A Refined Energy Efficiency Management Method for Railway Passenger Stations Based on Train-Passenger Flow Events

Abstract

This invention discloses a refined energy efficiency management method for railway passenger stations based on train-passenger flow event-driven approaches, belonging to the field of building energy consumption control technology. The method includes: collecting multi-source operational data to construct a train-passenger flow event sequence anchored to train physical times and passenger flow response characteristics; using this sequence to drive a regional load prediction model, explicitly separating the baseline operating condition prediction data reflecting a stable state and the event disturbance load data reflecting train impacts through dual-path deduction; dividing the event control time window accordingly, constructing and calling a collaborative optimization model with independent suppression weights for the event disturbance loads, and solving to generate the equipment control setpoint sequence. This invention can quantify and separate the load impacts brought about by train operation, solving the problems of lag and coarseness in traditional control response, and achieving collaborative optimization of railway passenger station comfort assurance and energy consumption peak shaving.

Description

Technical Field

[0001] This invention belongs to the field of building energy consumption control technology, and in particular, it is a method for fine control of railway station energy efficiency based on train-passenger flow event-driven approach. Background Technology

[0002] As typical large-scale public transportation buildings, railway passenger stations are characterized by their large ceilings, high population density, and high mobility. The energy consumption of their HVAC and lighting systems accounts for a significant portion of their operating costs. Reducing the energy consumption of railway passenger stations can not only reduce carbon emissions but also enable precise control of station energy efficiency, contributing to the construction of green railways and promoting the sustainable development of the transportation industry.

[0003] Currently, energy consumption control in railway passenger stations largely follows the management model of large public buildings, primarily relying on timed start-stop based on schedules or conventional PID (proportional-integral-derivative) feedback control based on indoor temperature and carbon dioxide concentration. While some systems incorporate load forecasting technology, they employ general regression models based on historical data, treating the passenger station as a constant or slowly varying load object. Existing control strategies passively adjust equipment frequency or valve opening based on real-time environmental parameters, or simply pre-adjust cooling capacity based on coarse-grained passenger flow forecasts, lacking integration with the specific train operation logic of railways.

[0004] Existing technologies suffer from several drawbacks, including a disconnect between transportation operations and energy consumption control logic, and an inability to distinguish between steady-state loads and impact loads. Therefore, further research and innovation are needed to address these issues in existing technologies. Summary of the Invention

[0005] Purpose of the invention: In view of the above-mentioned problems of the prior art, this application provides a method for fine control of railway station energy efficiency based on train-passenger flow event-driven approach.

[0006] Technical solution: According to one aspect of this application, a method for refined energy efficiency management of railway passenger stations based on train-passenger flow events includes:

[0007] Based on train timetable data and real-time passenger flow monitoring data, a spatiotemporally aligned train passenger flow event sequence is constructed, which anchors the physical arrival and departure times of trains and passenger flow response characteristics as discrete event nodes.

[0008] Using a train passenger flow event sequence-driven regional load forecasting model, scenario-based multi-region load forecasting data is generated. Among them, a dual-path extrapolation is used to explicitly separate the baseline operating condition forecasting data reflecting the stable operating state and the event disturbance load data reflecting the impact of train events.

[0009] Based on the train passenger flow event sequence, the event control time window is divided to cover the events before and after the events. Within the event control time window, the collaborative optimization model with independent suppression weights for the event disturbance load data is invoked.

[0010] Solve the collaborative optimization model to generate an event-driven control setpoint sequence for adjusting the operating status of equipment within the station.

[0011] Beneficial Effects: This invention introduces train passenger flow event sequencing and scenario-based labeling, dual-path extrapolation, and load decomposition techniques to address the disconnect between transportation operations and energy consumption control logic, as well as the inability to distinguish between steady-state loads and impact loads. It achieves synergistic optimization of railway station comfort assurance and energy consumption peak shaving. The related technical effects will be described in detail below with reference to specific embodiments. Attached Figure Description

[0012] Figure 1 A flowchart illustrating a method for refined energy efficiency management of railway passenger stations based on train-passenger flow events, provided as an embodiment of this application.

[0013] Figure 2 The flowchart provided in this application illustrates the explicit separation of baseline operating condition prediction data reflecting a stable operating state and event disturbance load data reflecting train event impacts using a dual-path extrapolation method in an embodiment of this application.

[0014] Figure 3 A flowchart illustrating the training of the regional load prediction model provided in this application embodiment.

[0015] Figure 4 The flowchart illustrates how real-time passenger flow monitoring data provided in this application is generated by multi-source weighted fusion of ticketing passenger flow data, entry and exit gate data, and video passenger flow statistics.

[0016] Figure 5 A flowchart illustrating the calculation of confidence scores for characterizing the reliability of data quality, provided in an embodiment of this application. Detailed Implementation

[0017] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0018] It should be noted that the terms "first," "second," etc., in the specification and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "including" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0019] To address the aforementioned issues, the applicant conducted in-depth searches and analyses, and discovered:

[0020] Specifically, traditional methods ignore the strong deterministic correlation between railway station load fluctuations and train timetables, treating the pulse-like passenger flow shocks caused by train arrivals and departures as random disturbances, resulting in significant lag in the control system.

[0021] Furthermore, existing forecasting models typically output general total load forecasts, failing to separate the baseline load determined by environmental conditions from the instantaneous disturbance load caused by train events. This makes it difficult for control strategies to design independent suppression weights for short-term, high-intensity passenger flow shocks, leading to an oversupply of global cooling capacity in an attempt to smooth out local fluctuations or resulting in uncontrolled comfort during peak periods. Consequently, energy efficiency control cannot be achieved while ensuring the instantaneous passenger experience.

[0022] To solve these problems, combined with Figures 1 to 5 The present invention will be specifically described through the following embodiments.

[0023] In some embodiments, a method for refined energy efficiency management of railway passenger stations based on train-passenger flow events is provided. In particular, its overall operation process in railway passenger stations solves the problem that traditional energy management systems cannot predict and control train timetables and sudden passenger flows. This provides an energy consumption perception and control framework with train events as the core anchor point. Specifically, this application includes the following steps:

[0024] Step S101: Based on train timetable data and real-time passenger flow monitoring data, construct a spatiotemporally aligned train passenger flow event sequence. The train passenger flow event sequence anchors the physical arrival and departure times of trains and passenger flow response characteristics as discrete event nodes.

[0025] The train timetable data mainly comes from the electronic timetable or train timetable interface of the railway passenger dispatching system. It contains basic information such as the train number, planned arrival time, planned departure time, stopping track, and train type for each train.

[0026] The real-time passenger flow monitoring data comes from various sensing devices deployed inside the station, including gate counting data at ticket gates, infrared counting data at entrance and exit channels, video analysis data from high-definition cameras in the waiting hall and platform areas, and real-time ticketing data from the ticketing system.

[0027] Furthermore, constructing a spatiotemporally aligned train passenger flow event sequence refers to segmenting and correlating continuously changing real-time passenger flow data along a timeline, using key time points in the train timetable as a benchmark. For example, the system might extract all passenger flow monitoring data from 30 minutes before a train's scheduled arrival to 20 minutes after arrival and package them into event data units.

