Shared bicycle multi-station demand intelligent prediction method based on graph convolution and sequential network cooperation

Abstract

The invention relates to a shared bicycle multi-site demand intelligent prediction method based on graph convolution and sequential network cooperation, and the method comprises the steps: obtaining multi-source data of a shared bicycle, carrying out the preprocessing and feature engineering of the multi-source data, and obtaining a static feature, a dynamic feature and an environment feature; the time sequence diagram network receives the dynamic characteristics and the environment characteristics and outputs a dynamic adjacency matrix; the graph convolution network receives the static features and a dynamic adjacency matrix output by the time sequence graph network, and outputs time-space fusion features; the bidirectional long-short-term memory network receives the space-time fusion features of the graph convolution network and outputs a bidirectional hidden state; and the time domain convolutional network receives the bidirectional hidden state output by the bidirectional long short-term memory network, and outputs demand prediction results of future S time steps. The dynamic adjacency matrix is generated through the TGN, and is cascaded with the DGCN-BiLSTM, so that the problem of insufficient modeling of space-time coupling phenomena such as tidal effect in the traditional method is effectively solved, and the required space propagation and time evolution laws can be accurately captured.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation technology, and in particular to an intelligent prediction method for multi-site demand of shared bicycles based on graph convolution and temporal network collaboration. Background Technology

[0002] As an important solution for short-distance urban travel, the operational efficiency of shared bicycle systems depends heavily on the rational allocation of bicycle resources, and accurately predicting the bicycle demand at each station is key to optimizing scheduling.

[0003] The demand for shared bicycles has significant temporal and spatial correlation characteristics. However, existing intelligent prediction methods mostly rely on "historical demand + simple spatiotemporal features", which do not make full use of multi-source data that are strongly correlated with the demand for shared bicycles and cannot adapt to dynamic scenarios. Furthermore, existing prediction models are complex, take a long time to predict, and have poor robustness due to missing data, which cannot meet the real-time scheduling needs of multiple sites at the city level.

[0004] Based on the above technical problems, this application proposes an intelligent prediction method for multi-site demand of shared bicycles based on graph convolution and temporal network collaboration. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent demand prediction method for shared bicycles at multiple stations based on graph convolution and temporal network collaboration, in order to solve the technical problems mentioned in the background art. This objective is achieved through the following technical solutions: A method for intelligent prediction of demand at multiple stations for shared bicycles based on the collaboration of graph convolution and temporal networks includes the following steps: Step S1. Data Acquisition and Preprocessing: Acquire multi-source data of shared bicycles, perform preprocessing and feature engineering on the multi-source data to obtain static features, dynamic features and environmental features; Step S2. Dynamic Topology Generation: The temporal graph network receives the dynamic features and the environmental features, and outputs a dynamic adjacency matrix. ; Step S3. Spatial Feature Aggregation: The graph convolutional network receives static features and the dynamic adjacency matrix output by the time-series graph network. Output spatiotemporal fusion features; Step S4. Temporal Dependency Capture: The bidirectional long short-term memory network receives the spatiotemporal fusion features of the graph convolutional network and outputs the bidirectional hidden state; Step S5. Multi-step prediction output: The temporal convolutional network receives the bidirectional hidden state from the bidirectional long short-term memory network and outputs the demand prediction results for the next S time steps.

[0006] Furthermore, the multi-source data includes historical order data, real-time OD traffic, holiday and special event annotations, meteorological data, and site information.

[0007] Furthermore, static features include the site's geographical location, the attributes of surrounding facilities, and the site's physical attributes; dynamic features include historical traffic, real-time OD interactions, and short-term trends; environmental features include meteorological data, time attributes, and special events.

[0008] Furthermore, the preprocessing of multi-source data includes: filling missing data with a combination of bidirectional interpolation and spatiotemporal neighborhood mean; and identifying and correcting outlier data using density clustering algorithms.

[0009] Furthermore, step S2, dynamic topology generation, specifically includes: Step S21. Input the dynamic features and the environmental features; Step S22. Update node state via event-driven update: In the formula, Let v be the hidden state of node v at time t; Let v be the hidden state of node v at time t-1; Let be the edge characteristics from node u to node v at time t; Step S23. Output the dynamic adjacency matrix: In the formula, Input features for the node at time t; Let t be the hidden state of node u at time t.

[0010] Furthermore, step S3, spatial feature aggregation, specifically includes: Step S31. Input the dynamic adjacency matrix and static features; Step S32. Perform graph convolution calculation. The formula for graph convolution is: In the formula, For activation functions; for The corresponding degree matrix, The node features input at time t; This is a learnable weight matrix.

