A method for predicting the distribution of road transport capacity by integrating long and short-term fluctuation trends

Abstract

The application provides a road transport capacity distribution prediction method which comprehensively considers long-term and short-term fluctuation trends, carries out multistage decoupling on original data, respectively extracts transport capacity space-time characteristics under different scales, simultaneously, dynamically updates transport capacity correlation information by using a dynamic regularization algorithm, and on this basis, captures transport capacity space-time dependence by using a graph attention mechanism and a diffusion dilated convolution network, and generates a multiscale fused comprehensive space-time feature representation by using a dynamic weight. The method comprehensively considers space-time dependence of transport capacity under different scales, and is beneficial to accurate prediction of road transport capacity distribution.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of highway transportation capacity distribution prediction, and in particular to a highway transportation capacity distribution prediction method integrating long-term and short-term fluctuation trends. BACKGROUND

[0002] Highway transportation capacity distribution is a key indicator for measuring the highway transportation capacity of a specific geographical area within a specific time range, covering both freight and passenger transportation. Its changes are influenced by multiple factors such as regional attributes, geographical environment, climate conditions, and human factors. With the rapid development of China's economy and the acceleration of urbanization, the demand for freight transportation between regions is growing, and the demand for highway freight capacity prediction is increasingly urgent.

[0003] At present, with the optimization of logistics networks and the widespread application of intelligent technology, how to effectively predict the highway transportation capacity in a specific region and accurately perceive the changes in transportation capacity has become a core issue to ensure regional logistics efficiency and optimize resource allocation. Currently, the research on highway transportation capacity distribution prediction methods is relatively scarce, and existing researches are mostly focused on road traffic prediction, resulting in a lack of strong technical support for precise scheduling and management of regional highway transportation capacity. Due to the complex transformation characteristics of transportation capacity, traditional prediction methods based on statistics are difficult to fully capture the changes in regional transportation capacity, leading to low resource allocation efficiency, excess or insufficient transportation capacity in actual transportation management. Therefore, there is an urgent need for a prediction method specifically for highway transportation capacity distribution to improve the accuracy and scientificity of logistics and transportation management.

[0004] However, the prediction of regional highway transportation capacity faces multiple challenges: first, regional highway transportation capacity data presents complex time dependence and spatial heterogeneity, and its spatio-temporal correlation is highly dynamic in different regions and time periods. Second, transportation capacity distribution is also affected by dynamic factors such as weather and unexpected events, and the data shows complex distribution shift phenomena, increasing the difficulty of prediction. How to model and capture event information and individual spatio-temporal correlation to effectively deal with distribution shift and spatio-temporal heterogeneity of transportation capacity distribution in the prediction problem is of great significance to improve regional freight efficiency and resource utilization. SUMMARY

[0005] The present application aims to overcome the shortcomings of the prior art and provide a highway transportation capacity distribution prediction method integrating long-term and short-term fluctuation trends, which can accurately predict the auxiliary highway transportation capacity distribution.

[0006] To achieve the above-mentioned purpose, the present application adopts the following technical solutions:

[0007] The present application provides a highway transportation capacity distribution prediction method integrating long-term and short-term fluctuation trends, comprising the following steps:

[0008] S1. Data acquisition and preprocessing, defining highway transport capacity evaluation indicators, and determining regional highway transport core points based on kernel density estimation and centroid analysis;

[0009] S2. After determining the coordinates of the regional transport core points, comprehensively consider traffic conditions, geographical distances, and traffic flow data, and build an initial highway transport capacity association network based on a path planning algorithm for subsequent spatiotemporal feature extraction and network updates.

[0010] S3. Split the historical highway transport capacity association network into association networks under overlapping time windows with a moving step size of Combining the moving average filter with the wavelet transform to separate the long-term and short-term trend characteristics of transport capacity, the multi-scale trend decoupling of regional transport capacity is performed;

[0011] S4. Use trigonometric functions to encode external events and periodic information. At the same time, for the data within the input time window, use the dynamic regularization algorithm to update the capacity fluctuation correlation matrix. Based on the capacity fluctuation correlation matrix, use the diffusion dilation convolution and graph attention mechanism to extract the cross-scale and multi-period spatiotemporal characteristics of regional highway transportation capacity.

[0012] S5. Adaptive weight learning and attention mechanism are used to fuse multi-scale spatiotemporal features and external dynamic influencing factors, and a multi-layer perceptron is used as the prediction head to realize the prediction of highway transport capacity distribution.