[0028] The train passenger flow event sequence is composed of a series of event data units arranged chronologically. Each event node includes not only the physical operating status of the train but also corresponding response characteristics such as passenger flow aggregation and dissipation. This mechanism can establish a direct causal link between isolated passenger flow fluctuations and specific train operation behaviors, providing a data foundation for subsequent scenario-based predictions.

[0029] Optionally, the system can also access meteorological monitoring data, such as outdoor temperature, humidity, and light intensity, and integrate environmental parameters as additional attributes into the event nodes of the train passenger flow event sequence, so that each event node can fully reflect the operating environment status at that time.

[0030] Step S102: Use the train passenger flow event sequence to drive the regional load prediction model and generate scenario-based multi-region load prediction data. Among them, the baseline operating condition prediction data reflecting the stable operating state and the event disturbance load data reflecting the impact of train events are explicitly separated by dual-path inference.

[0031] In other words, the train passenger flow event sequence is input into the regional load prediction model to generate scenario-based multi-region load prediction results. Among them, a dual-path extrapolation mechanism is executed to split it into two parts: one is the baseline operating condition prediction data that reflects the stable operation status, and the other is the event disturbance load data that reflects the impact of train events.

[0032] In this embodiment, the regional load prediction model is a machine learning model capable of processing time-series data and categorical features; for example, it can employ a long short-term memory network or a transformer model architecture. Furthermore, the model is driven by train passenger flow event sequences, using scene labels, time features, and passenger flow features from the sequences as input variables.

[0033] Based on this, the system executes dual-path inference logic to quantify the impact of train events on energy consumption.

[0034] One approach is the real-world projection path, which involves inputting a sequence containing real train event information and predicted passenger flow characteristics into the model to obtain prediction results that include the impact of the events.

[0035] The second is the baseline extrapolation path, which involves artificially masking or removing train event information from the input data. For example, the current time is marked as an idle state with no trains stopping, and stable background passenger flow data is input. The same model is then used to predict the baseline operating condition prediction data assuming no events occur.

[0036] The baseline operating condition prediction data represents the energy consumption and thermal / humidity load resulting solely from basic environmental maintenance and routine operations. Event disturbance load data is obtained by calculating the difference between the actual simulation results and the baseline simulation results. This difference reflects the additional thermal / humidity load and energy consumption increases caused by passenger influx and equipment operation due to train arrivals and departures. Based on this, the system can identify load spikes that need to be suppressed.

[0037] Step S103: Divide the event control time window before and after the event according to the train passenger flow event sequence, and build a collaborative optimization model with independent suppression weights for the event disturbance load data within the event control time window.

[0038] Specifically, an event control time window refers to a specific control period set for each train passenger flow event, and its length is usually longer than the train's stop time at the station. For example, for a high-speed train that stops for 15 minutes, the system may set an event control time window that starts 15 minutes before arrival and ends 15 minutes after departure, covering the entire process of platform pre-cooling, passenger evacuation, and environmental restoration.

[0039] Within this time window, the system constructs a collaborative optimization model designed to minimize operational energy consumption while ensuring passenger comfort. This model assigns independent suppression weights to the energy consumption components corresponding to event disturbance load data within its objective function.

[0040] In other words, the model assigns higher priority to the additional energy consumption caused by reduction events. For example, if an event disrupts the load and causes an increase in energy consumption, the optimization model will adjust the weight parameters so that the control strategy tends to prioritize mitigating that incremental increase rather than reducing overall energy consumption on an average basis. This mechanism allows the control strategy to respond to short-term high load shocks, avoiding environmental runaway or energy waste caused by drastic load fluctuations.

[0041] Step S104: Solve the collaborative optimization model to generate an event-driven control setpoint sequence for adjusting the operating status of equipment within the station.

[0042] In this embodiment, the collaborative optimization model is typically solved using numerical optimization algorithms, such as sequential quadratic programming or genetic algorithms. During the solution process, a set of control variables can be found that minimizes the constructed objective function value. These control variables specifically include the air conditioning unit's supply air temperature setpoint, the opening percentage of the fresh air valve, and the variable frequency of the water pump. The output obtained is the event-driven control setpoint sequence, which is a dataset that changes over time and specifies the specific operating parameters that each device should execute at each discrete time step within the event control time window.

[0043] For example, as a train approaches the station, the sequence might instruct the air conditioning supply temperature in the platform area to decrease by 2°C while increasing the fresh air volume to prepare for the heat load impact from the upcoming surge in passenger traffic. This setpoint sequence is converted into standard industrial control commands and sent to on-site controllers and actuators via the building automation system. These commands directly adjust the physical operating status of equipment within the station, achieving refined closed-loop management of railway station energy efficiency.

[0044] Other embodiments describe optional technical solutions for confidence fusion and missing data compensation methods for multi-source passenger flow data. They illustrate how to handle multi-source heterogeneous passenger flow data in the complex environment of railway stations, addressing issues of inconsistent data quality and local missing data, and also providing input data for upper-level prediction models. This embodiment specifically includes:

[0045] Step S201: Map ticketing passenger flow data, entry and exit gate data, and video passenger flow statistics onto a unified time axis to generate multi-source time-slice passenger flow data.

[0046] Specifically, since ticketing systems, turnstile systems, and video analytics systems typically have different data update frequencies and timestamp formats, time alignment is essential. The system sets a uniform time slice length, such as 5 minutes. For each type of raw data, the system merges its timestamps into the corresponding 5-minute time slice based on proximity.

[0047] The ticketing passenger flow data includes the number of pre-sold tickets and the estimated ticket checking time, which the system maps to the theoretical passenger flow distribution for each future time slot. The entry and exit gate data records the actual time points when passengers pass through the gates, and the system counts the number of people passing through each time slot.

[0048] The video passenger flow statistics are generated by using image recognition algorithms to output real-time estimates of the number of people in the area, with the system taking the average or maximum value within each time slice. After mapping processing, the system generates multi-source time-slice passenger flow data. This data structure contains passenger flow observations from three different sources simultaneously in each time slice, providing a data foundation for subsequent fusion calculations.

[0049] Step S202: For each time slice and each data source, calculate the confidence score that characterizes the reliability of data quality.

[0050] Specifically, the passenger flow growth rate of the current time slot relative to the previous time slot is calculated, as well as the ticketing deviation rate of the current time slot relative to the ticketed passenger flow data.

[0051] The passenger flow growth rate and ticketing deviation rate indicators are compared with preset anomaly detection thresholds.

[0052] When the passenger flow growth rate or ticketing deviation rate exceeds the anomaly judgment threshold, the confidence score of the data source in the current time slice is reduced by a preset penalty factor until it is marked as an anomaly.

[0053] In this step, the system quantitatively evaluates the quality of each data source. Specifically, for each data source i, the system calculates its passenger flow growth rate index r in the current time slice t. _i It is the ratio of the difference between the current value and the previous value to the value at the previous time.

[0054] Furthermore, the system uses ticketing data as a benchmark to calculate the ticketing deviation rate, which is the ratio of the difference between the current data source value and the theoretical ticketing value to the theoretical ticketing value. Next, the system introduces preset anomaly detection thresholds, such as a growth rate threshold R. _max And the deviation rate threshold. If the absolute value of the calculated index exceeds the threshold, it indicates that there is an unreasonable abrupt change or deviation in the data.