[0011] Furthermore, the calculation formula for timing dependency capture in step S4 is as follows: In the formula, This represents the hidden state of the forward LSTM at time t; This represents the hidden state of the backward LSTM at time t; The node features input at time t; This is the predicted output at time t.

[0012] Furthermore, the multi-step prediction output of step S5 specifically includes: Step S51. Stack four layers of dilated convolutions using dilated causal convolutions: In the formula, This is the output of the l-th convolutional layer; For activation functions; The weights of the convolutional kernel in layer l; For a convolution operation with an inflation rate of d; This is the input for the (l-1)th layer; This is the bias term for the l-th layer; Step S52. Capture the evolution of node features over time using a temporal convolutional network: In the formula, For node i in The state at any given moment; for Edge features from time i to j; Step S53. Connect the multi-step output projector: In the formula, Let S be the predicted demand at the S-th time step; S is the total prediction step size. This is the output of the last temporal convolution layer; The weights of the output projection layer; This is the bias for the output projection layer.

[0013] The technical solutions provided in this application have at least the following technical effects or advantages: 1. By generating a dynamic adjacency matrix through TGN and cascading it with DGCN-BiLSTM, the problem of insufficient modeling of spatiotemporal coupling phenomena such as "tidal effect" by traditional methods is effectively solved, and the spatial propagation and temporal evolution of demand can be accurately captured.

[0014] 2. By fusing multi-source heterogeneous data and combining bidirectional temporal modeling with dilated causal convolution, prediction accuracy is significantly improved. The data preprocessing and model structure employed enhance robustness in the event of missing data, with an error increase target of ≤15% at a 20% missing data rate.

[0015] 3. Through lightweight cascading design and efficient model components (such as TCN and GCN), the single-site prediction latency is compressed to less than 20ms, which meets the needs of real-time scheduling of multiple sites at the city level and has good prospects for practical application. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of an embodiment of this application. Detailed Implementation

[0018] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0019] like Figure 1 The method for intelligent prediction of demand at multiple sites for shared bicycles, based on the collaboration of graph convolution and temporal networks, includes the following steps: Step S1. Data Acquisition and Preprocessing: Collect historical order data of shared bicycles (including origin and destination stations, timestamps, riding duration, etc.), real-time OD traffic data (the number of rides from station i to station j in time period t), meteorological data (temperature, precipitation, wind speed, etc.), holiday and special event annotation data, and station information.

[0020] As shown in Table 1, the multi-source data is preprocessed. Missing data is processed by a combination of bidirectional interpolation and spatiotemporal neighborhood mean filling. Abnormal data (such as abnormal order time and site coordinate deviation) are identified and corrected by density clustering algorithm.

[0021] Table 1. Multi-source data and their preprocessing methods Data categories Specific fields Data format / dimension Obtain source Preprocessing methods Historical order data Origin and end station IDs, timestamps, and riding duration Table (rows: number of orders; columns: 5 fields) Shared bicycle operation backend database Deleting abnormal orders (incorrect timestamps / incorrect site IDs) using density clustering. Real-time OD flow <![CDATA[t - period cycling volume from i to j (OD t ᵢⱼ)]]> Matrix (dimension: N×N, where N is the number of stations) Real-time data transmission interface (MQTT protocol) Missing values ​​were filled with "the average of the previous 3 time periods + the average of neighboring sites". Meteorological data Temperature, precipitation, wind speed Table (rows: time steps; columns: 3 fields) Third-party weather APIs (such as Gaode Weather) Standardization (mean minus / standard deviation divided) Site Information Latitude and longitude, capacity, surrounding facilities Table (rows: number of sites; columns: 5 fields) Urban Transportation GIS System Latitude and longitude are converted to UTM coordinates, and surrounding facilities are coded using 0-1. Characteristic engineering yields static, dynamic, and environmental features, including: Problem Definition and Graph Structure Initialization The problem of predicting the demand for shared bicycles involves forecasting the future traffic conditions at each of N shared bicycle stations, given historical traffic data. To accurately predict the travel demand for shared bicycles, this application defines the following related concepts.

