-spatial hierarchical exclusive quantitative feature allocation information, which includes four types of core data: accuracy level of each partition, deviation correction coefficient, initial energy-saving potential, and exclusive influence factor weight, which serve as input for subsequent fusion operations.

[0063] Based on load forecasting accuracy calibration standards, energy-saving potential spatial quantification rules, and load heat map mapping specifications, multi-dimensional weighted fusion calculations are performed on hierarchical quantified feature allocation information to complete single-zone time-series accuracy verification and station-wide energy-saving potential spatial characterization. After weight allocation and deviation verification meet the standards, load heat maps are mapped, and the time-segmented load forecast curves and energy-saving spatial time-series results for the entire station are output. (The second paragraph is a repetition of the first and can be omitted.)

[0064] The multi-dimensional weighted fusion iterative calibration algorithm is executed in three steps, with two rounds of calibration and iterative verification to ensure the accuracy of the results. The first step performs zonal precision calibration, multiplying the initial minute-by-minute energy-saving potential value of each zone by the corresponding zone's deviation correction coefficient to obtain the zone's energy-saving potential value after one calibration. The energy-saving potential value corresponding to abnormal deviation points is further multiplied by a reduction factor of 0.8 to reduce the interference of abnormal data. For example, the initial energy-saving potential of the south zone of the station hall is 18% for a certain minute, with a deviation correction coefficient of 0.71; after one calibration, the energy-saving potential is 12.78%. The second step performs station-wide spatial weighted aggregation, multiplying the calibrated energy-saving potential value of each zone by the corresponding zone's corrected spatial weight, and summing element-wise to obtain the overall average energy-saving rate of the entire station. Simultaneously, the weighted average load of all zones is calculated using the same method to obtain the overall load value of the entire station. The third step involves iterative verification and weight optimization. The calculated station-wide energy-saving rate is compared with the measured average energy-saving rate under the same operating conditions over the past 30 days. If the relative deviation is less than 8%, the verification is considered successful. If the deviation exceeds 8%, iterative weight adjustment is triggered, prioritizing the adjustment of the weight ratios of passenger flow factors and meteorological factors. The adjustment increment is 0.02 per iteration. After adjustment, the fusion calculation is re-executed until the deviation meets the standard or the maximum number of iterations (10) is reached. For example, if the initial calculated station-wide average energy-saving rate is 15.6%, and the measured average energy-saving rate over the past 30 days is 14.2%, the relative deviation is 9.86%, exceeding the 8% threshold. After adjusting the weight of the passenger flow factor by 0.02 and the weight of the meteorological factor by 0.02, the calculation is re-executed. The second calculation yields a station-wide energy-saving rate of 14.7%, with a relative deviation of 3.5%, meeting the verification requirements. The current weights and calculation results are then locked.

[0065] The load heat map mapping algorithm employs linear color-gradient mapping and spatial boundary smoothing to transform numerical results into a visualized spatial distribution, specifically through three steps. The first step completes spatial coordinate mapping, using the station concourse and platform layout as the base map. Each air conditioning control zone corresponds to a set of closed pixel coordinate regions on the base map, establishing a one-to-one mapping relationship between zone numbers and pixel regions. The second step performs linear color-gradient mapping. For load data, a four-color gradient of blue, cyan, yellow, and red is used, with load values ​​ranging from 0 to 25 kW. Lower values ​​are more bluish, and higher values ​​are more reddish. The load values ​​are converted to corresponding RGB color values ​​using linear interpolation rules. For energy-saving potential data, a four-color gradient of green, yellow, orange, and red is used, with values ​​ranging from 0 to 30%. Lower energy-saving potential is more greenish, and higher potential is more reddish. The predicted load value and calibrated energy-saving potential value for each zone are matched to the corresponding color gradient and filled into the corresponding pixel region of the zone. The third step involves boundary smoothing, where a 1-pixel-width color transition interpolation is applied to the boundary lines of adjacent zones to eliminate harsh color boundaries between different zones and improve the visual coherence of the heat map. Load heat maps and energy-saving potential heat maps are then generated minute by minute to form a complete time-series heat map sequence.

[0066] A dual result verification rule is set up: first, the maximum relative deviation of the prediction for all zones does not exceed 10%; second, the deviation between the weighted energy saving rate of the entire station and the measured data of the same period in the past 30 days is less than 8%. If both conditions are met simultaneously, the weight allocation and deviation verification are deemed to have passed, and the final result is output. The final output consists of three types of structured results: the first type is the time-sharing and zone-based load prediction curve for the entire station, which stores the minute-by-minute load prediction value independently for each zone, forming a continuous time-series curve with a time granularity of 1 minute, covering the prediction window for the next 15 minutes; the second type is the energy-saving spatial time-series result, including the minute-by-minute calibrated energy-saving potential value of each zone, and the time-series sequence of the station's minute-by-minute average load and average energy saving rate; the third type is the load heat map and energy-saving potential heat map at the corresponding time, arranged in chronological order to form a dynamic distribution result, supporting zoned energy-saving management decisions.