[0055] Furthermore, the individual confidence score can be calculated using the following formula:

[0056] s _i =max(0, 1-|r _i | / R _max );

[0057] Above, s _i For the individual confidence score of the i-th data source, max(.) is the maximum value function to ensure that the score is not less than 0. _i |r represents the growth rate of passenger flow in the current time slot relative to the previous time slot. _i | represents the absolute value of the growth rate indicator, R _max This is the preset threshold for judging abnormal growth rates. The formula indicates that when the growth rate indicator r... _i The closer the absolute value is to the abnormal threshold R _max The lower the confidence score, the lower it becomes; once the threshold is exceeded, the score is reduced to zero.

[0058] Similarly, another scoring component can be calculated based on the deviation rate index, and the weighted average of the two can be taken as the final confidence score of the data source. Therefore, the system can automatically identify data anomalies caused by equipment failure, network latency, or sensor obstruction, and accordingly reduce their confidence level.

[0059] Step S203: Calculate the dynamic fusion weight of each data source in the time slice based on the confidence score, wherein the data source with the lower confidence score is assigned a lower dynamic fusion weight.

[0060] Specifically, the system allocates the contribution ratio of each data source to the final fusion result based on the confidence score of each data source. Given N data sources, the confidence score of the i-th data source is s. _i Then its unnormalized weight w _raw_i It can be set to s _i To ensure that the sum of all weights is 1, the system performs a normalization operation and calculates the dynamic fusion weights, which can be described by the following formula:

[0061] w _i =w _raw_i / (Σw _raw_j );

[0062] Among them, w _i w is the normalized dynamic fusion weight of the i-th data source. _raw_i The unnormalized weight of the i-th data source is taken as the confidence score s of that source in this embodiment. _i , Σw _raw_j This is the sum of the unnormalized weights of all N data sources; in other words, the summation is performed on all N data sources.

[0063] Based on this, higher-quality, higher-rated data sources will receive greater weight and dominate the final passenger flow estimation results; while lower-quality, lower-rated data sources will be assigned less weight, and their impact on the results will be suppressed. This mechanism can be dynamically adjusted, making the fusion algorithm highly robust and able to cope with fluctuations and interference from different data sources.

[0064] Step S204: Use dynamic fusion weights to perform weighted summation on multi-source time-slice passenger flow data to generate fused passenger flow estimation data, which constitutes real-time passenger flow monitoring data.

[0065] In this embodiment, for each time slice, the system multiplies the observations from each data source with their corresponding dynamic fusion weights, and sums all the products to obtain the fused passenger flow estimation data for that time slice. This process performs a weighted average of the multi-source observations, which can smooth out random noise and comprehensively utilize the information advantages of each data source. The sequence formed by concatenating the fusion results of each consecutive time slice constitutes the real-time passenger flow monitoring data, which can be used as input to the downstream event sequence construction module and load prediction model.

[0066] In some optional embodiments, after generating real-time passenger flow monitoring data, the system further extracts multi-scale features. These include short-scale features and long-scale features. Short-scale features may include passenger flow values ​​for the current 5-minute time slice and the first-order difference of the passenger flow, used to reflect instantaneous changes. Long-scale features may include the cumulative passenger flow over the past 30 minutes and the linear regression slope of the passenger flow curve, used to characterize long-term growth trends and overall pressure.

[0067] According to one aspect of this application, when anomalies occur in multiple source data, historical template compensation can be performed, specifically as follows:

[0068] When the confidence scores of all data sources are lower than the preset confidence threshold within a certain period of time, a set of historical passenger flow curve templates are retrieved from the historical database based on the current train type and time period type.

[0069] Extract the fused passenger flow estimation data generated before and after the time period as reference data segments, and calculate the morphological similarity between the reference data segments and each historical passenger flow curve template;

[0070] The historical passenger flow curve template with the highest morphological similarity is selected as the best matching template. The best matching template is scaled according to the average amplitude of the reference data segment to generate a template to compensate for passenger flow estimation data and fill in the data gaps in the time period.

[0071] Correspondingly, when all sensors malfunction or the data quality is poor, the system cannot obtain satisfactory results through weighted fusion. In this case, the system turns to historical experience data. Based on the train type (e.g., high-speed rail, regular train) and time period (e.g., weekday morning rush hour, holiday off-peak), the system can search for a corresponding standard passenger flow curve template in a pre-built historical database. This template reflects the passenger flow patterns under similar operating conditions.

[0072] Next, to determine which template best reflects the current situation, the system extracts a segment of known merged passenger flow data before and after the missing data period as a reference data segment. The system calculates the morphological similarity between the reference data segment and each candidate template within the corresponding time interval, using methods such as Euclidean distance or Pearson correlation coefficient. The template with the highest morphological similarity, meaning its fluctuation trend most closely resembles the current passenger flow, is selected.

[0073] Based on this, after selecting the best matching template, the system calculates the ratio of the average value of the reference data segment to the average value of the corresponding segment of the template. This ratio is used as a scaling factor to scale the overall amplitude of the values ​​of the best matching template within the missing time period. The scaled template data is then used as template-compensated passenger flow estimation data and filled into the gaps in the real-time passenger flow monitoring sequence, ensuring the continuity and integrity of the data stream.

[0074] Based on this, the embodiment describes a missing data imputation mechanism in extreme cases, which adopts multi-source confidence fusion and historical template compensation to solve the implicit problem of unstable input data sources and ensure the robustness of the algorithm in the case of missing data in the field.

[0075] In other embodiments, optional implementations of methods for constructing and contextualizing train passenger flow event sequences are provided. This illustrates how to extract structured event features that can be understood by computer models from raw train operation data and passenger flow monitoring data. This embodiment solves the problem of logically binding physical train operation times with actual passenger flow behavior, and provides a tagging system with business semantics through contextualized annotation.

[0076] For example, this embodiment can be performed in the following manner:

[0077] Step S301: Extract the train arrival time and train departure time from the train timetable data as anchor points for key train events.

[0078] Specifically, the system reads digitized train timetables or real-time dispatch instructions to identify the operational time nodes of each train at the current station. The train arrival time represents the moment the train comes to a complete stop and its doors open, while the train departure time represents the moment the train's doors close and it starts moving away. These two times constitute the physical trigger points for dynamic changes in passenger flow, i.e., the anchor points of key train events. These anchor points are definite and discrete points on the timeline. The system will use these as centers to extend forward and backward along the timeline to search for and correlate relevant passenger flow data.

[0079] Step S302: Based on the preset train passenger flow linkage parameters, define a time range before and after each key train event anchor point, and extract passenger flow characteristic segments within that time range from the real-time passenger flow monitoring data.

[0080] In other words, by clearly defining the preset train passenger flow linkage parameters, determining the time range before and after each key train event, and extracting relevant data segments of passenger flow characteristics within that time range from the real-time collected passenger flow monitoring data.

[0081] In this step, train passenger flow linkage parameters can be used to define the event impact boundary. These parameters specifically include buffer time settings for different train types. For example, for high-speed trains, due to their large passenger capacity and rapid boarding and alighting, the system can set the pre-arrival buffer time to 15 minutes to cover the period of passenger gathering and platform pre-cooling; and the post-arrival buffer time to 20 minutes to cover the periods of passenger disembarkation, exiting the station, and platform clearing. For regular long-distance trains, considering that passengers carry more luggage and move slower, the buffer time can be extended accordingly.

