Container generation quantity prediction method and device based on multi-source service data, equipment and medium

Abstract

The application discloses a container generation quantity prediction method and device based on multi-source business data, equipment and medium, relates to the technical field of logistics supply chain management, and comprises the following steps: obtaining a container generation quantity sequence by performing abnormal elimination and unit conversion on multi-source container business data of highways and railways; obtaining each intrinsic mode component and residual component with clear physical meaning by adding paired positive and negative white noise for multiple empirical mode decompositions and integrating the average; combining external driving factors to screen and construct a multi-dimensional feature tensor, inputting a prediction model of an encoder containing a probabilistic sparse attention layer and a distillation layer and a generative decoder to obtain each component prediction result, and finally superimposing reconstruction and outputting a prediction sequence after inverse normalization. The application eliminates modal aliasing interference, reduces the calculation complexity of a long sequence, realizes high-precision long-sequence prediction of container generation quantity, and provides data support for logistics planning.

Description

Technical Field

[0001] This invention relates to the field of logistics supply chain management technology, and in particular to a method, apparatus, equipment and medium for predicting container production volume based on multi-source business data. Background Technology

[0002] Currently, container volume forecasting mainly employs the following technical means: at the data level, it relies on macroeconomic statistical data and combines it with transportation business data such as highway toll revenue and railway loading and unloading ledgers for analysis; at the feature mining level, it uses empirical mode decomposition methods to decompose non-stationary volume sequences into several intrinsic mode functions to mine the intrinsic features at different time scales; at the predictive modeling level, it uses autoregressive integral moving average models, convolutional neural networks, long short-term memory networks, or transformer models based on self-attention mechanisms to model and predict volume sequences.

[0003] However, existing practices have the following drawbacks: First, the statistical standards are inconsistent and the spatial granularity is coarse. The data systems for highways and railways are isolated from each other and it is difficult to integrate and sink them down to the county-level administrative units. Second, the empirical mode decomposition process is often accompanied by mode aliasing, and the physical meaning of the decomposed components is ambiguous, which interferes with the prediction accuracy. Third, traditional linear models are difficult to handle non-stationary sequences. Recurrent neural networks are prone to gradient vanishing or exploding in long sequence prediction. The self-attention computational complexity of traditional transformer models increases quadratically with the sequence length, resulting in huge memory consumption, which is difficult to meet the needs of real-time prediction of long sequences.

[0004] Therefore, how to eliminate mode aliasing interference in the decomposition of non-stationary freight volume sequences under the influence of multiple coupled factors, and achieve high-precision prediction of long-sequence container generation with low computational complexity, has become an urgent problem to be solved. Summary of the Invention

[0005] The main objective of this application is to provide a method, apparatus, equipment, and medium for predicting container production volume based on multi-source business data, aiming to solve the technical problem of how to achieve high-precision prediction of long-sequence container production volume with low computational complexity.

[0006] To achieve the above objectives, this application proposes a method for predicting container generation volume based on multi-source business data, including: Acquire multi-source container business data, and remove abnormal data and unify the units of the multi-source container business data to obtain a container generation sequence, wherein the multi-source container business data includes highway toll flow data and railway loading and unloading data. Pairs of positive and negative white noise are added to the container generation sequence to construct multiple auxiliary sequences; Each of the auxiliary sequences is subjected to empirical mode decomposition, and the corresponding components obtained from multiple decompositions are integrated and averaged to obtain multiple intrinsic mode components and residual components. Obtain external driving factor data, and perform driving feature screening and standardization processing on each intrinsic mode component based on the external driving factor data to obtain the multidimensional feature tensor corresponding to each intrinsic mode component; Each of the multidimensional feature tensors is input into the probabilistic sparse attention prediction model for processing to obtain the subsequence prediction results corresponding to each intrinsic mode component and the residual prediction results corresponding to the residual component. The probabilistic sparse attention prediction model includes an encoder and a generative decoder. The encoder includes a preset encoder layer and several processing units connected in sequence. Each processing unit includes a probabilistic sparse attention layer and a distillation layer. The probabilistic sparse attention layer is used to filter query vectors whose attention contribution satisfies the preset contribution condition for attention calculation. The prediction results of each subsequence and the prediction results of the residuals are superimposed and reconstructed, and the superimposed and reconstructed sequence is denormalized to obtain the container generation prediction sequence.

[0007] Furthermore, to achieve the above objectives, this application also proposes a container production volume prediction device based on multi-source business data, the container production volume prediction device based on multi-source business data comprising: The data acquisition module is used to acquire multi-source container business data, and to remove abnormal data and unify the units of the multi-source container business data to obtain a container generation sequence. The multi-source container business data includes highway toll data and railway loading and unloading data. A noise addition module is used to add pairs of positive and negative white noise to the container generation sequence to construct multiple auxiliary sequences; The sequence decomposition module is used to perform empirical mode decomposition on each of the auxiliary sequences, and to integrate and average the corresponding components obtained from multiple decompositions to obtain multiple intrinsic mode components and residual components. The feature construction module is used to acquire external driving factor data, and perform driving feature screening and standardization processing on each intrinsic mode component based on the external driving factor data to obtain the multidimensional feature tensor corresponding to each intrinsic mode component. The model prediction module is used to input the multidimensional feature tensors into the probabilistic sparse attention prediction model for processing, and to obtain the subsequence prediction results corresponding to each intrinsic mode component and the residual prediction results corresponding to the residual components. The probabilistic sparse attention prediction model includes an encoder and a generative decoder. The encoder includes a preset encoder layer and several processing units connected in sequence. The processing unit includes a probabilistic sparse attention layer and a distillation layer. The probabilistic sparse attention layer is used to filter query vectors whose attention contribution satisfies the preset contribution condition for attention calculation. The reconstruction output module is used to superimpose and reconstruct the prediction results of each subsequence and the prediction results of the residuals, and to perform inverse standardization on the superimposed and reconstructed sequence to obtain the container generation prediction sequence.

[0008] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the container production prediction method based on multi-source business data as described above.

[0009] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the container production volume prediction method based on multi-source business data as described above.

[0010] This application obtains a container generation sequence by anomaly removal and unit standardization of multi-source container business data from highways and railways. It then performs multiple empirical mode decompositions (EMDs) by adding paired positive and negative white noise and integrating the results to obtain physically meaningful intrinsic mode components and residual components. A multidimensional feature tensor is constructed by selecting external driving factors and input into a prediction model containing a probabilistic sparse attention layer and a distillation layer, along with a generative decoder, to obtain prediction results for each component. Finally, the predicted sequence is superimposed, reconstructed, and denormalized to output the final prediction sequence. This application eliminates mode aliasing interference, reduces the computational complexity of long sequences, and achieves high-precision long-sequence prediction of container generation, providing data support for logistics planning. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a flowchart illustrating the first embodiment of the container generation prediction method based on multi-source business data in this application; Figure 2 This is a flowchart illustrating the second embodiment of the container generation prediction method based on multi-source business data in this application; Figure 3 This is a schematic diagram of the module structure of the container generation prediction device based on multi-source business data in this application. Figure 4 This is a schematic diagram of the equipment structure of the hardware operating environment involved in the container generation prediction method based on multi-source business data in the embodiments of this application.