[0067] S105, relying on the prediction curve and model feature weights, extracts the load contribution of a single zone and a single channel through feature gradient attribution, uses clustering algorithm to explore the coupling law between passenger flow tides and load fluctuations, precipitates typical operating condition control paradigms, and forms a time-sharing and zoned load prediction model and energy-saving control strategy library adapted to the tidal characteristics of the metro.

[0068] In one implementation, combining the temporal characteristics of time-sharing and zone-specific load forecast curves with the spatial distribution specifications of model feature weights, a feature gradient attribution analysis rule and a single-zone, single-channel contribution decomposition logic are introduced. The attribution dimension decomposition is performed on the forecast curve output by the benchmark model and the intermediate layer feature weights to generate zone-specific, single-channel load contribution adaptation information. Combining the temporal characteristics of time-sharing and zone-specific load forecast curves with the spatial distribution specifications of model feature weights, a gradient multiplication input feature attribution algorithm is used. Relying on the backpropagation mechanism of the tidal scenario load forecast benchmark model, the marginal contribution of each input feature to the load forecast result is quantified. Combined with the single-zone, single-channel independent decomposition logic, the attribution dimension decomposition is performed on the forecast curve output by the benchmark model and the intermediate layer feature weights to generate zone-specific, single-channel load contribution adaptation information.

[0069] The feature gradient attribution algorithm is executed in five steps to fully quantify the strength and direction of the impact of each type of input feature on the load prediction result. The first step involves forward propagation to lock the attribution target. All weight parameters of the baseline model are fixed, and a six-channel feature tensor of a single partition and single window is input. This tensor is then processed through a shared feature layer and a temporal convolution generation branch to obtain the minute-by-minute load prediction value for the next 15 minutes of the corresponding window. The arithmetic mean of all prediction points within the window is taken as the attribution target variable. Backpropagation directly performs a gradient calculation on this average target variable. The gradient is automatically propagated along the averaging operation of the 15 prediction points to the generation branch and the shared feature layer, ultimately mapping to each feature dimension of the input layer. This eliminates the need for point-by-point gradient calculation, ensuring that the attribution result reflects the overall feature contribution of the window and that the attribution targets the overall load level of the entire prediction window, rather than random fluctuations at a single moment.

[0070] The second step involves backpropagation to solve for the input gradient. Starting from the attribution target variable, the algorithm backpropagates layer by layer along the generation branch and shared feature layer to the input layer, calculating the partial derivative of the target value with respect to each dimension of the input feature, i.e., the gradient value. The sign of the gradient represents the direction of the feature's influence on the load; a positive value indicates that an increase in the feature value will push up the predicted load value, while a negative value indicates that it will pull down the predicted load value. The absolute value of the gradient represents the marginal influence strength of the feature. To avoid gradient explosion and outlier interference, a gradient pruning operation is performed simultaneously. The upper limit of the absolute value of the gradient is set to 1.0, and all gradients exceeding the threshold are truncated to the threshold boundary. At the same time, the confidence level of the pruned gradient is marked, and it is multiplied by a confidence reduction factor of 0.9 in subsequent contribution calculations to reduce the interference of outlier gradients on the results.

[0071] The third step is to calculate the original contribution score by multiplying element-wise. The normalized input value of each feature is multiplied by the absolute value of its corresponding gradient to obtain the original contribution score for that feature. This calculation method considers both the actual numerical magnitude of the feature and the strength of its marginal influence, avoiding numerical magnitude bias caused by using only the gradient. For example, if the normalized value of passenger flow density at a certain moment is 0.7, and the corresponding absolute gradient value is 0.6, then the original contribution score for this feature is 0.7 multiplied by 0.6, resulting in 0.42.

[0072] The fourth step is to perform channel aggregation and global normalization. The original contribution scores of all sub-features belonging to the same channel are summed to obtain the total contribution value of a single channel; then the total contribution value of the six channels is globally normalized, that is, the contribution value of each channel is divided by the sum of the contribution values ​​of the six channels to obtain the contribution ratio of a single channel, and the sum of the ratios of all channels is 1.