[0082] Based on this, the system reads the corresponding parameter values ​​according to the train type at the current anchor point, and extracts a data subset of a specific length from the continuous real-time passenger flow monitoring data stream, namely the passenger flow characteristic segment. This segment completely records the entire process of passenger flow during this train operation, from the initial gathering to the final dissipation.

[0083] Step S303: Calculate the passenger flow intensity index and growth rate index of the passenger flow characteristic segment, and determine the passenger flow response level label of the event node based on the passenger flow intensity index and growth rate index.

[0084] Specifically, the system performs statistical analysis on the captured passenger flow characteristic segments. Passenger flow intensity indicators, which can be represented by the peak passenger flow density or cumulative passenger volume within that time period, are used to measure the absolute scale of passenger flow. Growth rate indicators, which can be represented by the slope of the passenger flow curve at its rising edge or the increment per unit time, are used to measure the rapidity of the passenger flow influx.

[0085] The system pre-defines a series of threshold standards, such as low passenger flow threshold, medium passenger flow threshold, and high passenger flow threshold. When the passenger flow intensity index exceeds the high passenger flow threshold and the growth rate index exceeds the preset rate threshold, the system determines the passenger flow response level label for that event node as either a rapid influx level or a high response level; when the passenger flow intensity index is below the low passenger flow threshold, it is determined as a smooth flow level or a low response level. Through this classification, the system abstracts waveform characteristics into level labels, facilitating subsequent rule matching.

[0086] Step S304: Associate and store the key train event anchor points, passenger flow characteristic segments and passenger flow response level labels, and arrange them in chronological order to form a train passenger flow event sequence.

[0087] In this step, the system establishes structured data objects, packaging the train's physical attributes, such as train number and stopping track, with passenger flow characteristics, such as characteristic segment data and response level labels. These data objects are indexed and stored in the database according to the chronological order of train arrivals and departures, forming a train passenger flow event sequence that incorporates both railway scheduling logic and passenger flow fluctuation logic. This provides input for subsequent prediction and control algorithms.

[0088] Step S304: Use multi-dimensional rules to perform contextualized annotation of train passenger flow event sequences, specifically including:

[0089] Optionally, a mapping relationship can be established between each event node in the train passenger flow event sequence and the set of functional area identifiers within the station space.

[0090] Alternatively, a mapping relationship is established between each event node and the set of functional area identifiers; where the event nodes are taken from the train passenger flow event sequence, and the set of functional area identifiers corresponds to each functional zone of the station space.

[0091] Specifically, different trains stop at different tracks, corresponding to different ticket gates, waiting areas, and exit passages. The system maintains a spatial mapping table that automatically identifies the spatial range affected by a train event based on track information and ticket gate allocation information in the train timetable. For example, when train G1234 stops at platform 1, the system identifies its associated functional area set as the first waiting room, ticket gate A1, platform 1, and the north exit passage. This mapping allows subsequent scene labels to be accurately attached to specific physical spaces.

[0092] Optionally, for each event node, retrieve scene tag rules that include event type conditions, passenger flow response level conditions, and time density conditions;

[0093] Optionally, when the attributes of an event node meet the scene tag rules, a corresponding scene tag is generated and attached to the train passenger flow event sequence to identify the business operation status of the functional area at that moment.

[0094] In this embodiment, the scenario labeling rules are a digital expression of the experience of business experts. The system stores a rule base, and each rule defines the triggering conditions for a specific scenario. For example, the triggering rule logic for the delayed and stranded scenario can be set as follows: if the event type is train arrival, and the difference between the actual arrival time and the planned arrival time is greater than 30 minutes, and the real-time average passenger flow density in the area is greater than the medium passenger flow threshold, then the triggering condition is met.

[0095] For example, the triggering rule logic for peak-hour ticket checking scenarios can be set as follows: if the current time is during the morning rush hour, and the number of departure events associated with this ticket gate is greater than or equal to 2 within the next 15 minutes, and the passenger flow response level is high, then the triggering condition is met. When the attribute data of an event node matches all the conditions of a certain rule, the system generates the corresponding scenario label.

[0096] Tags such as peak ticket checking, delays and congestion, and concentrated boarding at the platform are attached to the event sequence, so that the subsequent load forecasting model can clearly perceive the specific business situation that the area is in at the current moment and call the corresponding forecasting logic.

[0097] It should be understood that the embodiment employs methods for constructing train passenger flow event sequences and contextualized annotation. The hard anchor points (arrival and departure times) in the train timetable are spatiotemporally aligned with the passenger flow characteristics after multi-source fusion, and event time and location coding technology is introduced, enabling the control system to transform from a passive response to an active adjustment based on event prediction, capturing the start and end times of passenger flow pulses.

[0098] In other embodiments, optional implementations of a scenario-event jointly driven prediction model construction and training method are provided. This embodiment introduces feature encoding technology and a hierarchical modeling strategy, which solves the problem that traditional models have difficulty handling discrete events and nonlinear mutations, and improves the model's sensitivity to load fluctuations driven by train timetables.

[0099] Step S401: Map the scene labels in the train passenger flow event sequence to numerical scene encoding vectors.

[0100] Specifically, scene labels are typically discrete symbols in text form, which cannot be directly processed by mathematical models. This embodiment uses embedding encoding technology to transform discrete labels into continuous numerical vectors. The system predefines an embedding matrix, where the number of rows corresponds to the total number of categories of scene labels, and the number of columns corresponds to the dimension of the embedding vector, such as 16-dimensional or 32-dimensional.

[0101] For each scene label, such as peak ticket checking scene, the system finds the corresponding row in the matrix and uses it as the scene encoding vector for that scene. Compared to one-hot encoding, this method can better capture the potential semantic relationships between different scenes. For example, the distance between a large passenger flow entry scene and a peak ticket checking scene in the feature space may be closer than the distance between them and an off-peak waiting scene.

[0102] Step S402: Calculate the time difference between the current predicted time and the nearest key train event in the train passenger flow event sequence, and use a periodic function to transform the time difference to generate an event time location code.

[0103] In this step, to enable the model to perceive which stage of the event cycle the current moment falls within, the system calculates the time difference Δt between the current moment and the most recent train event. To capture the periodicity of time and the relative positional relationship, the system uses a periodic function to encode Δt, generating an event time position code. Specifically, this code can be derived from a sinusoidal component p. _1 and cosine component p _2 composition.

[0104] For example, p _1 =sin(2×π×Δt / T), p _2 =cos(2×π×Δt / T).

[0105] Where, p _1 The sinusoidal component that encodes the time and location of the event, p _2 The cosine component is used to encode the time and location of the event, where π is the mathematical constant pi, Δt is the time difference between the current predicted time and the most recent critical train event, and T is a set time scale parameter, such as 120 minutes.

[0106] Therefore, the model can distinguish the differences in time structure between 10 minutes before a train arrives at the station and 10 minutes after the train departs, and more accurately predict the rising or falling trend of the load.