[0013] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0014] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0015] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0016] Current methods for predicting container generation primarily rely on macroeconomic statistics and road and rail transport data. They employ Empirical Mode Decomposition (EMD) to mine intrinsic features of the sequence and utilize autoregressive integral moving average (IMM), long short-term memory (LSTM) networks, or transformer models for modeling and prediction. However, these existing technologies have significant drawbacks: inconsistent statistical standards and coarse spatial granularity among multi-source data make it difficult to integrate and scale down to the county-level administrative unit; EMD is prone to mode aliasing, leading to ambiguity in the physical meaning of the decomposed components; traditional linear models struggle to handle non-stationary sequences, recurrent neural networks are susceptible to gradient vanishing or exploding, and the computational complexity of transformer models' self-attention increases quadratically with sequence length, failing to meet the real-time prediction requirements for long sequences. These combined shortcomings make it difficult for current technologies to achieve high-precision prediction of long-sequence container generation in non-stationary, multi-factor coupled scenarios.

[0017] Based on the above, this application also provides a method for predicting container production volume based on multi-source business data, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the container generation prediction method based on multi-source business data in this application.

[0018] In this embodiment, the container generation prediction method based on multi-source business data includes steps S10 to S60: Step S10: Obtain multi-source container business data, and remove abnormal data and unify the units of the multi-source container business data to obtain the container generation sequence.

[0019] It should be noted that multi-source container business data refers to raw data from different transportation business systems that record container transportation activities. Multi-source container business data includes highway toll data and railway loading and unloading data. Highway toll data refers to vehicle passage data recorded by the highway toll system, which may include license plate number, vehicle type, total vehicle and cargo weight, entry station, exit station, and passage time. Container generation volume refers to the volume of containerized cargo formed by goods shipped out due to production, trade, and other activities, usually measured in 20-foot equivalent units. A container generation volume sequence refers to a time series consisting of container generation volumes for each statistical period arranged in chronological order.

[0020] Further, step S10 includes: First, collecting multi-source container business data. Specifically, accessing the highway toll system to extract the passage flow data for toll-charging vehicle types of two-axle and six-axle container trucks, and simultaneously accessing the railway freight system to obtain container loading and unloading ledgers for each freight station, summarizing to form multi-source container business data, and accessing the port scheduling system to obtain monthly throughput data for major ports. Among these, highway toll flow data and railway loading and unloading data serve as the basis for calculating container generation volume. Port throughput data does not participate in the merging calculation of the generation volume sequence, but is used as independent external reference data to verify the subsequently obtained container generation volume sequence. This is because highway and railway data record the details of container transportation for each shipment, which can directly reflect the generation of goods, while port throughput data is the loading and unloading statistics of the port link, and has a transit relationship with the generation volume in the hinterland. It is suitable as a verification benchmark rather than a calculation source. Highway toll flow data includes toll-charging vehicle type, total vehicle and cargo weight, entry station, exit station, and passage time. Secondly, abnormal data in the highway toll data flow is removed, such as the same entrance and exit station, or the same vehicle appearing at two toll stations with a distance greater than a preset distance threshold within a preset travel time, to obtain the first filtered data. Specifically, logical anomaly removal is performed on the highway toll data flow: the passage records of the same vehicle are collected according to the license plate number, and data with the same entrance and exit station in the same record is removed; the passage records of the same vehicle are compared between two adjacent passage records, and if the time interval between the two passage records is less than the preset travel time, and the distance between the two toll stations is greater than the preset distance threshold, it is determined that the vehicle has experienced a physically impossible "teleportation", and the corresponding passage data is removed, with the retained data constituting the first filtered data.

[0021] Next, the weight of the first screening data is verified. The total weight of the vehicle and cargo is compared with the empty vehicle weight and the preset legal weight limit. Records where the total weight of the vehicle and cargo is less than the empty vehicle weight indicate an error in the weighing process. Records where the total weight of the vehicle and cargo is greater than the preset legal weight limit do not comply with the road transport regulations. Data in the first screening data that are either less than the empty vehicle weight or greater than the preset legal weight limit are removed to obtain the second screening data.

[0022] Then, the total vehicle and cargo weight in the second-filtered data is converted into the number of standard containers (TCLs) according to a preset mass conversion benchmark. Non-integer TCLs are rounded up to obtain the highway TCL data. Specifically, the second-filtered data undergoes a unit conversion; the total vehicle and cargo weight of each record is divided by the preset mass conversion benchmark to obtain the number of TCLs transported by that vehicle. The specific formula is as follows: in This indicates the number of TEUs (twenty-foot equivalent units) transported by the vehicle; This indicates the rounding up operation; This indicates the total weight of the vehicle and cargo in this charge transaction record; This indicates the vehicle's unloaded weight. This represents the preset mass conversion benchmark, i.e., the average cargo weight corresponding to one standard container (TCV), expressed in kilograms. This value is determined based on statistics of the average cargo weight in container transportation. Since in actual logistics operations, containers that are not fully loaded are still counted as one TCV, the calculated non-integer TCV counts are rounded up to obtain the TCV data for road container transport.

[0023] Finally, the highway container TEU data and railway loading and unloading data are spatiotemporally merged according to timestamps and geographical units to obtain a container generation sequence. Specifically, the highway container TEU data and railway loading and unloading data are aligned to a unified statistical period according to timestamps, and the two data streams are assigned to corresponding geographical units according to administrative division codes. After merging the two data streams, a container generation sequence covering all administrative units in the target area is formed. Then, port throughput data corresponding to the same statistical period and the same geographical unit is retrieved, and the deviation rate between the container generation sequence and the port throughput data is calculated. If the deviation rate is greater than a preset verification threshold, such as 30%, the highway and railway data of that geographical unit are backtracked to check whether there are any omissions or duplicate statistics. If the deviation rate is within the preset verification threshold, the container generation sequence is determined to have passed the verification. The value of the preset verification threshold takes into account the normal difference between the transshipment container volume included in port throughput and the hinterland generation volume.

[0024] Step S20: Add pairs of positive and negative white noise to the container generation sequence to construct multiple auxiliary sequences.

[0025] It should be noted that positive white noise refers to a random noise sequence that follows a Gaussian distribution, while negative white noise refers to a noise sequence with the same amplitude but opposite sign as positive white noise; the two constitute a pair of complementary noises. The auxiliary sequence refers to the sequence to be decomposed after adding noise. The number of integration iterations refers to the number of rounds of noise addition and decomposition operations, and its value can range from 50 to 200. A higher value results in more thorough noise cancellation, but also increases computational overhead; a balance should be struck within this range based on the data size.

[0026] Further, step S20 includes: First, setting the number of integrations and the noise amplitude, wherein the number of integrations is the number of construction rounds of the auxiliary sequence. The number of integrations can be 100. Generally, when the number of integrations reaches 50 or more, the noise residue can be reduced to a low level. Further increasing the number of integrations has limited improvement on the decomposition results and will instead linearly increase the amount of computation.

[0027] Secondly, in each round of construction, original white noise following a Gaussian distribution is generated, and complementary white noise with the same amplitude but opposite sign as the original white noise is generated. Specifically, in each round of construction, random sampling is performed according to a Gaussian distribution to generate original white noise of the same length as the container generation sequence, and then the original white noise is inverted point by point to obtain complementary white noise.

[0028] Then, the original white noise is added to the container generation sequence to obtain the positive noise sequence; complementary white noise is added to the container generation sequence to obtain the negative noise sequence. The specific formula is as follows: in Indicates a positive noise sequence; This represents a sequence of container production volumes; This represents the original white noise generated in the i-th round; Indicates a negative noise sequence; This represents the complementary white noise generated in the i-th round.

[0029] Finally, the positive and negative noise sequences obtained from each round of construction are summarized to obtain multiple auxiliary sequences. Specifically, the above construction process is repeated until the integration number is reached, and the positive and negative noise sequences obtained from each round are summarized to form multiple auxiliary sequences. For example, when the integration number is 100, a total of 200 auxiliary sequences are obtained.

