Double-path pulse encoding and frequency domain perception prediction method for express cargo flow

Abstract

The application discloses a method for double-path pulse coding and frequency domain perception prediction of express volume, and aims at the problems of coexistence of stable period and local mutation in express volume sequence, disturbance change of frequency domain structure and difficulty in joint modeling of multi-scale time dependence, and proposes the following scheme: collecting and preprocessing logistics time series data, dividing and embedding into time segments; extracting stable state and differential dynamic information and fusing by a double-path pulse coding module; performing frequency domain transformation by a frequency domain adaptive gating module, generating a frequency perception gating signal and modulating a dynamic response; jointly modeling frequency guidance, cross-time segment long-range dependence and pulse simulation time step evolution by a multi-dimensional dependence modeling module; and finally predicting and outputting future volume. The application can simultaneously retain stable state and local mutation, enhance frequency domain perception and multi-dimensional dependence modeling capability, and take into account sparse computing efficiency.

Description

Technical Field

[0001] This invention belongs to the fields of smart logistics, time series prediction, spiking neural networks and frequency domain feature modeling, specifically involving a dual-path pulse coding and frequency domain perception prediction method for express delivery traffic. Background Technology

[0002] Forecasting express delivery volume is a crucial foundation for express delivery companies in areas such as network operation, staff scheduling, vehicle dispatching, warehousing resource allocation, trunk transportation arrangements, and early warning of abnormal risks. With the continuous expansion of e-commerce, express delivery volume exhibits significant cyclical, sudden, and regional variations. For example, there are significant fluctuations between weekdays and holidays, and promotional activities, weather changes, and unforeseen events can cause drastic changes in volume within a short period. Therefore, accurately predicting express delivery volume over multiple time steps has become a key technical challenge in intelligent logistics scheduling and refined operations.

[0003] Existing methods for predicting express delivery volume typically include traditional statistical models, machine learning models, and deep learning models. Traditional statistical models mainly rely on linear assumptions and fixed periodic characteristics, making it difficult to fully express the nonlinear changes and sudden fluctuations in express delivery volume. While conventional machine learning models can use multidimensional features for prediction, their ability to capture long-term dependencies and local dynamic changes is limited. Deep learning models such as recurrent neural networks, convolutional neural networks, and Transformers can improve prediction accuracy, but they usually involve large computational costs and still have shortcomings in simultaneously modeling the state-preserving characteristics, short-term change characteristics, and frequency domain periodic characteristics of the volume.

[0004] Furthermore, the time series data of express delivery and logistics volume not only exhibits trends in the time domain but also contains significant frequency domain periodic information, such as daily cycles, weekly cycles, promotional cycles, and seasonal fluctuations. Existing methods often model directly in the time domain or only use frequency domain features as auxiliary inputs, lacking adaptive selection and dynamic modulation mechanisms for different frequency components. This results in insufficient predictive stability when faced with complex periodic overlaps, abnormal peaks, and short-term drastic fluctuations.

[0005] Therefore, it is necessary to propose a method for predicting express logistics volume that can simultaneously capture the original sequence state information, local differential dynamic information, and frequency domain periodic information, so as to improve prediction accuracy, enhance adaptability to sudden fluctuations and complex periodic changes, and reduce the difficulty of applying the model in actual logistics scheduling scenarios. Summary of the Invention

[0006] To overcome the problem that existing express delivery volume prediction methods struggle to simultaneously consider long-term state maintenance, local dynamic changes, and frequency domain periodic characteristics, this invention provides a dual-path pulse coding and frequency domain sensing prediction method for express delivery volume. This method extracts the original sequence state features and time-difference dynamic features using a dual-path pulse coding module, and combines this with a frequency domain adaptive gating module to select and modulate different frequency components. Finally, a multi-dimensional dependency modeling module fuses frequency guidance, cross-time segment dependencies, and pulse simulation time-step dependencies, thereby enabling the prediction of express delivery volume for multiple future time steps.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: A dual-path pulse coding and frequency domain sensing prediction method for express delivery volume includes the following steps: Obtain time series data of express delivery and logistics business volume, perform missing value processing, outlier processing, normalization processing or standardization processing on the time series data of express delivery and logistics business volume, and divide the preprocessed data into multiple time segments along the time dimension; perform linear mapping and normalization processing on the time segments to obtain the time segment embedded representation.

[0008] The time segment embedding representation is input into a dual-path pulse coding module. The dual-path pulse coding module includes an original sequence state coding branch and a temporal difference dynamic coding branch. The original sequence state coding branch accumulates membrane potentials and fires pulses on the time segment embedding representation using LIF spiking neurons, obtaining a state-preserving pulse representation that characterizes the continuous state of the traffic volume. The temporal difference dynamic coding branch calculates the first-order difference along the original time axis, performs time segmentation, linear mapping, and normalization on the difference sequence, and generates a continuous local dynamic activation representation through a smoothing threshold function. The state-preserving pulse representation and the local dynamic activation representation are fused using learnable fusion parameters to obtain a fused pulse representation.

[0009] The fused pulse representation is input into the frequency domain adaptive gating module, and a frequency domain transformation is performed along the time segment dimension to obtain the spectral representation. Frequency-aware selection weights are generated based on the fused pulse representation, and these weights are applied to the spectral representation to enhance frequency components relevant to the prediction task and suppress interference frequency components. Subsequently, an inverse frequency domain transformation is performed to obtain the frequency-modulated time segment representation, and a frequency-aware gating signal is generated based on the frequency-aware selection weights.

[0010] Based on the frequency-aware gating signal, the dynamic response modulation guided by frequency is applied to the dynamic changes of the pulse representation in the pulse simulation time step dimension, enabling the model to adaptively adjust the pulse response intensity according to the periodic characteristics and frequency changes in the traffic sequence.