[0107] Step S403: Concatenate the scene coding vector, event time and location coding, and passenger flow response features to construct scene-event joint coding features, and feed these (scene-event joint coding features) into the regional load prediction model as input vectors.

[0108] Specifically, the system concatenates the generated scene encoding vector, event time and location encoding vector, and numerical passenger flow response features extracted from passenger flow feature segments along the feature dimension to form a long vector, namely the scene-event joint encoding feature. This feature integrates qualitative scene semantics, quantitative relative time and location, and specific passenger flow intensity information, serving as the input layer data for the regional load prediction model. The model learns the nonlinear mapping relationship between these features and historical energy consumption and environmental parameters to achieve accurate predictions of future loads.

[0109] Among them, the regional load forecasting model includes: a subset of scenario-specific models built for high-disturbance load scenarios;

[0110] A subset of regionally general models built for everyday off-peak scenarios;

[0111] Furthermore, the training of the regional load forecasting model includes:

[0112] A sample dataset for scenario-driven multi-region load modeling, containing scenario labels, was constructed based on historical operational data.

[0113] The scenario-driven multi-region load modeling sample dataset was used to iteratively train the scenario-specific model subset and the region-general model subset, respectively.

[0114] The regional load forecasting model is not a single, unified model, but rather a set of sub-models designed for different operating conditions. Specifically, it includes a subset of scenario-specific models built for high-disturbance load scenarios, such as models specifically designed to predict extreme situations like the Spring Festival travel rush, summer travel rush, or large-scale delays; and a subset of general-purpose regional models built for daily off-peak scenarios, used to predict load during regular operating periods.

[0115] Furthermore, the system cleans and labels historical data to construct a sample dataset with scene labels. The system then splits the samples according to the scene labels and trains corresponding sub-models for each. This strategy avoids the mediocrity tendency of a single model when dealing with imbalanced data, enabling the model to maintain high prediction accuracy even in extreme scenarios.

[0116] Furthermore, during the training process, a scenario-weighted loss function is applied. Compared with the samples of daily off-peak scenarios, the samples corresponding to large passenger flow events in the scenario-driven multi-region load modeling sample dataset are given higher loss weights, thereby improving the regional load prediction model's sensitivity to event-induced load disturbances.

[0117] During training, the system can construct a weighted loss function. For each training sample, the system checks its corresponding scene label. If the sample belongs to a high-traffic event scenario, such as peak ticket checking or periods of high train arrival and departure, the system will assign it a larger weight coefficient, such as 2.0 or 5.0; while for ordinary off-peak scenario samples, the weight coefficient is 1.0. The loss function calculates the weighted sum of squares of the deviations between the predicted and actual values.

[0118] In this way, the optimization algorithm will pay more attention to reducing the prediction error of high-weight samples during the iteration process, enabling the model to learn and adapt to short-term and drastic load fluctuation characteristics, and improving the model's sensitivity to the load data of event disturbances.

[0119] In some optional implementations, the method also supports online incremental updates of the model. After the system has been running for a period of time and accumulated new measured data, or after obtaining closed-loop feedback, the system can select samples with large prediction deviations from the latest historical data and fine-tune the existing scenario-specific model, instead of retraining from scratch. This mechanism allows the model to adapt more quickly to seasonal changes or long-term evolution of passenger flow structures.

[0120] According to one aspect of this application, an alternative implementation of the dual-path prediction and event-driven disturbance load decomposition method is described. It illustrates how to quantify the instantaneous load impact caused by train passenger flow events by using the same prediction model architecture and differentiating input data. This embodiment solves the technical problem that traditional energy consumption prediction methods cannot distinguish between basic environmental load and sudden passenger load, providing a quantitative basis for subsequent control.

[0121] Step S501: At each prediction time step, extract the current scene label and event features from the train passenger flow event sequence, input them into the regional load prediction model, and generate event scene prediction result data containing the impact of train events.

[0122] Specifically, the system determines the current predicted time point and retrieves the complete context information associated with that moment from the train passenger flow event sequence. This includes the current scenario label, such as a peak ticket checking scenario, and specific event characteristics, such as 5 minutes having passed since the last train arrived and the current passenger flow response level being high. The system then inputs the feature vector reflecting the current business status into the trained regional load prediction model.

[0123] Based on learned historical patterns, the model outputs predicted values ​​for environmental parameters and energy load under a specific scenario. For example, the model might predict that within the next 10 minutes, the temperature in the waiting area will rise to 26°C, and the power load of the air conditioning system will reach 150kW. This prediction result is called the event scenario prediction result data, reflecting the system state under the combined influence of train events and environmental factors.

[0124] Step S502: Simultaneously replace the event states in the train passenger flow event sequence with preset stable baseline states, input the regional load prediction model, and generate baseline operating condition prediction data that shields the impact of train events.

[0125] In this step, while keeping background features such as outdoor meteorological parameters and basic equipment status unchanged, the system manually modifies the event-related features in the input vector. Specifically, the system forcibly changes the scene label to an off-peak idle scene, modifies the passenger flow intensity feature to a preset base value, such as zero passenger flow or passenger flow maintaining the minimum number of on-duty personnel, and sets the event time and location codes to zero or invalid values.

[0126] The cleaned or shielded feature vectors are input into the same regional load prediction model. Since the model is unaware of train events, its output predictions will only reflect the load generated by heat transfer from the building envelope, basic lighting, and basal metabolic rate. For example, at the same moment mentioned above, the model might predict a temperature of only 24°C and an electrical load of only 100kW under baseline conditions. This result is the baseline load prediction data.

[0127] Step S203: Calculate the difference between the event scenario prediction result data and the baseline operating condition prediction data, and decompose the difference into environmental load disturbance component and energy consumption load disturbance component, which together constitute the event disturbance load data.

[0128] Alternatively, the event disturbance load data is calculated, the difference between the predicted data and the baseline prediction is calculated and decomposed into a first load disturbance component and a second load disturbance component, which are then combined to form the load disturbance load data.

[0129] Among them, the predicted data refers to the prediction results of the event scenario, the baseline prediction refers to the prediction data of the baseline operating condition, and the first load disturbance component and the second load disturbance component correspond to the environmental load disturbance component and the energy consumption load disturbance component, respectively.

[0130] In this step, the system performs differential calculations on the outputs of the two paths mentioned above. By subtracting the baseline operating condition prediction data from the event scenario prediction data, the system separates the incremental portion caused solely by the train passenger flow event. Specifically, the system calculates the difference in environmental parameters, for example, 26℃-24℃=2℃, i.e., temperature rise, which is defined as the environmental load disturbance component, characterizing the degree of impact of the event on environmental stability. Simultaneously, the system calculates the difference in energy consumption parameters, for example, 150kW-100kW=50kW, i.e., power increment, which is defined as the energy consumption load disturbance component, characterizing the additional energy cost required to cope with the event.

[0131] Based on this, the two components together constitute the event disturbance load data. This method allows the control system to know what proportion of the current energy consumption surge is necessary to cope with passenger flow, and what proportion can be reduced through optimization strategies, providing a targeted objective for subsequent accurate control.

[0132] It should be understood that in this embodiment, the event disturbance load data is obtained by inputting real event characteristics and virtual stationary characteristics respectively during the prediction stage and performing a difference calculation.

