Express delivery time prediction method and system
By using the Transformer architecture and machine learning methods, we constructed full-segment and partial time consumption models, which solved the problem of insufficient accuracy of existing express delivery time prediction schemes in complex scenarios, and achieved high-precision, stable and interpretable express delivery time prediction.
Patent Information
- Application Number
- CN202511456112.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-13
- Publication Date
- 2026-02-27
AI Technical Summary
Existing express delivery time prediction schemes rely too heavily on historical average delivery times or simple rules, failing to effectively integrate real-time dynamic factors. This results in insufficient accuracy of predictions in complex scenarios, a lack of interpretability and self-learning ability, and difficulty in adapting to changes in the logistics network.
By employing the Transformer architecture and machine learning methods, and acquiring the full lifecycle trajectory data of express delivery tracking numbers, we construct full-segment and local time consumption models. Combining DBSCAN and GMM models, we perform regional division and feature extraction, and establish a dual-line network point route selection time consumption compensation model to achieve real-time feature deep extraction and closed-loop optimization.
It improves prediction accuracy and system stability, quantifies prediction uncertainty, provides an interpretable time-consuming prediction process, achieves high-precision express delivery time prediction, and adapts to changes in logistics networks.
Smart Images
Figure CN121580347A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of machine learning technology, specifically to a method and system for predicting express delivery time. Background Technology
[0002] Existing express delivery time prediction solutions often rely excessively on historical average delivery times or simple rule-based judgments, failing to effectively integrate real-time dynamic factors such as weather and traffic. This results in insufficient accuracy of predictions in complex scenarios. Furthermore, these methods typically only output a single time point, failing to quantify the uncertainty of the prediction and lacking interpretability, making it difficult for operators to assess their reliability. More importantly, existing systems are mostly static models, lacking the ability to learn autonomously from anomalies and adapt to the continuous changes in the logistics network. This leads to a degradation in prediction performance over time, making it difficult to meet the demands of modern logistics for precise and intelligent predictions. Summary of the Invention
[0003] To solve the above-mentioned technical problems, a method and system for predicting express delivery time are provided. This technical solution solves the above-mentioned problems.
[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for predicting express delivery time includes: S1. Obtain the full lifecycle trajectory dataset of express delivery tracking numbers in several historical regions, mark the historical express delivery tracking number network point routing information in each region, establish a network point route selection time model for the entire segment, and initialize the global time distribution of neighboring network point routes in each associated region. S2. Based on the Transformer architecture, the historical express delivery tracking number network routing information of each region is divided according to the various task stages of express delivery at each network point. A local time consumption model of each task stage of express delivery is constructed to determine the feature vector of the time consumption of each task stage of express delivery between neighboring network points in each region. S3. Based on the time consumption and impact feature vector of each task stage of express delivery between neighboring network points in each region, establish a dual-line network point route selection time compensation model, and compensate for the global time consumption distribution of the route between neighboring network points in each region to obtain the standard time consumption compensation value of the route between neighboring network points in each region. S4. Obtain the real-time express tracking number and branch routing information for each region. Based on the standard time compensation value of the neighboring branch routes for each region, verify whether the location of the real-time express tracking number and branch routing information is within the normal range. If yes, output the total standard time. If no, send out the collection of time consumption factors of abnormal neighboring branch routes, substitute them into the corresponding branch route selection time consumption model, and update the standard time consumption distribution of neighboring branch routes for each region.
[0005] Preferably, step S1 specifically includes: Based on the historical full lifecycle trajectory dataset of express tracking numbers in several regions, the historical express tracking number network point routing information of each region is marked; the historical express tracking number network point routing information includes: region code, express tracking number code, network point number, reachable routes between neighboring network points, and operation type (pickup timestamp, transportation timestamp, transit timestamp, and signature timestamp). Based on the historical express tracking number and outlet routing information of each region, the outlets in each region are used as nodes and the reachable routes between nodes are used as edges to generate the historical express tracking number and outlet routing sequence of each region. Based on the historical express tracking number and outlet routing sequence in each region, a binary scatter plot of historical express tracking number and outlet routing in each region is established, with the horizontal axis representing the signing outlet, the vertical axis representing the pickup outlet, and the historical express tracking number and outlet routing information as attributes. Based on the DBSCAN density clustering function, for the binary scatter plot of historical express delivery tracking number network points within a region, the regions of historical express delivery tracking number network point routes are used as cluster centers to divide and cluster the binary scatter plots of historical express delivery tracking number network point routes, thus obtaining a structured dataset of historical express delivery tracking number network point routes in each associated region.