[0073] The fifth step is to perform multi-window sliding smoothing calibration. For the contribution results of consecutive prediction windows, an arithmetic mean smoothing is performed using a sliding window of length 5. This means the final contribution of the current window is the average of its own contribution and the contributions of the two windows before and after it. Smoothing eliminates contribution bias caused by random fluctuations in single windows, improving the stability of the time series dimension. For example, if the passenger flow channel contributions of five consecutive windows in the south section of the station hall are 47%, 45%, 49%, 46%, and 48%, respectively, the final contribution of the middle window after smoothing is the average of the five values, which is 47%.

[0074] The single-zone, single-channel contribution breakdown logic performs attribution calculations independently for each zone, generating a unique contribution result for each zone without cross-zone averaging, ensuring accurate matching between contribution and zone physical attributes and passenger flow characteristics. The window-by-window contribution data, confidence level indicators, and smoothing flags for all zones together constitute the zone-specific single-channel load contribution adaptation information, serving as standardized input for subsequent pattern mining.

[0075] Combining the needs for mining the coupling patterns of passenger flow tides and load fluctuations with the standards for refining energy-saving control paradigms, a two-stage progressive modeling logic is used for allocating load contribution information for single-channel load distribution in different zones. This involves dividing the data into a single-factor contribution precision quantification link and a multi-factor coupling pattern clustering mining link, configuring a passenger flow-load tide correlation matching mechanism, and generating unique feature allocation information for pattern mining and paradigm sedimentation. The logic involves sequentially performing calculations through the single-factor contribution precision quantification link and the multi-factor coupling pattern clustering mining link, and using the passenger flow-load tide correlation matching mechanism to complete business semantic mapping, thereby generating unique feature allocation information for pattern mining and paradigm sedimentation.

[0076] The single-factor contribution quantification method employs a grouped statistical analysis approach, cross-grouping by operating condition and zoning type to quantify the independent influence weights of each characteristic channel under different scenarios. The first step involves dividing the data into statistical units, categorizing operating conditions into peak, off-peak, and low-peak periods, and zoning types into three areas: station concourse public area, platform public area, and equipment management area, resulting in nine independent statistical units. The second step involves sample aggregation, assigning all historical samples to their corresponding units based on their operating condition and zoning type, and removing low-quality samples that failed attribution accuracy verification to ensure the reliability of the statistical basis. The third step calculates statistical characteristic values, calculating the mean, standard deviation, and 95th percentile of the contribution of each of the six channels for each unit. The mean represents the typical influence intensity of the channel under that scenario, the standard deviation represents the degree of fluctuation in contribution, and the 95th percentile represents the upper limit of influence under extreme scenarios. For example, in the peak-hour station concourse public area unit, which contains 1200 valid samples, the average contribution of passenger flow channels is 45%, the standard deviation is 6%, and the 95th percentile is 55%. In the off-peak equipment management area unit, the average contribution of equipment channels is 28%, and the standard deviation is 4%. Through the statistical results of 9 units, the influence weight variation patterns of each factor under different operating conditions and different zones can be clarified, providing a quantitative basis for the scenario adaptation of subsequent control strategies.

[0077] The multi-factor coupling pattern clustering mining link uses the K-means clustering algorithm to identify typical multi-factor combination influence patterns and explore the inherent coupling law between passenger flow tides and load fluctuations. This process is divided into five steps. The first step constructs a clustering feature set, with each sample corresponding to an 8-dimensional feature vector. The first six dimensions represent the contribution ratio of the six channels; the seventh dimension is the partition type code, with the concourse, platform, and equipment area corresponding to values ​​of 0, 0.5, and 1 respectively; and the eighth dimension is the operating condition type code, with peak, off-peak, and trough corresponding to values ​​of 0, 0.5, and 1 respectively. All features have been normalized to the range of 0 to 1 to eliminate the interference of dimensional differences on the clustering distance.

[0078] The second step initializes the cluster centers using a distributed initial point selection strategy. First, the first sample is randomly selected from the entire sample set as the initial center. Subsequent initial centers are selected from the sample furthest from any existing center, for a total of six initial cluster centers. This method avoids the local optima problem caused by excessive clustering of initial centers, thus improving the stability of the clustering results.

[0079] The third step involves iterative sample cluster allocation. The Euclidean distance from each sample to the six cluster centers is calculated. The Euclidean distance is the square root of the sum of the squares of the differences in features between the sample and the centers across all dimensions. The sample is then assigned to the cluster containing the nearest cluster center. After all samples are allocated, the mean feature value of all samples within each cluster is recalculated and used as the new cluster center for that cluster. This allocation and update process is repeated, recording the total offset of all cluster centers after each round.

[0080] The fourth step involves convergence testing, with two convergence conditions: first, the average offset of the cluster centers in two consecutive iterations is less than 0.001, where the average offset is the arithmetic mean of the offset distances of all centers; second, the maximum number of iterations (50) is reached. The clustering operation terminates when either condition is met, and the final six clusters and their corresponding center vectors are output.

