Arrival time prediction method, device and equipment of logistics vehicle and storage medium

Abstract

The application discloses a logistics vehicle arrival time prediction method, device and equipment and a storage medium, relates to the technical field of logistics, and the method is characterized in that: the method collects multi-source heterogeneous data of a logistics vehicle in real time, pre-processes the data, and constructs a continuously updated time series data set; feature engineering is performed based on a Pearson correlation coefficient and principal component analysis to obtain a standardized feature vector; an attention mechanism is used to adaptively focus on high-correlation features, a long short-term memory network is combined to model time series dependency, and an initial prediction value is output; a sliding window mechanism is triggered by data change weight detection to dynamically update the prediction result; a prediction deviation threshold is used to determine a model update instruction; and model parameters are adaptively optimized through average absolute error back propagation in response to the update instruction. Through multi-source data fusion perception, dynamic feature updating and model adaptive optimization, the application solves the technical problems that the prediction deviation of the existing method is continuously accumulated and cannot be dynamically updated according to real-time data changes.

Description

Technical Field

[0001] This application relates to the field of logistics technology, and in particular to a method, apparatus, equipment and storage medium for predicting the arrival time of logistics vehicles. Background Technology

[0002] The rapid development of the logistics industry has placed higher demands on transportation efficiency. Vehicle arrival time prediction, as a core component of logistics scheduling and route planning, directly impacts the allocation of loading and unloading resources within the park, vehicle queue management, and delivery timeliness. Existing vehicle arrival time prediction technologies mainly employ a simple calculation method of dividing static road network distance by average vehicle speed, or rely on fixed travel time estimates provided by map APIs. These methods cannot detect real-time changes in road congestion and vehicle queue conditions within the park, resulting in significant discrepancies between predicted and actual arrival times, which fails to meet the business needs of precise logistics scheduling.

[0003] Currently, the prediction of arrival time for existing logistics vehicles mainly relies on a simple calculation of static road network distance divided by average vehicle speed. The prediction model has a single input feature and outputs a one-time static value. It cannot perceive changes in road congestion and queuing in the park in real time. When the transportation environment changes suddenly, the prediction results cannot be automatically corrected, and the prediction deviation continues to accumulate, making it difficult to meet the needs of refined logistics scheduling. Summary of the Invention

[0004] The main objective of this application is to provide a method, apparatus, equipment, and storage medium for predicting the arrival time of logistics vehicles in order to solve the above-mentioned problems.

[0005] To achieve the above objectives, this application provides the following technical solution: A method for predicting the arrival time of logistics vehicles, the method comprising: Step S1: Obtain real-time road condition data, vehicle operation data, park operation data, and task data of the logistics vehicle, and perform data preprocessing to form a continuously updated time-series dataset with vehicle identifier and timestamp as the key. Step S2: After extracting multidimensional features from the time series dataset, dimensionality reduction is performed by screening with Pearson correlation coefficient and principal component analysis to obtain standardized feature vectors. Step S3: Input the standardized feature vector into the fusion prediction model of the attention mechanism and the long short-term memory network. After weighting the standardized feature vector through the Softmax function of the attention layer, it is processed by the Tanh activation function of the LSTM layer to obtain the initial predicted value of the vehicle arrival time. Step S4: Detect whether the change of the latest update of the time series dataset compared with the previous update meets the trigger update condition. If it does, repeat steps S2 to S3 with the latest updated time series dataset as input to obtain the updated predicted value of vehicle arrival time. Step S5: Calculate the prediction deviation between the updated predicted value and the initial predicted value. When the prediction deviation exceeds a preset deviation threshold, generate a model update instruction. Step S6: In response to the model update instruction, calculate the average error between the predicted arrival time and the actual arrival time of the historical completed trips using the mean absolute error, and update the feature weights and parameter configuration of the fusion prediction model according to the average error to obtain the optimized prediction model parameters.

[0006] Beneficial effects of steps S1 to S6: This method forms a complete closed loop of logistics vehicle arrival time prediction and dynamic optimization through steps S1 to S6, effectively solving the technical defects of existing methods such as single input features, inability to dynamically update prediction results according to real-time data changes, and continuous accumulation of prediction bias.

[0007] The process involves several steps: Step S1 acquires multi-source real-time data on logistics vehicles and preprocesses it, constructing a continuously updated time-series dataset using vehicle identifiers and timestamps to ensure the timeliness and completeness of the prediction input; Step S2 uses Pearson correlation coefficient filtering and principal component analysis to extract highly correlated feature vectors and filter out redundant noise; Step S3 uses an attention mechanism to fuse the LSTM time-series model output to achieve joint modeling of multi-dimensional features and time-series dependencies; Step S4 triggers a sliding window mechanism by detecting weighted average change and re-executes the prediction to achieve dynamic updates of the prediction results; Step S5 calculates the deviation between the updated prediction value and the initial prediction value and generates a model update instruction when the deviation exceeds the limit; Step S6 responds to the update instruction by adjusting the feature weights and LSTM model parameters through backpropagation of MAE loss to ensure that the model continuously optimizes as the data distribution evolves.

[0008] As a further improvement to this application, step S1 involves acquiring real-time road condition data, vehicle operation data, park operation data, and task data of the logistics vehicles, and performing data preprocessing to form a continuously updated time-series dataset with vehicle identifiers and timestamps as keys, including: Step S1.1: Obtain real-time road condition data, vehicle operation data, park operation data, and task data of the logistics vehicles, and integrate them into a multi-source raw dataset; Step S1.2: Data points exceeding three times the standard deviation of the mean in the original multi-source dataset are removed using the 3σ criterion to obtain the anomaly-filtered dataset. Step S1.3: Fill in the missing values ​​of the anomaly-filtered dataset using the forward imputation method to obtain the complete dataset; Step S1.4: Map the data of each dimension in the complete dataset to the interval of 0 to 1 through Min-Max standardization to obtain a standardized dataset; Step S1.5: Perform linear interpolation and time synchronization on the standardized dataset with a sampling interval of 10 seconds to obtain a time-aligned dataset; Step S1.6: Group the time-aligned dataset by vehicle identifier and arrange it in ascending order of timestamp to form a continuously updated time-series dataset with vehicle identifier and timestamp as the key.

[0009] Beneficial effects of steps S1.1 to S1.6: This series of steps, through six progressive data processing steps, constructs a high-quality, time-aligned, and continuously updated time-series dataset, laying a reliable data foundation for subsequent feature extraction and predictive modeling, and effectively ensuring the accuracy, completeness, and timeliness of the input data for the predictive model.

[0010] The process involves the following steps: Step S1.1: Acquiring real-time road conditions, vehicle operation, park operations, and task data of logistics vehicles and integrating them into a multi-source raw dataset to achieve multi-dimensional information coverage; Step S1.2: Removing outliers using the 3σ criterion to ensure the statistical reliability of the data; Step S1.3: Filling in missing values ​​using forward imputation to ensure data integrity; Step S1.4: Eliminating differences in the dimensions of each dimension through Min-Max standardization to achieve data comparability; Step S1.5: Performing linear interpolation and time synchronization with a sampling interval of 10 seconds to solve the problem of time misalignment in multi-source data; and Step S1.6: Arranging the data in ascending order of timestamp using vehicle identifier and timestamp as keys to form a continuously updated time-series dataset, supporting the time-series modeling needs of subsequent feature extraction and prediction models.

[0011] As a further improvement to this application, in step S2, after extracting multidimensional features from the time-series dataset, dimensionality reduction is performed using Pearson correlation coefficient screening and principal component analysis to obtain a standardized feature vector, including: Step S2.1: Extract the congestion index sequence, vehicle speed change rate, and sudden road condition markers from the time-series dataset as time-series road condition features; Step S2.2: Extract vehicle type weight, task priority coefficient, and remaining distance as vehicle and task features; Step S2.3: Extract the predicted queuing time and the deviation of the channel restriction speed as dynamic features of the park; Step S2.4: Extract the energy replenishment demand markers and the urgency of the unloading window as constraint features; Step S2.5: Align and stitch the time-series road condition features, vehicle and task features, park dynamic features, and constraint features according to vehicle identifiers and timestamps to form a multi-dimensional feature matrix; Step S2.6: Calculate the Pearson correlation coefficient between each pair of features in the multidimensional feature matrix to construct a correlation coefficient matrix; Step S2.7: Identify feature pairs whose absolute values ​​of correlation coefficients are greater than a preset absolute value threshold from the correlation coefficient matrix, and remove features with smaller variances from each pair to obtain the filtered feature matrix; Step S2.8: Calculate the eigenvalues ​​and variance contribution rates of each principal component in the filtered feature matrix, and mark the first few principal components whose cumulative variance contribution rates exceed the preset contribution rate threshold according to the eigenvalues ​​from largest to smallest. Project the filtered feature matrix onto the subspace formed by the first few principal components to obtain the standardized feature vector.

[0012] Beneficial effects of steps S2.1 to S2.8: This series of steps constructs a four-dimensional feature system covering road conditions, vehicle tasks, park dynamics, and constraints. After correlation screening and principal component analysis for dimensionality reduction, standardized feature vectors are obtained, which effectively improves feature quality and model input efficiency, providing highly relevant and low-redundancy feature inputs for subsequent prediction modeling.

[0013] The process involves several steps: Step S2.1 extracts the congestion index sequence, vehicle speed change rate, and sudden road condition markers to quantitatively represent road traffic conditions; Step S2.2 extracts vehicle type weights, task priority coefficients, and remaining distances to incorporate task urgency into prediction; Step S2.3 extracts predicted queuing time and lane restriction speed deviations to quantitatively perceive the internal scheduling status of the park; Step S2.4 extracts energy replenishment demand markers and unloading window urgency to incorporate hard constraints into the feature system; Step S2.5 aligns and splices the temporal road condition features, vehicle and task features, park dynamic features, and constraint features in a temporal sequence to form a unified multidimensional feature matrix; Step S2.6 quantifies the linear correlation between features using the Pearson correlation coefficient matrix; Step S2.7 eliminates redundant features through variance analysis to effectively reduce feature dimensionality and eliminate multicollinearity interference; and Step S2.8 selects principal components based on cumulative variance contribution rate and projects them to a low-dimensional subspace to achieve dimensionality reduction while retaining key information, resulting in a standardized feature vector.