[0022] Definition 1 (Graph): Due to the non-Euclidean topology of traffic flow data, traditional modeling methods based on regular grids are no longer applicable. Graphs, however, can effectively represent this non-Euclidean topology. Therefore, this application employs a weighted directed graph. Define a shared bicycle system, in which, Let N represent the set of sites, where N is the number of sites; and E represents the edges between sites. Represents the adjacency matrix between sites. For a connected graph, the adjacency matrix This is the adjacency matrix of the interaction graph; Definition 2 (trip): Using a single shared bicycle order =( , , , , Define a trip, where, Indicates the travel order ID. , Indicates the return time of the shared bicycle. Indicates the shared bicycle rental station ID. Indicates the ID of the shared bicycle return station; Definition 3 (Demand): The demand for shared bicycles is divided into two types: the number of bicycles borrowed (O) and the number of bicycles returned (D). Assuming there are T historical time periods, the demand for shared bicycle stations... Lending amount and the amount returned They can be represented as follows: Definition 4 (Prediction Problem): The purpose of this application is to learn a demand prediction function by analyzing the historical demand for shared bicycles at various sites, thereby enabling the prediction of future demand for shared bicycles. That is, through... The vehicle demand X at one historical time period is used to predict the vehicle demand Y over the next S time intervals, where: The model can be represented as: Static base graph construction and fusion The number of interactions between shared bike stations reveals the spatial correlation between each station. Assuming from station... Arrive at the station If there is a large amount of interaction between the shared bikes, it indicates that the interaction matrix between these two sites has a large weight value. Assuming from the site... Arrive at the station All shared bike trips are collected as follows Then the adjacency matrix of the interaction graph It can be defined as: In the formula, O represents the number of edges, L represents the number of nodes, and L represents the number of links.

[0023] Then, a connected graph adjacency matrix is ​​constructed, which is used to characterize the distance correlation between stations, road connectivity, and road condition characteristics. Generally speaking, the connectivity strength between stations increases as the path distance decreases, but factors such as terrain undulations, differences in road conditions, and vegetation distribution may lead to special cases. Therefore, assessing connectivity solely based on path distance is obviously limited, and the impact of actual road conditions must be comprehensively considered.

[0024] Therefore, this application adopts a method that does not require prior knowledge of road space information, characterizing connectivity by calculating the average riding time of all shared bicycle trips between stations. This method fully incorporates the impact of actual riding experience on traffic efficiency, and compared to evaluation methods that solely rely on geographical distance, it can more comprehensively reflect the true connectivity characteristics between stations. The specific calculation definition is shown in the following formula: in, Indicates from the site Arrive at the station A collection of all shared bike trips. Indicates from the site Arrive at the station The average time for all shared bike trips For the time a shared bicycle is rented out, K represents the time for returning the shared bicycle. The number of trips in the trip set Indicates the result after calculation Arrive at the station connectivity, The larger the value, the stronger the connectivity; conversely, the smaller the value, the weaker the connectivity.

[0025] To enable the predefined graph input to the adaptive graph structure learning module to simultaneously possess both interactive and connectivity dependencies related to shared bicycles, the constructed interactive graph and connected graph are fused to generate a graph containing multiple long-term spatial dependencies.

[0026] Since the data in the two graphs have different meanings, they are normalized before being merged, as shown in the equation: Where the denominator is the Frobenius norm, ; The resulting fused graph will be used as a predefined input matrix for adaptive graph structure learning.

[0027] Multi-source feature construction First, define the following symbols: T: The number of historical time steps entered (e.g., the past 24 hours, with one time step per hour). Ds: Static feature dimension, which is the number of features inherent to the site. Dd: Dynamic feature dimension, which is the number of traffic-related features that change over time. De: Environmental characteristic dimension, that is, the number of characteristics of external environmental factors.

[0028] Among them, static characteristics are inherent attributes of the site, independent of time, and are denoted as... , of which: a single site The static feature vector is It includes the station's geographical location (latitude and longitude), surrounding facilities attributes (whether it is near the subway, shopping mall and school) and station physical attributes (capacity, station type).

[0029] Dynamic features are flow-related characteristics that change over time, denoted as... Among them, time step Under ) site The dynamic feature vector is This includes historical traffic (lending volume at time t). Quantity returned Real-time OD interaction (time t) (Ride volume compared to surrounding stations) and short-term trends ( , ),therefore: In the formula, Station at time t With surrounding stations The amount of interaction during cycling.

[0030] Environmental characteristics are external factors that change over time. (Global) and (Site). The environmental feature vector at time step t is: (Global) or (Site) includes meteorological data (temperature, precipitation, wind speed, etc.), time attributes (hour, whether it is a weekday or a holiday), and special events.

[0031] Multi-source feature concatenation fuses three types of features along the feature dimension, ultimately outputting a spatiotemporal feature tensor F with the following dimensions: Defined as .