[0006] Preferably, step S1 further includes: Based on the Gaussian mixture model (GMM), a time consumption model for route selection of all network points is established. Using the density probability function, the maximum prior probability of route selection for network point routes of historical express tracking numbers in the structured dataset of network point routes in each associated area is calculated and marked, and the optimal route for network point routes of historical express tracking numbers in each associated area is determined. Based on the expectation-maximization algorithm, and utilizing distributed maximum likelihood estimation, the mean and standard deviation of the time consumption between each network point in the optimal route for historical express tracking numbers within each associated region are calculated. The global time consumption distribution of the routes between neighboring network points in each associated region per unit time is then calculated as follows:
[0007] in, The global time consumption distribution of the neighboring network points of each associated region in the t-th unit of time. Given the structured routing data of the i-th historical express tracking number within each associated region, we observe the total prior probability density of the k-th route choice for the route to the network point. Choose the prior probability weight for the k-th route. Given the optimal route k for historical express tracking numbers within each associated region, we have a normal distribution of the time consumption of the i-th historical express tracking number route structured data, considering the mean and variance of time consumption among all network points. Let k be the average time taken for the k-th line. Let be the standard deviation of the k-th line. For standardized constants, It is an exponential function. For the function that maximizes the value, The function is a Gaussian distribution. This represents the total number of lines.
[0008] Preferably, step S2 specifically includes: The structured dataset of historical express tracking numbers and network point routes in each associated region is divided according to the various task stages of express delivery at each network point, resulting in a structured data array of historical express tracking numbers and network point routes for each task stage in each associated region. Based on the structured data array of historical express tracking number outlets in each task stage within each associated region, the global time consumption of each task stage within each associated region is used as the observation window, and the structured data of historical express tracking number outlets in each task stage is used as the observation object to obtain the structured time series data of historical express tracking number outlets in each task stage within each associated region. The temporal characteristics of the historical express tracking number and outlet routing structured time series data of each task stage in each associated region are statistically analyzed. Substituted into the STL-based time series decomposition, the trend, seasonality and residual of each task stage are extracted to obtain the static features and dynamic context features of the historical express tracking number and outlet routing of each task stage in the associated region. Using the Pearson correlation coefficient, the attention value between the time consumption of each task stage and the static and dynamic context features of historical express tracking number network point routing in each task stage within the associated region is calculated as follows:
[0009] in, The focus value is the static feature of the routing between the time spent in the a-th task stage within the associated region and the i-th historical express tracking number network point in the a-th task stage. The time consumption of the a-th task stage within the associated region is the dynamic context feature attention value of the route for the i-th historical express tracking number network point in the a-th task stage. For the i-th historical express tracking number of the network point routing static feature in the a-th task phase, Let be the mean of the static routing characteristics of the ith historical express tracking number at the network point in the ith task phase a. For the i-th historical express tracking number of the network point routing dynamic context feature in the a-th task stage, Let be the mean of the dynamic context features of the route for the ith historical express tracking number at the network point in the ith task phase a. The time taken for the a-th task stage within the associated region. Let L be the average time taken for the a-th task stage within the associated region, and L be the total number of historical express tracking number routes. Using Principal Component Analysis (PCA), the covariance matrix of the attention values between the static and dynamic context features of the historical express tracking number network point routing for each task stage in each region is extracted. The static and dynamic context features are then decomposed and sorted. The contribution load of the principal component with the largest static and dynamic context features is calculated, and weights are assigned to the attention values between the static and dynamic context features of the historical express tracking number network point routing for each task stage in each region.
[0010] Preferably, step S2 further includes: Based on the Transformer architecture, independent encoders are established for each task stage within each region. The attention value weights between the static features and dynamic context features of the historical express tracking number and network point routing of each task stage within each region are used as the benchmark values for the local self-attention mechanism of the corresponding model. Using an RNN recurrent convolutional neural network, temporal position encoders for each task stage within each region are established. The structured temporal data of the historical express tracking number and network point routing of each task stage within each associated region are used as input, and the position encoding of the static features and dynamic context features of the historical express tracking number and network point routing of each task stage are used as output. The time consumption of each task stage is used as the input of the static feature location encoding and dynamic context feature location encoding of the historical express tracking number and the network point routing of each task stage in each task stage. The time consumption of each task stage in each region is predicted as the output, and the context-aware hidden state sequence of the last layer output of the independent encoder of each task stage is marked. The context-aware hidden state sequences output from the final layer of the independent encoders at each task stage are substituted into the global self-attention mechanism of the Transformer. The most influential feature factors of each context-aware hidden state sequence relative to the time consumption of each task stage are fused in an explicit weighted manner to obtain the feature vector of the time consumption attention influence of each task stage of express delivery between neighboring network points in each region.
[0011] Preferably, step S3 specifically includes: Using LDA linear discriminant analysis, the dimensionality of the feature vectors of time consumption, attention and impact of each task stage in the express delivery of neighboring network points in each region is reduced to obtain the dimensionality-reduced feature vectors of time consumption, attention and impact of each task stage in the express delivery of neighboring network points in each region. Based on linear regression, a mean-compensated regression model is established, with the dimensionality-reduced vector of the time consumption and impact features of each task stage of express delivery between neighboring network points in each region as the input, and the mean of the time consumption distribution of each task stage of express delivery between neighboring network points in each region as the output. Based on linear regression, a variance scaling regression model is established. The reduced-dimensional vector of the time consumption, attention, and impact characteristics of each task stage of express delivery between neighboring network points in each region is used as input, and the comprehensive scaling factor of the time consumption, attention, and impact of each task stage of express delivery between neighboring network points in each region is used as output. Substituting these values into the sigmoid function, the standard deviation of the time consumption distribution of each task stage of express delivery between neighboring network points in each region is generated.