[0011] The frequency-modulated time segment representation is input into a multidimensional dependency modeling module, which includes a frequency-guided branch, a cross-time segment dependency branch, and a pulse simulation time step dependency branch. The frequency-guided branch is used to model the global periodic dependency after frequency modulation, the cross-time segment dependency branch is used to model the temporal correlation between different time segments, and the pulse simulation time step dependency branch is used to model the dynamic response relationship in the internal simulation process of the spiking neural network. The outputs of the three branches are residually fused with the current layer input representation to obtain an enhanced representation.

[0012] The enhanced representation is input into the prediction layer, pooled and aggregated along the time step dimension of the impulse simulation, and then linearly mapped to output the predicted express logistics volume for multiple future time steps. The express logistics volume includes at least one of order volume, pickup volume, delivery volume, inbound volume, and outbound volume. The prediction results can be used for branch staff scheduling, vehicle dispatching, warehousing resource allocation, trunk transportation arrangements, or abnormal risk warnings.

[0013] Step 1: Collect and construct time series data of express delivery and logistics volume Historical business data is collected from the express delivery and logistics system. This historical business data includes, but is not limited to, order volume, pickup volume, delivery volume, inbound volume, outbound volume, holiday markers, and weather information. The data is then aligned according to a fixed time granularity, which is one day or other time intervals set according to actual business needs. Missing value imputation, outlier handling, duplicate data removal, and normalization are performed on the collected data to obtain a time series of express delivery and logistics business volume.

[0014] Let the processed historical input sequence be: in, Indicates batch size, Indicates the length of a historical time step. Indicates the number of variables.

[0015] Predicting the target is the future. The sequence of express delivery volume at each time step: in, Indicates the predicted length. This indicates the number of target business metrics that need to be predicted. Depending on actual business needs, the prediction targets can be changes in business volume such as order volume, pickup volume, and delivery volume over the next few days.

[0016] The input for this step is the original express logistics business data, and the output is the preprocessed historical input sequence. and corresponding prediction targets .

[0017] Step 2: Time Segment Partitioning and Embedding Representation In order to transform long-duration express logistics sequences into local time structures suitable for subsequent pulse coding and frequency domain analysis, this step processes the input sequence along the time dimension. Divide the time into segments.

[0018] Let the length of each time segment be... The number of time segments is: in, This indicates the number of time segments obtained from the division. When Cannot be When divisible, the length can be kept consistent by padding with zeros or truncating.

[0019] Time segmentation operation , input sequence Divided into multiple consecutive time segments: in, This represents the set of local time segments obtained by dividing the original express logistics sequence. Then, linear mapping and normalization are performed on the divided time segments to obtain the local embedding representation: in, This indicates a trainable linear mapping layer. This indicates a normalization operation. This represents the potential representation corresponding to a segment of the original express delivery and logistics volume. The dimensions of this representation can be expressed as: ,in This represents the hidden feature dimension obtained after linear mapping of each time segment.

[0020] The input for this step is the preprocessed historical input sequence. The output is the original time segment embedding representation. This output will be fed into the subsequent dual-path pulse code module.

[0021] Step 3: Dual-path pulse coding module The dual-path pulse coding module includes a raw sequence state coding branch and a time difference dynamic coding branch, which are used to process the steady-state information and local dynamic change information in the express logistics business volume, respectively, and obtain the fused pulse representation through learnable fusion parameters.

[0022] (1) Original sequence state coding branch The original sequence state encoding branch is embedded in the original time segment obtained in step two. As input, membrane potential accumulation and pulse firing are performed through LIF spiking neurons, thereby preserving continuous state information and basic trends in express logistics business volume. Let the spiking neuron at simulation time step... The input is The membrane potential is The membrane potential update and pulse firing process of LIF neurons are as follows: in, Indicates the membrane potential leakage factor. Indicates the pulse firing threshold. Indicates the resting potential. This represents the intermediate state after the current input and the accumulated membrane potential from the previous moment. This represents the pulse output at the current simulation time step. The step function is also included. Defined as: When the membrane potential exceeds a threshold, the neuron fires a pulse; otherwise, it remains silent. Through this membrane potential accumulation process, the original sequence state encoding branch can retain continuous state information in the express delivery logistics business volume.

[0023] The output of the original sequence state coding branch is: , This indicates the number of simulation time steps within the spiking neural network. It should be noted that... It is not equivalent to the actual timeline of the original express delivery logistics sequence, but rather represents the simulated number of steps in which the spiking neuron accumulates membrane potential internally.

[0024] (2) Temporal Differential Dynamic Coding Branch The temporal difference dynamic coding branch is used to capture local variations in express delivery volume along the original real timeline. This branch first calculates the first-order difference along the original time dimension: in, Indicates the first The observed value of express delivery volume at a real time step. This represents the change between adjacent time steps. The difference value of the first time step is zero-padded to keep the sequence length constant.

[0025] The difference sequence is obtained by combining the difference values ​​of all time steps: Then, the difference sequence is divided into time segments, linearly mapped, and normalized: in, A potential representation of a difference time segment.

[0026] To avoid the loss of local dynamic information caused by hard pulse discretization, this branch uses a smooth threshold function to generate a continuous dynamic activation representation: in, This represents the Sigmoid function. The dynamic threshold representing the differential branch. Indicates temperature parameter, This represents a continuous local dynamic activation representation.

[0027] Unlike the hard impulses generated by the original sequence state coding branch, the temporal difference dynamic coding branch outputs continuous dynamic activations, which can more finely preserve local variations in express logistics volume. For example, the rapid increase in order volume caused by e-commerce promotional activities such as "618" and "Double Eleven", the decrease in delivery volume caused by the Spring Festival holiday, and short-term business fluctuations caused by extreme weather can all be enhanced by this branch.

[0028] (3) Dual-path fusion To simultaneously preserve both stable state information and local dynamic change information, this invention introduces learnable fusion parameters. For state-preserving pulse representation and local dynamic activation representation To merge: in, These are learnable parameters used to adaptively balance the contributions of the original sequence state coding branch and the temporal difference dynamic coding branch. The fused output is denoted as: in, This output serves as the initial input for the subsequent frequency domain adaptive gating module and multidimensional dependency modeling module.