[0032] Step S2. Dynamic Topology Generation: The temporal graph network receives the dynamic features and the environmental features, and outputs a dynamic adjacency matrix. ; Step S21. Input dynamic features and environmental features; Step S22. Update node state via event-driven update: In the formula, Let v be the hidden state of node v at time t; Let v be the hidden state of node v at time t-1; Let be the edge characteristics from node u to node v at time t; Step S23. Output the dynamic adjacency matrix: In the formula, Input features for the node at time t; Let t be the hidden state of node u at time t.

[0033] Step S3. Spatial Feature Aggregation: The graph convolutional network receives static features and the dynamic adjacency matrix output by the time-series graph network. Output spatiotemporal fusion features; Step S3, spatial feature aggregation, specifically includes: Step S31. Input the dynamic adjacency matrix and static features; Step S32. Perform graph convolution calculation. The formula for graph convolution is: In the formula, For activation functions; for The corresponding degree matrix is ​​used to achieve dynamic space aggregation; The node features input at time t; This is a learnable weight matrix.

[0034] Step S4. Temporal Dependency Capture: The bidirectional long short-term memory network receives the spatiotemporal fusion features of the graph convolutional network and outputs the bidirectional hidden state; The calculation formula for timing dependency capture in step S4 is as follows: In the formula, This represents the hidden state of the forward LSTM at time t; This represents the hidden state of the backward LSTM at time t; The node features input at time t; This is the predicted output at time t.

[0035] Bidirectional Long Short-Term Memory (Bi-LSTM) networks consist of a forward-biased LSTM and a backward-biased LSTM, which process forward and backward information respectively. Compared to traditional LSTMs, Bi-LSTMs consider past and future contextual information in the sequence data at each time step, capturing more comprehensive contextual information and long-term dependencies in the sequence data.

[0036] Step S5. Multi-step prediction output: The temporal convolutional network receives the bidirectional hidden state from the bidirectional long short-term memory network and outputs the demand prediction results for the next S time steps.

[0037] Step S5, the multi-step prediction output, specifically includes: Step S51. Stack four layers of dilated convolutions using dilated causal convolutions, with dilation rates d=1,2,4,8: In the formula, This is the output of the l-th convolutional layer; For activation functions; The weights of the convolutional kernel in layer l; For a convolution operation with an inflation rate of d; This is the input for the (l-1)th layer; This is the bias term for the l-th layer.

[0038] TCN utilizes dilated convolutions to make the time window size grow exponentially with the number of layers, thus flexibly adjusting the size of its receptive field. By adjusting the dilation rate in the dilated convolutions, long-range dependencies are captured, making it adaptable to input sequences of different lengths and improving the applicability of model training.

[0039] Step S52. Capture the evolution of node features over time using a temporal convolutional network: In the formula, For node i in The state at any given moment; for Edge features from time i to j; Step S53. Connect the multi-step output projector: In the formula, Let S be the predicted demand at the S-th time step; S is the total prediction step size. This is the output of the last temporal convolution layer; The weights of the output projection layer; This is the bias for the output projection layer.

[0040] The model consists of an adaptive graph structure learning module and a demand prediction module working together. The adaptive graph structure learning module uses time series correlation calculation combined with a multi-head graph attention network (GAT) to extract site association features from two dimensions: long-term dependency (fusion of interaction graph, connected graph, and Pearson correlation coefficient between sites) and short-term dependency (responding to dynamic fluctuations such as weather / congestion), generating an optimal dynamic adjacency matrix that fuses long-term and short-term spatial information. The demand prediction module uses a graph convolutional network (GCN) to significantly improve the efficiency of spatial feature extraction, while retaining the causal convolution and dilated convolution mechanism of the temporal convolutional network (TCN) to capture local temporal patterns, and coupling a bidirectional LSTM (BiLSTM) to model bidirectional long-term dependencies. Finally, the demand prediction result is output through a fully connected layer.

[0041] Compared to the original ChebNet solution, the GCN module reduces computational complexity from Down to This eliminates polynomial approximation errors while enhancing adaptability to dynamic topologies.

[0042] After completing steps 1-3 above, the predicted values ​​are obtained. The final core output is the predicted results of the number of shared bicycles lent out (O) and returned (D) at all target shared bicycle stations (N stations) in the city within a specified future time step (denoted as S time steps).

[0043] The technical solutions provided in this application have at least the following technical effects or advantages: 1. By generating a dynamic adjacency matrix through TGN and cascading it with DGCN-BiLSTM, the problem of insufficient modeling of spatiotemporal coupling phenomena such as "tidal effect" by traditional methods is effectively solved, and the spatial propagation and temporal evolution of demand can be accurately captured.