[0012] Preferably, step S3 further includes: Based on the mean-compensated regression model and the variance-scaled regression model, a dual-line network route selection time compensation model is constructed. By using the mean of the time distribution of each task stage of express delivery between neighboring network points in each region, and using the weighted average method, the global time distribution average of the route between neighboring network points in each region is compensated to obtain the optimal estimate of the standard time of the route between neighboring network points in each region. By using the standard deviation of the time distribution of each task stage of express delivery between neighboring network points in each region, the actual time distribution range of the optimal estimate of the standard time of the route between neighboring network points in each region is verified, and the standard time fluctuation value of the route between neighboring network points in each region is obtained. The optimal estimate of the standard time of the neighboring network points in each region is added to the standard time fluctuation value of the neighboring network points in each region to obtain the time compensation value of the neighboring network points in each region.
[0013] Furthermore, a predictive delivery time system for express delivery, used to implement the express delivery time prediction method described above, includes: The module includes a global time consumption module, a local influencing factor screening module, a global time consumption compensation module, and a monitoring module. The global time consumption module is used to obtain the full life cycle trajectory dataset of express delivery tracking numbers in several historical regions, mark the historical express delivery tracking number network point routing information of each region, establish a network point route selection time consumption model for the entire segment, and initialize the global time consumption distribution of neighboring network point routes in each associated region. The local influencing factor screening module is used to divide the historical express delivery number network routing information of each region based on the Transformer architecture, and construct a local time consumption model for each task stage of express delivery, and determine the time consumption and influence feature vector of each task stage of express delivery between neighboring network points in each region. The global time consumption compensation module is electrically connected to the local influencing factor screening module and the global time consumption module. The global time consumption compensation module is used to establish a dual-line network point route selection time consumption compensation model based on the time consumption attention influence feature vector of each task stage of express delivery between neighboring network points in each region. It compensates for the global time consumption distribution of the neighboring network point routes in each region to obtain the standard time consumption compensation value of the neighboring network point routes in each region. The monitoring module is electrically connected to the global time consumption compensation module. The monitoring module is used to obtain the real-time express tracking number network point routing information of each region. According to the standard time consumption compensation value of the neighboring network point lines in each region, it verifies whether the location of the real-time express tracking number network point routing information is within the normal range. If so, it outputs the total standard time consumption. If not, it sends out the time consumption factors of abnormal neighboring network point lines, substitutes them into the corresponding network point line selection time consumption model, and updates the standard time consumption distribution of neighboring network point lines in each region.
[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention proposes a scheme for predicting the estimated delivery time of express packages. Through historical benchmark initialization, real-time deep feature extraction, dual-channel probability compensation, and closed-loop self-optimization steps, it achieves a comprehensive improvement in prediction accuracy, system stability, and intelligent decision-making. This scheme not only integrates multi-dimensional data from both short and long-term periods to quantify prediction uncertainties and significantly improve forecast accuracy, but also ensures output stability in the event of data anomalies through a layered architecture. It provides a fully interpretable time-consuming prediction process and confidence interval, achieving high-precision estimated delivery time prediction and providing logistics companies with reliable and accurate timeliness prediction services. Attached Figure Description
[0015] Figure 1 A flowchart of a method for predicting express delivery time; Figure 2 This is a framework diagram of a predictive delivery time system for express delivery. Detailed Implementation
[0016] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.
[0017] Reference Figure 1 As shown, a method for predicting express delivery time includes: S1. Obtain the full lifecycle trajectory dataset of express delivery tracking numbers in several historical regions, mark the historical express delivery tracking number network point routing information in each region, establish a network point route selection time model for the entire segment, and initialize the global time distribution of neighboring network point routes in each associated region. Step S1 specifically includes: Based on the historical full lifecycle trajectory dataset of express tracking numbers in several regions, the historical express tracking number network point routing information of each region is marked; the historical express tracking number network point routing information includes: region code, express tracking number code, network point number, reachable routes between neighboring network points, and operation type (pickup timestamp, transportation timestamp, transit timestamp, and signature timestamp). Based on the historical express tracking number and outlet routing information of each region, the outlets in each region are used as nodes and the reachable routes between nodes are used as edges to generate the historical express tracking number and outlet routing sequence of each region. Based on the historical express tracking number and outlet routing sequence in each region, a binary scatter plot of historical express tracking number and outlet routing in each region is established, with the horizontal axis representing the signing outlet, the vertical axis representing the pickup outlet, and the historical express tracking number and outlet routing information as attributes. Based on the DBSCAN density clustering function, for the binary scatter plot of historical express delivery tracking number network points within a region, the regions of historical express delivery tracking number network point routes are used as cluster centers to divide and cluster the binary scatter plots of historical express delivery tracking number network point routes, thus obtaining a structured dataset of historical express delivery tracking number network point routes in each associated region.