[0029] The input for this step is the original time segment embedding representation. and differential time segment embedding representation The output is represented as a fused pulse. .

[0030] Step 4: Frequency Domain Adaptive Gating Module The frequency domain adaptive gating module is used to map the fused pulse representation to the frequency domain, extract the periodic structure and frequency domain change information in the express logistics business volume, and generate a frequency-aware gating signal to guide subsequent local dynamic response modulation.

[0031] Let the first The input to the multidimensional dependency modeling module is represented as follows: in Indicates the layer number. When hour, That is, the fused pulse representation output by the dual-path pulse coding module.

[0032] (1) Perform frequency domain transformation along the time segment dimension The frequency domain adaptive gating module first moves along the time segment dimension Input representation Performing a real-number fast Fourier transform yields the spectral representation: in, This represents the frequency domain transformation performed along the time segment dimension. Represents the Fast Fourier Transform of real numbers. It represents the frequency domain characteristics.

[0033] In the express delivery and logistics scenario, different frequency components correspond to different business patterns. For example, the difference in business volume between Monday to Friday and the weekend can be reflected as a weekly frequency component.

[0034] (2) Generate frequency-aware selection weights In order to generate adaptive modulation weights based on the frequency domain structure of the current input sequence, this invention maps the input representation to frequency-selective features and generates sparse frequency-aware weights through spiking neurons.

[0035] Specifically, the input representation for each simulation time step Learnable frequency projection and normalization are performed, and then sequentially input into generative LIF neurons and aggregated LIF neurons: in, The projection parameters from the learnable time segment to the frequency space are represented. This represents a spiking neuron used to generate sparse frequency-selective features. This represents a spiking neuron used to aggregate the frequency selection results of multiple simulation time steps. This represents the frequency-aware selection weight.

[0036] (3) Spectrum modulation and time-domain reconstruction Frequency-aware selection weights Action on frequency domain representation The modulated frequency domain features are obtained, and then reconstructed back to the time segment domain through inverse Fourier transform: Among them, among them, Frequency dimension and The frequency bins must be consistent to ensure that the dimensions of subsequent element-wise multiplications match. This represents element-wise multiplication. Indicates the inverse Fourier transform. This represents the frequency-modulated time segment representation. This representation retains the original time segment structure while incorporating the frequency domain context of the express delivery logistics sequence.

[0037] (4) Generate frequency-sensing gating signals To further transform the frequency domain context into a gated signal that can be used for dynamic response modulation, this invention employs frequency-aware weighting. Standardization and average pooling are performed along the frequency bin dimension: in, This represents the average pooling operation along the frequency bin dimension. and These represent the mean and variance along the frequency dimension, respectively. To prevent extremely small positive numbers with a denominator of zero, This represents a frequency-aware gating signal. This gating signal can be broadcast to subsequent dynamic response modulation processes in the time segment dimension.

[0038] The input for this step is the first... Layer input representation The output is a frequency-modulated time segment representation. and frequency-sensing gating signals .

[0039] Step 5: Frequency-guided dynamic response modulation This step explains the internal calculations of the frequency-guided dynamic unit, which serves as a fundamental component in the frequency-domain adaptive gating module and the multidimensional dependency modeling module for dynamic response modulation. Frequency-guided dynamic response modulation is used to sense and gate signals based on frequency. The adjustment pulse represents the dynamic change in the time step dimension of the pulse simulation.

[0040] It should be noted that the dynamic difference in this step differs from the time difference in step three. The time difference in step three is calculated along the true timeline of the original express logistics sequence to capture changes between adjacent dates, adjacent hours, or adjacent business cycles; the dynamic difference in this step follows the time step dimension of the spiking neural network simulation. The calculation is used to characterize the internal evolution of the pulse state during multiple membrane potential accumulation and discharge processes.

[0041] For the Layer input representation Its difference along the pulse simulation time step dimension is defined as: in, Indicates the first The first in the layer Representation of each simulation time step This represents the internal state changes between adjacent simulation time steps.

[0042] Using frequency sensing gating signals Modulate the above simulation time step differential: in, For learnable gain parameters, via broadcast Dimensional alignment, This indicates the dynamic changes in the simulation time step after frequency guidance.

[0043] Based on the above-mentioned dynamic changes in the simulated time step after modulation, the output of the frequency-guided dynamic unit is denoted as: .in, This represents the dynamic impulse response modulated using a frequency-sensing gated signal.

[0044] The input for this step is the first... Layer representation and frequency-sensing gating signals The output is a frequency-guided dynamic response. .

[0045] Step Six: Multidimensional Dependency Modeling Module The multidimensional dependency modeling module enhances long-term prediction capabilities from three dimensions: First, the frequency-guided branch preserves the frequency modulation representation generated by the frequency-domain adaptive gating module; second, the cross-time-segment dependency branch models long-range temporal relationships between different time segments; and third, the pulse simulation time-step dependency branch models the state evolution relationships on the pulse simulation time step of the spiking neural network.

[0046] Let the first Layer input is The multidimensional dependency modeling module sequentially generates frequency-guided representations. Time segment representation Simulation time step representation The next layer representation is obtained by fusing the data using a residual method. .

[0047] (1) Frequency-guided branch Frequency-guided branch in frequency-modulated representation and frequency-sensing gating signals As input, the fundamental impulse response and the frequency-guided dynamic response are fused: in, For learnable fusion parameters, This represents the output obtained by performing basic LIF pulse processing on the frequency modulation representation. This represents the dynamic response output guided by the frequency-sensing gating signal.

[0048] Output of frequency-guided branch This output is used to preserve frequency domain periodicity information and event-driven change information in the express logistics business volume sequence, such as weekly cycle, holiday cycle, promotional activity cycle, and frequency domain structure changes caused by sudden events.