[0014] As a further improvement to this application, step S3 involves inputting the standardized feature vector into the fusion prediction model of the attention mechanism and the long short-term memory network. After weighting the standardized feature vector through the Softmax function of the attention layer, it is processed by the Tanh activation function of the LSTM layer to obtain the initial predicted value of the vehicle arrival time, including: Step S3.1: The standardized feature vector is mapped to a query vector, a key vector, and a value vector respectively through a preset linear mapping matrix, and the matrix product of the query vector and the key vector is calculated. After normalization by the Softmax function, the attention weight vector is obtained. Step S3.2: Perform an element-wise weighted summation operation on the attention weight vector and the value vector to obtain the attention weighted feature vector; Step S3.3: Input the attention-weighted feature vector into the long short-term memory network layer, and update the cell state through the collaborative calculation of the forget gate, input gate, and output gate to obtain the updated cell state; Step S3.4: Process the updated cell state using the Tanh activation function to obtain the activated cell state vector; Step S3.5: Multiply the activated cell state vector by the output gate activation value of the long short-term memory network layer to obtain the final hidden state vector of the long short-term memory network. Step S3.6: Input the final hidden state vector into the fully connected layer, perform a linear transformation, and then map it using the Tanh function to obtain the initial predicted value of the vehicle arrival time.

[0015] Beneficial effects of steps S3.1 to S3.6: This series of steps uses a fusion model of attention mechanism and long short-term memory network to gradually transform standardized feature vectors into initial predicted values ​​of vehicle arrival time. The attention layer achieves adaptive focusing on highly relevant features, and the LSTM layer achieves effective modeling of temporal dependencies. Together, they output accurate initial predicted values.

[0016] In step S3.1, the standardized feature vector is mapped to a query vector, a key vector, and a value vector using a preset linear mapping matrix. The matrix product of the query vector and the key vector is calculated and then normalized using Softmax to obtain the attention weight vector, achieving adaptive focusing on key features. Step S3.2, the attention weight vector and the value vector are summed element-wise to obtain the attention-weighted feature vector, achieving feature fusion representation. Step S3.3, the attention-weighted feature vector is input into the LSTM layer, and the cell state is updated through the collaborative calculation of the forget gate, input gate, and output gate, effectively preserving long-term temporal dependency information. Step S3.4, the updated cell state is processed by the Tanh activation function to complete the nonlinear feature transformation. Step S3.5, the activated cell state vector is multiplied by the output gate activation value to obtain the final hidden state vector of the fusion model. Step S3.6, the final hidden state vector is input into the fully connected layer for linear transformation and then mapped by the Tanh function to obtain the initial predicted value of the vehicle arrival time.

[0017] As a further improvement of this application, step S4 involves detecting whether the change in the latest update of the time series dataset compared to the previous update satisfies the triggering update condition. If it does, steps S2 to S3 are repeated with the latest updated time series dataset as input to obtain the updated predicted value of the vehicle arrival time, including: Step S4.1: Extract all the latest data points corresponding to the latest update from the time series dataset; Step S4.2: Calculate the change between each latest data point and the corresponding data point updated in the previous step to obtain a set of data change amounts; Step S4.3: Multiply each data change in the set of data changes by the corresponding feature weight in the standardized feature vector and sum them to obtain the weighted average change. Step S4.4: Determine whether the weighted average change meets the trigger update condition. If the weighted average change exceeds the preset trigger threshold, it is determined that the trigger update condition is met, and step S4.5 is executed. Step S4.5: Extract window data containing all the latest data points from the time series dataset with a preset window length as the step size to form a sliding window data; Step S4.6: Based on the sliding window data, re-extract the temporal features, and concatenate the re-extracted temporal features with the historical temporal features in chronological order to form a reconstructed temporal feature vector; Step S4.7: Input the reconstructed temporal feature vector into the fusion prediction model to obtain the updated predicted value of the vehicle arrival time.

[0018] Beneficial effects of steps S4.1 to S4.7: This series of steps, through a weighted change detection trigger mechanism, sliding window data extraction, and temporal feature reconstruction, enables dynamic updates of the prediction results, allowing the predicted vehicle arrival time to respond in real time to changes in the transportation environment and remain synchronized with the latest environmental status.

[0019] The process involves the following steps: Step S4.1 extracts the latest updated data point from the time-series dataset to provide a foundation for change detection; Step S4.2 calculates the change amount of each data point to quantify the actual magnitude of environmental changes; Step S4.3 weights and sums the change amount with the corresponding feature weights to obtain a comprehensive change evaluation index, avoiding interference from single-dimensional abrupt changes; Step S4.4 determines whether the weighted average change amount exceeds a preset trigger threshold, and triggers the prediction update process when it does, achieving accurate response to changes in the transportation environment; Step S4.5 extracts window data containing the latest data point with a preset window length to form a stable context input; Step S4.6 concatenates the newly extracted time-series features with historical time-series features in chronological order to form a reconstructed time-series feature vector, ensuring the continuity of feature expression; and Step S4.7 inputs the reconstructed time-series feature vector into the fusion prediction model to obtain the updated vehicle arrival time prediction value.

[0020] As a further improvement to this application, step S5 involves calculating the prediction deviation between the updated predicted value and the initial predicted value. When the prediction deviation exceeds a preset deviation threshold, a model update instruction is generated, including: Step S5.1: Obtain the initial predicted value of the vehicle arrival time and the updated predicted value of the vehicle arrival time; Step S5.2: Calculate the absolute difference between the updated predicted value and the initial predicted value to obtain the prediction deviation; Step S5.3: Determine whether the prediction deviation exceeds a preset deviation threshold. If it does, proceed to the update instruction generation process. Step S5.4: Construct an instruction parameter set containing update type and priority based on the prediction deviation; Step S5.5: Generate and output model update instructions based on the instruction parameter set.

[0021] Beneficial effects of steps S5.1 to S5.5: This series of steps forms a complete deviation detection and model update triggering mechanism through quantitative calculation and condition judgment of prediction deviation. When the prediction deviation exceeds the threshold, the model update instruction is accurately generated, which effectively ensures the self-correction and continuous optimization capabilities of the fusion prediction model.

[0022] Specifically, step S5.1 simultaneously acquires the initial and updated predicted values, providing complete input for bias calculation; step S5.2 calculates the absolute difference between the updated and initial predicted values ​​to quantify the degree of bias in the prediction results; step S5.3 determines whether the prediction bias exceeds a preset bias threshold, triggering an update only when the bias is significant, thus avoiding model oscillations caused by frequent updates; step S5.4 constructs an instruction parameter set containing update type and priority based on the prediction bias, providing parameterized configuration for the update instructions; step S5.5 generates structured model update instructions based on the instruction parameter set and outputs them to step S6 for execution, achieving a seamless connection between bias detection and model optimization.

[0023] As a further improvement to this application, step S6, in response to the model update instruction, calculates the average error between the predicted arrival time and the actual arrival time of historically completed trips using the mean absolute error, and updates the feature weights and parameter configuration of the fusion prediction model based on the average error to obtain optimized prediction model parameters, including: Step S6.1: In response to the model update instruction, filter historical completed trip records from the time series dataset to form a historical trip dataset; Step S6.2: Extract the predicted arrival time and actual arrival time of each completed trip from the historical trip dataset to form a time pair sequence; Step S6.3: Calculate the absolute difference between the predicted arrival time and the actual arrival time for each pair of time pairs in the time pair sequence, and the mean of all absolute differences, to obtain the average error. Step S6.4: Calculate the gradient of the average error with respect to all feature weights of the fusion prediction model, and adjust the feature weights along the negative gradient direction to obtain the updated feature weights; Step S6.5: Use the average error and the updated feature weights as model hyperparameters to adjust the number of hidden units and the learning rate of the LSTM layer to obtain the updated parameter configuration; Step S6.6: Combine the updated feature weights with the updated parameter configuration to obtain the optimized prediction model parameters.

[0024] Beneficial effects of steps S6.1 to S6.6: This series of steps responds to model update instructions, quantifies prediction errors from historical travel data, and adjusts feature weights and LSTM model parameters through gradient backpropagation mechanism to form a complete model adaptive optimization closed loop, ensuring that the fusion prediction model can continuously optimize as the data distribution evolves.

[0025] The process involves several steps: Step S6.1: Filtering historical completed trip records from the time-series dataset to provide real-world label data for error analysis; Step S6.2: Extracting time pairs between predicted and actual arrival times to form the input sequence for error calculation; Step S6.3: Quantifying the overall prediction bias of the model using the mean absolute error (MAE) to provide an optimization target for parameter updates; Step S6.4: Calculating the gradient of the MAE with respect to the feature weights and updating the weights along the negative gradient direction to achieve adaptive adjustment of feature importance; Step S6.5: Using the MAE and updated feature weights as the basis for hyperparameter adjustment, dynamically adjusting the number of LSTM hidden units and the learning rate to enhance the model's ability to capture complex patterns; and Step S6.6: Combining the updated feature weights with the parameter configuration and outputting the optimized prediction model parameters, thus completing the entire adaptive optimization process.