[0018] As a further point, existing express delivery time predictions typically divide regions into geographical areas such as East China, North China, South China, and Central China. This results in East China including both the extremely dense Yangtze River Delta urban cluster (Shanghai, Suzhou, Hangzhou, Ningbo, etc., with network distances potentially less than 100 kilometers) and sparsely networked mountainous or rural areas (such as southern Anhui and western Zhejiang mountainous areas, with network distances potentially exceeding 200 kilometers). This results in two regions with drastically different network densities being grouped into the same region for time-consuming modeling, severely obscuring the model's accuracy. Consequently, the variance in time-consuming distribution in East China becomes very large, as it includes both high-frequency intercity transport and low-frequency remote routes, leading to an overly wide prediction range and a loss of precision.
[0019] Therefore, DBSCAN can automatically cluster dense points in different regions into a related region, while identifying points in sparse regions as noise or separate clusters, thus ensuring that the network density within each region is similar and providing accurate parameters for subsequent modeling.
[0020] Step S1 also includes: Based on the Gaussian mixture model (GMM), a time consumption model for route selection of all network points is established. Using the density probability function, the maximum prior probability of route selection for network point routes of historical express tracking numbers in the structured dataset of network point routes in each associated area is calculated and marked, and the optimal route for network point routes of historical express tracking numbers in each associated area is determined. Based on the expectation-maximization algorithm, and utilizing distributed maximum likelihood estimation, the mean and standard deviation of the time consumption between each network point in the optimal route for historical express tracking numbers within each associated region are calculated. The global time consumption distribution of the routes between neighboring network points in each associated region per unit time is then calculated as follows:
[0021] in, The global time consumption distribution of the neighboring network points of each associated region in the t-th unit of time. Given the structured routing data of the i-th historical express tracking number within each associated region, we observe the total prior probability density of the k-th route choice for the route to the network point. Choose the prior probability weight for the k-th route. Given the optimal route k for historical express tracking numbers within each associated region, we have a normal distribution of the time consumption of the i-th historical express tracking number route structured data, considering the mean and variance of time consumption among all network points. Let k be the average time taken for the k-th line. Let be the standard deviation of the k-th line. For standardized constants, It is an exponential function. For the function that maximizes the value, The function is a Gaussian distribution. This represents the total number of lines; When using it, please refer to the steps outlined above: As a further development, DBSCAN density clustering is used to automatically divide closely connected logistics areas based on the actual flow of express delivery, replacing subjective geographical division. Then, a Gaussian mixture model (GMM) is used to model the route selection between network points as a probabilistic generation process. The optimal path and its time distribution (mean and standard deviation) are fitted by the expectation-maximization (EM) algorithm, thereby automatically and intelligently constructing a global time probability benchmark for each neighboring route. This achieves the automatic and objective extraction of key logistics network features and stable time benchmarks from raw data without relying on human experience, improving the robustness of subsequent accurate predictions.
[0022] S2. Based on the Transformer architecture, the historical express delivery tracking number network routing information of each region is divided according to the various task stages of express delivery at each network point. A local time consumption model of each task stage of express delivery is constructed to determine the feature vector of the time consumption of each task stage of express delivery between neighboring network points in each region. Step S2 specifically includes: The structured dataset of historical express tracking numbers and network point routes in each associated region is divided according to the various task stages of express delivery at each network point, resulting in a structured data array of historical express tracking numbers and network point routes for each task stage in each associated region. Based on the structured data array of historical express tracking number outlets in each task stage within each associated region, the global time consumption of each task stage within each associated region is used as the observation window, and the structured data of historical express tracking number outlets in each task stage is used as the observation object to obtain the structured time series data of historical express tracking number outlets in each task stage within each associated region. The temporal characteristics of the historical express tracking number and outlet routing structured time series data of each task stage in each associated region are statistically analyzed. Substituted into the STL-based time series decomposition, the trend, seasonality and residual of each task stage are extracted to obtain the static features and dynamic context features of the historical express tracking number and outlet routing of each task stage in the associated region. As a further development, the static characteristics of historical express tracking numbers and outlet routes in each task stage within the associated area include: route distance, outlet type, etc., while the dynamic contextual characteristics of historical express tracking numbers and outlet routes in each task stage within the associated area include: real-time weather, real-time traffic index, real-time business volume, etc. Using the Pearson correlation coefficient, the attention value between the time consumption of each task stage and the static and dynamic context features of historical express tracking number network point routing in each task stage within the associated region is calculated as follows:
