Method and system for predicting carbon emissions during large sporting events based on machine learning

Abstract

The application discloses a large-scale sports event carbon emission prediction method and system based on machine learning, comprising collecting multi-source data of a preset activity, preprocessing the multi-source data, and constructing a carbon emission space-time graph according to the multi-source data; performing space-time convolution coding on the carbon emission space-time graph to obtain node-level space-time features, obtaining node carbon emission prediction values according to the node-level space-time features and emission factors, performing regional split mapping according to the venue data to obtain regional clusters, obtaining carbon emission baseline clusters of the regional clusters through clustering based on historical event carbon emission data; aggregating the node carbon emission prediction values to obtain regional cluster carbon emissions, obtaining a region to be analyzed and a fixed region, and performing dynamic regulation and analysis in the region to be analyzed to obtain a regulation ratio; constructing a large-scale sports event carbon emission prediction model according to the carbon emission baseline clusters and the regulation ratio, and outputting a prediction result.

Description

Technical Field

[0001] This invention relates to the field of carbon emission prediction technology, and in particular to a method and system for predicting carbon emissions during major sporting events based on machine learning. Background Technology

[0002] As global climate change becomes increasingly severe, carbon emission control has become a focus of international attention. Large-scale sporting events are characterized by a large number of participants, wide distribution of venues, complex energy consumption types, and long durations. Their carbon emissions are enormous and highly dynamic in terms of spatiotemporal distribution. Accurately predicting carbon emissions during events is of great significance for formulating emission reduction strategies.

[0003] Existing carbon emission prediction methods mainly include emission factor-based accounting methods, life cycle assessment methods, and statistical regression models. The emission factor method calculates carbon emissions by multiplying the activity level by the emission factor. It is simple to calculate but difficult to capture spatiotemporal dynamic changes. The life cycle assessment method assesses carbon emissions throughout the entire life cycle from raw material acquisition to waste disposal. It is suitable for ex-post accounting but difficult to achieve real-time prediction. The statistical regression model establishes statistical relationships between variables based on historical data. It fits linear relationships well but is difficult to handle high-dimensional nonlinear correlations of multi-source heterogeneous data.

[0004] In recent years, spatiotemporal graph neural networks have been widely used in traffic flow prediction, energy load prediction, and other fields, modeling spatial relationships through graph structures and capturing temporal dynamics through time-series networks. Pointer generation networks have demonstrated the advantages of combining historical pattern replication and generation in text summarization and sequence prediction. However, these technologies have not yet been systematically applied to carbon emission prediction scenarios for large-scale sporting events.

[0005] Carbon emission prediction for large-scale sporting events faces the following unique challenges: First, there is the issue of multi-regional coordination, as venues are scattered across different geographical areas, and carbon emission linkages arise from energy network connections and personnel movement between regions. Second, there is the issue of spatiotemporal dynamism, as carbon emission patterns vary significantly across different stages of the event (preparation, competition, and dismantling), and are also affected by real-time factors such as weather conditions and spectator traffic. Third, there is the issue of historical data migration, as the venue layout, event scale, and climate conditions differ for each event, and directly applying historical data results in limited prediction accuracy. Fourth, there is the issue of a closed-loop system for prediction and control, as existing methods mostly focus on the prediction itself and lack a closed-loop mechanism to feed the prediction results back to real-time control strategies.

[0006] Therefore, there is an urgent need for a method to predict carbon emissions during large-scale sporting events that can integrate multi-source heterogeneous data, model spatiotemporal dynamic correlations, utilize historical pattern transfer, and support a prediction-regulation closed loop. Summary of the Invention

[0007] The purpose of this invention is to provide a method for predicting carbon emissions during large-scale sporting events based on machine learning.

[0008] To achieve the above objectives, the present invention is implemented according to the following technical solution: This invention includes the following steps: Collect multi-source data for pre-planned activities, preprocess the multi-source data, and construct a spatiotemporal map of carbon emissions based on the multi-source data; the multi-source data includes venue IoT data, event data, and meteorological data; Spatiotemporal convolutional encoding is performed on the carbon emission spatiotemporal map to obtain node-level spatiotemporal features. Based on the node-level spatiotemporal features and emission factors, the predicted values ​​of node carbon emissions are obtained. Based on the venue data, regional splitting and mapping are performed to obtain regional clusters. Based on the carbon emission data of historical events, carbon emission baseline clusters of regional clusters are obtained through clustering. The predicted carbon emissions of the nodes are aggregated into regional clusters to obtain regional cluster carbon emissions. Regional clusters whose carbon emissions exceed the confidence interval of the carbon emission baseline cluster are designated as regions to be analyzed, while regional clusters whose carbon emissions do not exceed the confidence interval of the carbon emission baseline cluster are designated as fixed regions. Dynamic regulation analysis is performed within the regions to be analyzed to obtain the regulation ratio. A carbon emission prediction model for large-scale sporting events is constructed based on the carbon emission baseline cluster and the control ratio. The data to be predicted is input into the carbon emission prediction model for large-scale sporting events, and the prediction results are output.

[0009] Furthermore, the method for constructing the aforementioned spatiotemporal map of carbon emissions includes: Based on the requirements and guidelines of the event sustainability management system, the carbon emission sources of large-scale sports events are abstracted as graph nodes to construct a spatiotemporal graph of the event's carbon emissions. The graph describes the spatial relationships between nodes; the nodes include basic energy consumption nodes, temporary energy consumption nodes, special energy consumption nodes, transportation nodes, accommodation nodes, renewable energy nodes, and carbon sequestration nodes; among them For a set of nodes, Let be the set of edges. The node feature matrix, The adjacency matrix is ​​used; the edge set includes physical connection edges, functional coupling edges, and causal influence edges; physical connection edges: pipeline connections between venues and transportation hubs; functional coupling edges: functional similarity between nodes of the same type; causal influence edges: based on Granger causality test, if the historical carbon emissions of node i have predictive power for node j, then a directed edge is established. Based on the risk and distance dual-factor edge weighting mechanism, dynamic edge weights for the carbon emission spatiotemporal graph are designed. Historical carbon emission sequences of nodes i and j , Perform Granger causality testing: ; If you refuse Then a directed edge is established. causal strength is ;in Let i be the physical distance between node i and node j. Let the angle be the functional similarity angle between node i and node j. Let be the weight of the physical connection edge between node i and node j. Let be the carbon emissions of node j at time t. The regression coefficient vector of the lag term at node j is the object of the overall significance test. The lag order is... It is a lagging index. Let be the autoregressive coefficient of node i after lags by k steps. Let be the cross-regression coefficient of node j with respect to node i, lagged by k steps. For the residual term, Let be the causal strength of node j to node i; Introducing a carbon emission propagation resistance factor to adjust dynamic side weights in real time ;in As a carbon emission propagation resistance factor, ,when ,but , Let be the real-time power flow rate from node i to j at time t. Let be the rate of change of population flow density from node i to j at time t. Let t represent the fluctuation of meteorological conditions from node i to j at time t. As the weight for power flow change, Weighting for changes in population mobility density Weighting for changes in meteorological conditions, The causal sensitivity coefficient This is the resistance penalty coefficient.