[0026] To achieve the above objectives, this application also provides the following technical solutions: An arrival time prediction device for logistics vehicles, the arrival time prediction device being applied to the arrival time prediction method described above, the arrival time prediction device comprising: The time-series dataset acquisition and update module is used to acquire real-time road condition data, vehicle operation data, park operation data, and task data of the logistics vehicles, and perform data preprocessing to form a continuously updated time-series dataset with vehicle identifier and timestamp as the key. The time series dataset analysis and dimensionality reduction module is used to extract multidimensional features from the time series dataset and then perform dimensionality reduction processing through Pearson correlation coefficient screening and principal component analysis to obtain standardized feature vectors. The initial prediction module for vehicle arrival time is used to input the standardized feature vector into the fusion prediction model of the attention mechanism and the long short-term memory network. After the standardized feature vector is weighted by the Softmax function of the attention layer, it is processed by the Tanh activation function of the LSTM layer to obtain the initial prediction value of the vehicle arrival time. The vehicle arrival time prediction update module is used to detect whether the change of the latest update of the time series dataset compared with the previous update meets the trigger update condition. If it meets the condition, the time series dataset analysis and dimensionality reduction module is repeatedly executed to the vehicle arrival time initial prediction module with the latest updated time series dataset as input to obtain the updated predicted value of the vehicle arrival time. The prediction deviation calculation and model update module is used to calculate the prediction deviation between the updated prediction value and the initial prediction value, and generate a model update instruction when the prediction deviation exceeds a preset deviation threshold. The prediction model parameter iterative optimization module is used to respond to the model update command by calculating the average error between the predicted arrival time and the actual arrival time of historical completed trips through the mean absolute error, and updating the feature weights and parameter configuration of the fusion prediction model according to the average error to obtain the optimized prediction model parameters.

[0027] To achieve the above objectives, this application also provides the following technical solutions: An electronic device includes a processor and a memory coupled to the processor, the memory storing program instructions executable by the processor; when the processor executes the program instructions stored in the memory, it implements the logistics vehicle arrival time prediction method described above.

[0028] To achieve the above objectives, this application also provides the following technical solutions: A computer-readable storage medium storing program instructions that, when executed by a processor, enable the implementation of the aforementioned method for predicting the arrival time of logistics vehicles. Attached Figure Description

[0029] Figure 1 This is a flowchart illustrating the steps of one embodiment of the method for predicting the arrival time of a logistics vehicle according to this application. Figure 2 This is a schematic diagram of the functional modules of an embodiment of a logistics vehicle arrival time prediction device according to this application; Figure 3 This is a schematic diagram of the structure of an embodiment of the electronic device of this application; Figure 4 This is a schematic diagram of the structure of one embodiment of the storage medium of this application. Detailed Implementation

[0030] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0031] The terms "first," "second," and "third" in this application are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first," "second," or "third" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified. All directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of this application are only used to explain the relative positional relationships and movements between components in a specific orientation (e.g., as shown in the figures). If the specific orientation changes, the directional indications also change accordingly. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0032] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0033] It should be noted that, due to the limited types and number of symbols or letters that can represent specific meanings, for embodiments with many formulas or codes, there may be situations where symbols or letters cannot meet the usage requirements. Therefore, the interpretation of formula symbols in the steps or sub-steps of the embodiments is only valid for the current step or sub-step.

[0034] If the same symbol has different interpretations in different steps or sub-steps, the interpretation in the current step or sub-step shall prevail; if the same symbol appears in different steps or sub-steps, but no interpretation is given in subsequent steps or sub-steps after its first appearance, the interpretation in the first step or sub-step shall be used.

[0035] like Figure 1 As shown, this embodiment provides an example of a method for predicting the arrival time of logistics vehicles. In this embodiment, the method for predicting the arrival time includes the following steps: Step S1: Obtain real-time road condition data, vehicle operation data, park operation data, and task data of the logistics vehicles, and perform data preprocessing to form a continuously updated time-series dataset with vehicle identifiers and timestamps as keys.

[0036] Furthermore, step S1 specifically includes the following steps: Step S1.1: Obtain real-time road condition data, vehicle operation data, park operation data, and task data of the logistics vehicles, and integrate them into a multi-source raw dataset.

[0037] Preferably, the core of this step is to complete the real-time acquisition and format unification of multi-source heterogeneous data. Through parallel calls to multiple interfaces and data formatting, data from different sources and with different structures are integrated into a multi-source raw dataset with a unified format.

[0038] Specifically, for real-time traffic data, this step collects the real-time congestion index between the vehicle's current location and its destination via the map web service API. The congestion index ranges from 0 to 10, with higher values ​​indicating more severe congestion. For vehicle operation data, this step obtains the vehicle's current speed in real time via the CAN bus interface, with speed data in kilometers per hour and a sampling frequency of no less than once per second. For park operation data, this step receives the current number of vehicles in the queue and the estimated waiting time from the logistics park management system via a message queue, with the waiting time in seconds. For task data, this step queries the task management system for the priority coefficient and remaining transportation distance of the current task based on the vehicle identifier.

[0039] Preferably, the real-time traffic data acquisition interval is 30 seconds; the CAN bus data sampling frequency is 1Hz; the park operation data push frequency is 10 seconds; and the task data query frequency is 60 seconds.

[0040] For example, taking a certain transport vehicle as an example, the current congestion index obtained through the Gaode Map API is 4.5 (moderate congestion), the current vehicle speed collected by the CAN bus is 45km / h, the park system pushes that the current number of vehicles in the queue is 3, the estimated waiting time is 120 seconds, and the task system returns a task priority coefficient of 0.8 and a remaining transport distance of 25 kilometers.

[0041] Step S1.2: Data points exceeding three times the standard deviation of the mean in the original multi-source dataset are removed using the 3σ criterion to obtain the anomaly-filtered dataset.

[0042] Preferably, the core of this step is to identify and remove outliers. The statistical rationality of data points is judged dimension by dimension using the 3σ criterion, and data points that exceed the reasonable range are marked as outliers and removed.

[0043] Specifically, for the 3σ criterion, this step calculates the mean μ and standard deviation σ of each data dimension within the historical window. For any data point x at the current time, if |x-μ| is greater than 3σ, the data point is determined to be an outlier and is removed. Specifically, for the sliding window, this step uses a sliding window with a length of N sampling points to calculate the mean and standard deviation, and the number of data points in the window is no less than 10. Specifically, for the outlier handling logic, data points determined to be outliers are not directly deleted, but missing value filling is performed in the subsequent step S1.3.

[0044] Preferably, the σ multiple threshold is 3; the sliding window length N is 100 sampling points; and the minimum effective window data points are 10.

[0045] For example, in the vehicle speed data dimension, if the mean within the sliding window is 60 km / h and the standard deviation is 5 km / h, then the upper limit of the threshold is 75 km / h and the lower limit is 45 km / h. Vehicle speed data points that exceed this range (such as 85 km / h or 35 km / h) will be judged as abnormal.

[0046] Step S1.3: Fill in the missing values ​​of the anomaly-filtered dataset using the forward imputation method to obtain the complete dataset.

[0047] Preferably, the core of this step is to handle missing values ​​by using the most recent valid data to fill in the missing values ​​at the current time using the forward imputation method, thus ensuring the integrity and continuity of the data sequence.

[0048] Specifically, for the forward filling rule, if a data point for a vehicle, a timestamp, or a dimension has been marked as missing or abnormal in step S1.2, the value of the most recent valid data point for that vehicle or dimension before that timestamp is used as the filling value at the current moment. Specifically, for continuous missing data handling, when a dimension is missing at multiple consecutive timestamps, it is continuously filled with the most recent valid value until a new valid data point is encountered.

[0049] For example, if vehicle speed data is removed due to an anomaly at timestamp t2, the vehicle speed value at timestamp t2 is filled with the valid vehicle speed at timestamp t1; if vehicle speeds are missing for three consecutive times from t2 to t4, then the vehicle speeds at timestamps t2, t3, and t4 are filled with the valid vehicle speed value at timestamp t1.

[0050] Step S1.4: Map the data of each dimension in the complete dataset to the interval between 0 and 1 through Min-Max standardization to obtain a standardized dataset.

[0051] Preferably, the core of this step is to complete the normalization process of multi-dimensional data, and eliminate the differences in the dimensions of each dimension through Min-Max standardization, so that data from different sources and of different magnitudes are comparable.

[0052] For the standardized formula, the standardized value The calculation method is as follows: Where x is the original data, and These are the minimum and maximum values ​​of this dimension within the historical statistical period, respectively; where, for dynamic range updates, and The statistics are updated once per statistical period, which is 24 hours by default.

[0053] Preferably, the standardized interval is [0,1]; and The statistical period is 24 hours; if there is no historical data during the initial standardization, the typical value range of each dimension is used for initialization.

[0054] For example, the original range of the congestion index is 0 to 10. After standardization, 0 corresponds to smooth traffic, 5 corresponds to moderate congestion, and 10 corresponds to severe congestion. The original range of vehicle speed is 0 to 120 km / h. After standardization, it corresponds to the range of 0 to 1.

[0055] Step S1.5: Perform linear interpolation and time synchronization on the standardized dataset with a sampling interval of 10 seconds to obtain a time-aligned dataset.

[0056] Preferably, the core of this step is to complete the time alignment of multi-source data. By using linear interpolation, the sampling points of each data source are unified to a 10-second reference time axis, thus solving the time sequence misalignment problem caused by inconsistent timestamps of multi-source data.

[0057] For the linear interpolation method, for any non-integer time t on the target time axis, if there exists If both ends have valid data, then the interpolation result is: For time base alignment, this step uses the system startup time. A unified timeline is established with zero point as the reference point, and the timestamps of each data source are mapped to this timeline. The sampling interval is fixed at 10 seconds.

[0058] Preferably, the sampling interval is 10 seconds; the interpolation method is linear interpolation; and the time reference is the system startup time.

[0059] For example, the CAN bus vehicle speed data sampling frequency is 1Hz (1 second interval), the park queuing data push frequency is 0.1Hz (10 second interval), and the task data query frequency is 0.017Hz (60 second interval). After step S1.5, all data are unified to a 10-second time axis and aligned to the nearest 10-second integer multiple of time.

[0060] Step S1.6: Group the time-aligned dataset by vehicle identifier and arrange it in ascending order of timestamp to form a continuously updated time-series dataset with vehicle identifier and timestamp as the key.

[0061] Preferably, the core of this step is to complete the structured organization of time-series data, grouping and sorting the data with vehicle identifiers and timestamps as the primary keys, and constructing a time-series dataset to support subsequent feature extraction and time-series modeling.

[0062] For the primary key structure, the dataset is indexed using a combination of vehicle identifier and timestamp as the primary key. Each record uniquely corresponds to the complete data of a vehicle at a certain moment, and the data of multiple vehicles are stored in parallel. For the continuous update mechanism, when new data arrives, it is appended to the time series of the corresponding vehicle, maintaining the ascending order of timestamps, without the need for a full reordering.