[0023] in, The focus value is the static feature of the routing between the time spent in the a-th task stage within the associated region and the i-th historical express tracking number network point in the a-th task stage. The time consumption of the a-th task stage within the associated region is the dynamic context feature attention value of the route for the i-th historical express tracking number network point in the a-th task stage. For the i-th historical express tracking number of the network point routing static feature in the a-th task phase, Let be the mean of the static routing characteristics of the ith historical express tracking number at the network point in the ith task phase a. For the i-th historical express tracking number of the network point routing dynamic context feature in the a-th task stage, Let be the mean of the dynamic context features of the route for the ith historical express tracking number at the network point in the ith task phase a. The time taken for the a-th task stage within the associated region. Let L be the average time taken for the a-th task stage within the associated region, and L be the total number of historical express tracking number routes. Using PCA principal component analysis, the covariance matrix of the attention values between the static features and dynamic context features of the historical express tracking number network point routing of each task stage in each region is extracted. The static feature values and dynamic context feature values are decomposed and sorted. The contribution load of the principal component with the largest static feature value and dynamic context feature value is calculated. The attention values between the static features and dynamic context features of the historical express tracking number network point routing of each task stage in each region are weighted. Step S2 also includes: Based on the Transformer architecture, independent encoders are established for each task stage within each region. The attention value weights between the static features and dynamic context features of the historical express tracking number and network point routing of each task stage within each region are used as the benchmark values for the local self-attention mechanism of the corresponding model. Using an RNN recurrent convolutional neural network, temporal position encoders for each task stage within each region are established. The structured temporal data of the historical express tracking number and network point routing of each task stage within each associated region are used as input, and the position encoding of the static features and dynamic context features of the historical express tracking number and network point routing of each task stage are used as output. The time consumption of each task stage is used as the input of the static feature location encoding and dynamic context feature location encoding of the historical express tracking number and the network point routing of each task stage in each task stage. The time consumption of each task stage in each region is predicted as the output, and the context-aware hidden state sequence of the last layer output of the independent encoder of each task stage is marked. The context-aware hidden state sequences output from the final layer of the independent encoders at each task stage are substituted into the global self-attention mechanism of the Transformer. The most influential feature factors of each context-aware hidden state sequence relative to the time consumption of each task stage are fused in an explicit weighted manner to obtain the feature vector of the time consumption attention influence of each task stage of express delivery between neighboring network points in each region.
[0024] When using it, please refer to the steps outlined above: As a further development, this study leverages statistical priors to guide deep learning attention. First, it quantifies the influence weights of static and dynamic features on time consumption at each task stage using Pearson correlation coefficients and principal component analysis (PCA), injecting this as prior knowledge into the Transformer model. Then, it innovatively employs an RNN to generate dynamic contextual positional codes, which are fused with the features and input into an independent encoder for training. Finally, by extracting the context-aware hidden states from the encoder output and fusing them with global attention weights, a highly condensed and physically meaningful feature vector influencing time consumption is generated. The beneficial effects are a significant improvement in the model's interpretability, convergence speed, and noise resistance, enabling the final feature vector to accurately represent the core factors affecting time consumption, providing an extremely reliable and efficient input for subsequent time series prediction.
[0025] S3. Based on the time consumption and impact feature vector of each task stage of express delivery between neighboring network points in each region, establish a dual-line network point route selection time compensation model, and compensate for the global time consumption distribution of the route between neighboring network points in each region to obtain the standard time consumption compensation value of the route between neighboring network points in each region. Step S3 specifically includes: Using LDA linear discriminant analysis, the dimensionality of the feature vectors of time consumption, attention and impact of each task stage in the express delivery of neighboring network points in each region is reduced to obtain the dimensionality-reduced feature vectors of time consumption, attention and impact of each task stage in the express delivery of neighboring network points in each region. Based on linear regression, a mean-compensated regression model is established, with the dimensionality-reduced vector of the time consumption and impact features of each task stage of express delivery between neighboring network points in each region as the input, and the mean of the time consumption distribution of each task stage of express delivery between neighboring network points in each region as the output. Based on linear regression, a variance scaling regression model is established. The dimensionality reduction vector of the time consumption, attention and impact characteristics of each task stage of express delivery between neighboring network points in each region is used as input, and the comprehensive scaling factor of the time consumption, attention and impact of each task stage of express delivery between neighboring network points in each region is used as output. Substituting into the sigmoid function, the standard deviation of the time consumption distribution of each task stage of express delivery between neighboring network points in each region is generated. Step S3 also includes: Based on the mean-compensated regression model and the variance-scaled regression model, a dual-line network route selection time compensation model is constructed. By using the mean of the time distribution of each task stage of express delivery between neighboring network points in each region, and using the weighted average method, the global time distribution average of the route between neighboring network points in each region is compensated to obtain the optimal estimate of the standard time of the route between neighboring network points in each region. By using the standard deviation of the time distribution of each task stage of express delivery between neighboring network points in each region, the actual time distribution range of the optimal estimate of the standard time of the route between neighboring network points in each region is verified, and the standard time fluctuation value of the route between neighboring network points in each region is obtained. The optimal estimate of the standard time of the neighboring network points in each region is added to the standard time fluctuation value of the neighboring network points in each region to obtain the time compensation value of the neighboring network points in each region.