[0081] The fifth step is to verify the clustering results. The silhouette coefficient method is used to evaluate the clustering quality. The silhouette coefficient of each sample is calculated, with a value ranging from -1 to 1. The closer the coefficient is to 1, the higher the similarity between the sample and its cluster, and the better the distinction between the sample and other clusters. If the average silhouette coefficient of all samples is greater than 0.6, the clustering result is considered valid; if it is lower than 0.6, the number of cluster centers is adjusted and the calculation is repeated until the quality requirements are met.

[0082] A passenger flow load tidal correlation matching mechanism is configured. For each cluster, business semantic mapping is performed from three dimensions: operating condition distribution, zoning distribution, and core contribution channels, defining the business connotation of the corresponding coupling mode. First, the sample proportion of the three types of operating conditions within each cluster is counted, and the operating condition with the highest proportion exceeding 60% is determined as the typical applicable operating condition of the mode. Second, the sample proportion of the three types of zoning is counted, and the zoning with the highest proportion exceeding 50% is determined as the typical applicable zoning. Finally, the top two feature channels with the highest contribution in the cluster are extracted as the core driving factors of the mode. The mode naming and definition are completed by combining the three indicators. For example, in the first cluster, the peak operating condition sample proportion is 85%, the station hall zoning sample proportion is 78%, and the top two contributors are the passenger flow channel (48%) and the environmental channel (20%), which is defined as a passenger flow strongly dominant coupling mode, corresponding to the peak passenger flow period in the morning and evening, where load changes are mainly driven by fluctuations in personnel density. In the second cluster, the off-peak operating condition sample proportion is 72%, and the meteorological channel has an average contribution of 22%, ranking second, which is defined as a meteorological and environmental coupling mode, corresponding to stable passenger flow during off-peak periods, where load changes are mainly affected by changes in outdoor temperature and humidity. Ultimately, the six clusters correspond to six typical coupling patterns: passenger flow-driven, temperature and humidity coupled, outdoor weather-influenced, equipment status-driven, spatial attribute-based, and operation period switching. Each pattern corresponds to a specific applicable scenario and driving characteristic. After the single-factor quantitative benchmark and the results of the six coupling patterns are structurally integrated, information on pattern mining and paradigm sedimentation for the allocation of specific features is formed.

[0083] Based on the characteristic gradient attribution accuracy standard, coupling law clustering convergence rules, and energy-saving control strategy library construction specifications, pattern matching and paradigm extraction calculations are performed on the two-stage output results. This completes the quantification of single-channel load impact weights and the mining of passenger flow-load coupling patterns. After the typical operating condition control paradigms are finalized, a systematic encapsulation is performed to form a time-sharing and zone-based load prediction model and energy-saving control strategy library adapted to the tidal characteristics of the metro. Based on the characteristic gradient attribution accuracy standard, coupling law clustering convergence rules, and energy-saving control strategy library construction specifications, the reliability of the attribution and clustering results is first confirmed through double verification. Then, pattern matching and control paradigm extraction are performed. After the typical operating condition control paradigms are finalized, a systematic encapsulation and iterative mechanism configuration are executed to form a time-sharing and zone-based load prediction model and energy-saving control strategy library adapted to the tidal characteristics of the metro.

[0084] First, a double verification process is performed to confirm the validity of the results. The first verification is the feature gradient attribution accuracy verification, which uses a permutation test to verify the validity of the attribution results. The shuffling operation is performed along the sample dimension, keeping the temporal order within a single sample unchanged, and only shuffling the numerical correspondence of the feature between different samples. An error rise threshold of 10% is set; if the average prediction error rises by more than 10% after a single feature permutation, it is considered a high-importance feature. The ranking matching threshold is set to 90%. If the matching degree is insufficient, the sliding window length is expanded from 5 to 7, and the contribution is recalculated until the verification requirements are met. Specifically, the numerical order of all samples in a certain feature dimension is randomly shuffled, while keeping other features unchanged. The results are then re-entered into the baseline model to calculate the prediction results, and the rise in prediction error is statistically analyzed. The larger the rise in error, the higher the importance of the feature to the prediction result. The permutation operation is then performed on the 18-dimensional sub-features in sequence to obtain the feature importance ranking under the permutation method. The ranking is compared with the feature contribution ranking obtained from gradient attribution. Consistency is measured by ranking matching degree, which is the percentage of overlap in the top 9 features. If the matching degree is higher than 90% and the relative deviation of a single feature contribution value is less than 10%, the attribution accuracy is considered satisfactory. If the matching degree is insufficient, the sliding window length is increased and the contribution is recalculated until the accuracy requirement is met. The second step is to verify the convergence of the coupling pattern clustering. In addition to the iteration offset determination, the decrease in the sum of squares within the cluster is additionally verified. If the decrease in the sum of squares of the distances from samples within a cluster to the center is less than 0.1% after two consecutive iterations, the clustering result is considered to have stably converged. Simultaneously, the proportion of samples in each cluster to the total sample size is ensured to be no less than 5% to avoid insufficient pattern generalization caused by extremely small clusters.