[0063] Beneficial effects of steps S1.1 to S1.6: This series of steps, through six progressive data processing steps, constructs a high-quality, time-aligned, and continuously updated time-series dataset, laying a reliable data foundation for subsequent feature extraction and predictive modeling, and effectively ensuring the accuracy, completeness, and timeliness of the input data for the predictive model.

[0064] The process includes the following steps: Step S1.1: Acquire multidimensional real-time data of logistics vehicles and integrate it into a multi-source raw dataset to achieve unified coverage of multi-source information; Step S1.2: Remove outliers using the 3σ criterion to ensure the statistical reliability of data in each dimension; Step S1.3: Complete missing values ​​using forward imputation to ensure the integrity and continuity of the data sequence; Step S1.4: Eliminate dimensional differences using Min-Max standardization to achieve comparability of data in different dimensions; Step S1.5: Perform linear interpolation and time synchronization at 10-second intervals to solve the problem of inconsistent time axes in multi-source data; Step S1.6: Sort the data in ascending order by timestamp using vehicle identifier and timestamp as keys to form a continuously updated time-series dataset, supporting the time-series modeling needs of subsequent feature extraction and prediction models.

[0065] Step S2: After extracting multidimensional features from the time series dataset, dimensionality reduction is performed by screening with Pearson correlation coefficient and principal component analysis to obtain standardized feature vectors.

[0066] Furthermore, step S2 specifically includes the following steps: Step S2.1: Extract the congestion index sequence, vehicle speed change rate, and sudden road condition markers from the time-series dataset as time-series road condition features.

[0067] Preferably, the core of this step is to quantitatively extract road traffic status features and generate feature dimensions that reflect real-time road conditions through statistical analysis of the original time-series data.

[0068] Specifically, for the congestion index sequence, this step directly extracts the N most recent time points of the congestion index dimension from the time series dataset to form a one-dimensional time series, where N is the preset feature window length; for the vehicle speed change rate, this step calculates the first-order difference of vehicle speed between adjacent time points as the change rate sequence, with positive values ​​indicating acceleration and negative values ​​indicating deceleration; for the sudden road condition marking, this step marks a congestion index as 1 if the change exceeds a preset threshold within a single time step, and marks it as 0 otherwise.

[0069] Preferably, the feature window length N is 12 (i.e., the feature window of the most recent 120 seconds); the threshold for the change in the congestion index is 2.

[0070] For example, taking a certain vehicle as an example, the congestion index sequence within the feature window is [3.2,3.5,4.1,5.8,7.2], and the corresponding vehicle speed change rate sequence is [+2,+3,-5,-8] km / h. The sudden road condition marker is marked as 1 when the speed jumps from 4.1 to 5.8 (change amount 1.7) in the 4th time step.

[0071] Step S2.2: Extract vehicle type weight, task priority coefficient, and remaining distance as vehicle and task features.

[0072] Preferably, the core of this step is to quantitatively extract the urgency features of the transportation task, transforming vehicle and task attributes into calculable numerical features.

[0073] Specifically, for vehicle type weights, this step sets weight values ​​based on the average delay rate of historical trips for each vehicle type: 0.6 for heavy trucks, 0.8 for medium trucks, and 1.0 for light trucks. For task priority coefficients, this step directly extracts the priority coefficients from the task data, ranging from 0 to 1, with higher values ​​indicating more urgent tasks. For remaining distance, this step extracts the remaining transportation distance from the current vehicle to the destination from the task data, in kilometers.

[0074] Preferably, the vehicle type weights are (heavy 0.6, medium 0.8, light 1.0); the priority coefficient ranges from 0 to 1; and the remaining distance is in kilometers.

[0075] For example, if a light truck is performing an emergency task with a priority coefficient of 0.9 and the remaining transportation distance is 35 kilometers, then the vehicle type weight is 1.0, the task priority coefficient is 0.9, and the remaining distance is 35 kilometers.

[0076] Step S2.3: Extract the predicted queuing time and the deviation of the channel restriction speed as dynamic features of the park.

[0077] Preferably, the core of this step is to extract features of the internal scheduling status of the logistics park and transform the park's operational data into quantitative features to be incorporated into the prediction model.

[0078] Specifically, for the predicted queuing time, this step extracts the expected waiting time dimension from the park's operational data as a quantitative representation of the park's congestion level; for the deviation of the restricted speed in the passage, this step calculates the difference between the current restricted speed in the park's passage and the vehicle's current speed, with a positive value indicating that the vehicle is below the speed limit and a negative value indicating that the vehicle is above the speed limit.

[0079] Preferably, the waiting time is in seconds; the speed deviation is in km / h.

[0080] For example, if the park system pushes a predicted queuing time of 180 seconds, the park lane speed limit is 20 km / h, and the vehicle's current speed is 15 km / h, then the lane speed limit deviation is -5 km / h.

[0081] Step S2.4: Extract the energy replenishment demand marker and the urgency of the unloading window as constraint features.

[0082] Preferably, the core of this step is to extract features of transportation constraints, incorporating vehicle refueling needs and unloading time limits into the feature system.

[0083] Among them, for the energy replenishment demand marker, when the vehicle's current remaining fuel or battery level is lower than the preset threshold, it is marked as 1, indicating that mid-journey energy replenishment is required; otherwise, it is marked as 0. Among them, for the urgency of the unloading window, this step calculates the ratio of the remaining transportation distance to the remaining time of the unloading window in the park. The smaller the ratio, the more urgent the unloading window is. When the ratio is lower than the preset urgency threshold, it is judged as high urgency.

[0084] Preferably, the fuel level threshold is set according to the vehicle type (less than 50 liters for heavy trucks, less than 30 liters for medium trucks, and less than 15 liters for light trucks); the urgency threshold is 0.8 km / min.

[0085] For example, a vehicle has 25 liters of fuel remaining (medium-sized truck, threshold 30 liters), and its refueling demand is marked as 1; the remaining distance is 35 kilometers, the remaining time for unloading is 30 minutes, and the urgency ratio is 35÷30≈1.17km / minute, which is higher than the urgency threshold of 0.8, so the urgency of the unloading window is marked as low.

[0086] Step S2.5: Align and stitch the time-series road condition features, vehicle and task features, park dynamic features, and constraint features according to time sequence to form a multi-dimensional feature matrix.

[0087] Preferably, the core of this step is to complete the timeline alignment and splicing of four types of features, and to integrate features from different sources into a unified multi-dimensional feature matrix with vehicle identification and timestamp as the primary key.

[0088] For time alignment, all features are snapshots taken at the same time with the timestamp as the key, and features with missing times are filled with the historical mean of the feature; for matrix dimensions, the columns of the multidimensional feature matrix correspond to each feature dimension, and the rows correspond to each time point, with each row of data representing the complete feature vector of a vehicle at a certain time.

[0089] Preferably, missing features are filled with historical averages; the matrix storage format is a sparse matrix to save space.

[0090] Step S2.6: Calculate the Pearson correlation coefficients between each pair of features in the multidimensional feature matrix to construct a correlation coefficient matrix.

[0091] Preferably, the core of this step is to perform a quantitative analysis of the linear correlation between features, and to identify the degree of linear dependence between each feature dimension through the Pearson correlation coefficient matrix.

[0092] Specifically, for the Pearson correlation coefficient formula, the correlation coefficient r between the two feature dimensions X and Y is calculated as follows: ,in and , where are the means of the two features within the feature window, and n is the number of samples; where, for the construction of the correlation coefficient matrix, the correlation coefficient matrix is ​​a symmetric matrix with diagonal elements of 1.0.

[0093] Preferably, the sample size for calculating the correlation coefficient is the feature window length N=12; the update frequency of the correlation coefficient matrix is ​​synchronized with the feature window.

[0094] Step S2.7: Identify feature pairs whose absolute correlation coefficients are greater than a preset absolute value threshold from the correlation coefficient matrix, and remove features with smaller variances from each pair to obtain the filtered feature matrix.

[0095] Preferably, the core of this step is to identify and remove redundant features, eliminate multicollinearity based on correlation analysis, and reduce the redundancy of feature dimensions.

[0096] Specifically, for the determination of highly correlated feature pairs, when the absolute value of the Pearson correlation coefficient of two features is greater than the absolute value threshold, they are determined to be a pair of highly correlated features. Specifically, for the elimination rule, in each pair of highly correlated features, the variance of the two features within the feature window is calculated, and the feature with the larger variance is retained and the feature with the smaller variance is eliminated.

[0097] Preferably, the absolute value threshold is set to 0.7.

[0098] For example, if the correlation coefficient between the congestion index sequence and the predicted queuing time is 0.85 (greater than 0.7), the variance of the congestion index is 2.3, and the variance of the queuing time is 1.8, then the queuing time feature is removed, and the congestion index feature is retained.

[0099] Step S2.8: Calculate the eigenvalues ​​and variance contribution rates of each principal component in the filtered feature matrix, and mark the first few principal components whose cumulative variance contribution rates exceed the preset contribution rate threshold according to the eigenvalues ​​from largest to smallest. Project the filtered feature matrix onto the subspace formed by the first few principal components to obtain the standardized feature vector.

[0100] Preferably, the core of this step is to perform principal component analysis dimensionality reduction. After screening, the covariance matrix is ​​decomposed, and the principal components with a cumulative variance contribution rate exceeding the threshold are selected for projection, thereby achieving standardized dimensionality reduction of the feature vectors.

[0101] Specifically, for the covariance matrix, this step calculates the covariance matrix of the filtered feature matrices and performs eigenvalue decomposition to obtain the eigenvalues. For principal component selection, the variance contribution rate is accumulated from largest to smallest eigenvalue. The k principal components whose cumulative variance contribution rate first exceeds the contribution rate threshold are selected. For the projection formula, the selected feature matrix X is projected onto the projection matrix formed by the k principal components. To obtain the standardized feature vector .

[0102] Preferably, the contribution rate threshold is set to 0.85; the covariance matrix is ​​estimated using sample covariance; and the projected feature dimension is typically 3 to 8 dimensions.