[0010] Furthermore, the method for obtaining the node-level spatiotemporal features includes: Using graph convolution to aggregate neighbor node information Gated temporal convolutions are run in parallel on each node. It employs spatiotemporal graph convolution, alternatingly stacking graph convolutional layers and temporal convolutional layers to form a spatiotemporal joint representation. ;in For degree matrix, Let be the dynamic adjacency matrix containing self-loops at time t. For the first The input node feature matrix of the layer graph convolution, For the first The input node feature matrix of the layer graph convolution, It is a non-linear activation function. Let be the dynamic adjacency matrix without self-loops at time t. It is the identity matrix. For the first The learnable weight matrix of the layer, Let i be the temporal coding feature of node i at time t. The number of historical event steps in the temporal convolution. For the set of learnable parameters of the gated network, For gating networks, For node i from time... At the time The input feature sequence, For spatiotemporal convolutional networks, The length of the time series. For the input tensor, It is a time-varying adjacency matrix sequence; Introducing dual-axis attention (time and feature): Time-axis attention calculates the importance of each time step for temporal features. Output time-weighted features ;in Let i be the attention weight of node i at time t. Let be the attention score of node i at time t. For query vectors related to temporal attention, Here is the weight matrix for time attention. For the bias vector of temporal attention, The spatiotemporal convolutional network outputs features for node i at time t. For summation index variables; Feature axis attention: calculating the importance of each dimension of time-weighted features. Output feature weighted representation ,Will The output is node-level spatiotemporal features; dual-axis attention is fused. Automatically focus on key time steps and key features; in The feature attention score for the d-th feature dimension of node i. Let be the feature attention weight of the d-th feature dimension of node i. Here is the weight matrix for feature attention. For the bias of feature attention, The total number of feature dimensions. For summation index variables, The query vector for feature attention. Time-weighted features of node i The d-th component, Let d be the standard basis vector of the d-th dimension. For node-level spatiotemporal features, For timeline attention, For feature axis attention, Let i be the temporal characteristics of node i.

[0011] Furthermore, the method for obtaining the predicted carbon emissions of the node includes: Carbon emission factors and node-level spatiotemporal features are obtained using a learnable emission factor mapping gate. Adaptive selection of key emission factors, when When the standard emission factor is used as the emission factor, When the characteristic mapping emission factor is used as the emission factor, the emission factor is used as the emission factor. For adaptive emission factors, For activation function, For gated bias. For the gated weight matrix, For adaptive gating vectors, The feature-factor mapping matrix, Let be the emission factor of node i; Predicted activity levels Calculate carbon emission forecasts The carbon emission forecasts are compared with the baseline cluster center to calculate the deviation. The deviation is used as a feedback signal to correct the carbon emission prediction value through residual connection. ;in Let be the predicted activity level vector of node i at time t. This is the predicted value for the activity level of item f. Let f be the adaptive emission factor. The carbon emission forecast is corrected for time t. To correct the multilayer perceptron, Multilayer perceptron for predicting activities.

[0012] Furthermore, the method for obtaining the regional cluster and the carbon emission baseline cluster includes: The event's regional clusters are obtained by dual splitting and mapping based on the spatial and functional attributes of the venue data. Spatial attributes: Based on the venue's geographical coordinates, density-based clustering is used to group venues with a physical distance of less than or equal to 2km into one cluster. Functional attributes: Classified according to energy consumption type. Based on the transportation network topology, venues sharing major transportation hubs are grouped into one cluster. The intersection of the spatial clusters and functional clusters is taken to form a refined regional cluster under dual constraints of space and function, and the output is the regional cluster. Extract time-series carbon emission data at the regional cluster level from historical events, and use isolated forests to score event feature vectors for anomalies, removing events with scores below the scoring threshold τ; the event feature vectors include total carbon emissions, peak emissions, emission volatility, and meteorological deviation. Based on the overall characteristic vector of historical events Hierarchical clustering is performed to obtain event type clusters; dynamic time-warped distance is used for distance measurement. ;in To align the path, For local distance, This is the first time series. This is the second time series. Time series The a-th element, Time series The s-th element, For dynamic time-normalized distance, For the event's total carbon emissions, The peak carbon emissions for event m, The standard deviation of carbon emissions for the event. The number of days the event lasts. The scale of the event is m. Within each event type cluster, the carbon emission sequences of the regional clusters are further subdivided and clustered according to the event stages. Time windows are defined based on the event stages to obtain the regional cluster stage combinations. Shape features of the carbon emission series are extracted, and K-means clustering is performed on these features. The regional cluster stage combinations are then matched with the most similar baseline clusters to obtain the carbon emission baseline clusters, and the centers of these baseline clusters are defined. and baseline cluster boundary ;in The center of the s-th baseline cluster, Assign event type cluster numbers, Assign a cluster number to the region. This indicates the stage of the competition. Number the baseline clusters. For the baseline cluster set, The cardinality of the baseline cluster, For historical events In regional clusters ,stage The carbon emission sequence, The baseline is the standard deviation within the cluster. For the first 95% confidence intervals for each baseline cluster.

[0013] Furthermore, the method for obtaining the carbon emissions of the said regional cluster includes: The predicted carbon emissions of all nodes within a regional cluster are aggregated, and all regional clusters in the event are aggregated to obtain the initial regional cluster carbon emissions. We acquire spatiotemporal pattern samples of historical events and construct a pointer library. The historical pattern samples contain runaway context encoding, total carbon emission sequence, and metadata. The metadata includes event type, scale, climate conditions, and duration. Given the query code of the current event Calculate the similarity score between the current event and each historical pattern in the pointer library. Calculate the pointer probability distribution ;in This is the global query vector for the event. This is a global pooling operation. The number of event nodes, For the event and historical mode Attention score For the query vector of the pointer network, The weight matrix for encoding the event. The weight matrix encoding the historical pattern. This is the bias vector of the pointer network. This represents the total number of historical events in the pointer library. Encoding the spatiotemporal context of historical events; Calculate the generation probability of carbon emission predictions Self-predictive component generation Weighted replication of historical patterns Correcting initial regional cluster carbon emissions And output as regional cluster carbon emissions; among which For generation probability, To generate the gating weight matrix, To generate the bias for gating. To select historical mode The probability as a reference For the corrected initial regional cluster carbon emissions, This is the predicted value of carbon emissions. The weighted carbon emission values ​​are replicated from historical patterns. For historical mode Total carbon emissions at time t.