[0030] Step S30: Perform empirical mode decomposition on each auxiliary sequence, and integrate and average the corresponding components obtained from multiple decompositions to obtain multiple intrinsic mode components and residual components.

[0031] It should be noted that Empirical Mode Decomposition (EMD) refers to a method that decomposes a signal into several intrinsic mode functions (EMFs) based on its own time-scale characteristics by extracting the extremum envelope and successively sifting through it. EEMs are the components obtained through decomposition that satisfy the condition that the number of extrema is roughly equal to the number of zero-crossings and that the local mean is zero; they are simply called components, and their frequencies are arranged from high to low. Residual components are the parts remaining after stripping all components from the original sequence, representing the long-term evolution trend of the sequence.

[0032] Further, step S30 includes: First, performing empirical mode decomposition on each auxiliary sequence, extracting the extreme value envelopes of the auxiliary sequences and successively sieving them to obtain multiple initial mode components corresponding to each auxiliary sequence. It should be understood that the extreme value envelope refers to the upper and lower envelopes formed by connecting the local maxima and local minima of the sequence using cubic spline interpolation. Sieving refers to the iterative process of repeatedly subtracting the mean of the upper and lower envelopes from the sequence until the remaining part satisfies the intrinsic mode function conditions. The initial mode components refer to the components of each order obtained by performing one empirical mode decomposition on a single auxiliary sequence. Specifically, empirical mode decomposition is performed on each auxiliary sequence to extract all local maxima and local minima of the sequence. The upper and lower envelopes are constructed by cubic spline interpolation. The mean curves of the two envelopes are subtracted from the original sequence. The above envelope extraction and mean subtraction operations are repeated on the remaining part after subtraction until the stopping condition of the intrinsic mode function is met, and an initial mode component is obtained. This component is then subtracted from the sequence and decomposition is continued until the remaining part is monotonic or only one extreme point remains. Each auxiliary sequence is thus decomposed into multiple initial mode components.

[0033] Secondly, the initial modal components corresponding to each auxiliary sequence are aligned according to their order of frequency from high to low, resulting in modal component groups of each order. It's important to understand that a modal component group refers to a grouping of initial modal components of the same order from different auxiliary sequence decomposition results. Specifically, the initial modal components obtained from the decomposition of each auxiliary sequence are arranged according to their frequency from high to low, and initial modal components of the same order are grouped together to obtain modal component groups of each order. For example, the first-order components of all auxiliary sequences are grouped into the first group, the second-order components into the second group, and so on.

[0034] Next, the energy values ​​of each initial modal component in each modal component group are calculated. Abnormal initial modal components whose energy values ​​deviate from the average energy value of the modal component group by more than a preset energy deviation threshold are removed, resulting in effective component groups of each order. It is important to understand that the energy value refers to the sum of the squares of the amplitudes of each sampling point in the component, used to characterize the fluctuation intensity of the component. The preset energy deviation threshold is the relative deviation benchmark for judging abnormal components, which can be set to 0.5. That is, if the relative deviation of the energy value of an initial modal component from the average energy value of its modal component group exceeds 50%, it is removed. This value is determined with reference to the normal energy fluctuation range caused by noise in complementary set empirical mode decomposition, thus retaining normal fluctuation components while identifying abnormal components that fail to decompose. The effective component group refers to the remaining modal component group after removing abnormal initial modal components.

[0035] Then, the initial modal components in each effective component group are arithmetically averaged point by point to obtain the intrinsic modal components of each order. The specific formula for ensemble averaging is: in, Represents the j-th eigenmode component; This represents the j-th initial mode component obtained from the i-th round of positive noise sequence decomposition; This represents the j-th initial mode component obtained from the i-th round of negative noise sequence decomposition; Indicates the number of integrations. This represents the total number of initial modal components participating in the averaging, i.e., one component each from the positive and negative noise sequences in each round. Since the noise symbols added in each round are complementary, the noise is basically canceled out after averaging. The high-frequency intrinsic modal components correspond to short-term random fluctuations caused by holidays, while the low-frequency intrinsic modal components correspond to the long-term evolution trend of regional economic development.

[0036] Finally, the intrinsic modal components are extracted from the container generation sequence, and the remaining parts are identified as residual components.

[0037] Step S40: Obtain external driving factor data, and perform driving feature screening and standardization processing on each intrinsic mode component based on the external driving factor data to obtain the multidimensional feature tensor corresponding to each intrinsic mode component.

[0038] It should be noted that external driving factor data refers to the data on external factors affecting container generation, which may include indicators of transportation infrastructure, economic scale, and social structure. The multidimensional feature tensor is a matrix-like input data composed of historical container generation sequences and selected external driving factors concatenated according to time steps. Standardization refers to transforming the data to a uniform scale with a mean of 0 and a standard deviation of 1.

[0039] Further, step S40 includes: First, acquiring external driving factor data, which includes transportation infrastructure indicators, economic scale indicators, and social structure indicators. Transportation infrastructure indicators include the number of railway freight stations and highway length; economic scale indicators include regional GDP and total import and export value; and social structure indicators include year-end resident population, urbanization rate, and the proportion of secondary industry. These are summarized to constitute the external driving factor data.

[0040] Secondly, a spatial weight matrix is ​​constructed based on spatial adjacency relationships. The local spatial autocorrelation index of each administrative region is calculated using this matrix, and administrative regions whose local spatial autocorrelation indices meet preset clustering conditions are identified as spatially clustered regions. It's important to understand that the spatial weight matrix represents the adjacency relationships between administrative regions. Matrix elements can be determined based on the Queen's Adjacency Criterion: elements are set to 1 when two administrative regions share a boundary or vertex, and 0 otherwise. The local spatial autocorrelation index is a statistical measure of the degree of clustering or dispersion of container generation in an administrative region and its neighboring regions. A local Moran's index can be used; a positive and significant index indicates high or low clustering between the region and its neighboring regions. The preset clustering conditions are the criteria for identifying spatially clustered regions. These can be defined as a positive local Moran's index with a significance level less than 0.05, and the region's own generation exceeding the regional mean (i.e., high-high clustering). A significance level of 0.05 is a commonly used benchmark in spatial statistics. Spatially clustered regions are administrative regions that meet the preset clustering conditions; in this example, they represent regions with strong logistics radiation capabilities. Specifically, a spatial weight matrix covering all administrative regions is constructed based on the Queen Adjacency Criterion. The local Moran Index of each administrative region is calculated in combination with the spatial weight matrix. Administrative regions whose local Moran Index meets the preset agglomeration conditions are marked as high-high agglomeration regions, i.e. spatial agglomeration regions. These regions and their adjacent regions drive each other in terms of container generation and have stronger logistics radiation capabilities.

[0041] Next, the intrinsic modal components and external driving factor data are standardized to obtain standardized modal components and standardized driving factor data. Specifically, zero-mean standardization is performed on each intrinsic modal component and each external driving factor data to eliminate the weight imbalance caused by differences in dimensions and numerical magnitudes between indicators such as GDP and number of stations, thus obtaining standardized modal components and standardized driving factor data.

[0042] Next, regression equations are established between standardized modal components and standardized driving factor data. Redundant factors with regression coefficients lower than a preset significance threshold are removed to obtain key driving factor data. It's important to understand that the preset significance threshold is the benchmark for determining whether a regression coefficient is significant. A factor is considered insignificant and removed when the significance test probability value of the regression coefficient is greater than 0.05. 0.05 is a commonly used significance level in statistical tests. Specifically, for each standardized modal component, an ordinary least squares regression equation is established with it as the dependent variable and each standardized driving factor data as the independent variable. The regression coefficients of each factor are tested for significance, and redundant factors with regression coefficients lower than the preset significance threshold are removed. The retained factors constitute the key driving factor data. The key driving factors corresponding to different frequency components can be different; for example, high-frequency components may be mainly driven by short-term trade indicators, while low-frequency components may be mainly driven by economic scale indicators.