[0013] Furthermore, the S1 is specifically:

[0014] S101. Acquiring data: Acquiring historical data related to the highway transport capacity of the region to be predicted, including the types and quantities of trucks, regional road network information, truck transport record information, and event information at the corresponding time;

[0015] S102. Define highway transport capacity evaluation indicators: Highway transport capacity is determined by the types, number, carrying capacity, and actual utilization rate of trucks in a region;

[0016] The calculation formula for the highway transport capacity evaluation index is:

[0017] ;

[0018] in, The categories of truck types included in each region; For the region Middle Types of trucks; For the region Middle The single-vehicle carrying capacity of the truck; For the region The actual utilization rate of a truck; is time;

[0019] S103, complete the transport capacity data: determine the update frequency of highway transport capacity ; for regional highway transport capacity time series data ; if the data is missing at a time point, the spline interpolation method is used to complete it; is the regional transport capacity data at time ;

[0020] S104, determine the regional transport core point to represent the regional transport center: different data are rasterized according to a unified spatial resolution, divided into sub-regions, ensuring that regional layout and traffic network data can be analyzed on the same spatial scale;

[0021] In order to comprehensively consider the influence of adjacent regions on the region, the spatial density of population, economic activity and traffic activity of the sub-region is calculated by using kernel density estimation; and the population density, economic activity density and traffic flow density are weighted and superimposed, different weights are given to different attribute spatial densities, and the comprehensive density of the th sub-region is obtained :

[0022] ;

[0023] Among them, is the population density; is the economic activity density; is the traffic flow density; is the population density corresponding to the kernel density function; is the economic activity density corresponding to the kernel density function; is the traffic flow density corresponding to the kernel density function;

[0024] The expression of kernel density is:

[0025] ;

[0026] Among them, is the sub-region for which the kernel density is calculated; is the total number of sub-regions in the region ; is the bandwidth parameter to control the smoothness; is the th sub-region; is the Gaussian kernel function;

[0027] In the comprehensive density map, the centroid analysis method is used to identify the regional​ The comprehensive core transportation point, the core point coordinate point calculation formula is:

[0028]

[0029] Among them, Respectively, the coordinates of the regional transportation core point; The grid center point coordinates of each region in the region; The comprehensive density value of the first Sub-region; The total number of sub-regions;

[0030] Further, the S2 is specifically:

[0031] S201, node information construction: based on the data obtained by the S1, the highway transport capacity correlation network of the regional hub is constructed; in the correlation network, for the regional Determined by the core transportation point coordinates, at each time point, The attribute is the carrying capacity of the corresponding region;

[0032] S202, determine the weight of node edge: in the network, the edge represents the transport correlation between regions; the weight The weight Can reflect the transport flow between regional nodes And regional nodes , the weight of the edge is:

[0033] ;

[0034] Among them, The geographical distance between regional nodes And regional nodes , calculated by the Dijkstra algorithm applied to the road network; The transport time between regional nodes And regional nodes At time , obtained from historical data; The traffic flow between regional nodes And regional nodes At time , obtained from historical data; , , The weight coefficient;

[0035] S203, construct the adjacent matrix of the highway transport capacity correlation network: for each time , construct the adjacent matrix :

[0036] A t = [ a p q t ] ;

[0037] wherein, if there is an edge between the region node and the region node and the distance is less than a reached threshold , then , otherwise ; wherein, is the edge weight of the road capacity correlation network road adjacency matrix;

[0038] S204, integrating the historical road capacity correlation network: integrating the road capacity correlation network of adjacent time slices to form a complete historical road capacity correlation network ; wherein, is the node feature set of all time slices; is the adjacency matrix set of all time slices.

[0039] Further, the S3 is specifically:

[0040] S301, preliminary decoupling: for the node capacity data in the time window, a moving average filter with different scale factors is used to strip the long-term and short-term trends of the capacity; the original capacity data is decomposed into long-term trend data and short-term fluctuation data, for the node capacity sequence in the first time window, the long-term trend data is expressed as:

[0041] ;

[0042] wherein, is the original capacity data; is an adjustable factor of the moving average filter to control the decoupling degree; is the position index of the time ;

[0043] The short-term trend data is:

[0044] ;

[0045] S302, deep feature decoupling: through wavelet decomposition, the long-term trend data is further decoupled into multiple wave bands of different time scales to capture the capacity fluctuation characteristics under different time scales; the short-term fluctuation data is decomposed into different frequency bands in order to capture the small changes in a short time; for the long-term trend data, the low-frequency wave band after decomposition represents the long-term trend characteristics; for the short-term trend data, the high-frequency wave band after decomposition represents the short-term trend characteristics. ​

[0046] Further, the S4 is specifically:

[0047] S401, for the input time window, identify the external events affecting the transport capacity; if the event occurs, the corresponding code value is 1; if the event does not occur, the code value is 0;

[0048] Formalized as:

[0049] ;

[0050] Wherein, is the practice whether it happens or not;

[0051] In order to capture the change trend of transport capacity in the day, week and month cycle, the sine and cosine function is used to encode the time period characteristics, so as to map the time periodicity into the feature space, and ensure that the model can identify and use these periodic patterns;

[0052] Assuming that the period of a day is , the time is in the position of in a day, then the daily cycle code is:

[0053] ;

[0054] ;

[0055] Assuming that the period of a week is , the time is in the position of in a week, then the weekly cycle code is:

[0056] ;

[0057] ;

[0058] The periodicity encoding is integrated into a complete feature vector:

[0059] F c y c l e ( t ) = [ s i n _ d a y ( t ) , c o s _ d a y ( t ) , s i n _ w e e k ( t ) , c o s _ w e e k ( t ) ] ;

[0060] The external event encoding of time and the periodicity feature encoding are combined into a complete feature vector :

[0061] ;

[0062] Wherein, is the feature splicing.