[0014] Furthermore, the method for obtaining the control ratio by performing dynamic control analysis within the region to be analyzed includes: Construct state and action spaces, and use linear scalarization to transform the bi-objective Pareto optimization function into a single-objective optimization problem. The objective function is: Combine multiple objectives into a scalar reward function The data is input into a network of actors and critics for strategy optimization; among them... Let t be the predicted value of carbon emissions from the regional cluster. To account for the deviation from the baseline cluster, Let be the state space, and be the current cumulative carbon emission target difference. Let the meteorological conditions be vectors. For the action space, This is a real-time pedestrian density vector. For the proportion of shuttle bus frequency control, For air conditioner power control ratio, For lighting power control ratio, This represents the percentage of time offset in the competition schedule. Adjusting the indoor target temperature Given the current schedule, For carbon emission losses, To perform the action The predicted total carbon emissions are as follows. The target value for carbon emissions of the event, As a loss of the event experience, For the first The weighting coefficients of each experience dimension For experience dimension index, Let be the system state vector at time t. Let be the control action vector at time t. For motion smoothness loss term, As a weight for carbon emission losses, To reduce the weight of the event experience, The motion smoothing loss weights are used; the meteorological condition vector includes temperature, humidity, and precipitation; the experience dimensions are thermal comfort penalty, traffic waiting penalty, visual visibility penalty, schedule conflict penalty, and crowd safety penalty. Given an actor network, motion-constrained projection, and a critic network, a dominance function is given based on the temporal difference error. Update policy gradient The output is the adjusted action, and the adjustment ratio is output based on the adjusted action; whereby... For actors' networks, For motion constraint projection, For the critics' network, , Estimate the dominance function at time t. As a discount factor, The instant reward at time t, For state The expected cumulative return estimate, For value networks, For policy networks, The action mean vector, Let be the action covariance matrix. The distribution is a multivariate Gaussian distribution; the control ratios include the reduction ratio of shuttle bus frequency, the adjustment ratio of air conditioning power, the adjustment ratio of lighting power, the offset ratio of race schedule time, and the adjustment amount of target temperature.

[0015] Furthermore, the method for constructing a carbon emission prediction model during the aforementioned large-scale sporting event includes: The regulation ratio is used as a dynamic correction factor for the baseline cluster. Calibrate the control influence coefficient Based on baseline cluster confidence intervals Adjustments and corrections were made; among them Center of the original baseline cluster, The proportion of the c-th regulatory dimension. In order to regulate the influence coefficient, To control the action vector, For the number of historical control experiments, The carbon emission observation value before the nth experimental adjustment is given. The carbon emission observation value after the nth experimental adjustment is given. Let c be the magnitude of the regulation in the nth experiment. To adjust the standard deviation, To control the uncertainty amplification factor, To regulate the L2 norm of the action, For carbon emissions, To adjust the corrected baseline mean, Let be the static standard deviation of the baseline cluster at time t without adjustment. It is a positive minimum value; The carbon emission prediction model during major sporting events includes an input encoding layer, a spatiotemporal graph convolutional prediction layer, and a pointer generation network correction layer. Input coding layer: State coding ;in For fusion state encoding, For state vectors, It is a multilayer perceptron encoder. To introduce the adjustment and correction time t, given conditions The 95% confidence interval of the carbon emission baseline cluster; Spatiotemporal graph convolutional prediction layer: embeds the modified baseline clusters into graph node features. Convolutional dynamic graph Regulating the dynamic adjacency matrix of perception ;in For the enhanced features of node i at time t, For the feature at time node i, The cluster number of the region to which node i belongs. To regulate the synergistic attenuation coefficient, The control category to which node i belongs. For the first The node feature matrix of the layer, For the first +1 layer node feature matrix For degree matrix, For the first The trainable weight parameter matrix of layered graph convolution. It is a non-linear activation function. This is an indicator function, which takes the value 1 when the node carbon emission deviation exceeds the regulation correction confidence interval, and 0 otherwise. Pointer Generation Network Correction Layer: Extends the pointer library to a joint pointer library for regulation and baseline. Adjusting the probability of similarity-weighted pointers By generating and replicating a mixed output of carbon emission predictions ;in This refers to the sequence of control actions actually executed in historical events. The baseline cluster deformation of historical event m, To adjust the difference weight matrix, This represents the difference vector between current and historical regulations. Encoding the spatiotemporal context of historical event m, The total carbon emissions sequence of historical events m The total number of historical events in the pointer library. Weighted attention score is calculated based on the similarity between the current event and historical pattern m. For the query vector of the pointer network, This is the global query vector for the current event. The weight matrix for encoding the current event. The weight matrix encoding the historical pattern. This is the bias vector of the pointer network. The hyperbolic tangent activation function is used. This is the predicted regional cluster carbon emissions after correction at time t+1. For generation probability, To generate the network, For the replication probability, To select historical mode pointer probability, This represents the historical carbon emission value of pattern m at time t+1. To regulate the intensity ratio; Loss function for constructing a carbon emission prediction model during major sporting events A carbon emission prediction model for major sporting events is trained based on a loss function until the maximum number of iterations is reached, at which point training stops. To predict losses, For baseline cluster constraint loss, For L2 regularization, For baseline cluster weights, For regularization weights.

[0016] Secondly, a machine learning-based carbon emission prediction system for large-scale sporting events includes: Data acquisition, processing, and spatiotemporal map construction module: This module is used to collect multi-source data for preset activities, preprocess the multi-source data, and construct a carbon emission spatiotemporal map based on the multi-source data; the multi-source data includes venue IoT data, event data, and meteorological data. Carbon emission prediction and baseline cluster acquisition module: used to perform spatiotemporal convolutional encoding on the carbon emission spatiotemporal map to obtain node-level spatiotemporal features, obtain node carbon emission prediction values ​​based on the node-level spatiotemporal features and emission factors, obtain regional clusters by performing regional splitting and mapping based on the venue data, and obtain carbon emission baseline clusters of regional clusters by clustering based on historical event carbon emission data. The regional cluster aggregation and dynamic regulation analysis module is used to aggregate the predicted carbon emissions of the nodes into regional clusters to obtain regional cluster carbon emissions. Regional clusters whose carbon emissions exceed the confidence interval of the carbon emission baseline cluster are designated as regions to be analyzed, while regional clusters whose carbon emissions do not exceed the confidence interval of the carbon emission baseline cluster are designated as fixed regions. Dynamic regulation analysis is performed within the regions to be analyzed to obtain the regulation ratio. Model building and output module: used to build a carbon emission prediction model during large-scale sporting events based on the carbon emission baseline cluster and the control ratio, input the data to be predicted into the carbon emission prediction model during large-scale sporting events, and output the prediction results.

[0017] The beneficial effects of this invention are: This invention relates to a method and system for predicting carbon emissions during large-scale sporting events based on machine learning. Compared with existing technologies, this invention has the following technical advantages: This invention improves the accuracy of capturing multi-regional carbon emission correlation features by constructing a spatiotemporal graph of carbon emissions, modeling multi-dimensional relationships between nodes using physical connection edges, functional coupling edges, and causal influence edges, and introducing a carbon emission propagation resistance factor to dynamically adjust edge weights. It also enhances the spatiotemporal resolution of carbon emission prediction by automatically focusing on key time steps and feature dimensions through spatiotemporal graph convolutional coding combined with a time-feature dual-axis attention mechanism. Furthermore, it enhances the adaptability of emission factors by dynamically selecting standard emission factors or feature-mapped emission factors through an adaptive emission factor mapping gate. Finally, it establishes a refined regional division and historical benchmark reference system under spatial-functional dual constraints through dual regional cluster splitting mapping and baseline cluster hierarchical clustering, improving the interpretability and credibility of predictions. Additionally, it enhances the ability to transfer historical patterns by matching historical similar event patterns through a pointer generation network and using a hybrid generation and replication mechanism to correct initial predictions. Finally, it achieves closed-loop optimization of carbon emission prediction and emission reduction strategies by embedding the control ratio as a dynamic correction factor for baseline clusters into the prediction model through bi-objective Pareto optimization and reinforcement learning dynamic regulation analysis. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the steps of the machine learning-based carbon emission prediction method for large-scale sporting events according to the present invention. Detailed Implementation

[0019] The present invention will be further described below through specific embodiments. The illustrative embodiments and descriptions herein are used to explain the present invention, but are not intended to limit the present invention.

