Deep learning-based railway station carbon emission prediction method and system

Abstract

The invention provides a railway station carbon emission prediction method and system based on deep learning, and the method comprises the steps: collecting and preprocessing multi-source carbon emission influence factor data of a railway station, and forming a multi-dimensional data sample set; constructing a railway station space association network containing heterogeneous nodes and dynamic edge weights, and extracting dynamic space features of the data sample set by using the association network; performing multi-scale decomposition on time sequence data of the data sample set to obtain subsequences with different frequencies, extracting time sequence features of the subsequences, and fusing the time sequence features through a frequency attention mechanism to obtain multi-scale time sequence features; performing feature alignment and weight fusion on the dynamic space features and the multi-scale time sequence features through a fusion attention mechanism to obtain cross-modal space-time fusion features; and inputting the cross-modal space-time fusion features into a trained carbon emission prediction model for processing, and outputting a carbon emission prediction result and a corresponding uncertainty quantitative index thereof.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, and in particular relates to a method and system for predicting carbon emissions from railway passenger stations based on deep learning. Background Technology

[0002] As a key sector for carbon emissions, the transportation industry has received widespread attention for its efforts to reduce emissions and energy consumption.

[0003] As hubs of the railway transportation network, railway passenger stations are characterized by dense populations, long operating cycles, diverse equipment, and high energy intensity. Their carbon emission sources are complex, encompassing multiple dimensions such as building energy consumption, equipment operation energy consumption, and passenger flow-related energy consumption. Therefore, predicting the carbon emissions of railway passenger stations is a prerequisite for achieving refined carbon emission management and formulating scientific emission reduction strategies.

[0004] Currently, the relevant technologies for predicting carbon emissions from railway passenger stations can be mainly divided into two categories: traditional statistical prediction methods and preliminary machine learning prediction methods. However, both have significant technical shortcomings and are difficult to meet the accurate prediction needs of actual operation scenarios. On the one hand, traditional statistical forecasting methods are based on linear assumptions and can only capture simple time-series patterns of carbon emission data, which cannot adapt to the complex nonlinear characteristics of carbon emissions from railway passenger stations being affected by multiple coupled factors. On the other hand, while existing preliminary machine learning prediction methods have broken through the linear limitation to some extent, they still have many shortcomings. For example, they have not explored the spatiotemporal coupling characteristics of carbon emissions in the context of railway passenger stations, such as the spatial correlation of different waiting areas, the dynamic adaptation relationship between passenger flow time sequence changes and equipment operating status, etc., which leads to low model input quality. Secondly, the prediction model architecture is simple, lacks adaptive adjustment ability, and has weak generalization ability. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for predicting carbon emissions from railway passenger stations based on deep learning, so as to solve the technical problems mentioned in the background.

[0006] To achieve the above objectives, the present invention provides the following technical solution.

[0007] According to an embodiment of the present invention, a deep learning-based method for predicting carbon emissions from railway passenger stations is provided, comprising the following steps: Data on the multi-source carbon emission influencing factors of railway passenger stations were collected and preprocessed to form a multi-dimensional data sample set. A spatial association network for railway passenger stations, including heterogeneous nodes and dynamic edge weights, is constructed, and the dynamic spatial features of the data sample set are extracted using the association network. The time series data of the data sample set is decomposed into subsequences of different frequencies through multi-scale decomposition. The time series features of each subsequence are extracted, and the time series features are fused through a frequency attention mechanism to obtain multi-scale time series features. By using a fusion attention mechanism, the dynamic spatial features and the multi-scale temporal features are aligned and weighted to obtain the cross-modal spatiotemporal fusion features. The cross-modal spatiotemporal fusion features are input into a trained carbon emission prediction model for processing, and the carbon emission prediction results and their corresponding uncertainty quantification indicators are output; wherein, the carbon emission prediction model is an adaptive fusion deep learning model containing multiple feature processing branches.

[0008] Furthermore, the step of preprocessing to form a multi-dimensional data sample set includes: A missing value imputation strategy combining spatiotemporal interpolation and domain rule constraints is adopted, and an outlier detection mechanism that adapts to operational scenario thresholds is used to clean multi-source data; the cleaned data is aligned based on a unified spatiotemporal benchmark. A multi-dimensional data labeling system is constructed to label the data, and a combination of max-min normalization and domain-adaptive normalization is used to normalize the labeled data. A hybrid augmentation model combining generative adversarial networks and domain adaptation is adopted to augment the normalized data, thereby generating expanded data samples and forming a multi-dimensional data sample set.