[0063] S402: For data in the input time window: use the dynamic time normalization algorithm to calculate the similarity between the time series of the input time window nodes. , set the similarity threshold ,

[0064] If the node and The sequence similarity between , then update the weight If the regional node and regional nodes The sequence similarity between , then it remains unchanged;

[0065] S403, spatial feature extraction: for the capacity correlation network within a period of time, the graph attention mechanism is used to extract spatial dependency features from the capacity correlation network; in the In the layer graph attention mechanism, the node represents the vector By performing weighted summation with the features of adjacent nodes and updating the features through nonlinear transformation:

[0066] ;

[0067] in, For the The node feature matrix of the layer; is the capacity fluctuation correlation matrix; is the activation function; For the The learned weight matrix of the layer; For regional nodes and regional nodes The attention weight between them is calculated by the features of the node pair:

[0068] α p q ( l ) = e x p ( L e a k y R e L U ( o T [ W ( l ) H q ( l ) | | W ( l ) H p ( l ) ] ) ) ∑ p , q ∈ A c e x p ( o T [ W ( l ) H q ( l ) | | W ( l ) H p ( l ) ] ) ;

[0069] in, It is the vector concatenation operation; is the vector used to calculate attention;

[0070] S404, Time Feature Extraction: Use dilated convolution to extract time-dependent features in time series. Dilated convolution can capture long-term and short-term dependencies in time series by enhancing the receptive field of dilated convolution. Dilated convolution is used to operate on the input time series at each layer:

[0071] ;

[0072] in, For the The input sequence of the layer; is dilated convolution; is bias; is convolution calculation;

[0073] The calculation formula of dilated convolution is:

[0074] C o n v ( l ) = ∑ p o s p o s s − 1 W [ p o s ] ⋅ Z [ t − d ⋅ p o s ]

[0075] wherein, is the size of the convolution kernel; is the dilation rate; is the index of the convolution kernel on the input sequence.

[0076] Further, the S5 is specifically:

[0077] S501, fusion of space-time features: the last layer of time features and space features coded in the S4 are spliced for feature:

[0078]

[0079] wherein, is the last layer;

[0080] The spatial features and temporal features of the original sequence, the long-term sequence and the short-term sequence are dynamically hierarchically fused; an adaptive weight learning mechanism is introduced, and the fusion weights of different scale features are dynamically learned through a neural network, and the weight parameters are set , and correspond to the original features , long-term features and short-term features , and the multi-scale space-time features are:

[0081]

[0082] S502, comprehensive feature fusion: the multi-scale space-time features are fused with external influencing factor features by using an attention mechanism, specifically:

[0083]

[0084] wherein, is the final fused multi-scale space-time feature representation; is the encoded feature of external events and key time nodes; is the fusion feature dimension calculated by the attention mechanism;

[0085] S503, receiving the fused space-time feature vector as input, the space-time feature vector ​​​​The input is fed into the multi-layer perception to generate the final prediction value:

[0086] ;

[0087] wherein, P G Figure 1 = [ P G t , Figure 2 P G t + T Figure 3 , … P G t + u T Figure 4 ] is the prediction value of the road transport capacity for the spatio-temporal transport network , containing the transport capacity prediction value and relationship distribution characteristics of each regional node in the graph; is the forward prediction for a time step;

[0088] S504, model training and updating: the training target of the model is to minimize the mean square error between the prediction value and the actual transport capacity value , and the model parameters are optimized through training to generate the best prediction value for actual transport scheduling and management.

[0089] The beneficial effects of the present application are: the core points of regional road transport are identified by combining kernel density estimation and centroid analysis, and the multi-source heterogeneous data information such as population, economic activity and traffic flow is effectively integrated while considering the correlation within the region, ensuring the accuracy of the spatial information representative characteristics. At the same time, considering the traffic conditions, geographical distance and other information, the road transport correlation network is constructed based on the path planning algorithm, laying a foundation for subsequent effective spatio-temporal feature extraction;