[0044] 2. By fusing multi-source heterogeneous data and combining bidirectional temporal modeling with dilated causal convolution, prediction accuracy is significantly improved. The data preprocessing and model structure employed enhance robustness in the event of missing data, with an error increase target of ≤15% at a 20% missing data rate.

[0045] 3. Through lightweight cascading design and efficient model components (such as TCN and GCN), the single-site prediction latency is compressed to less than 20ms, which meets the needs of real-time scheduling of multiple sites at the city level and has good prospects for practical application.

[0046] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for intelligent prediction of demand at multiple sites for shared bicycles based on graph convolution and temporal network collaboration, characterized in that, Includes the following steps: Step S1. Data Acquisition and Preprocessing: Acquire multi-source data of shared bicycles, perform preprocessing and feature engineering on the multi-source data to obtain static features, dynamic features and environmental features; Step S2. Dynamic Topology Generation: The temporal graph network receives the dynamic features and the environmental features, and outputs a dynamic adjacency matrix. ; Step S3. Spatial Feature Aggregation: The graph convolutional network receives static features and the dynamic adjacency matrix output by the time-series graph network. Output spatiotemporal fusion features; Step S4. Temporal Dependency Capture: The bidirectional long short-term memory network receives the spatiotemporal fusion features of the graph convolutional network and outputs the bidirectional hidden state; Step S5. Multi-step prediction output: The temporal convolutional network receives the bidirectional hidden state from the bidirectional long short-term memory network and outputs the demand prediction results for the next S time steps.

2. The intelligent prediction method for multi-site demand of shared bicycles based on graph convolution and temporal network collaboration as described in claim 1, characterized in that, The multi-source data includes historical order data, real-time OD traffic, holiday and special event annotations, meteorological data, and site information.

3. The intelligent prediction method for multi-site demand of shared bicycles based on graph convolution and temporal network collaboration as described in claim 1, characterized in that, The static features include the site's geographical location, surrounding facility attributes, and physical attributes; the dynamic features include historical traffic, real-time OD interaction, and short-term trends; and the environmental features include meteorological data, time attributes, and special events.

4. The intelligent prediction method for multi-site demand of shared bicycles based on graph convolution and temporal network collaboration as described in claim 1, characterized in that, The preprocessing of the multi-source data includes: filling missing data by combining bidirectional interpolation with spatiotemporal neighborhood mean; and identifying and correcting abnormal data using density clustering algorithms.

5. The intelligent prediction method for multi-site demand of shared bicycles based on graph convolution and temporal network collaboration as described in claim 1, characterized in that, Step S2, dynamic topology generation, specifically includes: Step S21. Input the dynamic features and the environmental features; Step S22. Update node state via event-driven update: In the formula, Let v be the hidden state of node v at time t; Let v be the hidden state of node v at time t-1; Let be the edge characteristics from node u to node v at time t; Step S23. Output the dynamic adjacency matrix: In the formula, Input features for the node at time t; Let t be the hidden state of node u at time t.

6. The intelligent prediction method for multi-site demand of shared bicycles based on graph convolution and temporal network collaboration as described in claim 1, characterized in that, Step S3, spatial feature aggregation, specifically includes: Step S31. Input the dynamic adjacency matrix and static features; Step S32. Perform graph convolution calculation. The formula for graph convolution is: In the formula, For activation functions; for The corresponding degree matrix, The node features input at time t; This is a learnable weight matrix.

7. The intelligent prediction method for multi-site demand of shared bicycles based on graph convolution and temporal network collaboration as described in claim 1, characterized in that, The calculation formula for timing dependency capture in step S4 is as follows: In the formula, This represents the hidden state of the forward LSTM at time t; This represents the hidden state of the backward LSTM at time t; The node features input at time t; This is the predicted output at time t.

8. The intelligent prediction method for multi-site demand of shared bicycles based on graph convolution and temporal network collaboration as described in claim 1, characterized in that, Step S5, the multi-step prediction output, specifically includes: Step S51. Stack four layers of dilated convolutions using dilated causal convolutions: In the formula, This is the output of the l-th convolutional layer; For activation functions; The weights of the convolutional kernel in layer l; For a convolution operation with an inflation rate of d; This is the input for the (l-1)th layer; This is the bias term for the l-th layer; Step S52. Capture the evolution of node features over time using a temporal convolutional network: In the formula, For node i in The state at any given moment; for Edge features from time i to j; Step S53. Connect the multi-step output projector: In the formula, Let S be the predicted demand at the S-th time step; S is the total prediction step size. This is the output of the last temporal convolution layer; The weights of the output projection layer; This is the bias for the output projection layer.

Images