[0085] Subsequently, the control paradigm was refined, transforming the coupled modes into implementable energy-saving control rules. Each type of coupled mode corresponds to a dedicated control paradigm, and each paradigm includes five core components: applicable scenario determination conditions, core control parameter mapping rules, comfort constraint boundaries, anomaly fallback strategies, and expected energy-saving indicators, realizing the practical transformation from data patterns to control actions. The control parameter mapping adopts a two-level conversion mechanism of load and air volume. The first step converts the zoned predicted load values ​​into the required air volume, using the following conversion formula: Where Q represents the zoned predicted cooling load. air density, The specific heat of air at constant pressure. The default temperature difference for the supply air is 5℃. The second step is to convert the ratio of the supply air volume to the rated air volume of the fan into the corresponding operating frequency according to the fan performance curve. The opening of the air valve is adjusted linearly according to the supply air volume, and the opening of the surface cooling valve is matched synchronously according to the ratio of the load to the rated cooling load, so as to realize the complete conversion from the predicted load to the equipment execution parameters.

[0086] Taking the passenger flow following control paradigm corresponding to the strong passenger flow-dominated mode as an example, the applicable scenario judgment conditions are: the passenger flow density in a zone exceeds 2 people per square meter for more than 5 minutes, and it is during the peak operating period; the core control parameter mapping rule is to adopt linear following logic, with passenger flow density as input and the operating frequency of the zone air conditioning fan as output. The base frequency is 25 Hz. For every 0.5 people per square meter increase in passenger flow density, the fan frequency increases by 5 Hz accordingly, and the opening of the air valve is adjusted proportionally to ensure that the air supply volume matches the heat dissipation needs of the personnel; the comfort constraint boundary is that the upper limit of the zone dry bulb temperature is 26 degrees Celsius and the upper limit of the relative humidity is 70%. When the temperature reaches the threshold, the frequency reduction action stops, and environmental compliance is prioritized; the abnormal fallback strategy is that when passenger flow data is interrupted, it switches to the fixed parameter operation corresponding to the historical base load of the same period; the expected energy saving indicator is that the average energy saving rate is not less than 18% compared with the fixed frequency control mode.

[0087] The temperature and humidity coordinated control paradigm is applicable to station hall zoning under off-peak conditions. The core rule is: take the enthalpy value of the zone as the control target, when the outdoor enthalpy value is lower than the indoor enthalpy value, give priority to increasing the proportion of fresh air to utilize natural cold source, and adjust the opening of the surface cooling valve linearly with the difference between the indoor enthalpy value and the target enthalpy value. The temperature constraint range is 24-26℃, the humidity constraint range is 40%-70%, and the expected energy saving rate is 15%.

[0088] The outdoor weather-affected type is applicable to entrance and exit zones. The core rule is: dynamically adjust the opening of the fresh air valve according to the outdoor temperature and fresh air load, use 100% fresh air operation in the transition season, supply the minimum fresh air volume in winter and summer, and adjust the opening of the surface cooling valve in conjunction to compensate for the fresh air load. The expected energy saving rate is 20%.

[0089] The equipment status-driven type applies to equipment management areas. The core rule is: based on the equipment heat dissipation load, dynamically adjust the fan speed according to the equipment operating power to maintain the temperature of the equipment area not exceeding 30°C, and reduce the ventilation level to the lowest level during unattended periods, with an expected energy saving rate of 22%.

[0090] The basic spatial attribute type is suitable for off-peak nighttime conditions. The core rule is: set the basic insulation load according to the heat transfer coefficient of the zoned building envelope, maintain the minimum ventilation volume and the lower limit of the temperature, do not actively provide cooling, and only start a small adjustment when the temperature exceeds the threshold. The expected energy saving rate is 30%.