[0090] By multi-level decoupling of the original data, the spatio-temporal features of the transport capacity at different scales are extracted, and at the same time, the dynamic updating of the transport correlation information is realized by using the dynamic regularization algorithm; On this basis, the graph attention mechanism and the diffusion dilated convolution network are used to capture the spatio-temporal dependence of the transport capacity, and the dynamic weight is used to generate a comprehensive spatio-temporal feature representation of multi-scale fusion. This method comprehensively considers the spatio-temporal dependence of the transport capacity at different scales, which is beneficial to the accurate prediction of the distribution of road transport capacity;

[0091] By fine coding of external events and periodic information, the dynamic external influencing factors are effectively integrated into the model. Through the attention mechanism, the multi-scale spatio-temporal information and external information are fused, and the influence of external factors and spatio-temporal information on the change of transport capacity is adaptively learned to improve the prediction performance of the model. BRIEF DESCRIPTION OF DRAWINGS

[0092] Figure 1 is a flowchart of a comprehensive long-term and short-term fluctuation trend road transport capacity distribution prediction method;

[0093] Figure 2 is a diffusion dilated convolution diagram;

[0094] Figure 3A network architecture diagram for highway transport capacity forecast that integrates long-term and short-term trends;

[0095] o Regional transportation capacity data of Nanjing, Nantong, Suzhou, Wuxi, Xuzhou and Yancheng from November 2020 to July 2021. DETAILED DESCRIPTION

[0096] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0097] See also o A highway transport capacity distribution prediction method that integrates long-term and short-term fluctuation trends includes the following steps:

[0098] S1. Data acquisition and preprocessing, defining highway transport capacity evaluation indicators, and determining regional highway transport core points based on kernel density estimation and centroid analysis;

[0099] S2. After determining the coordinates of the regional transport core points, comprehensively consider traffic conditions, geographical distances, and traffic flow data, and build an initial highway transport capacity association network based on a path planning algorithm for subsequent spatiotemporal feature extraction and network updates.

[0100] S3. Split the historical highway transport capacity association network into association networks under overlapping time windows with a moving step size of Combining the moving average filter with the wavelet transform to separate the long-term and short-term trend characteristics of transport capacity, the multi-scale trend decoupling of regional transport capacity is performed;

[0101] S4. Use trigonometric functions to encode external events and periodic information. At the same time, for the data in the input time window, use the dynamic regularization algorithm to update the capacity fluctuation correlation matrix; based on the capacity fluctuation correlation moment, use the diffusion dilation convolution and graph attention mechanism to extract the cross-scale multi-period spatiotemporal characteristics of regional highway transportation capacity, such as Figure 4 As shown;

[0102] S5. Adaptive weight learning and attention mechanism are used to integrate multi-scale spatiotemporal features and external dynamic influencing factors, and a multi-layer perceptron is used as the prediction head to achieve highway transport capacity distribution prediction, such as ​ shown.

[0103] The S1 is specifically:

[0104] S101. Acquiring data: Acquiring historical data related to the highway transport capacity of the region to be predicted, including the types and quantities of trucks, regional road network information, truck transport record information, and event information at the corresponding time;

[0105] S102, define the highway transport capacity evaluation index: highway transport capacity is determined by the type, number, carrying capacity and actual utilization rate of trucks in the region;

[0106] The highway transport capacity evaluation index calculation formula is:

[0107] ;

[0108] Wherein, is the type of truck contained in each region; is the type of truck in the region ; is the carrying capacity of the truck in the region ; is the actual utilization rate of the truck in the region ; is the time;

[0109] S103, complete the transport capacity data: determine the update frequency of highway transport capacity ; for the time series data of regional highway transport capacity ; if the data is missing at the time point, the spline interpolation method is used to complete it; is the transport capacity data of the region at time ;

[0110] S104, determine the regional transport core point to represent the regional transport center: different data are rasterized according to the same spatial resolution, divided into sub-regions, to ensure that regional layout and traffic network data can be analyzed on the same spatial scale;

[0111] In order to comprehensively consider the influence of adjacent regions on the region, the kernel density estimation is used to calculate the spatial density of population, economic activity and traffic activity of the sub-region; and the population density, economic activity density and traffic flow density are weighted and superimposed, different weights are given to different attribute spatial densities, to obtain the comprehensive density of the th sub-region

[0112] ;

[0113] Wherein, is the population density; is the economic activity density; is the traffic flow density; is the population density corresponding to the kernel density function;​​​ is the kernel density function corresponding to the density of economic activities; is the kernel density function corresponding to the traffic flow density;

[0114] The expression for kernel density is:

[0115] ;

[0116] in, The sub-region where the kernel density is currently calculated; For the region the total number of neutron regions; is the bandwidth parameter to control the smoothness; For the sub-regions; is the Gaussian kernel function;

[0117] In the integrated density map, the centroid analysis method is used to identify the regions The calculation formula of the core transport point is:

[0118]

[0119] in, are the coordinates of the regional transportation core points respectively; are the coordinates of the grid center points of each area in the region; For the The comprehensive density value of each sub-region; is the total number of sub-regions;