[0133] According to another aspect of this application, a specific technical solution for a collaborative optimization and control execution method for event disturbance suppression is described, particularly based on the predicted disturbance load, constructing and solving a multi-objective optimization problem, and converting the optimization results into instructions executable by field equipment. This embodiment also provides a two-layer collaborative optimization architecture that can suppress energy consumption peaks caused by train events while maintaining passenger comfort within a red line constraint.

[0134] Furthermore, the method in this embodiment specifically includes:

[0135] Step S601: Divide the event control time window into a finite number of discrete time step sequences.

[0136] Specifically, the system discretizes the event control time window. For example, for a 40-minute window, the system can set the time step to 2 minutes, dividing it into 20 consecutive discrete time steps. Each time step has a unique index identifier, representing the specific control decision moment. Discretization ensures that the control actions have sufficient time resolution to cope with rapid changes in passenger flow, while keeping the size of the optimization problem within the solver's acceptable range.

[0137] Step S602: For each discrete time step and each functional region, define a control decision vector that includes the supply air temperature setpoint and the supply air volume setpoint. This clarifies the decision variables for the optimization problem.

[0138] For each discrete time step and each controlled functional area, the system defines a control decision vector. Specific elements of this vector may include the air conditioning unit's supply air temperature setpoint, the fan's supply air volume setpoint, and the fresh air valve opening ratio, etc. These variables are the solutions that the subsequent optimization algorithm needs to find, determining the operating state of the equipment. For example, in the 5th time step, the decision vector for waiting area A might include a supply air temperature of 18℃, a supply air volume of 80%, and a fresh air ratio of 30%.

[0139] Step S603: Using scenario-based multi-regional load forecast data, construct regional comfort constraints and equipment operation boundary constraints for discrete time step sequences.

[0140] Specifically, the system transforms the predicted data into constraint equations for an optimization problem. Among these, the regional comfort constraints stipulate that the predicted environmental parameters for each region must remain within a preset comfort range. For example, based on the heat balance equation, the system can establish a functional relationship between the predicted temperature and the control variables, requiring the temperature to be between 24℃ and 26℃. Equipment operation boundary constraints limit the physical feasible domain of the control variables; for example, the air volume delivered by the fan cannot exceed its rated maximum air volume, and the water supply temperature of the chiller unit cannot be lower than the antifreeze protection value.

[0141] The above constraints together constitute the feasible solution space, making the optimized control strategy physically executable and meeting the service quality standards.

[0142] Step S604: Integrate the control decision vector, regional comfort constraints, and equipment operating boundary constraints to generate a discrete-time series optimization problem description. This discrete-time series optimization problem description is the output of the collaborative optimization model.

[0143] Based on this, the objective function of the collaborative optimization model is constructed and solved, including:

[0144] The baseline energy consumption target item and the event disturbance energy consumption target item are invoked. The baseline energy consumption target item is associated with the baseline operating condition prediction data, and the event disturbance energy consumption target item is associated with the event disturbance load data.

[0145] The event disturbance suppression weights are used to weight the event disturbance energy consumption target item, and then combined with the baseline energy consumption target item to form the overall objective function; where the event disturbance suppression weights are variable.

[0146] During the iterative solution of the discrete time series optimization problem, the comfort constraint violation of the current solution is calculated.

[0147] The event disturbance suppression weight is adaptively adjusted based on the degree of comfort constraint violation. When the degree of comfort constraint violation exceeds the allowable range, the event disturbance suppression weight is reduced, thereby maximizing the suppression of train event impact energy consumption while ensuring comfort.

[0148] In other words, while ensuring comfort, we should suppress energy consumption that could impact train operations.

[0149] In this embodiment, a two-layer collaborative optimization solution architecture is constructed. In the lower layer, the system constructs the overall objective function J, which is expressed as:

[0150] J=λ×J _event +(1-λ)×J _base ;

[0151] Where J is the total objective function value of the collaborative optimization model, J _base This is the baseline energy consumption target, representing the energy consumption required to maintain basic operations; J _event λ is the event disturbance energy consumption target, representing the additional energy consumption generated in response to train events; λ can be the event disturbance suppression weight.

[0152] At the upper level, the system runs a master control loop to dynamically adjust the value of λ. Initially, the system may assign a large value to λ, such as 0.8, to suppress energy consumption impacts caused by events. The system then calls the lower-level solver, such as a quadratic programming solver, to obtain the control sequence under the current λ.

[0153] Furthermore, the upper-level system evaluates the violation of comfort constraints on this set of sequences, that is, checks whether the predicted temperature deviates from the comfort range. If the violation exceeds the allowable range, it indicates that the current energy-saving strategy is too aggressive, sacrificing too much comfort. In this case, the upper-level system will reduce the value of λ, for example, adjusting it to 0.6, increasing the focus on the baseline and comfort, and calling the lower-level solution again. Conversely, if the comfort requirements are met, the system can try to appropriately increase λ to pursue better energy efficiency. Through adaptive iteration, the system can eventually find a balance point that can both minimize the energy consumption of event disturbances and maintain the bottom line of comfort.

[0154] In some embodiments, after generating an event-driven control setpoint sequence for adjusting the operating state of equipment within the station, the method further includes:

[0155] Furthermore, the rate of change index of the event-driven control setpoint sequence between adjacent time steps is calculated, and the rate of change index exceeding the device response safety threshold is smoothed to generate a smooth control setpoint sequence.

[0156] Specifically, the control sequence obtained through direct optimization may contain abrupt changes. For example, requiring the fan to jump from 40% frequency to 90% frequency within 2 minutes is harmful to the mechanical equipment. Therefore, the optimization results undergo post-processing. The system calculates the difference between setpoints at adjacent time steps in the sequence, i.e., the rate of change index. If this index exceeds the equipment's response safety threshold, such as a fan frequency change rate limit of 10% per minute, the system will smooth the sequence. Moving average filtering or slope limiting algorithms can be used to generate a smooth control setpoint sequence, ensuring the stability of equipment operation.

[0157] Furthermore, the smooth control setpoint sequence is mapped to the pre-stored equipment operation topology data, transforming the setpoints for functional areas into equipment operation condition sequences for specific physical devices; the feasibility of the equipment operation condition sequences is verified, and instruction fragments that violate the physical start-stop constraints of the equipment are eliminated to obtain the final executable control instruction data.

[0158] Furthermore, the system transforms logical area control into physical device control. Device operation topology data describes which areas are served by which units. For example, waiting area A might be served jointly by fresh air units AHU-1 and AHU-2. Based on the topology, the system allocates the total air supply setpoint for waiting area A to both units.

[0159] During this process, the system executes specific conflict resolution and merging logic. For example, if, based on the topology, AHU-1 serves both waiting area A and waiting area B, and waiting area A requests a 60% fresh air ratio while waiting area B requests a 40% fresh air ratio, the system will execute the principle of prioritizing the highest demand, that is, issue a 60% fresh air ratio instruction to AHU-1 to meet the needs of the area with higher demand.

[0160] Based on this, the system performs feasibility checks, such as checking whether the compressor's minimum downtime protection logic is violated. After eliminating non-compliant instructions, the final generated equipment operating condition sequence is packaged into standard communication protocol messages and sent to the field controller for execution.