[0091] The shift-time operation mode is suitable for transitional periods before and after morning and evening peak hours. Its core rules are: pre-cooling / pre-heating the load 15 minutes in advance; gradually increasing cooling capacity at a steep rate during peak passenger flow; and gradually reducing output in advance during peak passenger flow to avoid energy waste caused by sudden load changes, with an expected energy saving rate of 12%. Other modes generate corresponding custom paradigms. For example, the temperature and humidity coupling mode corresponds to a temperature and humidity coordinated control paradigm, prioritizing the adjustment of the fresh air ratio and the opening of the cooling valve, simultaneously controlling temperature and humidity indicators to reduce redundant energy consumption during dehumidification.

[0092] After all six coupling modes have completed paradigm extraction, a systematic encapsulation is performed to build an energy-saving control strategy library. The strategy library adopts a two-layer index structure based on operating condition type and partition type. The upper layer is divided into first-level directories based on peak, off-peak, and low-peak periods, while the lower layer is divided into second-level directories based on station hall, platform, and equipment area. Each directory node stores four types of data for the corresponding scenario: control paradigm, contribution benchmark, typical sample characteristics, and historical execution results. A closed-loop iterative update mechanism is also configured, automatically triggering a full-process iterative update every 7 days of accumulated new operating data. The first step adds new operating data to the historical dataset, re-executes feature gradient attribution calculation, and updates the contribution benchmark for each scenario. The second step re-executes coupling mode clustering, updating cluster centers and mode boundaries. The third step compares the deviation between the actual operating energy-saving rate and the expected energy-saving index, adjusting the control parameter mapping coefficients in the paradigm; when the deviation exceeds 5%, the corresponding parameter mapping ratio is corrected. The fourth step supplements new samples to the training set, fine-tunes the weights of the load prediction benchmark model using incremental training, and learns new operating condition characteristics. After iteration, the version number of the strategy library is updated, and historical versions are retained for backtracking. The resulting time-sharing and zone-based load forecasting model and energy-saving control strategy library can automatically match the corresponding coupling mode based on real-time collected passenger flow and environmental data, call the appropriate control paradigm to output control parameters, and achieve dynamic energy-saving management and control adapted to the tidal characteristics of the subway.

[0093] In one implementation, such as Figure 2 As shown, this application also provides a multimodal sensing air conditioning load prediction and time-sharing and zone-based energy-saving control device adapted to the tidal characteristics of subway passenger flow, including: Data acquisition module 201 is used to collect six categories of multimodal sensing data, including passenger flow tidal distribution, station temperature and humidity, air conditioning operating parameters, outdoor weather, spatial location, and operating time period. It performs spatiotemporal regularization to unify minute-level time series benchmarks and station hall zoning spatial benchmarks, and generates a unified spatiotemporal sensing dataset for the subway air conditioning system. The feature construction module 202 is used to build six types of feature channels based on temporal interpolation completion, data normalization alignment, and static spatial attribute mapping. It calculates perception indicators with a single partition and a single time period as the smallest statistical unit, and constructs six feature tensors for passenger flow, environment, equipment, meteorology, space, and operation, and completes temporal window isolation of peak, off-peak, and low-peak tidal working condition samples. Model training module 203 is used to extract passenger flow-load coupling evolution features using a multimodal spatiotemporal prediction dual-branch network, output multi-step air conditioning load prediction sequence through temporal convolution, verify fitting accuracy with spatiotemporal discriminative branch, and solidify the load prediction benchmark model for tidal scenarios based on adversarial iterative convergence of tidal condition samples. The load calculation module 204 is used to build a mechanism for quantifying the zonal time-series deviation, the whole-station spatial weighted double-layer accuracy and energy-saving potential based on the prediction results of the benchmark model. It relies on intermediate features to model the load influence factor weights, and combines spatial weight allocation and load thermal mapping to output the time-series load prediction curves and energy-saving spatial time-series results of the whole station. The strategy sedimentation module 205 is used to extract the load contribution of a single channel in a zone based on the load forecast curve and model feature weights through feature gradient attribution, to mine the coupling law between passenger flow tides and load fluctuations by clustering, to precipitate typical operating condition control paradigms, and to output a time-sharing and zone-based load forecast model and energy-saving control strategy library adapted to the tidal characteristics of the metro.

[0094] The computer-readable storage medium provided in the above embodiments of this application and the multimodal sensing air conditioning load prediction and time-sharing and zone-based energy-saving control method adapted to the tidal characteristics of subway passenger flow provided in the embodiments of this application are based on the same inventive concept and have the same beneficial effects as the methods adopted, run or implemented by the application stored therein.

[0095] All embodiments in this application are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments of the multimodal sensing air conditioning load prediction and time-sharing energy-saving control method, electronic device, electronic device, and readable storage medium for evaluating and adapting to the tidal characteristics of subway passenger flow are basically similar to the embodiments of the multimodal sensing air conditioning load prediction and time-sharing energy-saving control method adapted to the tidal characteristics of subway passenger flow, so the description is relatively simple. Relevant parts can be referred to in the description of the embodiments of the multimodal sensing air conditioning load prediction and time-sharing energy-saving control method adapted to the tidal characteristics of subway passenger flow.