[0120] The S2 is specifically:

[0121] S201, node information construction: Based on the data obtained in S1, a road transport capacity association network of regional hubs is constructed; in the association network, for regional Determined by the coordinates of the core transportation point, at each time point, The attribute is the carrying capacity of the corresponding area;

[0122] S202. Determine the node edge weight: In the network, the edge represents the capacity relationship between regions; the edge Weight Responsive regional nodes and regional nodes The capacity flow between them, the weight of the edge is:

[0123] ;

[0124] in, For regional nodes and regional nodes geographical distance between the nodes, calculated by Dijkstra algorithm applied on the road network; is the regional node ; is the regional node ; is the traffic flow between the regional node and the regional node in time , derived from historical data; , , is the weight coefficient;

[0125] S203, constructing the highway capacity correlation network adjacency matrix: for each time, construct the adjacency matrix :

[0126] A t = [ a p q t ] ;

[0127] wherein, if there is an edge between the regional node and the regional node and the distance is less than the threshold value , then , otherwise ; wherein, is the highway capacity correlation network road adjacency matrix edge weight;

[0128] S204, integrating the historical highway capacity correlation network: integrating the highway capacity correlation networks of adjacent time slices to form a complete historical highway capacity correlation network ; wherein, is the node feature set of all time slices; is the adjacency matrix set of all time slices.

[0129] The S3 is specifically:

[0130] S301, preliminary decoupling: for the node capacity data within the time window, a moving average filter with different scale factors is used to strip the long-term and short-term trends of the capacity; the original capacity data is decomposed into long-term trend data and short-term fluctuation data, and for the node capacity sequence within the first time window, the long-term trend data is expressed as:

[0131] ;

[0132] wherein, is the original capacity data; a tunable factor to control the degree of decoupling for the moving average filter; a position index for time ;

[0133] the short-term trend data is:

[0134] ;

[0135] S302, deep feature decoupling: on the basis of preliminary decoupling, the long-term and short-term trend features are deeply decoupled based on wavelet transform to finely separate the multi-scale information in the transport capacity data. Specifically, through wavelet decomposition, the long-term trend data is further decoupled into wave bands of different time scales to capture the transport capacity fluctuation characteristics at different time scales; the short-term fluctuation data is decomposed into different frequency bands to capture the small changes in a short time; for the long-term trend data, the low-frequency wave band after decomposition represents the long-term trend feature; for the short-term trend data, the high-frequency wave band after decomposition represents the short-term trend feature.

[0136] The S4 is specifically:

[0137] S401, for the input time window, identify external events that affect transport capacity; such as holidays, major events, natural disasters, etc.; if the event occurs, the corresponding encoding value is 1; if the event does not occur, the encoding value is 0;

[0138] Formalized as:

[0139] ;

[0140] wherein, is whether the practice occurs;

[0141] In order to capture the change trend of transport capacity in the day, week and month cycle, the sine and cosine functions are used to encode the time period characteristics to map the time periodicity into the feature space, ensuring that the model can identify and utilize these periodic patterns;

[0142] Assuming that the cycle of a day is , the time is the position in a day , then the daily cycle encoding is:

[0143] ;

[0144] ;

[0145] Assuming that the cycle of a week is , the time is the position in a week , then the weekly cycle encoding is:

[0146] ;

[0147] ;

[0148] Combine the periodic codes into a complete eigenvector:

[0149] F c y c l e ( t ) = [ s i n _ d a y ( t ) , c o s _ d a y ( t ) , s i n _ w e e k ( t ) , c o s _ w e e k ( t ) ] ;

[0150] Time The external event code and periodic feature code are combined into a complete feature vector :

[0151] ;

[0152] in, Feature splicing.

[0153] S402: For data in the input time window: use the dynamic time normalization algorithm to calculate the similarity between the time series of the input time window nodes. , set the similarity threshold , if the regional node and regional nodes The sequence similarity between , then update the weight If the regional node and regional nodes The sequence similarity between , it remains unchanged;

[0154] S403, spatial feature extraction: for the capacity correlation network within a period of time, the graph attention mechanism is used to extract spatial dependency features from the capacity correlation network; in the In the layer graph attention mechanism, the node represents the vector By performing weighted summation with the features of adjacent nodes and updating the features through nonlinear transformation:

[0155] ;

[0156] in, For the The node feature matrix of the layer; is the capacity fluctuation correlation matrix; is the activation function; For the The learned weight matrix of the layer; For regional nodes and regional nodes The attention weight between them is calculated by the features of the node pair:

[0157] α p q ( l ) = e x p ( L e a k y R e L U ( ​ T [ W ( l ) H q ( l ) | | W ( l ) H p ( l ) ] ) ) ∑ p , q ∈ A c e x p ( ​ T [ W ( l ) H q ( l ) | | W ( l ) H p ( l ) ] ) ;