[0161] Building upon this, the embodiment constructs a multi-objective function with independent suppression weights, allowing the system to dynamically adjust its focus on disturbance energy consumption within an event window. This mechanism enables the control strategy to reduce additional energy consumption spikes caused by train events, not only lowering the overall average energy consumption.

[0162] According to another aspect of this application, an exemplary scheme for closed-loop evaluation and adaptive model update is provided. Specifically, after generating and executing the control setpoint sequence, the system evaluates the control effect through a closed-loop feedback mechanism and adaptively updates the front-end model parameters based on the evaluation results. This ensures long-term stable operation and continuous optimization of the system.

[0163] Accordingly, the system converts the generated equipment operating condition sequence into specific control commands and sends them to the field equipment control system. Subsequently, the air conditioning units, fresh air units, and other equipment in each area begin to execute actions according to the commands. During this process, the system does not stop working but enters a real-time monitoring state.

[0164] The system continuously collects updated environmental sensor data from the field, such as actual temperature, humidity, and carbon dioxide concentration, as well as actual equipment operation feedback data, such as real-time power and actual valve opening feedback. The system performs spatiotemporal alignment and comparative analysis of the real-time collected posterior data with the scenario-based multi-area load prediction data generated before control.

[0165] Specifically, the system calculates control effectiveness evaluation data and prediction deviation data. Control effectiveness evaluation data measures whether actual environmental parameters truly fall within the preset comfort range and whether actual energy consumption meets expected targets. Prediction deviation data focuses on the accuracy of the regional load prediction model, specifically the residual error between the predicted environmental load and the actually observed environmental load. For example, if the model predicts a waiting area temperature of 26℃ at a certain moment, while the actual sensor data indicates 27℃, a positive deviation of 1℃ is recorded.

[0166] Based on this, the system triggers an adaptive model update process. The system uses newly collected measured samples, especially those that produced significant prediction bias, to construct an incremental training set. The system then performs online adjustments to the model parameters in the regional load prediction model set, for example, by fine-tuning the weights of the neural network using gradient descent or updating the coefficients of the regression model.

[0167] Furthermore, the system can also modify key parameters in the event disturbance compensation mechanism based on the actual control response, such as adjusting the initial value of the event disturbance suppression weight λ mentioned in the previous embodiments or adjusting the step size. Through continuous closed-loop iteration, the system can automatically adapt to seasonal climate changes, equipment aging, and long-term drift in passenger flow patterns, achieving self-evolution of energy efficiency management strategies.

[0168] In other embodiments, anomaly correction logic for train operation data is also described. When constructing the basic time series data for train operation, the system performs consistency checks on the planned arrival time, planned departure time, actual arrival time, and actual departure time in the multi-source basic operation dataset. The system predefines train operation event rule parameters. For example, the system will only confirm a train arrival event when it detects that the actual arrival time field in a train's record is valid, but the actual departure time is not recorded.

[0169] In response to abnormal situations, the system performs rule-based corrections. For example, when the deviation between the actual arrival time and the planned arrival time exceeds a preset delay threshold, the system automatically marks the event as an abnormal arrival event and calculates the specific deviation. The cleaned and corrected train critical event data ensures the accuracy of subsequent scenario-based labeling and prevents control misjudgments caused by scheduling data errors.

[0170] In some embodiments, parts of the method of the present invention may also be:

[0171] Optionally, the system reads a sample dataset of scenario-driven multi-regional load modeling, which already contains the original scenario label information and train passenger flow event feature information. The system also reads predefined scenario and event coding parameters, including a list of scenario types, a list of event types, a passenger flow response level coding table, and a time and location coding method.

[0172] The system constructs an encoded representation for each sample according to the following logic:

[0173] The scene labels (such as peak ticket checking scene, delay and congestion scene) are converted into scene coded vector data through sparse coding or embedding coding.

[0174] The relationship between train event types (arrival, departure, start of ticket check, etc.) and sample time locations (e.g., the time difference from the nearest event) is encoded as event time location encoded data; discrete labels such as passenger flow response level and mode are encoded as passenger flow response encoded data.

[0175] Based on this, the system generates scene-event joint coding feature data for each sample, including scene coding vector, event time and location coding, and passenger flow response coding. This data, together with other input features in the sample, forms the model input feature vector. The output is a scene-event joint coding sample dataset.

[0176] Optionally, the system reads the scenario-event joint encoding sample dataset and reads predefined regional scenario modeling configuration parameters. These configuration parameters include a list of functional regions, scenario category division rules (e.g., off-peak scenario, peak ticket checking scenario, platform congestion boarding scenario), suggested model type for each subtask (e.g., time series model, tree model, or hybrid model), and initial hyperparameter settings.

[0177] Furthermore, the system is divided into multiple subsets according to functional areas, such as station hall, waiting area, ticket inspection area, and platform; within each area subset, scene subsets are divided according to scene tags, such as the off-peak scene subset and peak ticket inspection scene subset of the waiting area.

[0178] For each region-scene combination, the system defines a modeling subtask and assigns initial region-scene model structure description data to that subtask. This includes the model type (e.g., sequence network structure, ensemble regression structure), a list of input features (including scene-event encoding features), output variable definitions (e.g., environmental load, energy load), and initial hyperparameters. The model structure descriptions of all subtasks together constitute the region-scene model structure set data.

[0179] Optionally, the system reads updated model parameter data (which is empty or default if it is the first deployment), as well as regional scene model structure set data and sample subset partitioning data.

[0180] If a certain region-scene subtask has been trained in previous cycles, the system retrieves the historical model parameter snapshot data corresponding to that subtask from the model parameter update data and loads it into the current model structure to achieve model inheritance and incremental updates.

[0181] Furthermore, the system sets weighting strategies for training samples based on regional scenario modeling configuration parameters. For example, higher training weights are assigned to samples from scenarios with large load fluctuations, such as peak ticket checking and concentrated boarding at platforms, while lower weights or downsampling are used for samples from off-peak scenarios, forming scenario-weighted training configuration data. This configuration is used during training to construct a weighted loss function or sample resampling scheme, prioritizing the prediction accuracy of high-risk scenarios within the limited model capacity.

[0182] Based on this, the system reads the regional scene model training preparation data and the scene-event joint encoding sample dataset. For each regional-scene modeling subtask, the system extracts the corresponding modeling samples from the sample subset partitioning data, partitions the input features and output targets, and constructs training and validation sets.

[0183] During training, the system transforms the scenario-weighted training configuration data into a weighted loss function or a sample resampling strategy, so that high-weighted samples (usually corresponding to scenarios with large passenger flow) contribute more to the updating of model parameters.

[0184] During model training iterations, the system continuously evaluates performance metrics on the validation set and compares error distributions under different scenarios. If the error is found to be too large in certain scenarios (such as delays or congestion), the training weights or learning rate for that scenario are dynamically adjusted according to preset rules to form adaptive scenario weight adjustment data, further improving the fitting effect of the problem scenarios.

[0185] After training, the system saves the model parameters corresponding to each region-scenario subtask as regional load prediction model parameter data. The parameter set of all subtasks constitutes the final multi-level scenario-event jointly driven regional load prediction model set.