[0026] When using it, please refer to the steps outlined above: As a further step, dimensionality reduction of the high-dimensional feature vector is performed using LDA. The purpose is not only to reduce the amount of data but also to extract the most discriminative features that can distinguish different time consumption patterns (such as normal, delayed, and early). This provides the most informative input for the subsequent regression model, improving its performance. By establishing a bilinear regression channel (based on the principle of independent compensation for mean and variance), the mean compensation channel learns the linear mapping relationship between the dimensionality-reduced features and the time consumption deviation. The model weights intuitively reflect the contribution of different features (such as weather and traffic) to the time consumption variation. For example, for features with positive and large weights, an increase in their value will increase the predicted time consumption compensation value (Δμ). Secondly, the variance compensation channel uses the sigmoid function to constrain the linear regression output to a positive range (such as 0.5 to 2.0) as a scaling factor. This model learns the relationship between features and time consumption volatility. For example, when the "severe weather" feature value is high, the model outputs a scaling factor greater than 1, thus amplifying the standard deviation of the global baseline and reflecting higher uncertainty. Finally, a weighted average method is used to fuse the global baseline mean and the predicted mean from each stage, combining information from different sources (historical statistics vs. real-time predictions). This reduces the risk of a single model prediction and yields a more robust optimal estimate. The standard time-consuming fluctuation value is essentially the standard deviation of the final compensated distribution, forming a prediction interval used to quantify the uncertainty of the prediction. The final output time-consuming compensation value is not a single value but an interval ([optimal estimate - fluctuation value, optimal estimate + fluctuation value]). This provides a complete and confident prediction result for downstream systems.
[0027] S4. Obtain the real-time express tracking number and branch routing information for each region. Based on the standard time compensation value of the neighboring branch routes for each region, verify whether the location of the real-time express tracking number and branch routing information is within the normal range. If yes, output the total standard time. If no, send out the collection of time consumption factors of abnormal neighboring branch routes, substitute them into the corresponding branch route selection time consumption model, and update the standard time consumption distribution of neighboring branch routes for each region.
[0028] Reference Figure 2 As shown, a system for predicting the estimated delivery time of a courier service includes: The module includes a global time consumption module, a local influencing factor screening module, a global time consumption compensation module, and a monitoring module. The global time consumption module is used to obtain the full life cycle trajectory dataset of express delivery tracking numbers in several historical regions, mark the historical express delivery tracking number network point routing information of each region, establish a network point route selection time consumption model for the entire segment, and initialize the global time consumption distribution of neighboring network point routes in each associated region. The local influencing factor screening module is used to divide the historical express delivery number network routing information of each region based on the Transformer architecture, and construct a local time consumption model for each task stage of express delivery, and determine the time consumption and influence feature vector of each task stage of express delivery between neighboring network points in each region. The global time consumption compensation module is electrically connected to the local influencing factor screening module and the global time consumption module. The global time consumption compensation module is used to establish a dual-line network point route selection time consumption compensation model based on the time consumption attention influence feature vector of each task stage of express delivery between neighboring network points in each region. It compensates for the global time consumption distribution of the neighboring network point routes in each region to obtain the standard time consumption compensation value of the neighboring network point routes in each region. The monitoring module is electrically connected to the global time consumption compensation module. The monitoring module is used to obtain the real-time express tracking number network point routing information of each region. According to the standard time consumption compensation value of the neighboring network point lines in each region, it verifies whether the location of the real-time express tracking number network point routing information is within the normal range. If so, it outputs the total standard time consumption. If not, it sends out the time consumption factors of abnormal neighboring network point lines, substitutes them into the corresponding network point line selection time consumption model, and updates the standard time consumption distribution of neighboring network point lines in each region.
[0029] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.
Claims
1. A method for predicting express delivery time, characterized in that, include: S1. Obtain the full lifecycle trajectory dataset of express delivery tracking numbers in several historical regions, mark the historical express delivery tracking number network point routing information in each region, establish a network point route selection time model for the entire segment, and initialize the global time distribution of neighboring network point routes in each associated region. S2. Based on the Transformer architecture, the historical express delivery tracking number network routing information of each region is divided according to the various task stages of express delivery at each network point. A local time consumption model of each task stage of express delivery is constructed to determine the feature vector of the time consumption of each task stage of express delivery between neighboring network points in each region. S3. Based on the time consumption and impact feature vector of each task stage of express delivery between neighboring network points in each region, establish a dual-line network point route selection time compensation model, and compensate for the global time consumption distribution of the route between neighboring network points in each region to obtain the standard time consumption compensation value of the route between neighboring network points in each region. S4. Obtain the real-time express tracking number and branch routing information for each region. Based on the standard time compensation value of the neighboring branch routes for each region, verify whether the location of the real-time express tracking number and branch routing information is within the normal range. If yes, output the total standard time. If no, send out the time consumption factors of abnormal neighboring branch routes, substitute them into the corresponding branch route selection time model, and update the standard time distribution of neighboring branch routes for each region.