[0158] in, It is the vector concatenation operation; is the vector used to calculate attention;

[0159] S404, Time Feature Extraction: Use dilated convolution to extract time-dependent features in time series. Dilated convolution can capture long-term and short-term dependencies in time series by enhancing the receptive field of dilated convolution. Dilated convolution is used to operate on the input time series at each layer:

[0160] ;

[0161] in, For the The input sequence of the layer; is dilated convolution; is bias; It is the convolution calculation;

[0162] The calculation formula of dilated convolution is:

[0163] C o n v ( l ) = ∑ p o s p o s s − 1 W [ p o s ] ⋅ Z [ t − d ⋅ p o s ] ;

[0164] in, is the convolution kernel size; is the expansion rate; is the index of the convolution kernel on the input sequence.

[0165] The S5 is specifically:

[0166] S501, fusion of spatiotemporal features: feature splicing of the last layer of temporal features and spatial features encoded in S4:

[0167] ;

[0168] in, For the last layer;

[0169] Dynamically and hierarchically fuse the spatial and temporal features of the original sequence, long-term sequence, and short-term sequence; introduce an adaptive weight learning mechanism to dynamically learn the fusion weights of features of different scales through a neural network and set the weight parameters 、 and Corresponding to the original features , long-term characteristics and short-term characteristics , the multi-scale spatiotemporal features are:

[0170] ;

[0171] S502. Comprehensive feature fusion: Use the attention mechanism to fuse multi-scale spatiotemporal features with external influencing factor features, specifically:

[0172] ;

[0173] in, It is the multi-scale spatiotemporal feature representation after the final fusion; Coding features for external events and key time nodes; is the fusion feature dimension calculated using the attention mechanism;

[0174] S503, receiving the fused spatiotemporal feature vector As input, the spatiotemporal feature vector Input into the multi-layer perceptron to generate the final prediction value:

[0175] ;

[0176] in, P G ​ = [ P G t , ​ P G t + T ​ , … P G t + u T ​ ] For space-time capacity network The predicted value of highway transport capacity includes the predicted value of transport capacity and relationship distribution characteristics of each regional node in the graph; For forward prediction time steps;

[0177] S504, model training and updating: The training goal of the model is to minimize the predicted value and actual capacity The model parameters are optimized through training to generate the best prediction values ​​for actual capacity scheduling and management.

[0178] A method for predicting the distribution of road transport capacity by integrating long-term and short-term fluctuation trends: First, obtain the historical transport capacity data of the region to be tested, and preprocess and complete the data. Then, identify the regional transport core points based on kernel density estimation and centroid analysis, and construct an initial road transport capacity correlation network based on path planning algorithms and reachable thresholds, taking into account traffic conditions, geographic distances, and other data. Next, separate the long-term and short-term trend characteristics of the transport capacity sequence using a tunable moving average filter and wavelet transform, and decouple the regional node transport capacity sequence at multiple scales. At the same time, encode external events and key time nodes that affect transport capacity in detail, and dynamically update transport capacity correlation information using a dynamic regularization algorithm. On this basis, use graph attention mechanisms and diffusion dilation convolution networks to mine the spatiotemporal characteristics of transport capacity, and use dynamic weight fusion to integrate long-term and short-term trend transport capacity, and the spatiotemporal characteristics of the original transport capacity, to form a comprehensive multi-scale spatiotemporal feature. Finally, use the attention mechanism to fuse the multi-scale spatiotemporal features with external influencing factors and input them into a multilayer perceptron for feature mapping, to achieve accurate prediction of the distribution of road transport capacity.

[0179] Embodiment

[0180] In this paper, Jiangsu Province is taken as the object of prediction of the distribution of road transport capacity, and is divided into 13 regions according to the geographical boundaries of prefecture-level cities. According to the formula for evaluating the road transport capacity, the historical road transport capacity data of each region is calculated, and the transport capacity data of Nanjing, Nantong, Suzhou, Wuxi, Xuzhou, and Yancheng from November 2020 to July 2021 is intercepted, with a unit of ten thousand tons and a sampling interval of 2 hours, as shown in Table 1. ​

[0181] At the same time, the transport core points of each region are calculated as the regional transport center, an initial transport capacity correlation network is constructed, and a normalized adjacency matrix is generated, as shown in Table 1.

[0182] Table 1 Normalized Adjacency Matrix of Regional Transport Capacity Correlation Network

[0183]

[0184] The set of external factor events for each region is obtained, and the historical data and event data are input into the model for decoupling, feature encoding, and prediction, with a prediction step of 5 time steps. A long short-term memory neural network (LSTM) is used for comparative experiments, and the mean squared error (MSE) and root mean squared error (RMSE) of the two methods are calculated, as shown in Table 2, with the unit of MSE being ten thousand tons squared and the unit of RMSE being ten thousand tons. The comparison results show that the method for predicting the distribution of road transport capacity by integrating long-term and short-term fluctuation trends proposed in this paper has higher accuracy in predicting the distribution of regional transport capacity compared to the traditional LSTM model, especially in Changzhou and Suzhou.