[0043] Then, the key driver factor data, as exogenous covariates, are horizontally concatenated with the historical sequence of container generation according to a preset input step size to obtain the initial feature tensor. It's important to understand that exogenous covariates refer to covariates that participate in the model's prediction as external inputs; in this example, they are the selected key driver factors. In this example, when the input step size is 14, each sample consists of the historical generation amounts from 14 consecutive time points and various key driver factors, forming the initial feature tensor.

[0044] Finally, spatial weight coefficients are assigned to the features of samples located in spatially clustered regions in the initial feature tensor, resulting in multidimensional feature tensors corresponding to each intrinsic modal component. It's important to understand that the spatial weight coefficient is an amplification factor assigned to the features of samples in spatially clustered regions. It can be between 1.2 and 1.5; values ​​that are too small will not have an amplifying effect, while values ​​that are too large will suppress features from other regions. Cross-validation can be used to select a value within this range. The initial feature tensor is the concatenated feature tensor without spatial weight correction. Specifically, the administrative region to which each sample in the initial feature tensor belongs is queried. The features of samples belonging to spatially clustered regions are multiplied by the spatial weight coefficient to amplify them, while the features of samples from other regions remain unchanged, resulting in multidimensional feature tensors corresponding to each intrinsic modal component.

[0045] Step S50: Input each multidimensional feature tensor into the probabilistic sparse attention prediction model for processing to obtain the subsequence prediction results corresponding to each intrinsic mode component and the residual prediction results corresponding to the residual component.

[0046] It should be noted that the probabilistic sparse attention prediction model refers to a long sequence prediction model based on an improved transformer architecture, comprising an encoder and a generative decoder. The encoder consists of a predetermined number of sequentially connected processing units, each including a probabilistic sparse attention layer and a distillation layer. The probabilistic sparse attention layer is an attention layer that selects query vectors whose attention contribution meets a predetermined contribution condition by measuring the difference between the query vector and the key vector. Attention contribution is a measure of the uniformity of the attention distribution between the query vector and the key vector set; the sharper the distribution, the higher the contribution. The distillation layer is a network layer that sequentially performs one-dimensional convolution, activation, and max pooling on the output features of the attention layer, compressing the time axis dimension layer by layer. The generative decoder is a decoder that takes a merged vector containing known historical segments and placeholders to be predicted as input and outputs the predicted values ​​for the entire prediction period in parallel through a single forward computation. The predetermined number of encoder layers refers to the number of processing units in the encoder, which can be set to 4, as this value strikes a good balance between feature extraction capability and computational cost. The hyperparameters of the probabilistic sparse attention prediction model were determined using a grid search method combined with empirical features of logistics operations. The input step size was set to 14, corresponding to a bi-weekly operating cycle, which matches the statistical rhythm of container shipping operations. The number of training iterations was set to 50, at which the model has basically converged. The batch size was set to 18, balancing training stability and memory usage. The number of heads in the multi-head attention mechanism was set to 8, enabling the model to capture different fluctuation patterns in the freight volume sequence in parallel from multiple subspaces. The model's hidden feature dimension was set to 512, providing sufficient expressive power for complex fluctuation patterns. The number of encoder and decoder layers was set to 4 each. After four distillations and compressions, the time axis dimension of the features was compressed to one-sixteenth of its original size, achieving a balance between feature extraction capability and memory usage. The learning rate was set to 0.0001, ensuring smooth parameter updates. The dropout probability was set to 0.1, meaning that 10% of the neuron connections were randomly dropped during training to alleviate overfitting.

[0047] Specifically, a probabilistic sparse attention prediction model is deployed for each intrinsic mode component and residual component. Each model is trained and predicts independently. During prediction, the multidimensional feature tensor is input into the encoder. Each processing unit in the encoder sequentially performs sparse attention calculation and distillation compression on the input features. Query vectors whose attention contribution meets the preset contribution conditions are selected to participate in the attention calculation. The time axis dimension of the features is compressed layer by layer to obtain distilled features with high information density. Then, the generative decoder outputs the predicted values ​​of each time node in the prediction period at once based on the distilled features. After the model outputs of each component are summarized, the subsequence prediction results corresponding to each intrinsic mode component and the residual prediction results corresponding to the residual component are obtained.

[0048] Step S60: Overlay and reconstruct the prediction results of each subsequence and the residual prediction results, and perform inverse standardization on the overlay and reconstructed sequence to obtain the container generation prediction sequence.

[0049] It should be noted that overlay reconstruction refers to the process of aligning the prediction results of each component according to timestamps and then aggregating them into a complete prediction sequence. Destandardization refers to the process of using the mean and standard deviation of the historical training set to inversely transform the standardized data to restore its true dimensions. The container generation prediction sequence refers to the prediction results of container generation in future periods, expressed in TEUs (twenty-foot equivalent units), with clear physical meaning.

[0050] Further, step S60 includes: First, aligning the prediction results of each subsequence with the residual prediction results according to timestamps to obtain an aligned prediction component set. It is important to understand that the aligned prediction component set refers to the set of components formed by corresponding each subsequence prediction result with the residual prediction results point-by-point according to timestamps, where each component in the set has a predicted value at the same time node. Specifically, aligning the prediction results of each subsequence with the residual prediction results according to timestamps, checking whether each component has a predicted value at each time node within the prediction period, and if a component lacks a predicted value at certain time nodes, then using linear interpolation of the predicted values ​​at adjacent time nodes to fill in the missing value, thus obtaining the aligned prediction component set.

[0051] Secondly, the prediction error of each subsequence prediction result within the historical validation period is calculated, and the reconstruction weight corresponding to each subsequence prediction result is determined based on the prediction error. The smaller the prediction error, the larger the reconstruction weight. It's important to understand that the historical validation period refers to a reserved period from historical data that is not used for model training but is used to evaluate the prediction accuracy of each component. This period can be the last 30 consecutive statistical periods of the historical data; this length covers more than one seasonal fluctuation cycle without excessively crowding out training data. The reconstruction weight refers to the weight coefficient corresponding to each component when weighted summing, and in this example, it is determined based on the prediction error of the component. Specifically, the prediction process of each component is replayed within the historical validation period, and the prediction result of each subsequence is compared point-by-point with the actual component value of that period. The prediction error of each subsequence prediction result is calculated, and the reconstruction weight is determined according to the principle that the smaller the error, the larger the weight. The specific formula is as follows: in Represents the reconstruction weight of the prediction result of the j-th subsequence; This represents the prediction error of the j-th subsequence prediction result within the historical verification period. The denominator is the sum of the reciprocals of the errors of all components, used to normalize each weight to a sum of 1. This indicates the number of subsequence prediction results; This represents the prediction error of the k-th subsequence prediction result within the historical verification period.

[0052] Next, the components in the aligned prediction component set are weighted and summed according to the reconstruction weights, and the values ​​less than zero in the weighted summation are set to zero to obtain the reconstructed prediction sequence. Specifically, the components in the aligned prediction component set are weighted and summed according to the reconstruction weights. Since high-frequency components may exhibit significant negative predictions at certain time points, the values ​​at certain time points after weighted summation may be less than zero. However, the container generation volume cannot physically be negative. Therefore, the values ​​less than zero in the weighted summation are set to zero to obtain the reconstructed prediction sequence. The specific formula is as follows: in, This indicates the reconstruction of the predicted sequence; This represents the prediction result of the j-th subsequence; This represents the residual prediction result.

