Sales supply chain management methods and systems based on internet data

Abstract

This invention discloses a sales supply chain management method and system based on internet data, relating to the field of supply chain optimization technology. The method includes decomposition using discrete wavelet transform, calculation of approximate kernel distance using Gaussian random projection, generation of low-dimensional embeddings through low-rank decomposition and ADMM iterative optimization, initialization of predicted demand using a Poisson distribution, quantization using Shapley values, initialization of particle swarm optimization, obtaining individual and global optimal positions, performing local guided search, updating the optimal solution through neighborhood operations, and generating optimal logistics paths and scheduling schemes through SA perturbation and DE-GA mutation. This invention improves prediction accuracy by introducing Poisson distribution and Shapley value quantification of predicted demand, and generates efficient logistics paths and scheduling schemes through multi-objective particle swarm optimization and hybrid mutation strategies, thereby enhancing the adaptability and operational efficiency of the sales supply chain in dynamic market environments.

Description

Technical Field

[0001] This invention relates to the field of supply chain optimization technology, and in particular to a sales supply chain management method and system based on Internet data. Background Technology

[0002] With the rapid development of the Internet, big data, and cloud computing technologies, sales supply chain management methods based on Internet data have emerged and become a hot topic in the industry. Traditional supply chain management often relies on a single data source or human experience, which is difficult to meet the rapidly changing demands of the market environment. With the emergence of multi-source data such as e-commerce, social media, and the Internet of Things, researchers have begun to explore how to preprocess, fuse, and analyze multi-source heterogeneous data to improve prediction accuracy and decision-making efficiency. Discrete wavelet transform has been gradually introduced, showing advantages in multi-scale time series feature extraction. Meanwhile, efficient random projection and low-rank decomposition techniques provide new ideas for the real-time processing of massive data. In addition, demand forecasting models based on probability distribution and multi-objective optimization algorithms have made progress in logistics route planning and resource scheduling.

[0003] However, existing methods still have shortcomings. In terms of logistics route and supply scheduling optimization, most existing systems adopt a single particle swarm optimization algorithm, which lacks a collaborative multi-objective optimization mechanism that can adapt to complex constraints. At the same time, the optimization process often gets stuck in local optima, and it is difficult to dynamically introduce perturbation strategies when updating the particle swarm to break out of the limitations of local optima. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a sales supply chain management method and system based on Internet data, which solves the problem that in terms of logistics path and supply scheduling optimization, most existing systems adopt a single particle swarm optimization algorithm, which lacks a collaborative multi-objective optimization mechanism to adapt to complex constraints. At the same time, the optimization process often gets stuck in local optima, and it is difficult to dynamically introduce perturbation strategies when updating the particle swarm to escape the limitations of local optima.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a sales supply chain management method based on Internet data, which includes,

[0008] Multi-source data is collected and preprocessed to generate normalized multi-source data. It is then decomposed using discrete wavelet transform, detail coefficients are extracted and abrupt change points are marked, a three-dimensional feature vector is constructed, time alignment is performed using eDTW distance, approximate kernel distance is calculated using Gaussian random projection, low-dimensional embedding is generated through low-rank decomposition and ADMM iterative optimization, the prediction requirement is initialized with Poisson distribution and combined with Gamma prior, the prediction requirement is calculated using posterior distribution and prediction distribution, quantization is performed using objective function and Shapley value, and the final prediction requirement is generated.

[0009] The weighted average satisfaction is calculated based on multi-source data. The final allocation matrix is ​​obtained through the PSG objective function and stochastic gradient. A multi-objective function is defined based on the weighted average satisfaction, and constraints are set in combination with the final prediction requirements. The particle swarm is initialized. Through position and velocity updates, the optimal positions of individuals and the global optimum are determined, and a local guided search is performed. The optimal solution is output through neighborhood operations and the particle swarm is updated. The updated particle swarm is further seeded, and SA perturbation and DE-GA mutation are performed respectively to generate the optimal logistics path and scheduling scheme, which is then distributed through the API interface.

[0010] As a preferred embodiment of the sales supply chain management method based on Internet data described in this invention, the step of decomposing using discrete wavelet transform and aligning time using eDTW distance includes:

[0011] Based on normalized multi-source data, discrete wavelet transform is used for decomposition, and the mean of detail coefficients is calculated. If the detail coefficients are greater than the mean of detail coefficients, attack points are marked, a three-dimensional feature vector is constructed, the eDTW distance is calculated, and the normalized multi-source data is aligned on the time axis through linear interpolation to generate an aligned sequence.

[0012] As a preferred embodiment of the sales supply chain management method based on internet data described in this invention, the step of initializing the predicted demand based on a Poisson distribution and combining it with a Gamma prior, quantifying it using an objective function and a Shapley value, and generating the final predicted demand includes:

[0013] Based on the aligned sequence, it is concatenated into a high-dimensional feature matrix. A random projection matrix is ​​generated by Gaussian random projection, and the approximate kernel distance is calculated to generate a distance matrix. Low-rank decomposition moments are performed by LRI to obtain a low-rank coefficient matrix. ADMM is used for iterative optimization to output a low-dimensional embedding. The predicted demand is initialized to follow a Poisson distribution, and Gamma is set as a prior. Then the posterior distribution is calculated to obtain the predicted distribution.

[0014] Based on the predicted distribution, the mean is calculated to obtain the predicted demand. An objective function is defined, the Shapley value is calculated, and the final predicted demand is obtained through multiplication.

[0015] As a preferred embodiment of the sales supply chain management method based on Internet data described in this invention, the step of calculating the weighted average satisfaction based on multi-source data includes:

[0016] A satisfaction model is defined based on normalized delivery waiting time, normalized rating stars, and positive and negative evaluations, and a weighted average satisfaction rate is calculated.