[0020] The present invention provides a method and system for predicting carbon emissions during large-scale sporting events based on machine learning, comprising the following steps: like Figure 1 As shown, this embodiment includes the following steps: Collect multi-source data for pre-planned activities, preprocess the multi-source data, and construct a spatiotemporal map of carbon emissions based on the multi-source data; the multi-source data includes venue IoT data, event data, and meteorological data; In the actual assessment, a comprehensive sports event was used as the prediction target. This event lasted for 17 days and involved 3 host cities, 12 competition venues, 6 training venues, 3 athletes' villages and 2 media centers. The collected multi-source data includes: Venue IoT data: Real-time energy consumption data is collected at 15-minute intervals using smart meters, gas meters, water meters, and renewable energy metering devices deployed in various venues. Data fields include: venue ID, timestamp, electricity consumption, natural gas consumption, renewable energy generation, and cooling / heating supply. Data is transmitted to the edge computing gateway via the MQTT protocol and aggregated to the central data lake via a Kafka message queue. Event data: Event schedules, participant numbers, spectator ticketing data, and transportation plans are obtained from the event organizing committee's information system. Specifically, this includes: daily event sessions, estimated spectator numbers (by time slot), number of athletes and team officials, number of registered media, and official vehicle dispatch plans (shuttle routes, departure frequency, vehicle type, and passenger capacity). Meteorological data: Hourly weather forecast data for the area where each venue is located is obtained through access to the National Meteorological Information Center's API, including: ambient temperature, relative humidity, wind speed, solar radiation intensity, and precipitation probability. Preprocessing: Missing values ​​were filled using linear interpolation, and time periods with a missing value ratio exceeding 5% were marked as abnormal; peak smoothing was performed on power consumption data, and the 3σ criterion was used to remove abnormal readings caused by equipment failure; event data and energy consumption data were aligned by timestamp and unified to a 15-minute granularity; spatial interpolation was performed on meteorological data, and the inverse distance weighting method was used to map the station data to the coordinates of each venue. Spatiotemporal convolutional encoding is performed on the carbon emission spatiotemporal map to obtain node-level spatiotemporal features. Based on the node-level spatiotemporal features and emission factors, the predicted values ​​of node carbon emissions are obtained. Based on the venue data, regional splitting and mapping are performed to obtain regional clusters. Based on the carbon emission data of historical events, carbon emission baseline clusters of regional clusters are obtained through clustering. The predicted carbon emissions of the nodes are aggregated into regional clusters to obtain regional cluster carbon emissions. Regional clusters whose carbon emissions exceed the confidence interval of the carbon emission baseline cluster are designated as regions to be analyzed, while regional clusters whose carbon emissions do not exceed the confidence interval of the carbon emission baseline cluster are designated as fixed regions. Dynamic regulation analysis is performed within the regions to be analyzed to obtain the regulation ratio. A carbon emission prediction model for large-scale sporting events is constructed based on the carbon emission baseline cluster and the control ratio. The data to be predicted is input into the carbon emission prediction model for large-scale sporting events, and the prediction results are output.

[0021] In this embodiment, the method for constructing the spatiotemporal map of carbon emissions includes: Based on the requirements and guidelines of the event sustainability management system, the carbon emission sources of large-scale sports events are abstracted as graph nodes to construct a spatiotemporal graph of the event's carbon emissions. The graph describes the spatial relationships between nodes; the nodes include basic energy consumption nodes, temporary energy consumption nodes, special energy consumption nodes, transportation nodes, accommodation nodes, renewable energy nodes, and carbon sequestration nodes; among them For a set of nodes, Let be the set of edges. The node feature matrix, The adjacency matrix is ​​used; the edge set includes physical connection edges, functional coupling edges, and causal influence edges; physical connection edges: pipeline connections between venues and transportation hubs; functional coupling edges: functional similarity between nodes of the same type; causal influence edges: based on Granger causality test, if the historical carbon emissions of node i have predictive power for node j, then a directed edge is established. Based on the risk and distance dual-factor edge weighting mechanism, dynamic edge weights for the carbon emission spatiotemporal graph are designed. Historical carbon emission sequences of nodes i and j , Perform Granger causality testing: ; If you refuse Then a directed edge is established. causal strength is ;in Let i be the physical distance between node i and node j. Let the angle be the functional similarity angle between node i and node j. Let be the weight of the physical connection edge between node i and node j. Let be the carbon emissions of node j at time t. The regression coefficient vector of the lag term at node j is the object of the overall significance test. The lag order is... It is a lagging index. Let be the autoregressive coefficient of node i after lags by k steps. Let be the cross-regression coefficient of node j with respect to node i, lagged by k steps. For the residual term, Let be the causal strength of node j to node i; Introducing a carbon emission propagation resistance factor to adjust dynamic side weights in real time ;in As a carbon emission propagation resistance factor, Let be the real-time power flow rate from node i to j at time t. ,when ,but , Let be the rate of change of population flow density from node i to j at time t. Let t represent the fluctuation of meteorological conditions from node i to j at time t. As the weight for power flow change, Weighting for changes in population mobility density Weighting for changes in meteorological conditions, The causal sensitivity coefficient This is the resistance penalty coefficient; In actual evaluation, the lag order is automatically selected using the Bayesian information criterion. , The length of the time series. This is the residual variance estimate for the lag order m; Node set V contains a total of 89 nodes, categorized as follows: Basic energy consumption nodes (42): lighting systems, air conditioning systems, and elevator systems in each venue; Temporary energy consumption nodes (18): temporary broadcasting equipment, security equipment, and temporary catering facilities set up for the event; Special energy consumption nodes (6): ice-making systems (ice hockey rink, figure skating rink), and swimming pool temperature control systems; Transportation nodes (12): official shuttle bus routes between venues and subway connection stations; Accommodation nodes (8): athletes' village, media hotel, and technical officials' hotel; Renewable energy nodes (2): rooftop photovoltaic systems and wind power generation devices in the venues; Carbon sequestration node (1): the official carbon sink forest project for the event; Physical connection edges: Based on the urban energy pipeline network topology map, establish connections between the stadium and the substation and natural gas pressure regulating station; for example, if the physical distance between the main stadium and the 110kV substation is 2.3km and the functional similarity angle is 0, then the weight of the physical connection edge is 0.217. Functional coupling edge: Establish functional coupling edges between nodes of the same type; Causal influence edge: For example, if the functional similarity angle between the main stadium lighting system and the sub-venue lighting system is 15° and the physical distance is 8.5km, then the weight of the functional coupling edge is 0.115; Granger causality tests were performed on the carbon emission sequences at each stage of historical events (events of the same type). Taking the main stadium and media center as examples, with a lag order of 3, the test statistic was 4.82, p=0.003<0.05, therefore the null hypothesis was rejected. The causal strength of the directed edge is 0.24. Taking 14:00 on July 28, 2024 as an example, the real-time power flow change rate from the main stadium to the sub-venue was 0.15 (due to power load transfer caused by the broadcast of the event), the personnel flow density change rate was 0.32 (due to the transfer of spectators), and the meteorological condition fluctuation was 0.08 (due to localized thunderstorms); It is 0.4. It is 0.35. If it is 0.25, then The calculated value is 0.172, the causality sensitivity coefficient is 0.6, and the resistance penalty coefficient is 0.4. Therefore, the dynamic edge weight... It is 0.141.