[0053] Then, the reconstructed prediction sequence is inversely standardized based on the mean and standard deviation of the historical training set to obtain the container generation prediction sequence. Specifically, based on the mean and standard deviation recorded during the standardization transformation of the historical training set, the reconstructed prediction sequence is inversely standardized by multiplying it by the standard deviation and then adding the mean, thus restoring the dimensionless prediction result to TEU quantity, resulting in the container generation prediction sequence. The specific formula is as follows: in This represents the predicted container production volume after restoring its dimensions. This represents the standard deviation of the historical training set; This represents the mean of the historical training set.

[0054] Finally, the mean squared error (MSE) between the predicted container production volume sequence and the historical actual sequence is calculated. When the MSE exceeds a preset error threshold, a calibration command is sent to the probabilistic sparse attention prediction model to adjust its parameters. It's important to understand that the preset error threshold is the MSE benchmark for determining whether the prediction accuracy meets the standard; it can be set to 0.15. Specifically, the MSE between the predicted container production volume sequence and the historical actual sequence is calculated using the following formula: in Indicates mean square error; This represents the historical true value at the i-th time point; This represents the predicted value at the i-th time point; This indicates the number of time points involved in the evaluation. The smaller the mean squared error, the better the predicted curve fits the true curve. If the mean squared error is greater than the preset error threshold, it indicates that the model accuracy has degraded, for example, due to changes in distribution such as trade policy adjustments. In this case, a calibration command is sent to the probabilistic sparse attention prediction model, triggering the model to retrain using recent data or adjust hyperparameters until the backtest mean squared error falls back to within the preset error threshold.

[0055] This embodiment obtains a container generation sequence by anomaly removal and unit standardization of multi-source container business data from highways and railways. It then performs multiple empirical mode decompositions (EMDs) by adding paired positive and negative white noise and integrating the results to obtain physically meaningful intrinsic mode components and residual components. A multidimensional feature tensor is constructed by combining external driving factors and input into a prediction model containing a probabilistic sparse attention layer and a distillation layer, along with a generative decoder, to obtain prediction results for each component. Finally, the predicted sequence is superimposed, reconstructed, and denormalized to output the final prediction sequence. This embodiment eliminates mode aliasing interference, reduces the computational complexity of long sequences, and achieves high-precision long-sequence prediction of container generation, providing data support for logistics planning.

[0056] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in Embodiment 1 above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 2 The container generation prediction method based on multi-source business data, step S50, further includes steps S201 to S203: Step S201: Determine the time period to be predicted corresponding to the prediction task.

[0057] It should be noted that the forecast period is the future time range from which the predicted container generation value needs to be output. It can be taken as 30 consecutive statistical periods in the future, or it can be taken as far as the target year in the future, depending on the planning requirements. The length of the forecast period determines the length of the placeholder for forecast in the generative decoder. Specifically, the forecast period is determined according to the forecast task. For example, when making scheduling arrangements for next month, the forecast period is taken as 30 consecutive statistical periods in the future. When making long-term planning assessments, the forecast period is taken as all statistical periods corresponding to the target planning year. Once the forecast period is determined, the output length of the generative decoder is also determined.

[0058] Step S202: The encoder performs sparse attention calculation and layer-by-layer distillation compression on each multidimensional feature tensor, selects query vectors whose attention contribution meets the preset contribution conditions, and compresses the time axis dimension of the features to obtain the distilled features corresponding to each multidimensional feature tensor.

[0059] It should be noted that distilled features refer to the features output after the multidimensional feature tensor is processed by sparse attention calculation and distillation compression layer by layer by the processing units in the encoder. Its time axis dimension is lower than that of the input features, and its information density is higher.

[0060] Further, step S202 includes: First, mapping each multidimensional feature tensor to a query vector set, a key vector set, and a value vector set, respectively. It should be understood that the query vector, key vector, and value vector refer to the three sets of feature vectors obtained by multiplying the multidimensional feature tensor by three learnable linear mapping matrices; the query vector in the query vector set represents the feature state of the target to be predicted, the key vector in the key vector set represents the pattern information of the historical traffic sequence, and the value vector carries the feature content used for weighted output.

[0061] Secondly, the divergence metric between each query vector and each key vector in the key vector set is calculated through the first probabilistic sparse attention layer to obtain the attention contribution of each query vector. It's important to understand that the divergence metric measures the degree of difference between the attention distribution corresponding to a query vector and a uniform distribution. It can be calculated using relative entropy; a larger difference indicates a sharper attention distribution corresponding to the query vector, and thus a higher attention contribution. Specifically, the first probabilistic sparse attention layer calculates the divergence metric between each query vector and each key vector in the key vector set. Considering the sudden fluctuations in container generation during holidays, this example uses a probabilistic sparse attention mechanism to evaluate the importance of each query vector by calculating the divergence metric between the query vector and the key vector. The calculation formula is as follows: in This represents the attention contribution of the i-th query vector; This represents the i-th query vector in the query vector set; This represents the j-th key vector in the set of key vectors; express transpose; This represents the dimension of the feature vector input to the attention layer; Used to scale the dot product result to prevent the dot product value from becoming too large when d is large, causing the normalization function to enter the gradient saturation region; This represents the total number of key vectors and is positively correlated with the input step size. The first term of the formula is the maximum value of the dot product between the query vector and all key vectors, and the second term is the average value of the dot products. The larger the difference between the two, the sharper the attention distribution corresponding to the query vector, and the higher its attention contribution. Based on this metric, only the top u query vectors with the largest contributions can be retained for subsequent calculations.

[0062] Next, query vectors corresponding to historical periods within spatial clusters are given an initial attention sensitivity boost, and target query vectors that meet the preset contribution criteria are selected based on their attention contribution, from highest to lowest. It's important to understand that the initial attention sensitivity boost refers to multiplying the attention contribution of query vectors corresponding to historical periods within spatial clusters by an amplification factor before selection. This amplification factor can range from 1.1 to 1.3; a value that is too large will cause query vectors from clusters to crowd out too many selection slots, and should be selected within this range based on validation set performance. Target query vectors are those that, after being sorted by attention contribution from highest to lowest, meet the preset contribution criteria and are retained for attention calculation. Specifically, the initial attention sensitivity boost is applied to query vectors corresponding to historical periods within spatial clusters—that is, multiplied by the aforementioned amplification factor—before sorting. This makes it easier to retain fluctuation patterns in areas with strong logistics radiation capabilities. Subsequently, all query vectors are sorted from highest to lowest based on their boosted attention contribution, and target query vectors that meet the preset contribution criteria are selected.

[0063] Next, attention is calculated based on the target query vector and the set of key vectors to obtain sparse attention features. It's important to understand that sparse attention features refer to the output features obtained solely through attention calculation using the target query vector. Specifically, attention calculation is performed on the target query vector and the set of key vectors by weighting the sum of the value vectors based on the degree of matching between the target query vector and the key vectors. The outputs of query vectors not retained are filled with the mean of the value vectors, resulting in sparse attention features.