2. The method for predicting express delivery time according to claim 1, characterized in that, Step S1 specifically includes: Based on the historical full lifecycle trajectory dataset of express tracking numbers from several regions, the historical express tracking number network point routing information for each region is marked; the historical express tracking number network point routing information includes: region code, express tracking number code, network point number, reachable routes between neighboring network points, and operation type. Based on the historical express tracking number and outlet routing information of each region, the outlets in each region are used as nodes and the reachable routes between nodes are used as edges to generate the historical express tracking number and outlet routing sequence of each region. Based on the historical express tracking number and outlet routing sequence in each region, a binary scatter plot of historical express tracking number and outlet routing in each region is established, with the horizontal axis representing the signing outlet, the vertical axis representing the pickup outlet, and the historical express tracking number and outlet routing information as attributes. Based on the DBSCAN density clustering function, for the binary scatter plot of historical express delivery tracking number network points within a region, the regions of historical express delivery tracking number network point routes are used as cluster centers to divide and cluster the binary scatter plots of historical express delivery tracking number network point routes, thus obtaining a structured dataset of historical express delivery tracking number network point routes in each associated region.
3. The method for predicting express delivery time according to claim 2, characterized in that, Step S1 also includes: Based on the Gaussian mixture model (GMM), a time consumption model for route selection of all network points is established. Using the density probability function, the maximum prior probability of route selection for network point routes of historical express tracking numbers in the structured dataset of network point routes in each associated area is calculated and marked, and the optimal route for network point routes of historical express tracking numbers in each associated area is determined. Based on the expectation-maximization algorithm, and utilizing distributed maximum likelihood estimation, the mean and standard deviation of the time consumption between each network point in the optimal route for historical express tracking numbers within each associated region are calculated. The global time consumption distribution of the routes between neighboring network points in each associated region per unit time is then calculated as follows: ； in, Let be the global time distribution of the lines between neighboring network points in each associated region at time t. Given the structured routing data of the i-th historical express tracking number within each associated region, we observe the total prior probability density of the k-th route choice for the route to the network point. Choose the prior probability weight for the k-th route. Given the optimal route k for historical express tracking numbers within each associated region, we have a normal distribution of the time consumption of the i-th historical express tracking number route structured data, considering the mean and variance of time consumption among all network points. Let k be the average time taken for the k-th line. Let be the standard deviation of the k-th line. For standardized constants, It is an exponential function. For the function that maximizes the value, The function is a Gaussian distribution. This represents the total number of lines.
4. The method for predicting express delivery time according to claim 3, characterized in that, Step S2 specifically includes: The structured dataset of historical express tracking numbers and network point routes in each associated region is divided according to the various task stages of express delivery at each network point, resulting in a structured data array of historical express tracking numbers and network point routes for each task stage in each associated region. Based on the structured data array of historical express tracking number outlets in each task stage within each associated region, the global time consumption of each task stage within each associated region is used as the observation window, and the structured data of historical express tracking number outlets in each task stage is used as the observation object to obtain the structured time series data of historical express tracking number outlets in each task stage within each associated region. The temporal characteristics of the historical express tracking number and outlet routing structured time series data of each task stage in each associated region are statistically analyzed. Substituted into the STL-based time series decomposition, the trend, seasonality and residual of each task stage are extracted to obtain the static features and dynamic context features of the historical express tracking number and outlet routing of each task stage in the associated region. Using the Pearson correlation coefficient, the attention value between the time consumption of each task stage and the static and dynamic context features of historical express tracking number network point routing in each task stage within the associated region is calculated as follows: ； in, The focus value is the static feature of the routing between the time spent in the a-th task stage within the associated region and the i-th historical express tracking number network point in the a-th task stage. The time consumption of the a-th task stage within the associated region is the dynamic context feature attention value of the route for the i-th historical express tracking number network point in the a-th task stage. For the i-th historical express tracking number of the network point routing static feature in the a-th task phase, Let be the mean of the static routing characteristics of the ith historical express tracking number at the network point in the ith task phase a. For the i-th historical express tracking number of the network point routing dynamic context feature in the a-th task stage, Let be the mean of the dynamic context features of the route for the ith historical express tracking number at the network point in the ith task phase a. The time taken for the a-th task stage within the associated region. Let L be the average time taken for the a-th task stage within the associated region, and L be the total number of historical express tracking number routes. Using Principal Component Analysis (PCA), the covariance matrix of the attention values between the static and dynamic context features of the historical express tracking number network point routing for each task stage in each region is extracted. The static and dynamic context features are then decomposed and sorted. The contribution load of the principal component with the largest static and dynamic context features is calculated, and weights are assigned to the attention values between the static and dynamic context features of the historical express tracking number network point routing for each task stage in each region.