[0009] Furthermore, the steps for constructing a spatial association network of railway passenger stations that includes heterogeneous nodes and dynamic edge weights include: The nodes of the spatial network are divided into core nodes, associated nodes, and disturbance nodes; wherein, the core nodes correspond to the functional zones of the railway passenger station, the associated nodes correspond to the associated areas or facilities around the passenger station, and the disturbance nodes correspond to temporary operational disturbance areas. Based on the spatial association strength, real-time operational interaction degree and distance decay coefficient, a dynamic weight calculation model is used to assign weights to the connection edges between nodes, and the weights are updated according to real-time operational data. A spatial association network is constructed based on the defined heterogeneous nodes and the calculated dynamic edge weights.

[0010] Furthermore, the step of assigning weights to the connection edges between nodes using a dynamic weight calculation model includes: A multi-dimensional dynamic edge weight calculation model is constructed, and the edge weight is jointly determined by the basic strength of spatial association, the real-time operational interaction degree, and the distance decay coefficient. Among them, the basic strength of spatial association is calculated based on the long-term average of the association between different nodes based on historical operational data, the real-time operational interaction degree is dynamically updated through real-time passenger flow transfer volume and energy consumption coordination fluctuation value, and the distance decay coefficient is calculated using a non-linear decay function, which decays exponentially with the increase of spatial distance between nodes. For different operational scenarios, scenario correction coefficients are set to adaptively adjust the edge weights, as shown below: In the formula, Let be the edge weight at time t; , These are the weighting coefficients, and S represents the spatially correlated foundation strength. γ represents the real-time operational interaction at time t; D is the distance decay coefficient; and γ is the scene correction coefficient.

[0011] Furthermore, the step of performing multi-scale decomposition of the time-series data of the data sample set to obtain subsequences of different frequencies and extracting the time-series features of each subsequence includes: Trend separation and noise suppression are performed on time series data to obtain stationary time series data; A strategy of cascading empirical mode decomposition and variational mode decomposition is adopted to decompose stationary time series data into stationary subsequences of different frequencies; For subsequences of different frequencies, corresponding temporal neural network models are used to extract their temporal features. Weights are assigned based on the frequency contribution of each subsequence, and the temporal features are fused to obtain multi-scale temporal features.

[0012] Furthermore, the step of decomposing the stationary time series data into stationary subsequences of different frequencies includes: EMD is used to decompose stationary time series data into several intrinsic mode functions and residual components, and the IMF components containing less than the threshold of high-frequency noise are selected. The filtered IMF components and residual components are input into the VMD model, and stationary subsequences of different frequencies are obtained according to the preset number of decomposition modes. Based on power spectral density analysis, each subsequence is classified by frequency into high-frequency subsequence, mid-frequency subsequence, and low-frequency subsequence.

[0013] Furthermore, the step of extracting temporal features from subsequences of different frequencies using corresponding temporal neural network models includes: For high-frequency subsequences, a bidirectional gated recurrent unit combined with a local temporal attention mechanism is used to extract key time period features in short-term rapid fluctuations; For mid-frequency subsequences, a bidirectional long short-term memory network is used to extract periodic variation patterns. For low-frequency subsequences, a long short-term memory network combined with a global temporal attention mechanism is used to capture long-term trend features. In the feature extraction process of each subsequence, temporal position encoding is introduced to enhance the representation of positional information, and the temporal feature vector of each subsequence is output.

[0014] Furthermore, the adaptive fusion deep learning model comprising multiple feature processing branches includes: The feature processing module includes at least two parallel feature processing branches for receiving and processing the cross-modal spatiotemporal fusion features; An adaptive fusion layer is used to dynamically allocate weights and perform weighted fusion of the output features of each branch based on the correlation between the output features of each feature processing branch and the carbon emission data. The uncertainty quantification module is used to process the output of the adaptive fusion layer, output the carbon emission prediction results in probabilistic form, and generate corresponding uncertainty quantification indicators.