[0064] Then, the sparse attention features are sequentially subjected to one-dimensional convolution, activation processing, and max pooling through a distillation layer to obtain the pooled features. Specifically, the distillation layer sequentially performs one-dimensional convolution, activation processing, and max pooling on the sparse attention features. One-dimensional convolution aligns the feature dimensions, activation processing uses exponential linear units to enhance the nonlinear expression, and max pooling compresses the time axis dimension to half of its original value, resulting in the pooled features, as follows: in This represents the features after pooling; This represents the max pooling operation; Represents the activation function of the exponential linear unit; This represents a one-dimensional convolution operation; This indicates that the feature has been calculated using probabilistic sparse attention. This represents the sparse attention feature output by the probabilistic sparse attention layer in the i-th processing unit. With this setting, an input sequence of length L is compressed layer by layer into a representation with high information density, reducing the memory footprint when long sequence features are transferred between network layers.

[0065] Finally, the pooled features are passed to the next probabilistic sparse attention layer for repeated attention calculations. Layer by layer, the temporal dimension of the features is compressed to a preset multiple of the original temporal dimension, resulting in distilled features corresponding to each multidimensional feature tensor. It is important to understand that the preset multiple refers to the compression ratio of the temporal dimension by the distillation layer, which can be taken as half, meaning that the temporal length of the feature is halved for each processing unit.

[0066] Step S203: The generative decoder performs attention calculation on the merged vector based on the distillation features, and outputs the predicted values ​​of each time node in the prediction period in parallel through a single forward calculation, thereby obtaining the subsequence prediction results corresponding to each intrinsic mode component and the residual prediction results corresponding to the residual components.

[0067] It should be noted that the merged vector refers to the vector input to the generative decoder, including known historical segments and placeholders to be predicted. Known historical segments refer to the historical feature segments at the end of the multidimensional feature tensor corresponding to a preset historical step size. Placeholders to be predicted are feature vectors used to mark the time positions to be predicted; their length is the same as the length of the time period to be predicted. A single forward computation means that the data goes through only one complete forward propagation from the model input to the output, without iterative iterations across time nodes.

[0068] Further, step S203 includes: first, extracting the historical feature fragments corresponding to the last preset historical step size in the multidimensional feature tensor as known historical fragments to provide recent traffic context for the decoding process.

[0069] Secondly, the placeholders to be predicted are initialized based on the mean of features from historical time points in the historical feature segments, resulting in initialized placeholders. Specifically, the historical time points of the period to be predicted in the historical data are determined, features falling on the historical time points in the historical feature segments are extracted and their mean is calculated. This mean is used to fill the placeholders to be predicted, resulting in initialized placeholders. Compared to zero values ​​or random filling, initialized placeholders carry seasonal prior information corresponding to the period to be predicted. For example, if the period to be predicted covers the Spring Festival month, the initialized placeholders reflect the feature level of the historical Spring Festival trough.

[0070] Next, the known historical segments and the initialization placeholders are concatenated in chronological order to obtain the merged vector. Specifically, the known historical segments are placed first, followed by the initialization placeholders, and concatenated in chronological order to form a complete merged vector. The total length of the merged vector is the sum of the preset historical step size and the length of the period to be predicted.

[0071] Next, masked attention is performed on the known historical segments in the merged vector to obtain decoded self-attention features. Specifically, masked attention is performed on the known historical segments in the merged vector to mask the positional information after each time node, so that the features of each time node only interact with the time nodes in the known historical segments that precede it, preventing the positional information in the initial placeholder from leaking back to the historical part, thus obtaining decoded self-attention features.

[0072] The number of attention heads is then assigned to the corresponding generative decoder based on the frequency of each intrinsic mode component. It's important to understand that the higher the frequency of the intrinsic mode component, the more attention heads the generative decoder is assigned. Specifically, the number of attention heads assigned to the corresponding generative decoder is based on the frequency of each intrinsic mode component; the higher the frequency of the intrinsic mode component, the more attention heads the generative decoder is assigned. For example, when the total number of heads is 8, the decoder corresponding to high-frequency components can be assigned 4 to 6 attention heads to capture the variable local patterns caused by sudden fluctuations during holidays. The decoders corresponding to low-frequency components and residual components can be assigned 2 to 4 attention heads because the fluctuation patterns of long-term trends are relatively stable, and fewer heads are sufficient to meet the representation needs, while also reducing computational overhead.

[0073] Then, the decoded self-attention features are used as query features, and the distilled features are used as key and value features. Attention calculations are performed in parallel by each assigned attention head to obtain the head attention features corresponding to each attention head. Specifically, the decoded self-attention features are used as query features, and the distilled features output by the encoder are used as key and value features. Attention calculations are performed in parallel by each assigned attention head. Each attention head matches the decoded query with the historical patterns extracted by the encoder in its respective feature subspace to obtain the head attention features corresponding to each attention head.

[0074] Finally, the attention features of each head are concatenated and linearly mapped. The predicted values ​​for each time node within the prediction period are output in parallel through a single forward computation, yielding the subsequence prediction results for each intrinsic mode component and the residual prediction results for the residual components. Specifically, the attention features of each head are concatenated and linearly mapped to an output of a unified dimension. The predicted values ​​for each time node within the prediction period are output in parallel through a single forward computation. After summing the model outputs of each component, the subsequence prediction results for each intrinsic mode component and the residual prediction results for the residual components are obtained.

[0075] This embodiment allows the decoder output length to be flexibly configured according to the prediction task by pre-determining the time period to be predicted, taking into account both short-term scheduling and long-term planning requirements; the encoder extracts key dependency patterns in long sequences with low computational overhead through sparse attention filtering and layer-by-layer distillation compression; the generative decoder outputs the predicted values ​​of the entire time period to be predicted in a single parallel output based on the merged vector, avoiding the error accumulation caused by step-by-step iterative prediction and improving the stability of long-term prediction.

[0076] Based on the first embodiment of this application, this application also provides a container production volume prediction device based on multi-source business data. Please refer to... Figure 3 The device includes: The data acquisition module 10 is used to acquire multi-source container business data, and to remove abnormal data and unify the units of the multi-source container business data to obtain the container generation sequence. The multi-source container business data includes highway toll data and railway loading and unloading data.

[0077] The noise addition module 20 is used to add pairs of positive and negative white noise to the container generation sequence to construct multiple auxiliary sequences.

[0078] The sequence decomposition module 30 is used to perform empirical mode decomposition on each auxiliary sequence and to integrate and average the corresponding components obtained from multiple decompositions to obtain multiple intrinsic mode components and residual components.

[0079] The feature construction module 40 is used to acquire external driving factor data, and to perform driving feature screening and standardization processing on each intrinsic mode component based on the external driving factor data, so as to obtain the multidimensional feature tensor corresponding to each intrinsic mode component.

[0080] The model prediction module 50 is used to input the multidimensional feature tensors into the probabilistic sparse attention prediction model for processing, and to obtain the subsequence prediction results corresponding to each intrinsic mode component and the residual prediction results corresponding to the residual components. The probabilistic sparse attention prediction model includes an encoder and a generative decoder. The encoder includes several processing units connected in sequence in a preset encoder layer. The processing unit includes a probabilistic sparse attention layer and a distillation layer. The probabilistic sparse attention layer is used to filter query vectors whose attention contribution meets the preset contribution conditions for attention calculation.

[0081] The reconstruction output module 60 is used to overlay and reconstruct the prediction results of each subsequence and the residual prediction results, and to perform inverse standardization on the overlay and reconstructed sequence to obtain the container generation prediction sequence.

[0082] The container generation volume prediction device based on multi-source business data provided in this application adopts the container generation volume prediction method based on multi-source business data in the above embodiments, which can solve the technical problem of how to achieve high-precision prediction of long-sequence container generation volume with low computational complexity. Compared with the prior art, the beneficial effects of the container generation volume prediction device based on multi-source business data provided in this application are the same as the beneficial effects of the container generation volume prediction method based on multi-source business data provided in the above embodiments, and other technical features in the container generation volume prediction device based on multi-source business data are the same as the features disclosed in the methods of the above embodiments, and will not be repeated here.