[0185] Table 2 Prediction Error Table​

[0186]

[0187] The above-described embodiments only express the implementation of the present application, which is described in a more specific and detailed manner, but should not be understood as a limitation to the scope of the patent of the present application. It should be noted that for those skilled in the art, several modifications and improvements can be made without departing from the concept of the present application, which are all within the scope of protection of the present application. Therefore, the scope of protection of the patent of the present application should be subject to the appended claims.

Claims

1. A method for predicting highway transport capacity distribution based on comprehensive long-term and short-term fluctuation trends, characterized by: The following steps are involved: S1. Data acquisition and preprocessing, defining highway transport capacity evaluation indicators, and determining regional highway transport core points based on kernel density estimation and centroid analysis; S2. Based on the coordinates of the regional transport core points, a comprehensive consideration of traffic conditions, geographical distances, and traffic flow data is used to construct a highway transport capacity association network based on a path planning algorithm. This is used for subsequent spatiotemporal feature extraction and highway transport capacity association network updates. S3, split the highway transport capacity association network into highway transport capacity association networks under overlapping time windows with a moving step size of Combining the moving average filter with the wavelet transform to separate the long-term and short-term trend characteristics of transport capacity, the multi-scale trend decoupling of regional transport capacity is performed; S4. Use trigonometric functions to encode external events and periodic information. At the same time, for the data within the input time window, use the dynamic regularization algorithm to update the capacity fluctuation correlation matrix. Based on the capacity fluctuation correlation matrix, use the diffusion dilation convolution and graph attention mechanism to extract the cross-scale and multi-period spatiotemporal characteristics of regional highway transportation capacity. S5. Adaptive weight learning and attention mechanisms are used to integrate cross-scale, multi-period spatiotemporal features and external dynamic influencing factors, and a multi-layer perceptron is used as the prediction head to predict highway transport capacity distribution. The S1 is specifically: S101. Acquiring data: Acquiring historical data related to the highway transport capacity of the region to be predicted, including the types and quantities of trucks, regional road network information, truck transport record information, and event information at the corresponding time; S102. Define highway transport capacity evaluation indicators: Highway transport capacity is determined by the types, number, carrying capacity, and actual utilization rate of trucks in a region; The calculation formula for the highway transport capacity evaluation index is: ； in, The categories of truck types included in each region; For the region Middle Types of trucks; For the region Middle The single-vehicle carrying capacity of the truck; For the region Middle The actual utilization rate of each type of truck; For time; S103. Complete transport capacity data: Determine the update frequency of road transport capacity ; For regional highway transport capacity time series data ; If data are missing at a time point, spline interpolation is used to fill it; For the region At the moment capacity data; S104, determine the regional transport core point to represent the regional transport center: rasterize different data according to the unified spatial resolution, and divide them into sub-regions to ensure that regional layout and transportation network data can be analyzed at the same spatial scale; In order to comprehensively consider the impact of adjacent areas on the region, the spatial density of population, economic activities and traffic activities in the sub-region is calculated using kernel density estimation; and the population density, economic activity density and traffic flow density are weighted and superimposed, and different attribute spatial densities are given different weights to obtain the first The comprehensive density of sub-regions : ； in, is the population density; is the density of economic activities; is the traffic flow density; is the kernel density function corresponding to population density; is the kernel density function corresponding to the density of economic activities; is the kernel density function corresponding to the traffic flow density; In the integrated density map, the centroid analysis method is used to identify the regions Comprehensive core transportation point.

2. The highway transport capacity distribution prediction method based on comprehensive long-term and short-term fluctuation trends according to claim 1 is characterized in that: The S2 is specifically: S201, node information construction: Based on the data obtained in S1, a highway transport capacity association network of regional hubs is constructed; in the highway transport capacity association network, the regional The location is determined by the core transport point coordinates. At each time point, The attribute is the carrying capacity of the corresponding area; S202, determine the node edge weight: in the highway transport capacity association network, the edge represents the transport capacity association relationship between regions; the edge Weight Responsive regional nodes and regional nodes The capacity flow between them, the weight of the edge is: ； in, For regional nodes and regional nodes The geographical distance between them is calculated by Dijkstra's algorithm applied on the road network; For regional nodes and regional nodes Between time The shipping time is derived from historical data; Node area node and regional nodes Between time Traffic flow, derived from historical data; 、 、 is the weight coefficient; S203, construct the highway transport capacity association network adjacency matrix: for each time, construct the adjacency matrix : ； Among them, if the regional node and regional nodes There is an edge between Less than threshold ,but ,otherwise ;in, is the edge weight of the adjacency matrix of the highway transport capacity association network; S204. Integrate the historical highway transport capacity correlation network: Integrate the highway transport capacity correlation networks of adjacent time slices to form a complete historical highway transport capacity correlation network ;in, is the node feature set of all time slices; is the set of adjacency matrices for all time slices.