Claims

1. A multimodal sensing air conditioning load prediction and time-sharing and zone-based energy-saving control method adapted to the tidal characteristics of subway passenger flow, characterized in that, include: Collect six categories of multimodal sensing data, including passenger flow tidal distribution, station temperature and humidity, air conditioning operation parameters, outdoor weather, spatial location, and operating hours. Perform a spatiotemporal regularization process to unify the minute-level time series benchmark and the station hall zoning spatial benchmark, and generate a unified spatiotemporal sensing dataset for the subway air conditioning system. Based on temporal interpolation completion, data normalization alignment, and static spatial attribute mapping, six types of feature channels are constructed. Various perception indicators are calculated with a single partition and a single time period as the smallest statistical unit. Six feature tensors for passenger flow, environment, equipment, meteorology, space, and operation are constructed respectively. Simultaneously, temporal window isolation of peak, off-peak, and low-peak tidal working condition samples is completed. A multimodal spatiotemporal prediction dual-branch network is used to extract the passenger flow-load coupling evolution characteristics. A multi-step air conditioning load prediction sequence is generated by temporal convolution, and the prediction fitting accuracy is verified by spatiotemporal discrimination branch. Based on the adversarial iterative convergence of the tidal condition sample after windowing, a benchmark model for load prediction in tidal scenarios is solidified. Based on the output of the load forecasting benchmark model for tidal scenarios, a mechanism for configuring zonal time-series deviation, full-station spatial weighted double-layer forecasting accuracy and energy-saving potential quantification is configured. By modeling the load influence factor weights through the intermediate feature space of the model, combined with spatial weight allocation and load thermal mapping, the time-series and zonal load forecasting curves and energy-saving spatial time-series results of the entire station are output. Based on the prediction curve and model feature weights, the load contribution of a single zone and a single channel is extracted through feature gradient attribution. Clustering algorithms are used to explore the coupling pattern between passenger flow tides and load fluctuations, and typical operating condition control paradigms are precipitated to form a time-sharing and zoned load prediction model and energy-saving control strategy library adapted to the tidal characteristics of the metro.

2. The method as described in claim 1, characterized in that, Based on temporal interpolation completion, data normalization and alignment, and static spatial attribute mapping, six feature channels are constructed. Various perception indicators are calculated using a single partition and single time period as the smallest statistical unit. Six feature tensors are constructed for passenger flow, environment, equipment, meteorology, space, and operation. Simultaneously, temporal windowing isolation of peak, off-peak, and low-peak tidal operating condition samples is completed, including: Combining the requirements for constructing multimodal features of subway air conditioning load and dividing tidal condition samples, three types of feature processing methods are introduced: temporal interpolation completion, data normalization alignment, and static spatial attribute mapping, to build a six-channel standardized feature tensor construction system. Based on the single-zone, single-time period minimum statistical unit accounting standard, the multi-modal perception data of the whole domain is used as the basic input source to connect with the feature mapping rules, and the unified accounting and feature tensor construction of six types of perception indicators, namely passenger flow, environment, equipment, meteorology, space and operation, are completed. Based on the core requirement of accurate isolation of passenger flow tidal conditions, the constructed time-series feature tensor is processed by independent windowing for three types of conditions: peak, off-peak, and trough. The sample set is split according to the time window to achieve standardized isolation of tidal condition samples.

3. The method as described in claim 1, characterized in that, A multimodal spatiotemporal prediction dual-branch network is employed to extract passenger flow-load coupled evolution features. Temporal convolution generates multi-step air conditioning load prediction sequences from the branch outputs, and a spatiotemporal discriminative branch verifies the prediction fitting accuracy. Based on windowed tidal condition samples, adversarial iterative convergence is achieved, solidifying a benchmark model for tidal scenario load prediction, including: To meet the need for high-precision adaptation and prediction of air conditioning load under the scenario of strong tidal fluctuations in subway passenger flow, three core mechanisms are introduced: cross-modal coupling feature decoupling, generation-discrimination dual-track collaborative operation, and tidal condition adversarial knowledge distillation, to build a spatiotemporal dual-domain multimodal load prediction modeling system. Based on the standard for in-depth mining of passenger flow-load coupling evolution characteristics, the six-channel spatiotemporal feature tensor after windowing isolation is used as the basic input source to connect with the dual-branch convolution operation paradigm, and complete the bidirectional linkage adaptation of multi-step load prediction sequence generation and fitting accuracy closed-loop verification. Based on the core requirement of generalization solidification of the tidal scenario load prediction benchmark model, multi-condition cross-adversarial iteration and parameter distillation constraints are applied to the network weights. Parameter solidification is completed only when the fitting accuracy of the entire tidal cycle meets the standard and the loss function converges stably, thus completing the construction of a globally adaptable tidal scenario load prediction benchmark model.