[0083] This application provides a container production volume prediction device based on multi-source business data. The container production volume prediction device based on multi-source business data includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the container production volume prediction method based on multi-source business data in the above embodiment 1.

[0084] The following is for reference. Figure 4 This document illustrates a structural schematic diagram of a container generation prediction device based on multi-source business data suitable for implementing embodiments of this application. The container generation prediction device based on multi-source business data in the embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 4 The container generation prediction device based on multi-source business data shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0085] like Figure 4As shown, the container production volume prediction device based on multi-source business data may include a processing unit 1001 (e.g., a central processing unit, a graphics processor, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 1002 or a program loaded from a storage device 1003 into a random access memory (RAM) 1004. The RAM 1004 also stores various programs and data required for the operation of the container production volume prediction device based on multi-source business data. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following can be connected to I / O interface 1006: input devices 1007 including, for example, a touchscreen, touchpad, keyboard, mouse, image sensor, microphone, accelerometer, gyroscope, etc.; output devices 1008 including, for example, a liquid crystal display (LCD), speaker, vibrator, etc.; storage devices 1003 including, for example, magnetic tape, hard disk, etc.; and communication devices 1009. Communication device 1009 allows the container production volume prediction device based on multi-source business data to wirelessly or wiredly communicate with other devices to exchange data. Although various container production volume prediction devices based on multi-source business data are shown in the figures, it should be understood that it is not required to implement or possess all of them. More or fewer may be implemented alternatively.

[0086] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0087] The container generation volume prediction device based on multi-source business data provided in this application, employing the container generation volume prediction method based on multi-source business data in the above embodiments, can solve the technical problem of how to achieve high-precision prediction of long-sequence container generation volume with low computational complexity. Compared with the prior art, the beneficial effects of the container generation volume prediction device based on multi-source business data provided in this application are the same as those of the container generation volume prediction method based on multi-source business data provided in the above embodiments, and other technical features in this container generation volume prediction device based on multi-source business data are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0088] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0089] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0090] This application provides a computer-readable medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to execute the container production volume prediction method based on multi-source business data in the above embodiments.

[0091] The computer-readable medium provided in this application may be, for example, a USB flash drive, but is not limited to electrical, magnetic, optical, electromagnetic, infrared, or semiconductor devices, or any combination thereof. More specific examples of computer-readable media may include, but are not limited to: electrical connections with one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable medium may be any tangible medium containing or storing a program that can be executed by instructions, used by a device, or used in conjunction with it. The program code contained on the computer-readable medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0092] The aforementioned computer-readable medium may be included in a container generation prediction device based on multi-source business data; or it may exist independently and not be assembled into a container generation prediction device based on multi-source business data.

[0093] The aforementioned computer-readable medium carries one or more programs that, when executed by a container production volume prediction device based on multi-source business data, enable the device to write computer program code for performing the operations of this application in one or more programming languages ​​or a combination thereof. These programming languages ​​include object-oriented programming languages—such as Java, Smalltalk, and C++—and conventional procedural programming languages—such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0094] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and computer program products according to various embodiments of this application. In this regard, all blocks in the flowcharts or block diagrams may represent a module, segment, or portion of code containing one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that all blocks in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using dedicated hardware-based implementations that perform the specified functions or operations, or using a combination of dedicated hardware and computer instructions.

[0095] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0096] The readable medium provided in this application is a computer-readable medium, which stores computer-readable program instructions (i.e., a computer program) for executing the above-described container generation volume prediction method based on multi-source business data. This solves the technical problem of how to achieve high-precision prediction of long-sequence container generation volumes with low computational complexity. Compared with the prior art, the beneficial effects of the computer-readable medium provided in this application are the same as those of the container generation volume prediction method based on multi-source business data provided in the above embodiments, and will not be repeated here.

[0097] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the container production volume prediction method based on multi-source business data as described above.

[0098] The computer program product provided in this application solves the technical problem of how to achieve high-precision prediction of long-sequence container production volume with low computational complexity. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the container production volume prediction method based on multi-source business data provided in the above embodiments, and will not be repeated here.

[0099] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A method for predicting container production volume based on multi-source business data, characterized in that, The method includes: Acquire multi-source container business data, and remove abnormal data and unify the units of the multi-source container business data to obtain a container generation sequence, wherein the multi-source container business data includes highway toll flow data and railway loading and unloading data. Pairs of positive and negative white noise are added to the container generation sequence to construct multiple auxiliary sequences; Each of the auxiliary sequences is subjected to empirical mode decomposition, and the corresponding components obtained from multiple decompositions are integrated and averaged to obtain multiple intrinsic mode components and residual components. Obtain external driving factor data, and perform driving feature screening and standardization processing on each intrinsic mode component based on the external driving factor data to obtain the multidimensional feature tensor corresponding to each intrinsic mode component; Each of the multidimensional feature tensors is input into the probabilistic sparse attention prediction model for processing to obtain the subsequence prediction results corresponding to each intrinsic mode component and the residual prediction results corresponding to the residual component. The probabilistic sparse attention prediction model includes an encoder and a generative decoder. The encoder includes a preset encoder layer and several processing units connected in sequence. Each processing unit includes a probabilistic sparse attention layer and a distillation layer. The probabilistic sparse attention layer is used to filter query vectors whose attention contribution satisfies the preset contribution condition for attention calculation. The prediction results of each subsequence and the prediction results of the residuals are superimposed and reconstructed, and the superimposed and reconstructed sequence is denormalized to obtain the container generation prediction sequence.

2. The method as described in claim 1, characterized in that, The step of inputting each of the multidimensional feature tensors into the probabilistic sparse attention prediction model and processing them to obtain the subsequence prediction results corresponding to each intrinsic mode component and the residual prediction results corresponding to the residual components includes: Determine the forecast period corresponding to the forecast task, wherein the forecast period is the future time range from which the predicted value of container generation needs to be output; The encoder performs sparse attention calculation and layer-by-layer distillation compression on each of the multidimensional feature tensors, filters the query vectors whose attention contribution meets the preset contribution condition, and compresses the time axis dimension of the features to obtain the distilled features corresponding to each of the multidimensional feature tensors. The generative decoder performs attention calculation on the merged vector based on the distillation features, and outputs the predicted values ​​of each time node in the time period to be predicted in parallel through a single forward calculation, thereby obtaining the subsequence prediction results corresponding to each intrinsic mode component and the residual prediction results corresponding to the residual component. The merged vector includes known historical segments and placeholders to be predicted, and the length of the placeholders to be predicted is consistent with the length of the time period to be predicted.

3. The method as described in claim 2, characterized in that, The step of performing sparse attention calculation and layer-by-layer distillation compression on each of the multidimensional feature tensors through an encoder, selecting query vectors whose attention contribution satisfies a preset contribution condition, and compressing the time axis dimension of the features to obtain the distilled features corresponding to each of the multidimensional feature tensors includes: Each of the multidimensional feature tensors is mapped to a query vector set, a key vector set, and a value vector set, respectively, wherein the query vectors in the query vector set represent the feature state of the target to be predicted, and the key vectors in the key vector set represent the pattern information of the historical traffic sequence; The divergence metric between each query vector and each key vector in the key vector set is calculated through the first probabilistic sparse attention layer to obtain the attention contribution of each query vector; The query vectors corresponding to historical time periods in spatial clustering areas are given an initial attention sensitivity bonus, and target query vectors that meet the preset contribution conditions are filtered from high to low according to the attention contribution. Attention is calculated based on the target query vector and the set of key vectors to obtain sparse attention features; The sparse attention features are sequentially subjected to one-dimensional convolution, activation processing, and max pooling through the distillation layer to obtain pooled features. The pooled features are passed to the next probabilistic sparse attention layer for repeated attention calculation. Layer by layer, the time axis dimension of the features is compressed to a preset multiple of the original time axis dimension to obtain the distilled features corresponding to each multidimensional feature tensor.