[0049] (2) Cross-time segment dependency branches Cross-time-segment dependency branching is used to model long-range temporal relationships between different time segments. For express delivery volume forecasting, there are usually clear connections between multiple consecutive time segments. For example, promotional activities and holidays form consecutive phases. This invention utilizes the time-segment dimension... Perform convolution processing on the input representation: in, This represents a convolution operation performed along the time segment dimension. This represents the representation after processing across time segments.

[0050] Then The input consists of a parallel fusion structure composed of basic LIF branches and frequency-guided dynamic units: Output across time-segment dependent branches This branch can capture the continuous evolution of express logistics volume across multiple time segments, enhancing the model's ability to model long-term trends, changes before and after activities, and cross-cycle recurring patterns.

[0051] (3) Pulse simulation time step dependent branch The time-step dependent branch of the spiking simulation is used to model the state evolution along the simulation time-step dimension within the spiking neural network. Since spiking neurons accumulate membrane potential, leak, and fire pulses over multiple simulation time steps, the simulation time-step dimension can reflect the dynamic relationship of the internal spiking state as iterates.

[0052] This branch first adjusts the dimensions of the input representation so that the simulation time step dimension can serve as the primary modeling dimension. Then, convolution processing is performed along the simulation time step dimension: in, This represents a convolution operation performed along the time step dimension of the pulse simulation. This represents the simulation time-step dependent representation after processing.

[0053] Subsequently, fusion is performed through the basic LIF branch and frequency-guided dynamic unit: Output of time-step dependent branches in pulse simulation This branch can enhance the model's ability to express the evolution of the internal state of a pulse, thereby alleviating the problem of insufficient information retention caused by membrane potential decay in long sequence prediction.

[0054] (4) Multi-branch residual fusion The outputs of the above three branches are residually fused with the input representation of the current layer to obtain the enhanced representation of the next layer: in, By using residual summation instead of feature splicing or complex attention weighting, the sparsity of impulse representation can be maintained while reducing additional dense computational overhead, making the model more suitable for resource-constrained scenarios such as express delivery outlets, transfer centers or edge devices.

[0055] The above multidimensional dependency modeling process can be executed in a stacked manner. Layer. After After layer processing, the final enhanced representation is obtained: This step is represented by a fusion pulse. As the initial input, and combined with the corresponding frequency modulation representation at each layer. and frequency-sensing gating signals Perform stepwise residual enhancement to ultimately output multidimensional dependencies. Enhanced representation .

[0056] Step 7: Predict the output The representation output by the multidimensional dependency modeling module Input into the prediction module to obtain the future. The forecast results of express delivery and logistics volume at each time step.

[0057] First, the simulation time step dimension can be adjusted. Perform average pooling or weighted aggregation to obtain the compressed time segment representation: in, This represents a pooling operation along the simulation time step dimension. This represents the aggregated pulse timing representation.

[0058] Then, Flattened or rearranged before being input into the linear prediction layer: in, and These represent the learnable weights and biases of the prediction layer, respectively. This represents the model's output prediction of future express delivery and logistics volume. Indicates the prediction step size. This indicates the target quantity to be predicted. For example, when the prediction task is to forecast the order volume of a certain region over the next 7 days, .

[0059] The input for this step is a multidimensional dependency augmentation representation. The output is the forecast of future business volume. .

[0060] Step 8: Model Training and Error Optimization During the model training phase, the prediction results will be... Compared with the actual business volume sequence By comparing the results, a prediction loss function can be constructed. Mean squared error can be used as the training objective. in, Indicates the first The sample in the future The time step, the first The true value of each target indicator This represents the corresponding predicted value.

[0061] The mean absolute error can also be used as an auxiliary evaluation indicator according to business needs: All trainable parameters in the model are updated using backpropagation and gradient optimization methods.

[0062] Step Nine: Application of Prediction Results The predicted future express delivery volume output by the prediction module is applied to a smart logistics business system. This can be used for branch staff scheduling, vehicle and trunk line transportation dispatching, sorting and warehousing resource allocation, capacity warnings during promotional activities, and identification and emergency dispatching of abnormal fluctuations during special circumstances such as the Spring Festival holiday, extreme weather, or regional control measures. Through the above steps, this invention can simultaneously preserve the stable temporal state of the express delivery volume sequence, enhance the perception of local mutations, introduce frequency domain periodic information, and model cross-time segment and pulse simulation time step dependencies, thereby improving prediction accuracy, robustness, and computational efficiency.

[0063] The beneficial effects of this invention are: First, the present invention constructs a dual-path pulse coding structure by using the original sequence state coding branch and the time difference dynamic coding branch. This structure can not only preserve the continuous state information in the express logistics business volume sequence, but also capture the local change information between adjacent time steps, thereby improving the model's ability to express trend changes, short-term fluctuations and abnormal peaks.

[0064] Second, this invention utilizes LIF spiking neurons to perform spiking encoding on the embedded representation of time segments, enabling the model to express temporal state changes in an event-driven manner. This is beneficial for improving the sparsity and dynamic response capability of time series feature representation and provides a foundation for subsequent deployment of low-power intelligent logistics prediction.

[0065] Third, the present invention sets up a frequency domain adaptive gating module, which adaptively filters and enhances the periodic frequency components in the express logistics business volume through frequency domain transformation, frequency perception selection weight and inverse frequency domain transformation, so as to better adapt to multiple period superposition scenarios such as daily cycle, weekly cycle and promotion cycle.

[0066] Fourth, this invention modulates the dynamic changes in the time step dimension of pulse simulation based on frequency-sensing gating signals, so that the pulse response process can be guided by frequency domain information, thereby enhancing the model's adaptability to complex periodic changes and sudden fluctuations.

[0067] Fifth, this invention simultaneously models frequency-guided dependency, cross-time segment dependency, and impulse simulation time step dependency through a multi-dimensional dependency modeling module, and obtains an enhanced representation by using residual fusion. This can improve the model's ability to model multi-scale time series relationships and enhance the accuracy and stability of predicting express logistics business volume in multiple future time steps.