[0022] In this embodiment, the method for obtaining the node-level spatiotemporal features includes: Using graph convolution to aggregate neighbor node information Gated temporal convolutions are run in parallel on each node. It employs spatiotemporal graph convolution, alternatingly stacking graph convolutional layers and temporal convolutional layers to form a spatiotemporal joint representation. ;in For degree matrix, Let be the dynamic adjacency matrix containing self-loops at time t. For the first The input node feature matrix of the layer graph convolution, For the first The input node feature matrix of the layer graph convolution, It is a non-linear activation function. Let be the dynamic adjacency matrix without self-loops at time t. It is the identity matrix. For the first The learnable weight matrix of the layer, Let i be the temporal coding feature of node i at time t. The number of historical event steps in the temporal convolution. For the set of learnable parameters of the gated network, For gating networks, For node i from time... At the time The input feature sequence, For spatiotemporal convolutional networks, The length of the time series. For the input tensor, It is a time-varying adjacency matrix sequence; Introducing dual-axis attention (time and feature): Time-axis attention calculates the importance of each time step for temporal features. Output time-weighted features ;in Let i be the attention weight of node i at time t. Let be the attention score of node i at time t. For query vectors related to temporal attention, Here is the weight matrix for time attention. For the bias vector of temporal attention, The spatiotemporal convolutional network outputs features for node i at time t. For summation index variables; Feature axis attention: calculating the importance of each dimension of time-weighted features. Output feature weighted representation ,Will The output is node-level spatiotemporal features; dual-axis attention is fused. Automatically focus on key time steps and key features; among which The feature attention score for the d-th feature dimension of node i. Let be the feature attention weight of the d-th feature dimension of node i. Here is the weight matrix for feature attention. For the bias of feature attention, The total number of feature dimensions. For summation index variables, The query vector for feature attention. Time-weighted features of node i The d-th component, Let d be the standard basis vector of the d-th dimension. For node-level spatiotemporal features, For timeline attention, For feature axis attention, Let i be the temporal characteristics of node i; In the actual evaluation, a 3-layer spatiotemporal graph convolutional network was adopted, alternating between graph convolutional layers and temporal convolutional layers. The graph convolutional layers employed spectral domain graph convolution with normalized adjacency and degree matrices. Learnable weight matrices were used. The initial layer input feature dimension was 32 (including energy consumption, weather, and event activity levels), and the output dimension after the 3-layer graph convolution was 64. The temporal convolutional layers ran gated temporal convolution in parallel on each node, with 96 historical time steps (corresponding to 24 hours and 15-minute granularity). Dilated causal convolution was used, with a dilation factor d∈{1,2,4,8} and a kernel size of 3. Timeline attention: attention to nodes Calculate attention scores for the 96 time steps of the 5th day of the competition; Feature axis attention: [Further details needed for accurate translation] The importance of the nodes is calculated based on 64 dimensions; the node-level spatiotemporal features are output with 64 dimensions.

[0023] In this embodiment, the method for obtaining the predicted value of the node carbon emissions includes: Carbon emission factors and node-level spatiotemporal features are obtained using a learnable emission factor mapping gate. Adaptive selection of key emission factors, when When the standard emission factor is used as the emission factor, When the characteristic mapping emission factor is used as the emission factor, the emission factor is used as the emission factor. For adaptive emission factors, For activation function, For gated bias. For the gated weight matrix, For adaptive gating vectors, The feature-factor mapping matrix, Let be the emission factor of node i; Predicted activity levels Calculate carbon emission forecasts The carbon emission forecasts are compared with the baseline cluster center to calculate the deviation. The deviation is used as a feedback signal to correct the carbon emission prediction value through residual connection. ;in Let be the predicted activity level vector of node i at time t. This is the predicted value for the activity level of item f. Let f be the adaptive emission factor. The carbon emission forecast is corrected for time t. To correct the multilayer perceptron, For predicting activity, a multilayer perceptron; In actual assessment, nodes The standard emission factor is 0.5708 kgCO2 / kWh, the adaptive gating vector is 0.78, close to 1, and the adaptive emission level is 0.58 kgCO2 / kWh. Activity levels are predicted using MLP, and the output is a 5-dimensional vector (lighting, air conditioning, equipment, traffic, and others). Taking t=14:00 as an example... The predicted carbon emission value is 1374.5 kg CO2, with a deviation of 124.5 kg CO2. After correction by residual connection, the final predicted carbon emission value is 1356.2 kg CO2.

[0024] In this embodiment, the method for obtaining the regional cluster and the carbon emission baseline cluster includes: The event's regional clusters are obtained by dual splitting and mapping based on the spatial and functional attributes of the venue data. Spatial attributes: Based on the venue's geographical coordinates, density-based clustering is used to group venues with a physical distance of less than or equal to 2km into one cluster. Functional attributes: Classified according to energy consumption type. Based on the transportation network topology, venues sharing major transportation hubs are grouped into one cluster. The intersection of the spatial clusters and functional clusters is taken to form a refined regional cluster under dual constraints of space and function, and the output is the regional cluster. Extract time-series carbon emission data at the regional cluster level from historical events, and use isolated forests to score event feature vectors for anomalies, removing events with scores below the scoring threshold τ; the event feature vectors include total carbon emissions, peak emissions, emission volatility, and meteorological deviation. Based on the overall characteristic vector of historical events Hierarchical clustering is performed to obtain event type clusters; dynamic time-warped distance is used for distance measurement. ;in To align the path, For local distance, This is the first time series. This is the second time series. Time series The a-th element, Time series The s-th element, For dynamic time-normalized distance, For the event's total carbon emissions, The peak carbon emissions for event m, The standard deviation of carbon emissions for the event. The number of days the event lasts. The scale of the event is m. Within each event type cluster, the carbon emission sequences of the regional clusters are further subdivided and clustered according to the event stages. Time windows are defined based on the event stages to obtain the regional cluster stage combinations. Shape features of the carbon emission series are extracted, and K-means clustering is performed on these features. The regional cluster stage combinations are then matched with the most similar baseline clusters to obtain the carbon emission baseline clusters, and the centers of these baseline clusters are defined. and baseline cluster boundary ;in The center of the s-th baseline cluster, Assign event type cluster numbers, Assign a cluster number to the region. This indicates the stage of the competition. Number the baseline clusters. For the baseline cluster set, The cardinality of the baseline cluster, For historical events In regional clusters ,stage The carbon emission sequence, The baseline is the standard deviation within the cluster. For the first 95% confidence intervals for each baseline cluster; In actual assessment, the data structure of the carbon emission baseline cluster includes a center vector, standard deviation, confidence interval, cluster number of the event type, cluster number of the region, and event stage; the scoring threshold τ is adaptively determined based on the abnormal scoring distribution of the isolated forest, taking the 5th percentile of all historical event abnormal scores; Spatial attribute clustering: Based on the DBSCAN algorithm, the geographical coordinates (WGS-84 coordinate system) of 12 competition venues were clustered with a neighborhood radius of ϵ=2km and a minimum sample size of 2; 3 spatial clusters were obtained: City A core competition area (5 venues), City B coastal competition area (4 venues), and City C mountain competition area (3 venues). Functional attribute clustering: Based on energy consumption type, venues sharing major transportation hubs (airports / high-speed rail stations) are grouped into one cluster; all 5 venues in City A are connected to the airport via Metro Line 1, forming a functional cluster; taking the intersection, the core competition area of ​​City A is the refined area cluster R1, which includes 5 venues and 23 nodes; Carbon emission time-series data of R1 were extracted from two similar events in 2018 and 2022. Anomaly detection was performed based on isolated forests, with 100 isolated trees and an anomaly scoring threshold τ set at the 5th percentile (τ=−0.42). One event was removed due to excessive air conditioning load caused by extreme high temperatures. The remaining events are hierarchically clustered based on feature vectors, with DTW distance as the distance metric; two event type clusters are obtained: summer event cluster (g=1, 3 events) and transition season event cluster (g=2, 2 events). This event belongs to the summer event cluster. Within g=1, the carbon emission sequence of R1 is subdivided according to the event stage (preparation period, event period, and dismantling period). The time window for the event period is defined as the first day after the opening of the event to the closing day. The shape features of the carbon emission sequence (including peak position, peak-to-valley ratio, and daily fluctuation coefficient) are extracted, and K-means clustering (K=3) is used to obtain 3 baseline clusters. The baseline cluster s=2 is matched on the 5th day of the event period. The center of the baseline cluster is 1180kgCO2 / node / 15min, the standard deviation is 85kgCO2, and the 95% confidence interval is [1.13.4, 1346.6].