[0015] According to another embodiment of the present invention, a deep learning-based carbon emission prediction system for railway passenger stations is provided, comprising the following modules: The data acquisition module is used to collect data on the multi-source carbon emission influencing factors of railway passenger stations and preprocess the data to form a multi-dimensional data sample set. The spatial feature extraction module is used to construct a spatial association network of railway passenger stations containing heterogeneous nodes and dynamic edge weights, and to extract dynamic spatial features of the data sample set using the association network. The temporal feature extraction module is used to perform multi-scale decomposition of the temporal data of the data sample set to obtain subsequences of different frequencies, extract the temporal features of each subsequence, and fuse the temporal features through a frequency attention mechanism to obtain multi-scale temporal features. The feature fusion module is used to perform feature alignment and weight fusion on the dynamic spatial features and the multi-scale temporal features through a fusion attention mechanism to obtain the cross-modal spatiotemporal fusion features; The carbon emission prediction module is used to input the cross-modal spatiotemporal fusion features into a trained carbon emission prediction model for processing, and output carbon emission prediction results and their corresponding uncertainty quantification indicators; wherein, the carbon emission prediction model is an adaptive fusion deep learning model containing multiple feature processing branches.

[0016] According to another embodiment of the present invention, a non-volatile storage medium is provided, the non-volatile storage medium including a stored program, wherein, when the program is executed, the device where the non-volatile storage medium is located is controlled to execute the deep learning-based railway passenger station carbon emission prediction method described in the above embodiment.

[0017] Compared with existing technologies, the beneficial effects of the deep learning-based method and system for predicting carbon emissions from railway passenger stations in this invention are: This invention constructs a spatial association network for railway passenger stations, incorporating heterogeneous nodes and dynamic edge weights. The network is used to extract dynamic spatial features from a data sample set. The temporal data of the data sample set is decomposed into subsequences of different frequencies at multiple scales. Temporal features of each subsequence are extracted, and these features are fused using a frequency attention mechanism to obtain multi-scale temporal features. A fusion attention mechanism is then used to align the dynamic spatial features with the multi-scale temporal features and fuse their weights to obtain cross-modal spatiotemporal fusion features. These cross-modal spatiotemporal fusion features are then input into a trained carbon emission prediction model for processing, outputting carbon emission prediction results and their corresponding uncertainty quantification indicators. This invention effectively integrates full-dimensional data from railway passenger stations by constructing a dynamic heterogeneous spatial network, extracting multi-scale temporal features, introducing a deep cross-modal fusion mechanism, and an adaptive prediction model. This significantly improves the robustness and interpretability of carbon emission prediction and enables adaptive prediction and dynamic optimization for different operating scenarios, contributing to the carbon emission management of railway passenger stations. Attached Figure Description

[0018] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0019] In the attached diagram: Figure 1 A flowchart illustrating the implementation of the deep learning-based method for predicting carbon emissions from railway passenger stations, as provided in this embodiment of the invention. Figure 2 This is a sub-flowchart of the deep learning-based method for predicting carbon emissions from railway passenger stations according to the present invention. Figure 3 This is a block diagram of the deep learning-based railway passenger station carbon emission prediction system of the present invention; Figure 4 A structural block diagram of a computer device provided by the present invention. Detailed Implementation

[0020] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0021] It should be noted that, unless otherwise specified, all technical and scientific terms used in this application have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0022] The railway passenger station carbon emission prediction method of the present invention effectively integrates full-dimensional data of railway passenger stations by constructing a dynamic heterogeneous spatial network, extracting multi-scale time series features, introducing a deep cross-modal fusion mechanism and an adaptive prediction model, significantly improving the robustness and interpretability of carbon emission prediction, and realizing adaptive prediction and dynamic optimization for different operating scenarios, which is helpful for the carbon emission management of railway passenger stations.

[0023] The specific implementation of the present invention will be described below with reference to specific embodiments.