[0103] Beneficial effects of steps S2.1 to S2.8: This series of steps constructs a four-dimensional feature system covering road conditions, vehicle tasks, park dynamics, and constraints. After correlation screening and principal component analysis for dimensionality reduction, standardized feature vectors are obtained, which effectively improves feature quality and model input efficiency, providing highly relevant and low-redundancy feature inputs for subsequent prediction modeling.

[0104] Specifically, steps S2.1 to S2.4 extract time-series road condition features, vehicle and task features, park dynamic features, and constraint features, respectively, to achieve quantitative coverage of multi-dimensional information across the entire transportation chain; step S2.5 aligns and concatenates the four types of features according to time sequence to form a unified multi-dimensional feature matrix; step S2.6 quantifies the linear correlation between features using the Pearson correlation coefficient matrix; step S2.7 eliminates redundant features through variance analysis, effectively reducing feature dimensionality and eliminating multicollinearity interference; step S2.8 selects principal components based on cumulative variance contribution rate and projects them to a low-dimensional subspace, completing dimensionality reduction while retaining key information, and obtaining standardized feature vectors.

[0105] Step S3: Input the standardized feature vector into the fusion prediction model of the attention mechanism and the long short-term memory network. After weighting the standardized feature vector through the Softmax function of the attention layer, it is processed by the Tanh activation function of the LSTM layer to obtain the initial predicted value of the vehicle arrival time.

[0106] Furthermore, step S3 specifically includes the following steps: Step S3.1: The standardized feature vector is mapped to a query vector, a key vector, and a value vector respectively through a preset linear mapping matrix, and the matrix product of the query vector and the key vector is calculated. After normalization by the Softmax function, the attention weight vector is obtained.

[0107] Preferably, the core of this step is to complete the Query-Key-Value mapping of input features and the calculation of attention weights. By using a preset linear mapping matrix, the standardized feature vectors are mapped to query vectors, key vectors and value vectors respectively. The matrix product of the query vector and the key vector is calculated and then normalized by Softmax to obtain the attention weight vector, thereby achieving adaptive focusing on highly relevant features.

[0108] For the QKV mapping, the standardized feature vector The query vector, key vector, and value vector are mapped to three pre-defined linear mapping matrices, respectively. The specific formulas are as follows: , , ,in , , For the preset weight matrix, This is the hidden dimension; specifically, for attention weight calculation, the attention score is obtained by calculating the dot product of the query vector and the key vector. The attention score is then normalized using the Softmax function to convert it into a probability distribution, with the specific formula being... ,in Let be the attention weight in the i-th dimension, and the sum of all attention weights is 1.

[0109] Preferably, QKV hidden dimensions ; , , This is the learnable parameter matrix.

[0110] For example, using standardized feature vectors For example, after QKV mapping, we get (64-dimensional) (64-dimensional) (64-dimensional), calculate the dot product The attention weight vector is obtained by Softmax normalization.

[0111] Step S3.2: Perform an element-wise weighted summation operation on the attention weight vector and the value vector to obtain the attention weighted feature vector.

[0112] Preferably, the core of this step is to perform attention-weighted fusion of features by summing the attention weight vector and value vector obtained in step S3.1 element by element to obtain an attention-weighted feature vector that incorporates the importance information of the features.

[0113] Among them, for the weighted summation formula, the attention-weighted feature vector The calculation method is as follows ,in Let i be the attention weights in the i-th dimension. Let be the i-th element of the value vector; where, for the output dimension, the dimension of the attention-weighted feature vector is the same as the dimension of the value vector, which is 64 dimensions.

[0114] Step S3.3: Input the attention-weighted feature vector into the long short-term memory network layer, and update the cell state through the collaborative calculation of the forget gate, input gate, and output gate to obtain the updated cell state.

[0115] Preferably, the core of this step is to model the temporal dependency relationship. Through the triple gating mechanism (forget gate, input gate, and output gate) of the LSTM layer, the invalid information is selectively forgotten while retaining long-term temporal dependency information, thereby completing the update of the cell state.

[0116] Among them, for the calculation of the forget gate, the forget gate Controlling the cell state at the previous moment The proportion of information that needs to be forgotten; where, for input gate calculation, the input gate... Controlling new information at the current moment The proportion of cell states that need to be written; where, for cell state updates, the new cell state... ,in Candidate cell state, This is element-wise multiplication.

[0117] Preferably, the LSTM has 2 layers; the number of hidden units is... Forget Gate Bias Initialize to 1.0 (to enable the forget gate by default); the activation function is Sigmoid.

[0118] Step S3.4: Process the updated cell state using the Tanh activation function to obtain the activated cell state vector.

[0119] Preferably, the core of this step is to perform a nonlinear transformation of the cell state, and process the updated cell state through the Tanh activation function to generate activation values ​​in the range of -1 to 1, thereby enhancing the expressive power of the features.

[0120] For Tanh activation, ,in This is the updated cell state vector. The function compresses the cell state to the range of -1 to 1, and the output vector has the same dimension as the cell state, which is 64.

[0121] Step S3.5: Multiply the activated cell state vector by the output gate activation value of the long short-term memory network to obtain the final hidden state vector of the long short-term memory network.

[0122] Preferably, the core of this step is to generate the final output of the LSTM layer, and to control the amount of information that needs to be output in the cell state by controlling the output gate activation value, thereby generating the final hidden state vector to be passed to the next layer or the fully connected layer.

[0123] Among them, for output gate calculation, output gate Among them, for the final hidden state, ,in This is the final hidden state vector at the current moment, with a dimension of 64.

[0124] Step S3.6: Input the final hidden state vector into the fully connected layer, perform a linear transformation, and then map it using the Tanh function to obtain the initial predicted value of the vehicle arrival time.

[0125] Preferably, the core of this step is to output the prediction result, map the final hidden state vector of LSTM to the initial predicted value of vehicle arrival time through a fully connected layer, and output the normalized prediction result after Tanh activation.

[0126] For fully connected layers, ,in , Here are the weights and biases for the fully connected layer; where, for the Tanh output, The output value ranges from -1 to 1, and is then mapped to the [0,1] interval to obtain the normalized predicted value; where, for destandardization, the normalized predicted value is... Denormalize to the original time unit (minutes), the formula is as follows: ,in and These are the minimum and maximum historical travel times, respectively.

[0127] Preferably, the fully connected layer has an input dimension of 64 and an output dimension of 1; Tanh activation; and denormalized parameters. minute, minute.

[0128] For example, with a 64-dimensional final hidden state vector Through fully connected layers Activated by Tanh (Normalization), destandardization yields the initial predicted value. minute.

[0129] Beneficial effects of steps S3.1 to S3.6: This series of steps uses a fusion model of attention mechanism and long short-term memory network to gradually transform standardized feature vectors into initial predicted values ​​of vehicle arrival time. The attention layer achieves adaptive focusing on highly relevant features, and the LSTM layer achieves effective modeling of temporal dependencies. The two work together to output accurate initial predicted values, laying the foundation for subsequent dynamic updates.

[0130] In this process, step S3.1 calculates attention weights using QKV mapping and Softmax normalization to achieve adaptive focusing on key features; step S3.2 uses the attention weights to perform weighted summation of the value vectors to complete the fusion expression of features; step S3.3 updates the cell state using the LSTM triple-gating mechanism to effectively preserve long-term temporal dependency information; step S3.4 performs Tanh nonlinear transformation on the cell state to enhance feature expression capabilities; step S3.5 controls the generation of the final hidden state vector by controlling the output gate activation value; and step S3.6 uses a fully connected layer and Tanh activation to output normalized prediction values, which are then de-normalized to obtain the initial prediction value of the vehicle arrival time.

[0131] Step S4: Detect whether the change of the latest update of the time series dataset compared with the previous update meets the trigger update condition. If it does, repeat steps S2 to S3 with the latest updated time series dataset as input to obtain the updated predicted value of the vehicle arrival time.

[0132] Furthermore, step S4 specifically includes the following steps: Step S4.1: Extract all the latest data points corresponding to the latest update from the time series dataset.

[0133] Preferably, the core of this step is to extract the latest environmental status data, using vehicle identifiers and timestamps as the primary keys, and obtain a complete data snapshot of all vehicles at the latest moment from the time-series dataset.

[0134] For primary key queries, the most recent record for each vehicle is retrieved from the time-series dataset using the current system timestamp as the key, serving as the latest data point for each vehicle. For data integrity, the latest data point for each vehicle contains data from all dimensions of that vehicle, including real-time traffic conditions, vehicle operation, park operations, and task data.

[0135] Preferably, the query time window tolerance is 2 seconds (allowing for data to arrive late); batch querying of multiple vehicles is supported.

[0136] Step S4.2: Calculate the change between each latest data point and the corresponding data point updated in the previous step to obtain a set of data change amounts.

[0137] Preferably, the core of this step is to quantify the magnitude of environmental change by using the difference between the latest data point and the data point in the previous step as the change amount, thus providing a basis for subsequent weighted change amount calculation.

[0138] Among them, for the calculation of change, for numerical data, change ,in The value of the i-th dimension of the latest data point. Update the i-th dimension value of the corresponding data point in the previous step; where, for binary data, the change in the energy replenishment demand label and the sudden road condition label is calculated based on the numerical difference.

[0139] Step S4.3: Multiply each data change in the set of data changes by the corresponding feature weight in the standardized feature vector and sum them to obtain the weighted average change.

[0140] Preferably, the core of this step is to calculate the overall change. By weighting and summing the changes in each dimension with their corresponding feature weights, a comprehensive evaluation index reflecting the degree of change in the overall environment is obtained, thus avoiding interference from single-dimensional mutations in the judgment.

[0141] Among them, for the weighted summation formula, the weighted change is considered. ,in The standardization of the feature vector corresponds to the feature weights of the i-th dimension. Let be the absolute value of the change in the i-th dimension, and d be the feature dimension; where, for the feature weights, the feature weights are... Normalized weights of variance contribution rate of each dimension taken from the standardized feature vector output in step S2.8 during construction.