[0025] In this embodiment, the method for obtaining the carbon emissions of the regional cluster includes: The predicted carbon emissions of all nodes within a regional cluster are aggregated, and all regional clusters in the event are aggregated to obtain the initial regional cluster carbon emissions. We acquire spatiotemporal pattern samples of historical events and construct a pointer library. The historical pattern samples contain runaway context encoding, total carbon emission sequence, and metadata. The metadata includes event type, scale, climate conditions, and duration. Given the query code of the current event Calculate the similarity score between the current event and each historical pattern in the pointer library. Calculate the pointer probability distribution ;in This is the global query vector for the event. This is a global pooling operation. The number of event nodes, For the event and historical mode Attention score For the query vector of the pointer network, The weight matrix for encoding the event. The weight matrix encoding the historical pattern. This is the bias vector of the pointer network. This represents the total number of historical events in the pointer library. Encoding the spatiotemporal context of historical events; Calculate the generation probability of carbon emission predictions Self-predictive component generation Weighted replication of historical patterns Correcting initial regional cluster carbon emissions And output as regional cluster carbon emissions; among which For generation probability, To generate the gating weight matrix, To generate the bias for gating. To select historical mode The probability as a reference For the corrected initial regional cluster carbon emissions, This is the predicted value of carbon emissions. The weighted carbon emission values ​​are replicated from historical patterns. For historical mode Total carbon emissions at time t; In the actual assessment, the runaway mode is characterized by carbon emission deviations exceeding the baseline cluster confidence interval and persisting beyond the threshold duration. The corrected carbon emission predictions for the 23 nodes within R1 are summed to obtain an initial regional cluster carbon emission of wei = 31245.6 kgCO2. A pointer library P containing spatiotemporal pattern samples from 5 historical events is constructed. The current event query encoding is used, and a 128-dimensional vector is obtained through global average pooling. The similarity score with the historical pattern m=3 (event in the same city in 2022) is calculated to be 2.34, with a pointer probability of 0.42, a generation probability of 0.35, a generation component of 29800 kgCO2, and a replication component of 31420 kgCO2. The corrected regional cluster carbon emission is 30853 kgCO2.

[0026] In this embodiment, the method for obtaining the control ratio by performing dynamic control analysis within the region to be analyzed includes: Construct state and action spaces, and use linear scalarization to transform the bi-objective Pareto optimization function into a single-objective optimization problem. The objective function is: Combine multiple objectives into a scalar reward function The data is input into a network of actors and critics for strategy optimization; among them... Let t be the predicted value of carbon emissions from the regional cluster. To account for the deviation from the baseline cluster, Let be the state space, and be the current cumulative carbon emission target difference. Let the meteorological conditions be vectors. For the action space, This is a real-time pedestrian density vector. For the proportion of shuttle bus frequency control, For air conditioner power control ratio, For lighting power control ratio, This represents the percentage of time offset in the competition schedule. Adjusting the indoor target temperature Given the current schedule, For carbon emission losses, To perform the action The predicted total carbon emissions are as follows. The target value for carbon emissions of the event, As a loss of the event experience, For the first The weighting coefficients of each experience dimension For experience dimension index, Let be the system state vector at time t. Let be the control action vector at time t. For motion smoothness loss term, As a weight for carbon emission losses, To reduce the weight of the event experience, The motion smoothing loss weights are used; the meteorological condition vector includes temperature, humidity, and precipitation; the experience dimensions are thermal comfort penalty, traffic waiting penalty, visual visibility penalty, schedule conflict penalty, and crowd safety penalty. Given an actor network, motion-constrained projection, and a critic network, a dominance function is given based on the temporal difference error. Update policy gradient The output is the adjusted action, and the adjustment ratio is output based on the adjusted action; whereby... For actors' networks, For motion constraint projection, For the critics' network, , Estimate the dominance function at time t. As a discount factor, The instant reward at time t, For state The expected cumulative return estimate, For value networks, For policy networks, The action mean vector, Let be the action covariance matrix. The distribution is a multivariate Gaussian distribution; the control ratios include the reduction ratio of shuttle bus frequency, the adjustment ratio of air conditioning power, the adjustment ratio of lighting power, the offset ratio of race schedule time, and the adjustment amount of target temperature; In actual assessments, the percentage reduction in shuttle bus frequency = original frequency × (1 + ... Air conditioner power adjustment ratio = rated power × (1 + ... Lighting power adjustment ratio = original lighting × (1 + The percentage of time offset in the race schedule = original time + total duration × Target temperature adjustment amount: original set temperature + ; The value ranges for each dimension of the action space are determined based on the physical constraints of the venue equipment and the event operation specifications: the maximum reduction in shuttle bus frequency is -50% to ensure minimum transport capacity; the maximum adjustment of air conditioning power is +10% to prevent equipment overload; the maximum reduction in lighting power is -40% to meet safety illuminance standards; the maximum deviation of event schedule time is ±20% to ensure event continuity; and the indoor target temperature adjustment range is +1℃ to +3℃, in line with the human thermal comfort standard ISO 7730. On the fifth day of the competition, compared to previous events, the figure skating venue (medium-temperature ice surface cluster) for the target event implemented adjustments: lighting power was reduced by 12%, and the competition schedule was moved up by 2 hours to utilize off-peak electricity for ice making at night. The predicted carbon emissions were 18,500 kg CO2, while the actual observed emissions were 18,200 kg CO2, which was 1.6% lower than the prediction. The ice surface temperature was maintained at -2.5 ± 0.2 degrees Celsius, which met the competition standards.