[0024] Please refer to Figure 1 In one embodiment of the present invention, a method for predicting carbon emissions from railway passenger stations based on deep learning is provided, comprising the following steps: Step S10: Collect data on the multi-source carbon emission influencing factors of railway passenger stations and preprocess them to form a multi-dimensional data sample set; Specifically, in this embodiment of the invention, the multi-dimensional data sample set is obtained through the following steps: First, this invention employs a missing value imputation strategy that combines spatiotemporal interpolation with domain rule constraints, and uses an outlier detection mechanism that combines operational scenario threshold adaptive correction to clean multi-source data. Align the cleaned data based on a unified spatiotemporal reference; A multi-dimensional data labeling system is constructed to label the data, including: labeling each data sample with time dimension (such as time period, period, etc.), spatial dimension (such as functional zoning, station-city related area, etc.), influencing factor dimension (such as operation, environment, policy, etc.) and disturbance type dimension label information; After labeling, a combination of max-min normalization and neighborhood adaptive normalization is used to normalize the labeled data; A hybrid augmentation model combining generative adversarial networks and domain adaptation is used to augment normalized data to generate extended data samples and form a multi-dimensional data sample set. In this invention, the generative adversarial network is trained based on real historical data to generate extended samples for scarce and extreme scenarios such as Spring Festival travel rush, extreme weather, and large-scale event passenger flow. The generated samples need to pass the energy consumption rationality verification. In the hybrid enhancement model, this invention introduces a domain adversarial training mechanism to construct a domain adaptive module. Historical data from multiple mature operating passenger stations are used as source domain data, and a small amount of measured data from the target passenger station are used as target domain data. Domain adversarial training minimizes the data distribution offset between the source and target domains, enhancing the domain adaptability of the sample set. Finally, duplicate data is removed and validity is screened on the enhanced sample set to form a multi-dimensional data sample set that is universal, adaptable to different scenarios, and highly reliable.

[0025] Please continue to refer to Figure 1The deep learning-based method for predicting carbon emissions from railway passenger stations of the present invention further includes the following steps: Step S20: Construct a spatial association network for railway passenger stations that includes heterogeneous nodes and dynamic edge weights, and use the association network to extract dynamic spatial features of the data sample set; Specifically, step S20 of the present invention, the step of constructing a spatial association network of railway passenger stations including heterogeneous nodes and dynamic edge weights, includes: The nodes of the spatial network are divided into core nodes, associated nodes, and disturbance nodes. The core nodes of this invention correspond to the functional zones of the railway passenger station, including real-time energy consumption data, carbon emission intensity, passenger flow load, equipment operating status and area ratio of each zone; The associated nodes in this embodiment of the invention correspond to associated areas or facilities around the passenger station, such as transportation hubs, commercial complexes, green spaces and water bodies and municipal facilities around the passenger station. Their attribute parameters include the carbon emission contribution of each node, the spatial distance from the passenger station, the intensity of passenger flow interaction and the degree of matching of transport capacity. The disturbance nodes in this embodiment of the invention correspond to temporary operational disturbance areas, such as temporary connection points, construction areas, and temporary sites for large-scale events. Their attribute parameters include disturbance type, duration, impact range, and disturbance coefficient on carbon emissions of the passenger station. In this embodiment of the invention, a unique identifier and type label are assigned to each type of node to construct a heterogeneous node attribute matrix; Based on the spatial association strength, real-time operational interaction degree, and distance decay coefficient, a dynamic weight calculation model is used to assign weights to the connection edges between nodes, and the weights are updated according to real-time operational data; and a spatial association network is constructed based on the defined heterogeneous nodes and the calculated dynamic edge weights. In one implementation of the present invention, the step of assigning weights to the connection edges between nodes through a dynamic weight calculation model includes: constructing a multi-dimensional dynamic edge weight calculation model, wherein the edge weights are jointly determined by the spatial association basic strength, real-time operational interaction degree and distance attenuation coefficient. Among them, the spatial correlation basic strength is calculated based on historical operational data to determine the long-term correlation average between different nodes, the real-time operational interaction degree is dynamically updated through real-time passenger flow transfer volume and energy consumption synergy fluctuation value, and the distance attenuation coefficient is calculated using a non-linear attenuation function, which attenuates exponentially with the increase of spatial distance between nodes. This invention sets scenario correction coefficients for different operational scenarios and adaptively adjusts edge weights, as shown below: In the formula, Let be the edge weight at time t; , These are the weighting coefficients, and S represents the spatially correlated foundation strength. γ represents the real-time operational interaction at time t; D is the distance decay coefficient; and γ is the scene correction coefficient.

[0026] This invention constructs an initial spatial network of railway passenger stations based on heterogeneous node attribute matrices and dynamic edge weights using a graph structure modeling method. Specifically, by introducing a graph topology optimization algorithm, redundant edges with weights below a preset threshold are eliminated, simplifying the network structure and improving computational efficiency. Furthermore, a dynamic network update mechanism is established to collect operational data and status changes of each node in real time, recalculating edge weights every 15 minutes. When the change in node attributes exceeds a preset threshold or when nodes are added / deleted, the network topology is automatically updated. Finally, the present invention verifies the effectiveness of the constructed spatial association network to ensure that the network can accurately represent the spatial association relationships and dynamic change characteristics between different nodes, and outputs the optimized heterogeneous dynamic spatial association network and the corresponding graph structure data.