[0068] Sixth, the prediction results of this invention can be directly applied to business scenarios such as staff scheduling, vehicle dispatching, warehousing resource allocation, trunk transportation arrangements, and abnormal risk warnings at express logistics outlets, and have strong practical application value and industrial promotion significance. Attached Figure Description

[0069] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0070] Figure 1 is a schematic diagram of the non-stationary changes and frequency domain evolution of the express logistics business volume sequence.

[0071] Figure 2 is an overall flowchart of the present invention for predicting express logistics business volume.

[0072] Figure 3 is a schematic diagram of the frequency domain adaptive gating module structure of the present invention.

[0073] Figure 4 is a schematic diagram of the multidimensional dependency modeling module structure of the present invention. Detailed Implementation

[0074] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0075] Example 1 See Figures 1 to 4 This embodiment provides a dual-path pulse coding and frequency domain sensing prediction method for express delivery volume. This method takes the time series of express delivery volume as input and outputs prediction results for express delivery volume at multiple future time steps through a state coding branch of the original sequence, a time difference dynamic coding branch, a frequency domain adaptive gating module, and a multi-dimensional dependency modeling module. The express delivery volume can be one or more of the following: order volume, pickup volume, delivery volume, inbound volume, and outbound volume.

[0076] like Figure 1As shown, the volume of express delivery and logistics business typically includes both stable cyclical changes and local abrupt changes. For example, between ordinary weekdays and weekends, the business volume may exhibit relatively stable cyclical fluctuations; during e-commerce promotional events such as "Double Eleven" and "618", the number of orders, pickups, or deliveries may increase significantly in a short period of time; and during the Spring Festival holiday, extreme weather, or regional control periods, the business volume may experience a phased decline. Figure 1 The annual business volume curve, the local magnified window, and the frequency domain analysis results all indicate that the express delivery and logistics business volume not only experiences abrupt changes in the time domain, but these abrupt changes also cause changes in the frequency domain structure. Therefore, this embodiment does not only fit the original sequence in the time domain, but also simultaneously introduces differential dynamic modeling and frequency domain adaptive modulation.

[0077] like Figure 2 As shown, the method in this embodiment includes data acquisition and preprocessing, time segment division and embedding, dual-path pulse coding, frequency domain adaptive gating, multidimensional dependency modeling, and prediction output.

[0078] First, historical business data is collected from the express delivery and logistics system. This historical business data includes at least one of the following: order volume, pickup volume, delivery volume, inbound volume, and outbound volume. It may also further include auxiliary variables such as holiday markers, weather conditions, promotional activity markers, and branch area markers. The above data is then aligned according to a fixed time granularity, which can be daily, hourly, or a statistical period set by other business systems. This embodiment uses daily granularity prediction as an example, but this granularity does not constitute a limitation on the scope of protection of this invention.

[0079] The collected express logistics business data undergoes missing value imputation, outlier correction, duplicate data removal, and normalization. Missing values ​​can be handled using forward imputation, mean imputation, or mean imputation within the same period; outliers can be corrected based on business rules, box plot thresholds, or statistical quantile thresholds; normalization can employ max-min normalization or standardization methods. The processed historical input sequence is denoted as: Historical business data is collected from the express delivery and logistics system. This historical business data includes, but is not limited to, order volume, pickup volume, delivery volume, inbound volume, outbound volume, holiday markers, and weather information. The data is then aligned according to a fixed time granularity, which is one day or other time intervals set according to actual business needs. Missing value imputation, outlier handling, duplicate data removal, and normalization are performed on the collected data to obtain a time series of express delivery and logistics business volume.

[0080] Let the processed historical input sequence be: in, Indicates batch size, Indicates the length of a historical time step. Indicates the number of variables.

[0081] Predicting the target is the future. The sequence of express delivery volume at each time step: in, Indicates the predicted length. This indicates the number of target business metrics that need to be predicted. Depending on actual business needs, the prediction targets can be changes in business volume such as order volume, pickup volume, and delivery volume over the next few days.

[0082] The input for this step is the original express logistics business data, and the output is the preprocessed historical input sequence. and corresponding prediction targets .

[0083] In order to transform long-duration express logistics sequences into local time structures suitable for subsequent pulse coding and frequency domain analysis, this step processes the input sequence along the time dimension. Divide the time into segments.

[0084] Let the length of each time segment be... The number of time segments is: in, This indicates the number of time segments obtained from the division. When Cannot be When divisible, the length can be kept consistent by padding with zeros or truncating.

[0085] Time segmentation operation , input sequence Divided into multiple consecutive time segments: in, This represents the set of local time segments obtained by dividing the original express logistics sequence. Then, linear mapping and normalization are performed on the divided time segments to obtain the local embedding representation: in, This indicates a trainable linear mapping layer. This indicates a normalization operation. This represents the potential representation corresponding to a segment of the original express delivery and logistics volume. The dimensions of this representation can be expressed as: ,in This represents the hidden feature dimension obtained after linear mapping of each time segment.

[0086] The input for this step is the preprocessed historical input sequence. The output is the original time segment embedding representation. The output will be fed into the post-dual-path pulse encoder module, such as... Figure 2 As shown in the blue area on the left, this module includes a raw sequence state encoding branch and a temporal difference dynamic encoding branch, which are used to process the steady-state information and local dynamic change information in the express logistics business volume, respectively, and obtain the fused pulse representation through learnable fusion parameters.

[0087] The original sequence state encoding branch is embedded in the original time segment obtained in step two. As input, membrane potential accumulation and pulse firing are performed through LIF spiking neurons, thereby preserving continuous state information and basic trends in express logistics business volume. Let the spiking neuron at simulation time step... The input is The membrane potential is The membrane potential update and pulse firing process of LIF neurons are as follows: in, Indicates the membrane potential leakage factor. Indicates the pulse firing threshold. Indicates the resting potential. This represents the intermediate state after the current input and the accumulated membrane potential from the previous moment. This represents the pulse output at the current simulation time step. The step function is also included. Defined as: When the membrane potential exceeds a threshold, the neuron fires a pulse; otherwise, it remains silent. Through this membrane potential accumulation process, the original sequence state encoding branch can retain continuous state information in the express delivery logistics business volume.