4. The method as described in claim 2, characterized in that, The step of performing attention calculation on the merged vector based on the distillation features using the generative decoder, and outputting the predicted values ​​of each time node within the prediction period in parallel through a single forward computation to obtain the subsequence prediction results corresponding to each intrinsic mode component and the residual prediction results corresponding to the residual components includes: Extract the historical feature segment corresponding to the last preset historical step size from the multidimensional feature tensor, and use it as a known historical segment; The placeholder to be predicted is initialized based on the feature mean of the same time node in the historical feature segment to obtain the initialized placeholder; The known historical fragments and the initialization placeholders are concatenated in chronological order to obtain the merged vector; Masked attention calculation is performed on the known historical segments in the merged vector to obtain decoded self-attention features; The number of attention heads is assigned to the corresponding generative decoder based on the frequency of each intrinsic modal component, wherein the higher the frequency of the intrinsic modal component, the more attention heads are assigned to the generative decoder. The decoded self-attention features are used as query features, and the distillation features are used as key and value features. Attention calculations are performed in parallel through the assigned attention heads to obtain the head attention features corresponding to each attention head. The head attention features are concatenated and linearly mapped, and the predicted values ​​of each time node in the time period to be predicted are output in parallel through a single forward computation, so as to obtain the subsequence prediction results corresponding to each intrinsic mode component and the residual prediction results corresponding to the residual component.

5. The method as described in claim 1, characterized in that, The steps of performing empirical mode decomposition on each of the auxiliary sequences and ensemble averaging the corresponding components obtained from multiple decompositions to obtain multiple intrinsic mode components and residual components include: Empirical mode decomposition is performed on each of the auxiliary sequences. By extracting the extreme value envelope of the auxiliary sequences and sieving them one by one, multiple initial mode components corresponding to each of the auxiliary sequences are obtained. The initial modal components corresponding to each auxiliary sequence are aligned in descending order of frequency to obtain modal component groups of each order. Calculate the energy value of each initial modal component in each modal component group, and remove abnormal initial modal components whose energy values ​​deviate from the average energy value of the modal component group by more than a preset energy deviation threshold to obtain each effective component group; The initial modal components in each effective component group are arithmetically averaged to obtain the intrinsic modal components of each order. The high-frequency intrinsic modal components are used to characterize short-term random fluctuations caused by holidays, and the low-frequency intrinsic modal components are used to characterize the long-term evolution trend of regional economic development. The intrinsic modal components are stripped from the container generation sequence, and the remaining parts are identified as residual components.

6. The method as described in claim 1, characterized in that, The steps of acquiring external driving factor data, and performing driving feature filtering and standardization on each intrinsic mode component based on the external driving factor data to obtain the multidimensional feature tensor corresponding to each intrinsic mode component include: Acquire external driving factor data, including transportation infrastructure indicators, economic scale indicators, and social structure indicators. A spatial weight matrix is ​​constructed based on spatial adjacency relationships. The local spatial autocorrelation index of each administrative region is calculated based on the spatial weight matrix. Administrative regions whose local spatial autocorrelation index meets the preset clustering conditions are identified as spatial clustering regions. The intrinsic mode components and the external driving factor data are standardized to obtain standardized mode components and standardized driving factor data. Establish a regression equation between the standardized modal components and the standardized driving factor data, and remove redundant factors with regression coefficients less than a preset significance threshold to obtain key driving factor data; The key driving factor data are used as exogenous covariates and horizontally concatenated with the historical sequence of container generation according to a preset input step size to obtain the initial feature tensor. Spatial weight coefficients are assigned to the sample features in the spatial clustering region of the initial feature tensor to obtain the multidimensional feature tensor corresponding to each intrinsic mode component.

7. The method as described in claim 1, characterized in that, The step of superimposing and reconstructing the prediction results of each subsequence and the residual prediction results, and then performing denormalization on the superimposed and reconstructed sequence to obtain the container generation prediction sequence includes: Align the prediction results of each subsequence with the prediction results of the residual according to the timestamp to obtain an aligned prediction component set; Calculate the prediction error of each subsequence prediction result within the historical verification period, and determine the reconstruction weight corresponding to each subsequence prediction result based on the prediction error, wherein the smaller the prediction error, the larger the reconstruction weight corresponding to the subsequence prediction result; The components in the aligned prediction component set are weighted and summed according to the reconstruction weights, and the values ​​less than zero in the weighted summation result are set to zero to obtain the reconstructed prediction sequence. The reconstructed prediction sequence is inversely standardized based on the mean and standard deviation of the historical training set to obtain the container generation prediction sequence. The mean square error between the predicted sequence of container generation and the historical true sequence is calculated. When the mean square error is greater than a preset error threshold, a calibration command is sent to the probabilistic sparse attention prediction model so that the probabilistic sparse attention prediction model adjusts its model parameters according to the calibration command.

8. A container generation prediction device based on multi-source business data, characterized in that, The apparatus is applied to the container generation prediction method based on multi-source business data as described in any one of claims 1 to 7, and the apparatus comprises: The data acquisition module is used to acquire multi-source container business data, and to remove abnormal data and unify the units of the multi-source container business data to obtain a container generation sequence. The multi-source container business data includes highway toll data and railway loading and unloading data. A noise addition module is used to add pairs of positive and negative white noise to the container generation sequence to construct multiple auxiliary sequences; The sequence decomposition module is used to perform empirical mode decomposition on each of the auxiliary sequences, and to integrate and average the corresponding components obtained from multiple decompositions to obtain multiple intrinsic mode components and residual components. The feature construction module is used to acquire external driving factor data, and perform driving feature screening and standardization processing on each intrinsic mode component based on the external driving factor data to obtain the multidimensional feature tensor corresponding to each intrinsic mode component. The model prediction module is used to input the multidimensional feature tensors into the probabilistic sparse attention prediction model for processing, and to obtain the subsequence prediction results corresponding to each intrinsic mode component and the residual prediction results corresponding to the residual components. The probabilistic sparse attention prediction model includes an encoder and a generative decoder. The encoder includes a preset encoder layer and several processing units connected in sequence. The processing unit includes a probabilistic sparse attention layer and a distillation layer. The probabilistic sparse attention layer is used to filter query vectors whose attention contribution satisfies the preset contribution condition for attention calculation. The reconstruction output module is used to superimpose and reconstruct the prediction results of each subsequence and the prediction results of the residuals, and to perform inverse standardization on the superimposed and reconstructed sequence to obtain the container generation prediction sequence.

9. A container production volume prediction device based on multi-source business data, characterized in that, The device includes: a memory, a processor, and a container generation prediction program based on multi-source business data stored on the memory and running on the processor, the container generation prediction program based on multi-source business data being configured to implement the steps of the container generation prediction method based on multi-source business data as described in any one of claims 1-7.

10. A storage medium, characterized in that, The storage medium stores a container production volume prediction program based on multi-source business data. When the container production volume prediction program based on multi-source business data is executed by the processor, it implements the steps of the container production volume prediction method based on multi-source business data as described in any one of claims 1-7.

Images