[0027] In this embodiment, the method for constructing a carbon emission prediction model during the large-scale sporting event includes: The regulation ratio is used as a dynamic correction factor for the baseline cluster. Calibrate the control influence coefficient Based on baseline cluster confidence intervals Adjustments and corrections were made; among them Center of the original baseline cluster, The proportion of the c-th regulatory dimension. In order to regulate the influence coefficient, To control the action vector, For the number of historical control experiments, The carbon emission observation value before the nth experimental adjustment is given. The carbon emission observation value after the nth experimental adjustment is given. Let c be the magnitude of the regulation in the nth experiment. To adjust the standard deviation, To control the uncertainty amplification factor, To regulate the L2 norm of the action, For carbon emissions, To adjust the corrected baseline mean, Let be the static standard deviation of the baseline cluster at time t without adjustment. Positive minimum value; The carbon emission prediction model during major sporting events includes an input encoding layer, a spatiotemporal graph convolutional prediction layer, and a pointer generation network correction layer. Input coding layer: State coding ;in For fusion state encoding, For state vectors, It is a multilayer perceptron encoder. To introduce the adjustment and correction time t, given conditions The 95% confidence interval of the carbon emission baseline cluster; Spatiotemporal graph convolutional prediction layer: embeds the modified baseline clusters into graph node features. Convolutional dynamic graph Regulating the dynamic adjacency matrix of perception ;in For the enhanced features of node i at time t, For the feature at time node i, The cluster number of the region to which node i belongs. To regulate the synergistic attenuation coefficient, The control category to which node i belongs. For the first The node feature matrix of the layer, For the first +1 layer node feature matrix For degree matrix, For the first The trainable weight parameter matrix of layered graph convolution. It is a non-linear activation function. The indicator function is set to 1 when the node's carbon emission deviation exceeds the regulation correction confidence interval, and 0 otherwise. The regulation categories are divided according to the node's functional attributes, including energy consumption regulation, traffic regulation, lighting regulation, event schedule regulation, and temperature control. Nodes belonging to the same regulation category share the same regulation action dimension. Pointer Generation Network Correction Layer: Extends the pointer library to a joint pointer library for regulation and baseline. Adjusting the probability of similarity-weighted pointers By generating and replicating a mixed output of carbon emission predictions ;in This refers to the sequence of control actions actually executed in historical events. The baseline cluster deformation of historical event m, To adjust the difference weight matrix, This represents the difference vector between current and historical regulations. Encoding the spatiotemporal context of historical event m, The total carbon emissions sequence of historical events m The total number of historical events in the pointer library. Weighted attention score is calculated based on the similarity between the current event and historical pattern m. For the query vector of the pointer network, This is the global query vector for the current event. The weight matrix for encoding the current event. The weight matrix encoding the historical pattern. This is the bias vector of the pointer network. The hyperbolic tangent activation function is used. This is the predicted regional cluster carbon emissions after correction at time t+1. For generation probability, To generate the network, For the replication probability, To select historical mode pointer probability, This represents the historical carbon emission value of pattern m at time t+1. To regulate the intensity ratio; Loss function for constructing a carbon emission prediction model during major sporting events A carbon emission prediction model for major sporting events is trained based on a loss function until the maximum number of iterations is reached, at which point training stops. To predict losses, For baseline cluster constraint loss, For L2 regularization, For baseline cluster weights, For regularization weights; In actual assessments, the joint indicator library of regulation and baseline is expanded upon the existing indicator library. , The carbon emission prediction bias of node i at time t For node i, the event type cluster g and the region cluster are... On the baseline cluster confidence interval at time t; A pointer generation network is used for carbon emission prediction. The historical control experimental data comes from the actual control records of previous competitions or the A / B test data of the simulation platform. Each experiment records the carbon emission observation values ​​and control amplitude before and after control. Secondly, a machine learning-based carbon emission prediction system for large-scale sporting events includes: Data acquisition, processing, and spatiotemporal map construction module: This module is used to collect multi-source data for preset activities, preprocess the multi-source data, and construct a carbon emission spatiotemporal map based on the multi-source data; the multi-source data includes venue IoT data, event data, and meteorological data. Carbon emission prediction and baseline cluster acquisition module: used to perform spatiotemporal convolutional encoding on the carbon emission spatiotemporal map to obtain node-level spatiotemporal features, obtain node carbon emission prediction values ​​based on the node-level spatiotemporal features and emission factors, obtain regional clusters by performing regional splitting and mapping based on the venue data, and obtain carbon emission baseline clusters of regional clusters by clustering based on historical event carbon emission data. The regional cluster aggregation and dynamic regulation analysis module is used to aggregate the predicted carbon emissions of the nodes into regional clusters to obtain regional cluster carbon emissions. Regional clusters whose carbon emissions exceed the confidence interval of the carbon emission baseline cluster are designated as regions to be analyzed, while regional clusters whose carbon emissions do not exceed the confidence interval of the carbon emission baseline cluster are designated as fixed regions. Dynamic regulation analysis is performed within the regions to be analyzed to obtain the regulation ratio. Model building and output module: used to build a carbon emission prediction model during large-scale sporting events based on the carbon emission baseline cluster and the control ratio, input the data to be predicted into the carbon emission prediction model during large-scale sporting events, and output the prediction results.

[0028] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting carbon emissions during major sporting events based on machine learning, characterized in that, Includes the following steps: Collect multi-source data for pre-planned activities, preprocess the multi-source data, and construct a spatiotemporal map of carbon emissions based on the multi-source data; the multi-source data includes venue IoT data, event data, and meteorological data; Spatiotemporal convolutional encoding is performed on the carbon emission spatiotemporal map to obtain node-level spatiotemporal features. Based on the node-level spatiotemporal features and emission factors, the predicted values ​​of node carbon emissions are obtained. Based on the venue data, regional splitting and mapping are performed to obtain regional clusters. Based on the carbon emission data of historical events, carbon emission baseline clusters of regional clusters are obtained through clustering. The predicted carbon emissions of the nodes are aggregated into regional clusters to obtain regional cluster carbon emissions. Regional clusters whose carbon emissions exceed the confidence interval of the carbon emission baseline cluster are designated as regions to be analyzed, while regional clusters whose carbon emissions do not exceed the confidence interval of the carbon emission baseline cluster are designated as fixed regions. Dynamic regulation analysis is performed within the regions to be analyzed to obtain the regulation ratio. A carbon emission prediction model for large-scale sporting events is constructed based on the carbon emission baseline cluster and the control ratio. The data to be predicted is input into the carbon emission prediction model for large-scale sporting events, and the prediction results are output.

2. The method for predicting carbon emissions during major sporting events based on machine learning according to claim 1, characterized in that, The method for constructing the aforementioned spatiotemporal map of carbon emissions includes: Based on the requirements and guidelines of the event sustainability management system, the carbon emission sources of large-scale sports events are abstracted as graph nodes to construct a spatiotemporal graph of the event's carbon emissions. The graph describes the spatial relationships between nodes; the nodes include basic energy consumption nodes, temporary energy consumption nodes, special energy consumption nodes, transportation nodes, accommodation nodes, renewable energy nodes, and carbon sequestration nodes; among them For a set of nodes, Let be the set of edges. The node feature matrix, The adjacency matrix is ​​used; the edge set includes physical connection edges, functional coupling edges, and causal influence edges; physical connection edges: pipeline connections between venues and transportation hubs; functional coupling edges: functional similarity between nodes of the same type; causal influence edges: based on Granger causality test, if the historical carbon emissions of node i have predictive power for node j, then a directed edge is established. Based on the risk and distance dual-factor edge weighting mechanism, dynamic edge weights for the carbon emission spatiotemporal graph are designed. Historical carbon emission sequences of nodes i and j , Perform Granger causality testing: ; If you refuse Then a directed edge is established. causal strength is ; Introducing a carbon emission propagation resistance factor to adjust dynamic side weights in real time .