[0088] The output of the original sequence state coding branch is: , This indicates the number of simulation time steps within the spiking neural network. It should be noted that... It is not equivalent to the actual timeline of the original express delivery logistics sequence, but rather represents the simulated number of steps in which the spiking neuron accumulates membrane potential internally.

[0089] The temporal difference dynamic coding branch is used to capture local variations in express delivery volume along the original real timeline. This branch first calculates the first-order difference along the original time dimension: in, Indicates the first The observed value of express delivery volume at a real time step. This represents the change between adjacent time steps. The difference value of the first time step is zero-padded to keep the sequence length constant.

[0090] The difference sequence is obtained by combining the difference values ​​of all time steps: Then, the difference sequence is divided into time segments, linearly mapped, and normalized: in, A potential representation of a difference time segment.

[0091] To avoid the loss of local dynamic information caused by hard pulse discretization, this branch uses a smooth threshold function to generate a continuous dynamic activation representation: in, This represents the Sigmoid function. The dynamic threshold representing the differential branch. Indicates temperature parameter, This represents a continuous local dynamic activation representation.

[0092] Unlike the hard impulses generated by the original sequence state coding branch, the temporal difference dynamic coding branch outputs continuous dynamic activations, which can more finely preserve local variations in express logistics volume. For example, the rapid increase in order volume caused by e-commerce promotional activities such as "618" and "Double Eleven", the decrease in delivery volume caused by the Spring Festival holiday, and short-term business fluctuations caused by extreme weather can all be enhanced by this branch.

[0093] To simultaneously preserve both stable state information and local dynamic change information, this invention introduces learnable fusion parameters. For state-preserving pulse representation and local dynamic activation representation To merge: in, These are learnable parameters used to adaptively balance the contributions of the original sequence state coding branch and the temporal difference dynamic coding branch. The fused output is denoted as: in, This output serves as the initial input for the subsequent frequency domain adaptive gating module and multidimensional dependency modeling module.

[0094] The input for this step is the original time segment embedding representation. and differential time segment embedding representation The output is represented as a fused pulse. .

[0095] Frequency domain adaptive gating module, such as Figure 3 As shown, this module maps the fused pulse representation to the frequency domain, extracts the periodic structure and frequency domain variation information from the express logistics business volume, and generates a frequency-aware gating signal to guide subsequent local dynamic response modulation.

[0096] Let the first The input to the multidimensional dependency modeling module is represented as follows: in Indicates the layer number. When hour, That is, the fused pulse representation output by the dual-path pulse coding module.

[0097] The frequency domain adaptive gating module first moves along the time segment dimension Input representation Performing a real-number fast Fourier transform yields the spectral representation: in, This represents the frequency domain transformation performed along the time segment dimension. Represents the Fast Fourier Transform of real numbers. It represents the frequency domain characteristics.

[0098] In the express delivery and logistics scenario, different frequency components correspond to different business patterns. For example, the difference in business volume between Monday to Friday and the weekend can be reflected as a weekly frequency component.

[0099] In order to generate adaptive modulation weights based on the frequency domain structure of the current input sequence, this invention maps the input representation to frequency-selective features and generates sparse frequency-aware weights through spiking neurons.

[0100] Specifically, the input representation for each simulation time step Learnable frequency projection and normalization are performed, and then sequentially input into generative LIF neurons and aggregated LIF neurons: in, The projection parameters from the learnable time segment to the frequency space are represented. This represents a spiking neuron used to generate sparse frequency-selective features. This represents a spiking neuron used to aggregate the frequency selection results of multiple simulation time steps. This represents the frequency-aware selection weight.

[0101] Frequency-aware selection weights Action on frequency domain representation The modulated frequency domain features are obtained, and then reconstructed back to the time segment domain through inverse Fourier transform: Among them, among them, Frequency dimension and The frequency bins must be consistent to ensure that the dimensions of subsequent element-wise multiplications match. This represents element-wise multiplication. Indicates the inverse Fourier transform. This represents the frequency-modulated time segment representation. This representation retains the original time segment structure while incorporating the frequency domain context of the express delivery logistics sequence.

[0102] To further transform the frequency domain context into a gated signal that can be used for dynamic response modulation, this invention employs frequency-aware weighting. Standardization and average pooling are performed along the frequency bin dimension: in, This represents the average pooling operation along the frequency bin dimension. and These represent the mean and variance along the frequency dimension, respectively. To prevent extremely small positive numbers with a denominator of zero, This represents a frequency-aware gating signal. This gating signal can be broadcast to subsequent dynamic response modulation processes in the time segment dimension.

[0103] The input for this step is the first... Layer input representation The output is a frequency-modulated time segment representation. and frequency-sensing gating signals .

[0104] This step explains the internal calculations of the frequency-guided dynamic unit, which serves as a fundamental component in the frequency-domain adaptive gating module and the multidimensional dependency modeling module for dynamic response modulation. Frequency-guided dynamic response modulation is used to sense and gate signals based on frequency. The adjustment pulse represents the dynamic change along the time step dimension of the pulse simulation. This step corresponds to... Figure 3 The structure of the Frequency Guided Dynamic Unit (FGD Unit) in the text.

[0105] It should be noted that the dynamic difference in this step differs from the time difference in step three. The time difference in step three is calculated along the true timeline of the original express logistics sequence to capture changes between adjacent dates, adjacent hours, or adjacent business cycles; the dynamic difference in this step follows the time step dimension of the spiking neural network simulation. The calculation is used to characterize the internal evolution of the pulse state during multiple membrane potential accumulation and discharge processes.

[0106] For the Layer input representation Its difference along the pulse simulation time step dimension is defined as: in, Indicates the first The first in the layer Representation of each simulation time step This represents the internal state changes between adjacent simulation time steps.