5. The method for predicting express delivery time according to claim 4, characterized in that, Step S2 also includes: Based on the Transformer architecture, independent encoders are established for each task stage within each region. The attention value weights between the static features and dynamic context features of the historical express tracking number and network point routing of each task stage within each region are used as the benchmark values for the local self-attention mechanism of the corresponding model. Using an RNN recurrent convolutional neural network, temporal position encoders for each task stage within each region are established. The structured temporal data of the historical express tracking number and network point routing of each task stage within each associated region are used as input, and the position encoding of the static features and dynamic context features of the historical express tracking number and network point routing of each task stage are used as output. The time consumption of each task stage is used as the input of the static feature location encoding and dynamic context feature location encoding of the historical express tracking number and the network point routing of each task stage in each task stage. The time consumption of each task stage in each region is predicted as the output, and the context-aware hidden state sequence of the last layer output of the independent encoder of each task stage is marked. The context-aware hidden state sequences output from the final layer of the independent encoders at each task stage are substituted into the global self-attention mechanism of the Transformer. The most influential feature factors of each context-aware hidden state sequence relative to the time consumption of each task stage are fused in an explicit weighted manner to obtain the feature vector of the time consumption attention influence of each task stage of express delivery between neighboring network points in each region.
6. The method for predicting express delivery time according to claim 5, characterized in that, Step S3 specifically includes: Using LDA linear discriminant analysis, the dimensionality of the feature vectors of time consumption, attention and impact of each task stage in the express delivery of neighboring network points in each region is reduced to obtain the dimensionality-reduced feature vectors of time consumption, attention and impact of each task stage in the express delivery of neighboring network points in each region. Based on linear regression, a mean-compensated regression model is established, with the dimensionality-reduced vector of the time consumption and impact features of each task stage of express delivery between neighboring network points in each region as the input, and the mean of the time consumption distribution of each task stage of express delivery between neighboring network points in each region as the output. Based on linear regression, a variance scaling regression model is established. The dimensionality-reduced vector of the time consumption, attention and impact features of each task stage of express delivery between neighboring network points in each region is used as input, and the comprehensive scaling factor of the time consumption, attention and impact features of each task stage of express delivery between neighboring network points in each region is used as output. Substituting these factors into the sigmoid function, the standard deviation of the time consumption distribution of each task stage of express delivery between neighboring network points in each region is generated.
7. The method for predicting express delivery time according to claim 6, characterized in that, Step S3 also includes: Based on the mean-compensated regression model and the variance-scaled regression model, a dual-line network route selection time compensation model is constructed. By using the mean of the time distribution of each task stage of express delivery between neighboring network points in each region, and using the weighted average method, the global time distribution average of the route between neighboring network points in each region is compensated to obtain the optimal estimate of the standard time of the route between neighboring network points in each region. By using the standard deviation of the time distribution of each task stage of express delivery between neighboring network points in each region, the actual time distribution range of the optimal estimate of the standard time of the route between neighboring network points in each region is verified, and the standard time fluctuation value of the route between neighboring network points in each region is obtained. The optimal estimate of the standard time of the neighboring network points in each region is added to the standard time fluctuation value of the neighboring network points in each region to obtain the time compensation value of the neighboring network points in each region.
8. A system for predicting the estimated delivery time of express delivery, characterized in that, A method for predicting express delivery time as described in any one of claims 1-7, comprising: The module includes a global time consumption module, a local influencing factor screening module, a global time consumption compensation module, and a monitoring module. The global time consumption module is used to obtain the full life cycle trajectory dataset of express delivery tracking numbers in several historical regions, mark the historical express delivery tracking number network point routing information of each region, establish a network point route selection time consumption model for the entire segment, and initialize the global time consumption distribution of neighboring network point routes in each associated region. The local influencing factor screening module is used to divide the historical express delivery number network routing information of each region based on the Transformer architecture, and construct a local time consumption model for each task stage of express delivery, and determine the time consumption and influence feature vector of each task stage of express delivery between neighboring network points in each region. The global time consumption compensation module is electrically connected to the local influencing factor screening module and the global time consumption module. The global time consumption compensation module is used to establish a dual-line network point route selection time consumption compensation model based on the time consumption attention influence feature vector of each task stage of express delivery between neighboring network points in each region. It compensates for the global time consumption distribution of the neighboring network point routes in each region to obtain the standard time consumption compensation value of the neighboring network point routes in each region. The monitoring module is electrically connected to the global time consumption compensation module. The monitoring module is used to obtain the real-time express tracking number network point routing information of each region. According to the standard time consumption compensation value of the neighboring network point lines in each region, it verifies whether the location of the real-time express tracking number network point routing information is within the normal range. If so, it outputs the total standard time consumption. If not, it sends out the time consumption factors of abnormal neighboring network point lines, substitutes them into the corresponding network point line selection time consumption model, and updates the standard time consumption distribution of neighboring network point lines in each region.