[0027] Please continue to refer to Figure 1 The deep learning-based method for predicting carbon emissions from railway passenger stations of the present invention further includes the following steps: Step S30: Perform multi-scale decomposition on the time series data of the data sample set to obtain subsequences of different frequencies, extract the time series features of each subsequence, and fuse the time series features through a frequency attention mechanism to obtain multi-scale time series features; For details, please refer to Figure 2 The present invention includes the following steps: performing multi-scale decomposition of time-series data in a data sample set to obtain subsequences of different frequencies, and extracting the time-series features of each subsequence; Step S301: Perform trend separation and noise suppression on the time series data to obtain stationary time series data; In step S301, the present invention first separates the trend term of the time series data in the data sample set, extracts the long-term trend component in the time series data using a linear fitting method and temporarily stores it; then, it uses an adaptive wavelet threshold denoising algorithm to suppress noise in the detrended time series data, sets a threshold parameter that is dynamically adjusted according to the data fluctuation amplitude, removes high-frequency noise generated by sudden changes in passenger flow, instantaneous equipment failures, etc., and obtains stable time series data. Step S302: Employing a strategy of cascading Empirical Mode Decomposition (EMD) and Variational Mode Decomposition (VMD), the stationary time series data is decomposed into stationary subsequences of different frequencies. Specifically, this invention uses EMD to decompose the stationary time series data into several intrinsic mode functions (IMFs) and residual components, filtering out IMFs containing less than a threshold of high-frequency noise. The filtered IMFs and residual components are input into the VMD model, and stationary subsequences of different frequencies are obtained according to a preset number of decomposition modes. Based on power spectral density analysis, each subsequence is classified by frequency into high-frequency, mid-frequency, and low-frequency subsequences. Among them, the high-frequency subsequence corresponds to rapid changes such as intraday peaks and valleys and short-term passenger flow fluctuations; the medium-frequency subsequence corresponds to medium-term changes such as intraweek cycles and differences between weekdays and non-weekdays; and the low-frequency subsequence corresponds to long-term changes such as seasonal fluctuations and annual trends. Step S303: For subsequences of different frequencies, extract their temporal features using corresponding temporal neural network models, assign weights based on the frequency contribution of each subsequence, and fuse the temporal features to obtain multi-scale temporal features; Specifically, in step S303 of the present invention, for high-frequency subsequences, a bidirectional gated cyclic unit combined with a local temporal attention mechanism is used to extract key time period features in short-term rapid fluctuations. For mid-frequency subsequences, a bidirectional long short-term memory network is used to extract periodic variation patterns. For low-frequency subsequences, a long short-term memory network combined with a global temporal attention mechanism is used to capture long-term trend features; During the feature extraction process of each subsequence, temporal position encoding is introduced to enhance the representation of positional information, and the temporal feature vector of each subsequence is output.

[0028] Please continue to refer to Figure 1 The deep learning-based method for predicting carbon emissions from railway passenger stations of the present invention further includes the following steps: Step S40: By using a fusion attention mechanism, feature alignment and weight fusion are performed on the dynamic spatial features and the multi-scale temporal features to obtain the cross-modal spatiotemporal fusion features; Step S50: Input the cross-modal spatiotemporal fusion features into the trained carbon emission prediction model for processing, and output the carbon emission prediction results and their corresponding uncertainty quantification indicators; wherein, the carbon emission prediction model is an adaptive fusion deep learning model containing multiple feature processing branches.