[0142] Preferably, the feature weights are normalized weights based on variance contribution rate; the sum of all weights is 1.

[0143] Step S4.4: Determine whether the weighted average change meets the trigger update condition. If the weighted average change exceeds the preset trigger threshold, it is determined that the trigger update condition is met, and step S4.5 is executed.

[0144] Preferably, the core of this step is to determine the triggering condition. By comparing the weighted average change with a preset triggering threshold, it is determined whether to enter the prediction update process, thus avoiding frequent updates due to small fluctuations.

[0145] Among them, for the judgment logic, when Prediction updates are triggered at specific times, where The preset trigger threshold; when If the environmental change is determined to be insignificant, the current predicted value is maintained and the process exits step S4 directly.

[0146] Preferably, the trigger threshold .

[0147] For example, consider two dimensions: the change in the congestion index. Corresponding feature weights Change in vehicle speed km / h, corresponding feature weights The weighted average change If the value is greater than the trigger threshold of 0.15, a prediction update is triggered.

[0148] Step S4.5: Extract window data containing all the latest data points from the time series dataset with a preset window length as the step size to form sliding window data.

[0149] Preferably, the core of this step is to extract window data, extracting a fixed-length time window of data from the latest data point as the endpoint, forming a stable context input for feature reconstruction.

[0150] Specifically, for the window truncation logic, the latest data point timestamp is used as the endpoint, and the window length is used as the step size to truncate the data of L sampling points forward as the sliding window data, where L is the preset window length; and for the composition of the window data, the sliding window data includes all sampling points within the window and complete multi-dimensional data of each vehicle at each sampling point time.

[0151] Preferably, the window length L = 12 sampling points (corresponding to 120 seconds); the window truncation direction is from the latest data point backward.

[0152] Step S4.6: Based on the sliding window data, re-extract the temporal features, and concatenate the re-extracted temporal features with the historical temporal features in chronological order to form a reconstructed temporal feature vector.

[0153] Preferably, the core of this step is to perform incremental reconstruction of time-series features. Based on the sliding window data, time-series features such as congestion index sequence, vehicle speed change rate, and sudden road condition markers are re-extracted and concatenated with historical time-series features in chronological order to form a reconstructed time-series feature vector containing the latest context information.

[0154] Specifically, for new feature extraction, the same temporal feature extraction logic as in step S2.1 is performed on the sliding window data to generate the congestion index sequence, vehicle speed change rate sequence, and sudden road condition marker sequence within the window; specifically, for feature concatenation, the newly extracted temporal feature sequence is concatenated with the historical temporal feature sequence in chronological order, with the newly extracted feature located at the end of the time axis, forming a reconstructed temporal feature vector with the dimension of the original feature dimension (historical window length + L).

[0155] Step S4.7: Input the reconstructed temporal feature vector into the fusion prediction model to obtain the updated predicted value of the vehicle arrival time.

[0156] Preferably, the core of this step is to dynamically update the predicted value. The reconstructed time-series feature vector is input into the fusion prediction model trained in step S3, and after performing feature engineering processing (step S2), it is input into the prediction model to obtain the updated predicted value of the vehicle arrival time.

[0157] For the model input, the reconstructed temporal feature vector is processed in step S2 (feature concatenation, Pearson filtering, PCA dimensionality reduction) to obtain a new standardized feature vector, which is then input into the fusion prediction model. For the prediction output, the parameters of the fusion prediction model remain unchanged, and the updated predicted values ​​of vehicle arrival times are output. .

[0158] Beneficial effects of steps S4.1 to S4.7: This series of steps, through a weighted change detection trigger mechanism, sliding window data extraction, and temporal feature reconstruction, enables dynamic updates of the prediction results, allowing the predicted vehicle arrival time to respond in real time to changes in the transportation environment and remain synchronized with the latest environmental status.

[0159] The process includes the following steps: Step S4.1 extracts the latest data points to provide a basis for change detection; Step S4.2 calculates the change in each dimension to quantify the actual magnitude of environmental changes; Step S4.3 obtains a comprehensive change evaluation index by weighted summation of feature weights to avoid interference from single-dimensional abrupt changes; Step S4.4 achieves accurate response to changes in the transportation environment by triggering threshold judgments; Step S4.5 extracts sliding window data to form a stable context input; Step S4.6 reconstructs the temporal feature vector to ensure the continuity of feature expression; and Step S4.7 inputs the reconstructed temporal feature vector into the fusion prediction model to obtain the updated vehicle arrival time prediction value.

[0160] Step S5: Calculate the prediction deviation between the updated predicted value and the initial predicted value. When the prediction deviation exceeds a preset deviation threshold, generate a model update instruction.

[0161] Furthermore, step S5 specifically includes the following steps: Step S5.1: Obtain the initial predicted value of the vehicle arrival time and the updated predicted value of the vehicle arrival time.

[0162] Preferably, the core of this step is to obtain the consistency between the two predicted values, providing complete input data for deviation calculation.

[0163] Specifically, for the initial prediction value, the initial prediction value of the vehicle arrival time is read from the prediction result cache in step S3. Specifically, for the updated predicted value, the updated predicted value of the vehicle arrival time is read from the latest prediction result in step S4. .

[0164] Preferably, the units for both predicted values ​​are uniformly set to minutes; the timestamp recording precision is seconds.

[0165] Step S5.2: Calculate the absolute difference between the updated predicted value and the initial predicted value to obtain the prediction deviation.

[0166] Preferably, the core of this step is to quantify the prediction deviation by calculating the absolute difference between the updated prediction value and the initial prediction value, thereby quantifying the degree of deviation in the prediction result.

[0167] Among them, for the calculation of absolute difference, prediction bias ,in This is the updated forecast value (in minutes). This is the initial forecast value (in minutes). The unit is minutes.

[0168] Step S5.3: Determine whether the prediction deviation exceeds a preset deviation threshold. If it does, proceed to the update instruction generation process.

[0169] Preferably, the core of this step is to determine if the deviation exceeds the limit. When the prediction deviation exceeds the preset deviation threshold, the model update instruction generation process is triggered. When the prediction deviation does not exceed the threshold, step S5 is exited, and the current model parameters remain unchanged.

[0170] Among them, for the judgment logic, when When the process continues, proceed to step S5.4 to continue the update instruction generation process; when If the prediction deviation is determined to be within an acceptable range, the model update is not triggered, and the process exits directly from step S5.

[0171] Preferably, the deviation threshold minute.

[0172] Step S5.4: Construct a set of instruction parameters containing update type and priority based on the prediction deviation.

[0173] Preferably, the core of this step is to complete the parameterization configuration of the update instruction, determine the update type and priority based on the prediction deviation, and generate a structured instruction parameter set for use in step S5.5.

[0174] Among them, for the update type determination, when When the update type includes feature weight update; when At that time, the update type includes both feature weight update and LSTM parameter adjustment; among them, for priority determination, The time priority is high; The time priority is medium; The priority is low or skipped.

[0175] Preferably, the feature weight update trigger threshold Minutes; LSTM parameter adjustment trigger threshold is Minutes; Priority: High, Medium, Low.

[0176] For example, with the initial predicted value Minutes, Updated Predicted Value Taking minutes as an example, prediction bias For minutes (greater than the deviation threshold of 2 minutes, less than 4 minutes), the update type is feature weight update, and the priority is medium.

[0177] Step S5.5: Generate and output model update instructions based on the instruction parameter set.

[0178] Preferably, the core of this step is to generate and output the model update instruction, encapsulate the instruction parameter set constructed in step S5.4 into a structured model update instruction, and output it to step S6 for execution.

[0179] For the instruction structure, the model update instruction includes an update type field (feature weight update, LSTM parameter adjustment), a priority field (high, medium, low), and a prediction bias field (…). The data includes a numerical value and a timestamp field; for instruction output, the model update instruction is pushed to the instruction response module in step S6, triggering the model parameter update process in step S6.

[0180] Beneficial effects of steps S5.1 to S5.5: This series of steps forms a complete deviation detection and model update triggering mechanism through quantitative calculation and condition judgment of prediction deviation. When the prediction deviation exceeds the threshold, the model update instruction is accurately generated, which effectively ensures the self-correction and continuous optimization capabilities of the fusion prediction model.

[0181] Specifically, step S5.1 simultaneously acquires the initial and updated predicted values, providing complete input for bias calculation; step S5.2 calculates the absolute difference of the prediction bias to quantify the degree of deviation in the prediction results; step S5.3 determines whether to enter the update process based on the bias threshold to avoid model oscillation caused by frequent updates; step S5.4 constructs an instruction parameter set containing update type and priority based on the bias magnitude; step S5.5 encapsulates the instruction parameter set into a structured model update instruction and outputs it to step S6 for execution, achieving a seamless connection between bias detection and model optimization.

[0182] Step S6: In response to the model update instruction, calculate the average error between the predicted arrival time and the actual arrival time of the historical completed trips using the mean absolute error, and update the feature weights and parameter configuration of the fusion prediction model according to the average error to obtain the optimized prediction model parameters.

[0183] Furthermore, step S6 specifically includes the following steps: Step S6.1: In response to the model update instruction, filter historical completed trip records from the time series dataset to form a historical trip dataset.

[0184] Preferably, the core of this step is to filter historical training data, using vehicle identification and trip status as filtering criteria, to extract complete records of completed trips from the time-series dataset, providing historical samples with real labels for subsequent error analysis.

[0185] Specifically, regarding trip status determination, when a vehicle's trip status is marked as "completed" in the task management system, the vehicle's historical records are considered completed trips. Regarding the data range, the filtering time range is the most recent. Completed trip or recent Completed itinerary records within a calendar day are taken as the intersection of the two.

[0186] Preferably, the itinerary selection time window is the 50 most recent completed itineraries; or the 7 most recent calendar days, taking the larger time range.

[0187] Step S6.2: Extract the predicted arrival time and actual arrival time of each completed trip from the historical trip dataset to form a time pair sequence.

[0188] Preferably, the core of this step is to complete the pairing and extraction of predicted and actual values, aligning the historical predicted arrival time and actual arrival time of each completed trip with the trip identifier to form a time pair sequence that can be used for error analysis.