[0186] The optional embodiments of the present invention have been described in detail above. However, the present invention is not limited to the specific details of the above embodiments. Within the scope of the technical concept of the present invention, various equivalent transformations can be made to the technical solution of the present invention, and these equivalent transformations all fall within the protection scope of the present invention.

Claims

1. A train-passenger flow event-driven railway station energy efficiency fine control method, characterized in that, The method comprises: Based on train diagram data and real-time passenger flow monitoring data, a space-time aligned train passenger flow event sequence is constructed, and the train passenger flow event sequence anchors the train arrival and departure physical time and passenger flow response characteristics as discrete event nodes; Using the train passenger flow event sequence to drive the regional load prediction model, scenario-based multi-regional load prediction data is generated, wherein the dual-path deduction is used to explicitly separate the baseline working condition prediction data reflecting the smooth running state and the event disturbance load data reflecting the train event impact; According to the train passenger flow event sequence, an event control time window covering before and after the event is divided, and a collaborative optimization model for setting independent suppression weight for event disturbance load data is called in the event control time window; Solving the collaborative optimization model generates an event-driven control set point sequence for adjusting the running state of the station equipment; Wherein, the dual-path deduction is used to explicitly separate the baseline working condition prediction data reflecting the smooth running state and the event disturbance load data reflecting the train event impact, including: at each prediction time step, the current scene label and event characteristics are extracted from the train passenger flow event sequence, input into the regional load prediction model, and the event scene prediction result data containing the train event influence is generated; synchronously, the event state in the train passenger flow event sequence is replaced by the preset smooth baseline state, input into the regional load prediction model, and the baseline working condition prediction data shielding the train event influence is generated; the difference between the event scene prediction result data and the baseline working condition prediction data is calculated, and the difference is decomposed into an environmental load disturbance component and an energy consumption load disturbance component, which together constitute the event disturbance load data.

2. The method of claim 1, wherein, Using the train passenger flow event sequence to drive the regional load prediction model, comprising: Map the scene label in the train passenger flow event sequence to a numerical scene encoding vector; Calculate the time difference between the current prediction time and the nearest train key event in the train passenger flow event sequence, and use a periodic function to transform the time difference to generate an event time position encoding; Concatenate the scene encoding vector, event time position encoding and passenger flow response characteristics to construct a scene-event joint encoding feature, which is fed into the regional load prediction model as an input vector.

3. The method of claim 1, wherein, The regional load prediction model includes a scene-specific model subset constructed for high-disturbance load scenarios and a regional general model subset constructed for daily flat peak scenarios; the training of the regional load prediction model comprises: Based on historical operation data, a scene-driven multi-regional load modeling sample data set containing scene labels is constructed; Iteratively train the scene-specific model subset and the regional general model subset using the scene-driven multi-regional load modeling sample data set; In the training process, a scene weighting loss function is applied, which gives higher loss weight to the samples corresponding to the large passenger flow event scenarios in the scene-driven multi-regional load modeling sample data set than to the samples of the daily flat peak scenarios, thereby improving the fitting sensitivity of the regional load prediction model to the event disturbance load data.

4. The method of claim 1, wherein, Real-time passenger flow monitoring data is generated by multi-source weighted fusion of ticketing passenger flow data, entry and exit gate data and video passenger flow statistical data, and the generation process of real-time passenger flow monitoring data comprises: Map ticket passenger flow data, entry and exit gate data and video passenger flow statistics data to a unified time axis to generate multi-source time slice passenger flow data; Calculate a confidence score representing data quality reliability for each time slice and each data source; Calculate a dynamic fusion weight of each data source in the time slice based on the confidence score, wherein the lower the confidence score, the lower the dynamic fusion weight of the data source; Generate fusion passenger flow estimation data by weighted summation of the multi-source time slice passenger flow data using the dynamic fusion weight, and construct real-time passenger flow monitoring data based on the fusion passenger flow estimation data.

5. The method of claim 4, wherein, The confidence score representing data quality reliability is calculated, including: Calculate a passenger flow growth rate indicator of the current time slice relative to the previous time slice, and a ticket deviation rate indicator of the passenger flow value of the current time slice relative to the ticket passenger flow data; Compare the passenger flow growth rate indicator and the ticket deviation rate indicator with the preset abnormality determination threshold respectively; When the passenger flow growth rate indicator or the ticket deviation rate indicator exceeds the abnormality determination threshold, reduce the confidence score of the data source in the current time slice by a preset penalty factor until it is marked as an abnormal state.

6. The method of claim 4, wherein, The method further includes performing historical template compensation when all multi-source data are abnormal, specifically: When the confidence scores of all data sources in a certain time period are lower than a preset confidence threshold, retrieve historical passenger flow curve templates from a historical database according to the current train type and time period; Extract the generated fusion passenger flow estimation data before and after the time period as reference data segments, and calculate the shape similarity between the reference data segments and each historical passenger flow curve template; Select the historical passenger flow curve template with the highest shape similarity as the best matching template, scale the best matching template according to the average amplitude of the reference data segments, and generate template compensation passenger flow estimation data to fill in the data loss of the time period.

7. The method of claim 1, wherein, Invoke a collaborative optimization model for setting independent suppression weights for event disturbance load data within an event control time window, including: Divide the event control time window into a finite number of discrete time step sequences; For each discrete time step and each functional area, define a control decision vector containing the supply air temperature set point and the supply air volume set point; Use scenario-based multi-zone load prediction data to construct regional comfort constraint conditions and equipment operation boundary constraint conditions for the discrete time step sequence; Integrate the control decision vector, the regional comfort constraint condition and the equipment operation boundary constraint condition to generate a discrete time sequence optimization problem description, i.e. the output of the collaborative optimization model is the discrete time sequence optimization problem description.

8. The method of claim 7, wherein, The objective function construction and solution of the collaborative optimization model, including: Call the baseline energy consumption target item and the event disturbance energy consumption target item, wherein the baseline energy consumption target item is associated with the baseline working condition prediction data, and the event disturbance energy consumption target item is associated with the event disturbance load data; Weight the event disturbance energy consumption target item using the event disturbance suppression weight, combine it with the baseline energy consumption target item to form a total objective function; During the iterative solution process of the discrete time sequence optimization problem description, calculate the comfort constraint violation degree of the current solution; According to the comfort constraint violation degree, the event disturbance suppression weight is adaptively adjusted, the event disturbance suppression weight is reduced when the comfort constraint violation degree exceeds the allowed range, and the energy consumption of the train event impact is suppressed on the premise of ensuring the comfort.

9. The method of claim 1, wherein, Based on train operation diagram data and real-time passenger flow monitoring data, a train passenger flow event sequence aligned in time and space is constructed, including: From the train operation diagram data, the train arrival time and the train departure time are extracted as train key event anchors; According to the preset train passenger flow linkage parameters, a time range is defined before and after each train key event anchor, and the passenger flow feature segment within the time range is intercepted from the real-time passenger flow monitoring data; The passenger flow intensity index and the growth rate index of the passenger flow feature segment are calculated, and the passenger flow response level label of the event node is determined accordingly; The train key event anchor, the passenger flow feature segment and the passenger flow response level label are associated and stored, and arranged in time sequence to form the train passenger flow event sequence.

Images