[0029] Furthermore, the adaptive fusion deep learning model comprising multiple feature processing branches includes: The feature processing module includes at least two parallel feature processing branches for receiving and processing the cross-modal spatiotemporal fusion features; In one implementation, the branches in the feature processing module of the present invention include, but are not limited to, dynamic spatial feature branches, multi-scale temporal feature branches, cross-modal interaction feature branches, and disturbance compensation branches. Preferably, the input of the dynamic spatial feature branch in this embodiment of the invention is the graph structure data and dynamic spatial features corresponding to the heterogeneous dynamic spatial association network. Graph Sampling Aggregation (GraphSAGE) is used to achieve efficient sampling and feature aggregation of large-scale nodes. Graph Attention (GAT) is used to adaptively focus the contribution weight of highly correlated nodes to carbon emissions. Graph Convolution (GCN) is used to mine global spatial dependencies. Finally, the enhanced deep spatial feature vector is output. Preferably, the input of the multi-scale temporal feature branch of the present invention is the optimized multi-scale temporal feature. The long-short-term dependency relationship of the temporal features of different frequency subsequences is captured by the bidirectional long short-term memory network BiLSTM. The temporal attention mechanism dynamically allocates the weight of each time scale feature to characterize the temporal features of key periods such as peak passenger flow and seasonal transition, and outputs a deep temporal feature vector. Preferably, the input of the cross-modal interaction feature branch of the present invention is cross-modal spatiotemporal fusion features. The local interaction patterns of spatiotemporal features are extracted through convolutional layers, and the feature dimensionality is reduced while retaining key information by max pooling layers. This captures fine-grained features of spatiotemporal coupling and outputs a cross-modal interaction feature vector. Preferably, the input of the disturbance compensation branch of the present invention is non-steady-state disturbance features, such as temporary train additions, equipment failures, extreme weather, etc. The core impact information of the disturbance features is extracted through the fully connected layer, and the gating unit filters the effective disturbance features and suppresses redundant information to achieve accurate compensation for the impact of disturbances on carbon emissions and outputs a disturbance compensation feature vector.

[0030] Furthermore, the adaptive fusion deep learning model of the present invention also includes an adaptive fusion layer, which is used to dynamically allocate weights and perform weighted fusion of the output features of each branch according to the degree of correlation between the output features of each feature processing branch and the carbon emission data. The adaptive fusion layer of this invention adopts a Softmax dynamic weight allocation mechanism, introduces a scene recognition submodule to judge the current railway station operation scene in real time, such as daily operation, holiday peak, extreme weather, equipment maintenance, and other operation scenes. It calculates the feature contribution based on the mutual information entropy of the output features of each feature processing branch and carbon emission data, and dynamically adjusts the weight ratio of each branch feature vector in combination with the scene recognition results. The sum of the weights is 1, and the deep fusion of multi-dimensional features is achieved through weighted summation, outputting a fused feature vector. The adaptive fusion deep learning model of the present invention also includes: The uncertainty quantification module is used to process the output of the adaptive fusion layer, output the carbon emission prediction results in probabilistic form, and generate corresponding uncertainty quantification indicators.

[0031] For example, the fusion feature vector output by the adaptive fusion layer is input into the uncertainty quantification module. The model parameters are modeled by probability distribution using a Bayesian framework. The Monte Carlo dropout method is combined to randomly discard some network nodes multiple times during the model inference stage and repeat the prediction. The distribution characteristics of the multiple prediction results are statistically analyzed, and the 95% confidence interval, confidence level and prediction bias probability of the carbon emission prediction value are output. Finally, through a fully connected layer structure, the input of the fully connected layer is the feature vector processed by the uncertainty quantification module. The fully connected layer maps the features to the carbon emission prediction dimension. At the same time, it integrates a deviation warning unit. When the prediction deviation exceeds the preset threshold, it automatically triggers a warning. Finally, it outputs the carbon emission prediction result and the corresponding uncertainty quantification index.

[0032] Please refer to Figure 3 In another embodiment of the present invention, a deep learning-based carbon emission prediction system for railway passenger stations is provided, comprising the following modules: Data acquisition module 11 is used to collect data on the multi-source carbon emission influencing factors of railway passenger stations and perform preprocessing to form a multi-dimensional data sample set; The spatial feature extraction module 12 is used to construct a spatial association network of railway passenger stations containing heterogeneous nodes and dynamic edge weights, and to extract dynamic spatial features of the data sample set using the association network. The temporal feature extraction module 13 is used to perform multi-scale decomposition on the temporal data of the data sample set to obtain subsequences of different frequencies, extract the temporal features of each subsequence, and fuse the temporal features through a frequency attention mechanism to obtain multi-scale temporal features. Feature fusion module 14 is used to perform feature alignment and weight fusion on the dynamic spatial features and the multi-scale temporal features through a fusion attention mechanism to obtain the cross-modal spatiotemporal fusion features; The carbon emission prediction module 15 is used to input the cross-modal spatiotemporal fusion features into the trained carbon emission prediction model for processing, and output the carbon emission prediction results and their corresponding uncertainty quantification indicators; wherein, the carbon emission prediction model is an adaptive fusion deep learning model containing multiple feature processing branches.