[0107] Using frequency sensing gating signals Modulate the above simulation time step differential: in, For learnable gain parameters, via broadcast Dimensional alignment, This indicates the dynamic changes in the simulation time step after frequency guidance.

[0108] Based on the above-mentioned dynamic changes in the simulated time step after modulation, the output of the frequency-guided dynamic unit is denoted as: .in, This represents the dynamic impulse response modulated using a frequency-sensing gated signal.

[0109] The input for this step is the first... Layer representation and frequency-sensing gating signals The output is a frequency-guided dynamic response. .

[0110] Multidimensional dependency modeling module, such as Figure 4As shown, this module enhances long-term prediction capabilities in three dimensions: First, the frequency-guided branch preserves the frequency modulation representation generated by the frequency-domain adaptive gating module; second, the cross-time-segment dependency branch models long-range temporal relationships between different time segments; and third, the pulse simulation time-step dependency branch models the state evolution relationships of the spiking neural network at the pulse simulation time step.

[0111] Let the first Layer input is The multidimensional dependency modeling module sequentially generates frequency-guided representations. Time segment representation Simulation time step representation The next layer representation is obtained by fusing the data using a residual method. .

[0112] Frequency-guided branch in frequency-modulated representation and frequency-sensing gating signals As input, the fundamental impulse response and the frequency-guided dynamic response are fused: in, For learnable fusion parameters, This represents the output obtained by performing basic LIF pulse processing on the frequency modulation representation. This represents the dynamic response output guided by the frequency-sensing gating signal.

[0113] Output of frequency-guided branch This output is used to preserve frequency domain periodicity information and event-driven change information in the express logistics business volume sequence, such as weekly cycle, holiday cycle, promotional activity cycle, and frequency domain structure changes caused by sudden events.

[0114] Cross-time-segment dependency branching is used to model long-range temporal relationships between different time segments. For express delivery volume forecasting, there are usually clear connections between multiple consecutive time segments. For example, promotional activities and holidays form consecutive phases. This invention utilizes the time-segment dimension... Perform convolution processing on the input representation: in, This represents a convolution operation performed along the time segment dimension. This represents the representation after processing across time segments.

[0115] Then The input consists of a parallel fusion structure composed of basic LIF branches and frequency-guided dynamic units: Output across time-segment dependent branches This branch can capture the continuous evolution of express logistics volume across multiple time segments, enhancing the model's ability to model long-term trends, changes before and after activities, and cross-cycle recurring patterns.

[0116] The time-step dependent branch of the spiking simulation is used to model the state evolution along the simulation time-step dimension within the spiking neural network. Since spiking neurons accumulate membrane potential, leak, and fire pulses over multiple simulation time steps, the simulation time-step dimension can reflect the dynamic relationship of the internal spiking state as iterates.

[0117] like Figure 4 As shown, this branch first adjusts the dimensions of the input representation so that the simulation time step dimension can serve as the primary modeling dimension. Then, convolution processing is performed along the simulation time step dimension: in, This represents a convolution operation performed along the time step dimension of the pulse simulation. This represents the simulation time-step dependent representation after processing.

[0118] Subsequently, fusion is performed through the basic LIF branch and frequency-guided dynamic unit: Output of time-step dependent branches in pulse simulation This branch can enhance the model's ability to express the evolution of the internal state of a pulse, thereby alleviating the problem of insufficient information retention caused by membrane potential decay in long sequence prediction.

[0119] The outputs of the above three branches are residually fused with the input representation of the current layer to obtain the enhanced representation of the next layer: in, By using residual summation instead of feature splicing or complex attention weighting, the sparsity of impulse representation can be maintained while reducing additional dense computational overhead, making the model more suitable for resource-constrained scenarios such as express delivery outlets, transfer centers or edge devices.

[0120] The above multidimensional dependency modeling process can be executed in a stacked manner. Layer. After After layer processing, the final enhanced representation is obtained: This step is represented by a fusion pulse. As the initial input, and combined with the corresponding frequency modulation representation at each layer. and frequency-sensing gating signals The process involves progressive residual augmentation, ultimately outputting a multidimensional dependency augmentation representation. .

[0121] The representation output by the multidimensional dependency modeling module Input into the prediction module to obtain the future. The forecast results of express delivery and logistics volume at each time step.

[0122] First, the simulation time step dimension can be adjusted. Perform average pooling or weighted aggregation to obtain the compressed time segment representation: in, This represents a pooling operation along the simulation time step dimension. This represents the aggregated pulse timing representation.

[0123] Then, Flattened or rearranged before being input into the linear prediction layer: in, and These represent the learnable weights and biases of the prediction layer, respectively. This represents the model's output prediction of future express delivery and logistics volume. Indicates the prediction step size. This indicates the target quantity to be predicted. For example, when the prediction task is to forecast the order volume of a certain region over the next 7 days, .

[0124] The input for this step is a multidimensional dependency augmentation representation. The output is the forecast of future business volume. .

[0125] During the model training phase, the prediction results will be... Compared with the actual business volume sequence By comparing the results, a prediction loss function can be constructed. Mean squared error can be used as the training objective. in, Indicates the first The sample in the future The time step, the first The true value of each target indicator This represents the corresponding predicted value.

[0126] The mean absolute error can also be used as an auxiliary evaluation indicator according to business needs: All trainable parameters in the model are updated using backpropagation and gradient optimization methods.

[0127] The predicted future express delivery volume output by the prediction module is applied to a smart logistics business system. This can be used for branch staff scheduling, vehicle and trunk line transportation dispatching, sorting and warehousing resource allocation, capacity warnings during promotional activities, and identification and emergency dispatching of abnormal fluctuations during special circumstances such as the Spring Festival holiday, extreme weather, or regional control measures. Through the above steps, this invention can simultaneously preserve the stable temporal state of the express delivery volume sequence, enhance the perception of local mutations, introduce frequency domain periodic information, and model cross-time segment and pulse simulation time step dependencies, thereby improving prediction accuracy, robustness, and computational efficiency.