[0189] Specifically, for time pair extraction, for each completed trip record, the predicted arrival time of that trip is extracted. Compared with actual arrival time , constituting time pair Specifically, for sequence construction, all time pairs are arranged in chronological order of their completion times to form a time pair sequence. ,in This represents the number of historical trips.

[0190] Step S6.3: Calculate the absolute difference between the predicted arrival time and the actual arrival time for each pair of time pairs in the time pair sequence, and the mean of all absolute differences, to obtain the average error.

[0191] Preferably, the core of this step is to quantitatively evaluate the prediction error by calculating the average deviation between the predicted and actual historical travel values ​​using the mean absolute error, thus providing an optimization target for updating the model parameters.

[0192] Among them, for a single pair of absolute differences, the absolute difference of the i-th trip. Among them, for the mean absolute error, ,in The length of the time-to-sequence pair is given by MAE, which is in minutes.

[0193] Preferably, minimum sample size MAE calculation is performed when the number of samples is greater than 10; the default parameters are used when the number of samples is less than 10.

[0194] Step S6.4: Calculate the gradient of the average error with respect to all feature weights of the fusion prediction model, and adjust the feature weights along the negative gradient direction to obtain the updated feature weights.

[0195] Preferably, the core of this step is to complete the gradient backpropagation update of the feature weights. The mean absolute error (MAE) calculated in step S6.3 is used as the loss function. The partial derivatives of MAE with respect to each feature weight are calculated and updated along the negative gradient direction to achieve adaptive adjustment of feature importance.

[0196] Specifically, for gradient calculation, the QKV mapping weight matrix in step S3.1... , , ,gradient , , Calculated through backpropagation, specifically as follows: For the weight update rule, the weights are updated along the negative gradient direction, and the update formula is: ,in The learning rate is denoted as ; where, for the source of feature weights, the feature weights are the column normalized norms of the QKV mapping matrix in step S3.1, reflecting the degree of contribution of each input feature dimension to the prediction result.

[0197] Preferably, the learning rate α = 0.001; the weight update uses the Adam optimizer; and the maximum number of update rounds is 10.

[0198] Step S6.5: Use the average error and the updated feature weights as model hyperparameters to adjust the number of hidden units and the learning rate of the LSTM layer to obtain the updated parameter configuration.

[0199] Preferably, the core of this step is to adaptively adjust the hyperparameters of the LSTM model. Based on the MAE in step S6.3 and the updated feature weights in step S6.4, the number of hidden units and the learning rate of the LSTM are dynamically adjusted to enhance the model's ability to capture complex patterns.

[0200] Among them, regarding the adjustment of the number of hidden units, when When the model capacity is deemed insufficient, the number of LSTM hidden units is reduced. Increase to ;when At that time, the model capacity is determined to be redundant, and the number of LSTM hidden units is reduced. Reduce to The rest remain unchanged; specifically, regarding learning rate adjustment, if MAE does not decrease after K consecutive updates, the learning rate \eta is halved, with a minimum learning rate not lower than .

[0201] Preferably, the step size for adjusting the number of hidden units is 8; the maximum number of hidden units... Minimum number of hidden units The learning rate decay coefficient is 0.5; the threshold for consecutive epochs without a decrease in learning rate. Minimum learning rate .

[0202] Step S6.6: Combine the updated feature weights with the updated parameter configuration to obtain the optimized prediction model parameters.

[0203] Preferably, the core of this step is to encapsulate and output the model parameter update, combining the feature weights (QKV mapping matrix) updated in step S6.4 with the LSTM hyperparameters (number of hidden units, learning rate) adjusted in step S6.5 into a complete set of optimized prediction model parameters, replacing the original model parameters, and completing the entire model adaptive optimization process.

[0204] For parameter encapsulation, the optimized prediction model parameters include: the updated QKV mapping matrix. , , Number of hidden units in LSTM after update Updated learning rate And the fully connected layer parameters; among them, for parameter write-back, the optimized parameters are written back to the fusion prediction model in step S3 to complete the in-situ update of the model, and the updated model is used for subsequent prediction tasks.

[0205] Beneficial effects of steps S6.1 to S6.6: This series of steps responds to model update instructions, quantifies prediction errors from historical travel data, and adjusts feature weights and LSTM model parameters through gradient backpropagation mechanism to form a complete model adaptive optimization closed loop. This ensures that the fusion prediction model can continuously optimize as the data distribution evolves, and always maintains high prediction accuracy.

[0206] The steps are as follows: Step S6.1: Filter historical completed trip records to provide real label data for error analysis; Step S6.2: Extract the predicted and actual arrival time pairs to form the error analysis input sequence; Step S6.3: Quantify the overall prediction deviation of the model using the mean absolute error (MAE); Step S6.4: Adjust the feature weights, such as the QKV mapping matrix, using gradient backpropagation; Step S6.5: Adaptively adjust the number of hidden units and the learning rate of the LSTM based on the MAE and the updated feature weights; Step S6.6: Combine the updated feature weights with the LSTM parameters and output them to complete the in-situ optimization of the model.

[0207] In summary, the overall beneficial effects of steps S1 to S6 are as follows: This method forms a complete closed loop of logistics vehicle arrival time prediction and dynamic optimization through steps S1 to S6, effectively solving the technical defects of existing methods such as single input features, inability to dynamically update prediction results according to real-time data changes, and continuous accumulation of prediction bias.

[0208] The process involves several steps: Step S1 acquires multi-source real-time data on logistics vehicles and preprocesses it, constructing a continuously updated time-series dataset using vehicle identifiers and timestamps to ensure the timeliness and completeness of the prediction input; Step S2 uses Pearson correlation coefficient filtering and principal component analysis to extract highly correlated feature vectors and filter out redundant noise; Step S3 uses an attention mechanism to fuse the LSTM time-series model output to achieve joint modeling of multi-dimensional features and time-series dependencies; Step S4 triggers a sliding window mechanism by detecting weighted average change and re-executes the prediction to achieve dynamic updates of the prediction results; Step S5 calculates the deviation between the updated prediction value and the initial prediction value and generates a model update instruction when the deviation exceeds the limit; Step S6 responds to the update instruction by adjusting the feature weights and LSTM model parameters through backpropagation of MAE loss to ensure that the model continuously optimizes as the data distribution evolves.

[0209] An arrival time prediction device for logistics vehicles, the arrival time prediction device being applied to the arrival time prediction method described above, the arrival time prediction device comprising: The time-series dataset acquisition and update module is used to acquire real-time road condition data, vehicle operation data, park operation data, and task data of the logistics vehicles, and perform data preprocessing to form a continuously updated time-series dataset with vehicle identifier and timestamp as the key. The time series dataset analysis and dimensionality reduction module is used to extract multidimensional features from the time series dataset and then perform dimensionality reduction processing through Pearson correlation coefficient screening and principal component analysis to obtain standardized feature vectors. The initial prediction module for vehicle arrival time is used to input the standardized feature vector into the fusion prediction model of the attention mechanism and the long short-term memory network. After the standardized feature vector is weighted by the Softmax function of the attention layer, it is processed by the Tanh activation function of the LSTM layer to obtain the initial prediction value of the vehicle arrival time. The vehicle arrival time prediction update module is used to detect whether the change of the latest update of the time series dataset compared with the previous update meets the trigger update condition. If it meets the condition, the time series dataset analysis and dimensionality reduction module is repeatedly executed to the vehicle arrival time initial prediction module with the latest updated time series dataset as input to obtain the updated predicted value of the vehicle arrival time. The prediction deviation calculation and model update module is used to calculate the prediction deviation between the updated prediction value and the initial prediction value, and generate a model update instruction when the prediction deviation exceeds a preset deviation threshold. The prediction model parameter iterative optimization module is used to respond to the model update command by calculating the average error between the predicted arrival time and the actual arrival time of historical completed trips through the mean absolute error, and updating the feature weights and parameter configuration of the fusion prediction model according to the average error to obtain the optimized prediction model parameters.

[0210] like Figure 3 As shown, the electronic device 7 includes a processor 71 and a memory 72 coupled to the processor 71.

[0211] The memory 72 stores program instructions for implementing the logistics vehicle arrival time prediction method of any of the above embodiments.

[0212] The processor 71 is used to execute program instructions stored in the memory 72 to predict the arrival time of logistics vehicles.

[0213] The processor 71 can also be referred to as a CPU (Central Processing Unit). The processor 71 may be an integrated circuit chip with signal processing capabilities. The processor 71 can also be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. A general-purpose processor can be a microprocessor or any conventional processor.

[0214] Furthermore, Figure 4 This is a schematic diagram of the structure of a storage medium according to an embodiment of this application. See also: Figure 4 The storage medium 8 in this embodiment stores program instructions 81 capable of implementing all the above methods. These program instructions 81 can be stored in the storage medium as a software product and include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to perform all or part of the steps of the methods in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks, or terminal devices such as computers, servers, mobile phones, and tablets.

[0215] In the several embodiments provided in this application, it should be understood that the disclosed systems, methods, and approaches can be implemented in other ways. For example, the system embodiments described above are merely illustrative. For instance, the arrival time prediction of a unit is only a logical function of arrival time prediction. In actual implementation, there may be other arrival time prediction methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not performed. Furthermore, the mutual coupling or direct coupling or communication connection shown or discussed may be through some interface, and the indirect coupling or communication connection of the system or unit may be electrical, mechanical, signal, or other forms.