3. The method for predicting carbon emissions during major sporting events based on machine learning according to claim 1, characterized in that, The method for obtaining the node-level spatiotemporal features includes: Graph convolution is used to aggregate neighbor node information, and gated temporal convolution is run in parallel on each node. Spatiotemporal graph convolution is used, and graph convolutional layers and temporal convolutional layers are stacked alternately to form a spatiotemporal joint representation. Introduce dual-axis attention for time and features. Time-axis attention: calculates the importance of time-series features at each time step and outputs time-weighted features. Feature axis attention: Calculates the importance of each dimension of the time-weighted features, outputs a weighted representation of the features, and then... The output is node-level spatiotemporal features; it incorporates dual-axis attention to automatically focus on key time steps and key features.

4. The method for predicting carbon emissions during major sporting events based on machine learning according to claim 1, characterized in that, The method for obtaining the predicted carbon emissions of the node includes: Carbon emission factors and node-level spatiotemporal features are obtained using a learnable emission factor mapping gate. Adaptive selection of key emission factors, when When the standard emission factor is used as the emission factor, When using the characteristic mapping emission factor as the emission factor, the emission factor is used. The activity level is predicted, carbon emission forecasts are calculated, the carbon emission forecasts are compared with the baseline cluster center, the deviation is calculated, and the deviation is used as a feedback signal to correct the carbon emission forecasts through residual connection.

5. The method for predicting carbon emissions during major sporting events based on machine learning according to claim 1, characterized in that, The method for obtaining the regional cluster and the carbon emission baseline cluster includes: The event's regional clusters are obtained by dual splitting and mapping based on the spatial and functional attributes of the venue data. Spatial attributes: Based on the venue's geographical coordinates, density-based clustering is used to group venues with a physical distance of less than or equal to 2km into one cluster. Functional attributes: Classified according to energy consumption type. Based on the transportation network topology, venues sharing major transportation hubs are grouped into one cluster. The intersection of the spatial clusters and functional clusters is taken to form a refined regional cluster under dual constraints of space and function, and the output is the regional cluster. Extract time-series carbon emission data at the regional cluster level from historical events, and use isolated forests to score event feature vectors for anomalies, removing events with scores below the scoring threshold τ; the event feature vectors include total carbon emissions, peak emissions, emission volatility, and meteorological deviation. Hierarchical clustering is performed based on the overall feature vectors of historical events to obtain event type clusters; dynamic time-normalized distance is used for distance measurement. Within the event type cluster, the carbon emission sequence of the regional cluster is subdivided and clustered according to the event stage. The time window is defined according to the event stage to obtain the regional cluster stage combination. The shape features of the carbon emission series are extracted, and K-means clustering is performed on the shape features. The regional cluster stage combination is matched with the most similar baseline cluster to obtain the carbon emission baseline cluster. The center and boundary of the baseline cluster are defined.

6. The method for predicting carbon emissions during major sporting events based on machine learning according to claim 1, characterized in that, A method for obtaining the carbon emissions of the said regional cluster includes: The predicted carbon emissions of all nodes within a regional cluster are aggregated, and all regional clusters in the event are aggregated to obtain the initial regional cluster carbon emissions. We acquire spatiotemporal pattern samples of historical events and construct a pointer library. The historical pattern samples contain runaway context encoding, total carbon emission sequence, and metadata. The metadata includes event type, scale, climate conditions, and duration. Given the query code of the current event, calculate the similarity score between the current event and each historical pattern in the pointer library, and calculate the pointer probability distribution; Calculate the generation probability of carbon emission prediction, generate the self-generated prediction component, weight the replication component of historical patterns, correct the initial regional cluster carbon emissions, and output the regional cluster carbon emissions.

7. The method for predicting carbon emissions during major sporting events based on machine learning according to claim 1, characterized in that, A method for obtaining the control ratio by performing dynamic control analysis within the region to be analyzed includes: We construct state and action spaces, and use linear scalar quantization to transform the bi-objective Pareto optimization function into a single-objective optimization problem, merging the multiple objectives into a scalar reward function. The data is then fed into a network of actors and critics for strategy optimization. Given an actor network, motion constraint projection, and critic network, a dominance function is given based on temporal difference error. The policy gradient is updated to output the modulated action, and the modulated action is output based on the modulated action.

8. The method for predicting carbon emissions during major sporting events based on machine learning according to claim 1, characterized in that, The method for constructing a carbon emission prediction model during the aforementioned large-scale sporting event includes: The regulation ratio is used as a dynamic correction factor for the baseline cluster, the regulation influence coefficient is calibrated, and the regulation is corrected according to the confidence interval of the baseline cluster. The carbon emission prediction model during major sporting events includes an input encoding layer, a spatiotemporal graph convolutional prediction layer, and a pointer generation network correction layer. Input coding layer: State coding; Spatiotemporal graph convolutional prediction layer: embeds the modified baseline clusters into graph node features, convolves the dynamic graph, and modulates the perceived dynamic adjacency matrix; Pointer generation network correction layer: The pointer library is expanded to a joint pointer library with the baseline, and the similarity-weighted pointer probability is adjusted. The carbon emission prediction value is output by a mixture of generation and replication. Construct a loss function for a carbon emission prediction model during major sporting events. Train the model based on the loss function until the maximum number of iterations is reached, at which point training stops.

9. A machine learning-based carbon emission prediction system for large-scale sporting events, used to perform the method according to any one of claims 1-8, characterized in that, include: Data acquisition, processing, and spatiotemporal map construction module: This module is used to collect multi-source data for preset activities, preprocess the multi-source data, and construct a carbon emission spatiotemporal map based on the multi-source data; the multi-source data includes venue IoT data, event data, and meteorological data. Carbon emission prediction and baseline cluster acquisition module: used to perform spatiotemporal convolutional encoding on the carbon emission spatiotemporal map to obtain node-level spatiotemporal features, obtain node carbon emission prediction values ​​based on the node-level spatiotemporal features and emission factors, obtain regional clusters by performing regional splitting and mapping based on the venue data, and obtain carbon emission baseline clusters of regional clusters by clustering based on historical event carbon emission data. The regional cluster aggregation and dynamic regulation analysis module is used to aggregate the predicted carbon emissions of the nodes into regional clusters to obtain regional cluster carbon emissions. Regional clusters whose carbon emissions exceed the confidence interval of the carbon emission baseline cluster are designated as regions to be analyzed, while regional clusters whose carbon emissions do not exceed the confidence interval of the carbon emission baseline cluster are designated as fixed regions. Dynamic regulation analysis is performed within the regions to be analyzed to obtain the regulation ratio. Model building and output module: used to build a carbon emission prediction model during large-scale sporting events based on the carbon emission baseline cluster and the control ratio, input the data to be predicted into the carbon emission prediction model during large-scale sporting events, and output the prediction results.

Images