3. The highway transport capacity distribution prediction method based on comprehensive long-term and short-term fluctuation trends according to claim 2 is characterized in that: The S3 is specifically: S301, preliminary decoupling: for the node capacity data within the time window, a moving average filter with different scale factors is used to remove the long-term and short-term trends of the capacity; the original capacity data is decomposed into long-term trend data and short-term fluctuation data. Node capacity sequence within a time window , long-term trend data is expressed as: ； in, is the original capacity data; An adjustable factor that controls the degree of decoupling for the moving average filter; For time The position index of The short-term trend data is: ； S302. Deep feature decoupling: Through wavelet decomposition, long-term trend data is further decoupled into multiple bands of different time scales to capture the capacity fluctuation characteristics at different time scales; short-term fluctuation data is decomposed into different frequency bands to capture small changes in a short period of time; for long-term trend data, the low-frequency band after decomposition is taken to represent the long-term trend characteristics; for short-term trend data, the high-frequency band after decomposition is taken to represent the short-term trend characteristics.

4. The method for predicting highway transport capacity distribution based on comprehensive long-term and short-term fluctuation trends according to claim 1 is characterized in that: The S4 is specifically: S401. For the input time window, identify external events that affect transport capacity; if the event occurs, the corresponding code value is 1; if the event does not occur, the code value is 0; Formally: ； in, Whether the event occurred; To capture the changing trends of transport capacity over daily, weekly, and monthly cycles, sine and cosine functions are used to encode time period features to map time periodicity into feature space, ensuring that the model can identify and utilize these periodic patterns. Assume that the cycle of a day is ,time The position during the day is , then the daily cycle code is: ； ； Assume that the weekly cycle is ,time Position in the week is , then the weekly cycle code is: ； ； Combine the periodic codes into a complete eigenvector: ； Time The external event code and periodic feature code are combined into a complete feature vector : ； in, For feature splicing; S402: For data in the input time window: use the dynamic time normalization algorithm to calculate the similarity between the time series of the input time window nodes. , set the similarity threshold ; If the regional node and regional nodes The similarity between , then update the weight If the regional node and regional nodes The similarity between , then it remains unchanged; S403, spatial feature extraction: for the highway transport capacity correlation network within a period of time, the graph attention mechanism is used to extract spatial dependency features from the highway transport capacity correlation network; in the In the layer graph attention mechanism, the node represents the vector By performing weighted summation with the features of adjacent nodes and updating the features through nonlinear transformation: ； in, For the The node feature matrix of the layer; is the capacity fluctuation correlation matrix; is the activation function; For the The learned weight matrix of the layer; For regional nodes and regional nodes The attention weight between them is calculated by the features of the node pair: ； in, It is the vector concatenation operation; is the vector used to calculate attention; S404, Time Feature Extraction: Use dilated convolution to extract time-dependent features in time series. Dilated convolution can capture long-term and short-term dependencies in time series by enhancing the receptive field of dilated convolution. Dilated convolution is used to operate on the input time series at each layer: ； in, For the The input sequence of the layer; is dilated convolution; is bias; Calculates convolution.

5. The method for predicting highway transport capacity distribution based on comprehensive long-term and short-term fluctuation trends according to claim 4 is characterized in that: The S5 is specifically: S501, fusion of spatiotemporal features: feature splicing of the last layer of temporal features and spatial features encoded in S4: ； in, For the last layer; Dynamically and hierarchically fuse the spatial and temporal features of the original sequence, long-term sequence, and short-term sequence; introduce an adaptive weight learning mechanism to dynamically learn the fusion weights of features of different scales through a neural network and set the weight parameters 、 and Corresponding to the original features , long-term characteristics and short-term characteristics , the multi-scale spatiotemporal features are: ； S502. Comprehensive feature fusion: Use the attention mechanism to fuse multi-scale spatiotemporal features with external influencing factor features, specifically: ； in, is the final fused cross-scale multi-period spatiotemporal feature vector; Coding features for external events and key time nodes; is the fusion feature dimension calculated using the attention mechanism; S503, receiving the cross-scale multi-period spatiotemporal feature vector after final fusion As input, the final fused cross-scale multi-period spatiotemporal feature vector Input into the multi-layer perceptron to generate the final prediction value: ； in, For historical highway capacity correlation network The predicted value of highway transport capacity includes the predicted value of transport capacity and relationship distribution characteristics of each regional node in the graph; For forward prediction time steps; S504, model training and updating: The training goal of the model is to minimize the predicted value and actual capacity The model parameters are optimized through training to generate the best prediction values ​​for actual capacity scheduling and management.

Images