[0033] Please refer to Figure 4 This application also provides a computer device, including a memory and a processor, wherein the processor is used to run a program stored in the memory, wherein the program executes the above-described deep learning-based method for predicting carbon emissions from railway passenger stations.

[0034] According to another embodiment of the present invention, a non-volatile storage medium is provided, the non-volatile storage medium including a stored program, wherein, when the program is executed, the device where the non-volatile storage medium is located is controlled to execute the deep learning-based railway passenger station carbon emission prediction method described in the above embodiment.

[0035] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0036] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0037] In the above embodiments of this application, the information collected is information and data authorized by the user or fully authorized by all parties, and the collection, storage, use, processing, transmission, provision, disclosure and application of the relevant data all comply with relevant laws, regulations and standards, take necessary protective measures, do not violate public order and good morals, and provide corresponding operation entry points for users to choose to authorize or refuse.

[0038] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0039] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0040] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0041] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to related technologies, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0042] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A deep learning-based method for predicting carbon emissions from railway passenger stations, characterized in that, Includes the following steps: Data on the multi-source carbon emission influencing factors of railway passenger stations were collected and preprocessed to form a multi-dimensional data sample set. A spatial association network for railway passenger stations, including heterogeneous nodes and dynamic edge weights, is constructed, and the dynamic spatial features of the data sample set are extracted using the association network. The time series data of the data sample set is decomposed into subsequences of different frequencies through multi-scale decomposition. The time series features of each subsequence are extracted, and the time series features are fused through a frequency attention mechanism to obtain multi-scale time series features. By using a fusion attention mechanism, the dynamic spatial features and the multi-scale temporal features are aligned and weighted to obtain the cross-modal spatiotemporal fusion features. The cross-modal spatiotemporal fusion features are input into a trained carbon emission prediction model for processing, and the carbon emission prediction results and their corresponding uncertainty quantification indicators are output; wherein, the carbon emission prediction model is an adaptive fusion deep learning model containing multiple feature processing branches.

2. The method for predicting carbon emissions from railway passenger stations based on deep learning according to claim 1, characterized in that, The step of preprocessing to form a multi-dimensional data sample set includes: A missing value imputation strategy combining spatiotemporal interpolation and domain rule constraints is adopted, and an outlier detection mechanism that adapts to operational scenario thresholds is used to clean multi-source data; the cleaned data is aligned based on a unified spatiotemporal benchmark. A multi-dimensional data labeling system is constructed to label the data, and a combination of max-min normalization and domain-adaptive normalization is used to normalize the labeled data. A hybrid augmentation model combining generative adversarial networks and domain adaptation is adopted to augment the normalized data, thereby generating expanded data samples and forming a multi-dimensional data sample set.

3. The deep learning-based method for predicting carbon emissions from railway passenger stations according to claim 2, characterized in that, The steps for constructing a spatial association network of railway passenger stations that includes heterogeneous nodes and dynamic edge weights include: The nodes of the spatial network are divided into core nodes, associated nodes, and disturbance nodes; wherein, the core nodes correspond to the functional zones of the railway passenger station, the associated nodes correspond to the associated areas or facilities around the passenger station, and the disturbance nodes correspond to temporary operational disturbance areas. Based on the spatial association strength, real-time operational interaction degree and distance decay coefficient, a dynamic weight calculation model is used to assign weights to the connection edges between nodes, and the weights are updated according to real-time operational data. A spatial association network is constructed based on the defined heterogeneous nodes and the calculated dynamic edge weights.

4. The method for predicting carbon emissions from railway passenger stations based on deep learning according to claim 3, characterized in that, The step of assigning weights to the connection edges between nodes using a dynamic weight calculation model includes: A multi-dimensional dynamic edge weight calculation model is constructed, and the edge weight is jointly determined by the basic strength of spatial association, the real-time operational interaction degree, and the distance decay coefficient. Among them, the basic strength of spatial association is calculated based on the long-term average of the association between different nodes based on historical operational data, the real-time operational interaction degree is dynamically updated through real-time passenger flow transfer volume and energy consumption coordination fluctuation value, and the distance decay coefficient is calculated using a non-linear decay function, which decays exponentially with the increase of spatial distance between nodes. For different operational scenarios, scenario correction coefficients are set to adaptively adjust the edge weights, as shown below: In the formula, Let be the edge weight at time t; , These are the weighting coefficients, and S represents the spatially correlated foundation strength. γ represents the real-time operational interaction at time t; D is the distance decay coefficient; and γ is the scene correction coefficient.