[0216] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated units described above can be implemented in hardware or as software functional units. The above are merely embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made based on the description and drawings of this application, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for predicting the arrival time of logistics vehicles, characterized in that, The arrival time prediction method includes: Step S1: Obtain real-time road condition data, vehicle operation data, park operation data, and task data of the logistics vehicle, and perform data preprocessing to form a continuously updated time-series dataset with vehicle identifier and timestamp as the key. Step S2: After extracting multidimensional features from the time series dataset, dimensionality reduction is performed by screening with Pearson correlation coefficient and principal component analysis to obtain standardized feature vectors. Step S3: Input the standardized feature vector into the fusion prediction model of the attention mechanism and the long short-term memory network. After weighting the standardized feature vector through the Softmax function of the attention layer, it is processed by the Tanh activation function of the LSTM layer to obtain the initial predicted value of the vehicle arrival time. Step S4: Detect whether the change of the latest update of the time series dataset compared with the previous update meets the trigger update condition. If it does, repeat steps S2 to S3 with the latest updated time series dataset as input to obtain the updated predicted value of vehicle arrival time. Step S5: Calculate the prediction deviation between the updated predicted value and the initial predicted value. When the prediction deviation exceeds a preset deviation threshold, generate a model update instruction. Step S6: In response to the model update instruction, calculate the average error between the predicted arrival time and the actual arrival time of the historical completed trips using the mean absolute error, and update the feature weights and parameter configuration of the fusion prediction model according to the average error to obtain the optimized prediction model parameters.

2. The arrival time prediction method according to claim 1, characterized in that, Step S1: Obtain real-time road condition data, vehicle operation data, park operation data, and task data of the logistics vehicles, and perform data preprocessing to form a continuously updated time-series dataset with vehicle identifiers and timestamps as keys, including: Step S1.1: Obtain real-time road condition data, vehicle operation data, park operation data, and task data of the logistics vehicles, and integrate them into a multi-source raw dataset; Step S1.2: Data points exceeding three times the standard deviation of the mean in the original multi-source dataset are removed using the 3σ criterion to obtain the anomaly-filtered dataset. Step S1.3: Fill in the missing values ​​of the anomaly-filtered dataset using the forward imputation method to obtain the complete dataset; Step S1.4: Map the data of each dimension in the complete dataset to the interval between 0 and 1 through Min-Max standardization to obtain a standardized dataset; Step S1.5: Perform linear interpolation and time synchronization on the standardized dataset with a sampling interval of 10 seconds to obtain a time-aligned dataset; Step S1.6: Group the time-aligned dataset by vehicle identifier and arrange it in ascending order of timestamp to form a continuously updated time-series dataset with vehicle identifier and timestamp as the key.

3. The arrival time prediction method according to claim 1, characterized in that, Step S2: After extracting multidimensional features from the time-series dataset, dimensionality reduction is performed using Pearson correlation coefficient screening and principal component analysis to obtain a standardized feature vector, including: Step S2.1: Extract the congestion index sequence, vehicle speed change rate, and sudden road condition markers from the time-series dataset as time-series road condition features; Step S2.2: Extract vehicle type weight, task priority coefficient, and remaining distance as vehicle and task features; Step S2.3: Extract the predicted queuing time and the deviation of the channel restriction speed as dynamic features of the park; Step S2.4: Extract the energy replenishment demand markers and the urgency of the unloading window as constraint features; Step S2.5: Align and stitch the time-series road condition features, vehicle and task features, park dynamic features, and constraint features according to vehicle identifiers and timestamps to form a multi-dimensional feature matrix; Step S2.6: Calculate the Pearson correlation coefficient between each pair of features in the multidimensional feature matrix to construct a correlation coefficient matrix; Step S2.7: Identify feature pairs whose absolute values ​​of correlation coefficients are greater than a preset absolute value threshold from the correlation coefficient matrix, and remove features with smaller variances from each pair to obtain the filtered feature matrix; Step S2.8: Calculate the eigenvalues ​​and variance contribution rates of each principal component in the filtered feature matrix, and mark the first few principal components whose cumulative variance contribution rates exceed the preset contribution rate threshold according to the eigenvalues ​​from largest to smallest. Project the filtered feature matrix onto the subspace formed by the first few principal components to obtain the standardized feature vector.

4. The arrival time prediction method according to claim 1, characterized in that, Step S3: Input the standardized feature vector into the fusion prediction model of the attention mechanism and the long short-term memory network. After weighting the standardized feature vector through the Softmax function of the attention layer, it is processed by the Tanh activation function of the LSTM layer to obtain the initial predicted value of the vehicle arrival time, including: Step S3.1: The standardized feature vector is mapped to a query vector, a key vector, and a value vector respectively through a preset linear mapping matrix, and the matrix product of the query vector and the key vector is calculated. After normalization by the Softmax function, the attention weight vector is obtained. Step S3.2: Perform an element-wise weighted summation operation on the attention weight vector and the value vector to obtain the attention weighted feature vector; Step S3.3: Input the attention-weighted feature vector into the long short-term memory network layer, and update the cell state through the collaborative calculation of the forget gate, input gate, and output gate to obtain the updated cell state; Step S3.4: Process the updated cell state using the Tanh activation function to obtain the activated cell state vector; Step S3.5: Multiply the activated cell state vector by the output gate activation value of the long short-term memory network layer to obtain the final hidden state vector of the long short-term memory network. Step S3.6: Input the final hidden state vector into the fully connected layer, perform a linear transformation, and then map it using the Tanh function to obtain the initial predicted value of the vehicle arrival time.

5. The arrival time prediction method according to claim 1, characterized in that, Step S4: Detect whether the change in the latest update of the time series dataset compared to the previous update satisfies the trigger update condition. If it does, repeat steps S2 to S3 using the latest updated time series dataset as input to obtain the updated predicted value of the vehicle arrival time, including: Step S4.1: Extract all the latest data points corresponding to the latest update from the time series dataset; Step S4.2: Calculate the change between each latest data point and the corresponding data point updated in the previous step to obtain a set of data change amounts; Step S4.3: Multiply each data change in the set of data changes by the corresponding feature weight in the standardized feature vector and sum them to obtain the weighted average change. Step S4.4: Determine whether the weighted average change meets the trigger update condition. If the weighted average change exceeds the preset trigger threshold, it is determined that the trigger update condition is met, and step S4.5 is executed. Step S4.5: Extract window data containing all the latest data points from the time series dataset with a preset window length as the step size to form a sliding window data; Step S4.6: Based on the sliding window data, re-extract the temporal features, and concatenate the re-extracted temporal features with the historical temporal features in chronological order to form a reconstructed temporal feature vector; Step S4.7: Input the reconstructed temporal feature vector into the fusion prediction model to obtain the updated predicted value of the vehicle arrival time.

6. The arrival time prediction method according to claim 1, characterized in that, Step S5: Calculate the prediction deviation between the updated predicted value and the initial predicted value. When the prediction deviation exceeds a preset deviation threshold, generate a model update instruction, including: Step S5.1: Obtain the initial predicted value of the vehicle arrival time and the updated predicted value of the vehicle arrival time; Step S5.2: Calculate the absolute difference between the updated predicted value and the initial predicted value to obtain the prediction deviation; Step S5.3: Determine whether the prediction deviation exceeds a preset deviation threshold. If it does, proceed to the update instruction generation process. Step S5.4: Construct an instruction parameter set containing update type and priority based on the prediction deviation; Step S5.5: Generate and output model update instructions based on the instruction parameter set.

7. The arrival time prediction method according to claim 1, characterized in that, Step S6: In response to the model update instruction, calculate the average error between the predicted arrival time and the actual arrival time of historical completed trips using the mean absolute error, and update the feature weights and parameter configuration of the fusion prediction model based on the average error to obtain the optimized prediction model parameters, including: Step S6.1: In response to the model update instruction, filter historical completed trip records from the time series dataset to form a historical trip dataset; Step S6.2: Extract the predicted arrival time and actual arrival time of each completed trip from the historical trip dataset to form a time pair sequence; Step S6.3: Calculate the absolute difference between the predicted arrival time and the actual arrival time for each pair of time pairs in the time pair sequence, and the mean of all absolute differences, to obtain the average error. Step S6.4: Calculate the gradient of the average error with respect to all feature weights of the fusion prediction model, and adjust the feature weights along the negative gradient direction to obtain the updated feature weights; Step S6.5: Use the average error and the updated feature weights as model hyperparameters to adjust the number of hidden units and the learning rate of the LSTM layer to obtain the updated parameter configuration; Step S6.6: Combine the updated feature weights with the updated parameter configuration to obtain the optimized prediction model parameters.

8. An arrival time prediction device for logistics vehicles, wherein the arrival time prediction device is applied to the arrival time prediction method as described in any one of claims 1 to 7, characterized in that, The arrival time prediction device includes: The time-series dataset acquisition and update module is used to acquire real-time road condition data, vehicle operation data, park operation data, and task data of the logistics vehicles, and perform data preprocessing to form a continuously updated time-series dataset with vehicle identifier and timestamp as the key. The time series dataset analysis and dimensionality reduction module is used to extract multidimensional features from the time series dataset and then perform dimensionality reduction processing through Pearson correlation coefficient screening and principal component analysis to obtain standardized feature vectors. The initial prediction module for vehicle arrival time is used to input the standardized feature vector into the fusion prediction model of the attention mechanism and the long short-term memory network. After the standardized feature vector is weighted by the Softmax function of the attention layer, it is processed by the Tanh activation function of the LSTM layer to obtain the initial prediction value of the vehicle arrival time. The vehicle arrival time prediction update module is used to detect whether the change of the latest update of the time series dataset compared with the previous update meets the trigger update condition. If it meets the condition, the time series dataset analysis and dimensionality reduction module is repeatedly executed to the vehicle arrival time initial prediction module with the latest updated time series dataset as input to obtain the updated predicted value of the vehicle arrival time. The prediction deviation calculation and model update module is used to calculate the prediction deviation between the updated prediction value and the initial prediction value, and generate a model update instruction when the prediction deviation exceeds a preset deviation threshold. The prediction model parameter iterative optimization module is used to respond to the model update command by calculating the average error between the predicted arrival time and the actual arrival time of historical completed trips through the mean absolute error, and updating the feature weights and parameter configuration of the fusion prediction model according to the average error to obtain the optimized prediction model parameters.

9. An electronic device, characterized in that, The method includes a processor and a memory coupled to the processor, the memory storing program instructions that can be executed by the processor; when the processor executes the program instructions stored in the memory, it implements the arrival time prediction method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program instructions that, when executed by a processor, enable the arrival time prediction method as described in any one of claims 1 to 7.

Images