Claims

1. A dual-path pulse coding and frequency domain sensing prediction method for express delivery volume, characterized in that, Includes the following steps: Obtain time series data of express logistics business volume, preprocess the data, divide the preprocessed data into multiple time segments along the time dimension, perform linear mapping and normalization processing on the time segments, and obtain the time segment embedded representation; The time segment embedding representation is input into a dual-path pulse coding module, which includes an original sequence state coding branch and a temporal difference dynamic coding branch. The original sequence state coding branch accumulates membrane potentials and fires pulses on the time segment embedding representation through LIF spiking neurons, outputting a state-preserving pulse representation. The temporal difference dynamic coding branch calculates the first-order difference along the original time axis, performs time segmentation, linear mapping, and normalization on the difference sequence, and generates a continuous local dynamic activation representation through a smoothing threshold function. The state-preserving pulse representation and the local dynamic activation representation are fused using learnable fusion parameters to obtain a fused pulse representation. The fused pulse representation is input into the frequency domain adaptive gating module. The fused pulse representation is transformed in the frequency domain along the time segment dimension to obtain the spectrum representation. Frequency-aware selection weights are generated based on the fused pulse representation. The frequency-aware selection weights are applied to the spectrum representation and then transformed in the inverse frequency domain to obtain the frequency-modulated time segment representation. A frequency-aware gating signal is generated based on the frequency-aware selection weights. Based on the frequency-aware gating signal, frequency-guided dynamic response modulation is performed on the dynamic changes of the pulse representation in the pulse simulation time step dimension. The frequency-modulated representation is input into the multidimensional dependency modeling module, and modeling is performed through frequency-guided branch, cross-time segment dependency branch and pulse simulation time step dependency branch respectively. The enhanced representation is obtained through residual fusion. The enhanced representation is input into the prediction layer, which outputs the prediction results of express logistics business volume for multiple future time steps.

2. The dual-path pulse coding and frequency domain sensing prediction method for express delivery volume according to claim 1, characterized in that, In the original sequence state encoding branch, the membrane potential update and pulse firing process of the LIF spiking neuron is as follows: Where τ represents the membrane potential leakage factor, Vth represents the pulse firing threshold, Vrest represents the resting potential, U[t] represents the intermediate state after the current input and the membrane potential accumulated at the previous time step, S[t] represents the pulse output at the current simulation time step, and H(∙) is the step function.

3. The dual-path pulse coding and frequency domain sensing prediction method for express delivery volume according to claim 1, characterized in that, In the aforementioned time-difference dynamic coding branch, the first-order difference is calculated as follows: in, This represents the observed value of express delivery volume at the t-th real time step. This represents the change between adjacent time steps. The difference value of the first time step is zero-padded to maintain the sequence length. Let L represent the observed value of express delivery volume at the t-th real time step, and L represent the length of the historical time step. The smoothing threshold function is: in, This represents the Sigmoid function. The dynamic threshold representing the differential branch. Indicates temperature parameter, A potential representation of a difference time segment.

4. The dual-path pulse coding and frequency domain sensing prediction method for express delivery volume according to claim 1, characterized in that, In the dual-path pulse coding module, the learnable fusion parameters The fusion method is as follows: in, This indicates a state-holding pulse representation. This represents local dynamic activation.

5. The dual-path pulse coding and frequency domain sensing prediction method for express delivery volume according to claim 1, characterized in that, In the frequency domain adaptive gating module, the frequency domain transformation adopts real number fast Fourier transform; the frequency-aware selection weights are generated by performing learnable frequency projection and normalization on the input representation of each simulation time step, sequentially inputting the generative LIF spiking neuron and the aggregated LIF spiking neuron, and outputting the frequency-aware selection weights.

6. The dual-path pulse coding and frequency domain sensing prediction method for express delivery volume according to claim 1, characterized in that, The frequency-aware gating signal is generated by standardizing and averaging the frequency-aware selection weights along the frequency dimension to obtain the frequency-aware gating signal G.

7. The dual-path pulse coding and frequency domain sensing prediction method for express delivery volume according to claim 1, characterized in that, In the frequency-guided dynamic response modulation, the difference is calculated along the pulse simulation time step dimension and modulated using the frequency-sensing gating signal: ； Using frequency sensing gating signals Modulate the above simulation time step differential: in, For learnable gain parameters, via broadcast Dimensional alignment, ⊙ represents the dynamic change of the simulation time step after frequency guidance, ⊙ represents element-wise multiplication, and G is the frequency-sensing gating signal.

8. The dual-path pulse coding and frequency domain sensing prediction method for express delivery volume according to claim 1, characterized in that, In the multidimensional dependency modeling module: The frequency-guided branch takes the frequency-modulated representation and the frequency-aware gating signal as inputs, and fuses the basic LIF pulse processing output with the frequency-guided dynamic response output. The cross-time segment dependent branch performs convolution processing on the input representation along the time segment dimension, and then outputs the result after fusion with the basic LIF branch and the frequency-guided dynamic unit. The pulse simulation time step dependent branch performs convolution processing on the input representation along the pulse simulation time step dimension, and then outputs it after being fused by the basic LIF branch and the frequency-guided dynamic unit. The outputs of the three branches are summed and fused with the input representation of the current layer to obtain the enhanced representation of the next layer.

9. The dual-path pulse coding and frequency domain sensing prediction method for express delivery volume according to claim 1, characterized in that, The multidimensional dependency modeling module is stacked and executed in N layers, and the final enhanced representation is obtained after N layers of processing; the prediction layer performs pooling aggregation on the simulation time step dimension, and outputs the prediction results of express logistics business volume for the next H time steps through linear mapping.

10. The express delivery volume dual-path pulse coding and frequency domain sensing prediction method according to any one of claims 1-9, characterized in that, The express logistics volume includes at least one of the following: order volume, pickup volume, delivery volume, inbound volume, and outbound volume; the prediction results are used for branch staff scheduling, vehicle dispatching, warehousing resource allocation, trunk transportation arrangements, or abnormal risk warnings.

Images