4. The method as described in claim 3, characterized in that, Based on the output of the tidal load forecasting benchmark model, a mechanism is configured to implement zoned time-series deviation, station-wide spatial weighted two-layer forecasting accuracy, and energy-saving potential quantification. Load influence factor weights are modeled through intermediate feature space modeling, combined with spatial weight allocation and load thermal mapping, to output station-wide time-series and zoned load forecasting curves and energy-saving spatial time-series results, including: Combining the multi-step prediction sequence characteristics of the load prediction benchmark model for tidal scenarios with the dual-layer quantitative evaluation standard for subway air conditioning, we introduce the partitioned time-series deviation accounting rule and the spatial dimension weight mapping logic. We decompose and analyze the output results of the benchmark model in both time-series and spatial dimensions to generate temporal and spatial dual-domain load quantification adaptation information. Combining the time-sharing and zone-based energy-saving management requirements of subway stations with the dual-layer quantitative architecture specification of prediction accuracy and energy-saving potential, a hierarchical modeling logic for the allocation of load quantification adaptation information in the spatiotemporal domains is established. The system divides the zone time-series deviation accuracy verification link and the whole-station spatial weighted energy-saving calculation link, configures a load influence factor weight mining mechanism based on the intermediate feature space of the benchmark model, and generates time-series-spatial layered exclusive quantitative feature allocation information. Based on the load forecasting accuracy calibration standard, the spatial quantification rules for energy-saving potential, and the load heat map mapping specification, multi-dimensional weighted fusion calculations are performed on the hierarchical quantification feature allocation information to complete the time series accuracy verification of a single zone and the spatial characterization of the energy-saving potential of the entire station. After the weight allocation and deviation verification meet the standards, the load heat map is mapped, and the time-zone load forecast curves and energy-saving spatial time series results of the entire station are output.

5. The method as described in claim 4, characterized in that, Based on the prediction curve and model feature weights, the load contribution of a single zone and single channel is extracted through feature gradient attribution. Clustering algorithms are used to mine the coupling pattern between passenger flow tides and load fluctuations, and typical operating condition control paradigms are refined. This results in a time-sharing and zone-based load prediction model and energy-saving control strategy library adapted to the tidal characteristics of the metro, including: Combining the time-series characteristics of the time-sharing and zone-based load prediction curves with the spatial distribution specifications of model feature weights, we introduce feature gradient attribution analysis rules and single-zone single-channel contribution decomposition logic to decompose the attribution dimensions of the prediction curves output by the benchmark model and the intermediate layer feature weights, thereby generating zone-based single-channel load contribution adaptation information. By combining the coupling pattern of passenger flow tide and load fluctuation to explore the demand and refine the energy-saving control paradigm, a two-stage progressive modeling logic is used to adapt the information allocation of load contribution of single channel in the zone. The single-factor contribution precise quantification link and the multi-factor coupling mode clustering mining link are divided. The passenger flow-load tide correlation matching mechanism is configured to generate exclusive feature allocation information for pattern mining and paradigm sedimentation. Based on the feature gradient attribution accuracy standard, coupling law clustering convergence rule, and energy-saving control strategy library construction specifications, pattern matching and paradigm extraction calculations are performed on the two-stage output results to complete the quantification of single-channel load impact weight and the mining of passenger flow-load coupling law. After the typical operating condition control paradigm is settled, a systematic encapsulation is completed to form a time-sharing and zone-sharing load prediction model and energy-saving control strategy library adapted to the tidal characteristics of the subway.

6. A multimodal sensing air conditioning load prediction and time-sharing and zone-based energy-saving control device adapted to the tidal characteristics of subway passenger flow, characterized in that, The device is used to execute executable instructions to perform the multimodal sensing air conditioning load prediction and time-sharing and zone-based energy-saving control method adapted to the tidal characteristics of subway passenger flow as described in any one of claims 1 to 5.

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 multimodal sensing air conditioning load prediction and time-sharing and zone-based energy-saving control method adapted to the tidal characteristics of subway passenger flow as described in any one of claims 1 to 5 by executing the executable instructions.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the second processor, it implements the multimodal sensing air conditioning load prediction and time-sharing and zone-based energy-saving control method adapted to the tidal characteristics of subway passenger flow as described in any one of claims 1 to 5.

Images