5. The deep learning-based method for predicting carbon emissions from railway passenger stations according to claim 3 or 4, characterized in that, The step of performing multi-scale decomposition of the time-series data of the data sample set to obtain subsequences of different frequencies and extracting the time-series features of each subsequence includes: Trend separation and noise suppression are performed on time series data to obtain stationary time series data; A strategy of cascading empirical mode decomposition and variational mode decomposition is adopted to decompose stationary time series data into stationary subsequences of different frequencies; For subsequences of different frequencies, corresponding temporal neural network models are used to extract their temporal features. Weights are assigned based on the frequency contribution of each subsequence, and the temporal features are fused to obtain multi-scale temporal features.

6. The deep learning-based method for predicting carbon emissions from railway passenger stations according to claim 5, characterized in that, The step of decomposing stationary time series data into stationary subsequences of different frequencies includes: EMD is used to decompose stationary time series data into several intrinsic mode functions and residual components, and the IMF components containing less than the threshold of high-frequency noise are selected. The filtered IMF components and residual components are input into the VMD model, and stationary subsequences of different frequencies are obtained according to the preset number of decomposition modes. Based on power spectral density analysis, each subsequence is classified by frequency into high-frequency subsequence, mid-frequency subsequence, and low-frequency subsequence.

7. The deep learning-based method for predicting carbon emissions from railway passenger stations according to claim 6, characterized in that, The step of extracting the temporal features of subsequences of different frequencies using corresponding temporal neural network models includes: For high-frequency subsequences, a bidirectional gated recurrent unit combined with a local temporal attention mechanism is used to extract key time period features in short-term rapid fluctuations; For mid-frequency subsequences, a bidirectional long short-term memory network is used to extract periodic variation patterns. For low-frequency subsequences, a long short-term memory network combined with a global temporal attention mechanism is used to capture long-term trend features. In the feature extraction process of each subsequence, temporal position encoding is introduced to enhance the representation of positional information, and the temporal feature vector of each subsequence is output.

8. The method for predicting carbon emissions from railway passenger stations based on deep learning according to claim 7, characterized in that, The adaptive fusion deep learning model, which includes multiple feature processing branches, includes: The feature processing module includes at least two parallel feature processing branches for receiving and processing the cross-modal spatiotemporal fusion features; An adaptive fusion layer is used to dynamically allocate weights and perform weighted fusion of the output features of each branch based on the correlation between the output features of each feature processing branch and the carbon emission data. The uncertainty quantification module is used to process the output of the adaptive fusion layer, output the carbon emission prediction results in probabilistic form, and generate corresponding uncertainty quantification indicators.

9. A system for implementing the deep learning-based method for predicting carbon emissions from railway passenger stations as described in any one of claims 1 to 8, characterized in that, The system includes the following modules: The data acquisition module is used to collect data on the multi-source carbon emission influencing factors of railway passenger stations and preprocess the data to form a multi-dimensional data sample set. The spatial feature extraction module is used to construct a spatial association network of railway passenger stations containing heterogeneous nodes and dynamic edge weights, and to extract dynamic spatial features of the data sample set using the association network. The temporal feature extraction module is used to perform multi-scale decomposition of the temporal data of the data sample set to obtain subsequences of different frequencies, extract the temporal features of each subsequence, and fuse the temporal features through a frequency attention mechanism to obtain multi-scale temporal features. The feature fusion module is used to perform feature alignment and weight fusion on the dynamic spatial features and the multi-scale temporal features through a fusion attention mechanism to obtain the cross-modal spatiotemporal fusion features; The carbon emission prediction module is used to input the cross-modal spatiotemporal fusion features into a trained carbon emission prediction model for processing, and output carbon emission prediction results and their corresponding uncertainty quantification indicators; wherein, the carbon emission prediction model is an adaptive fusion deep learning model containing multiple feature processing branches.

10. A non-volatile storage medium, characterized in that, The non-volatile storage medium includes a stored program, wherein, when the program is executed, it controls the device containing the non-volatile storage medium to execute the deep learning-based method for predicting carbon emissions from railway passenger stations as described in any one of claims 1 to 8.

Images