[0017] As a preferred embodiment of the sales supply chain management method based on Internet data described in this invention, the step of obtaining the final allocation matrix through the PSG objective function and stochastic gradient includes:

[0018] Define a set of constraints based on the final predicted demand, define the PSG objective function, and calculate the stochastic gradient;

[0019] Initialize the allocation matrix, update the allocation matrix using stochastic gradient descent, and output the final allocation matrix.

[0020] As a preferred embodiment of the sales supply chain management method based on Internet data described in this invention, the initialization of the particle swarm, outputting the optimal solution and updating the particle swarm through domain operations, performing SA perturbation and DE-GA mutation, and generating the optimal logistics path and scheduling scheme include:

[0021] Based on the final allocation matrix and weighted average satisfaction, a multi-objective function is defined and constraints are set. The particle swarm is initialized, the initial multi-objective function value is calculated, sorted in descending order, and the particles corresponding to the top K multi-objective function values ​​are selected. Local guided search is performed, and the optimal domain solution is iteratively output through domain operations and guidance strategies. The initial scheduling scheme of the corresponding particles is replaced, the updated particle swarm is output, and initial partitioning is performed, including satisfaction seed group, cost seed group and reward seed group.

[0022] Based on the satisfaction seed group, the corresponding initial multi-objective function values ​​are sorted in ascending order, the top M values ​​are selected, and perturbation particles are generated through SA perturbation. Based on the cost seed group, adaptive differential weights are calculated, and DE-GA mutation is performed to generate mutated particles.

[0023] The corresponding selected particles are replaced by perturbation particles and mutation particles to generate a new particle swarm. The multi-objective function value is calculated to generate the optimal logistics path and scheduling scheme, which is then distributed through the API interface.

[0024] As a preferred embodiment of the sales supply chain management method based on Internet data described in this invention, the step of collecting and preprocessing multi-source data to generate normalized multi-source data includes:

[0025] Multi-source data is collected through API interfaces and preprocessed to generate standardized multi-source data. Normalized multi-source data is then generated through time series normalization.

[0026] Secondly, this invention provides a sales supply chain management system based on internet data, including:

[0027] The prediction collection module is used to collect and preprocess multi-source data, generate normalized multi-source data, decompose it through discrete wavelet transform, extract detail coefficients and mark abrupt change points, construct a three-dimensional feature vector, perform time alignment using eDTW distance, calculate the approximate kernel distance through Gaussian random projection, generate low-dimensional embeddings through low-rank decomposition and ADMM iterative optimization, initialize the prediction requirements of the Poisson distribution and combine them with the Gamma prior, calculate the prediction requirements through the posterior distribution and the prediction distribution, quantize it using the objective function and Shapley value, and generate the final prediction requirements.

[0028] The scheduling optimization module is used to calculate the weighted average satisfaction based on multi-source data. It obtains the final allocation matrix through the PSG objective function and stochastic gradient, defines a multi-objective function based on the weighted average satisfaction, sets constraints in combination with the final prediction requirements, initializes the particle swarm, determines the individual and global optimal positions through position and velocity updates, performs local guided search, outputs the best solution through neighborhood operations and updates the particle swarm, further seeds the updated particle swarm, performs SA perturbation and DE-GA mutation respectively, generates the optimal logistics path and scheduling scheme, and distributes them through the API interface.

[0029] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the sales supply chain management method based on Internet data as described in the first aspect of the present invention.

[0030] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the sales supply chain management method based on Internet data as described in the first aspect of the present invention.

[0031] The beneficial effects of this invention are as follows: This invention decomposes the data using discrete wavelet transform, calculates the approximate kernel distance using Gaussian random projection, generates low-dimensional embeddings through low-rank decomposition and ADMM iterative optimization, initializes the prediction requirements of the Poisson distribution, quantizes the data using Shapley values, initializes the particle swarm, obtains the individual and global optimal positions, performs local guided search, updates the data by outputting the best solution through neighborhood operations, and generates the optimal logistics path and scheduling scheme through SA perturbation and DE-GA mutation; thus improving prediction accuracy and enhancing the adaptability and operational efficiency of the sales supply chain in a dynamic market environment. Attached Figure Description

[0032] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0033] Figure 1 This is a flowchart of the sales supply chain management method based on Internet data in Example 1.

[0034] Figure 2 This is a schematic diagram of the sales supply chain management system based on Internet data in Example 1. Detailed Implementation

[0035] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0036] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0037] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0038] Example 1, referring to Figure 1 and Figure 2 This is the first embodiment of the present invention, which provides a sales supply chain management method based on Internet data, including the following steps:

[0039] S1. Collect and preprocess multi-source data to generate normalized multi-source data. Decompose the data using discrete wavelet transform, extract detail coefficients, and mark abrupt change points. Construct a three-dimensional feature vector. Use eDTW distance for time alignment. Calculate the approximate kernel distance using Gaussian random projection. Generate a low-dimensional embedding through low-rank decomposition and ADMM iterative optimization. Initialize the prediction requirements of the Poisson distribution and combine them with the Gamma prior. Calculate the prediction requirements using the posterior distribution and the prediction distribution. Quantize the data using the objective function and Shapley value to generate the final prediction requirements.

[0040] Specifically, multi-source data is collected and preprocessed to generate normalized multi-source data, including:

[0041] Collect data from multiple sources via API and perform preprocessing.

[0042] The multi-source data includes e-commerce platform data (real-time sales, prices, inventory, and user reviews), social media data (user sentiment data and trend tags), logistics platform data (delivery time, cost, and delivery efficiency), and macroeconomic indicator data (market purchasing power).

[0043] The preprocessing includes removing missing, duplicate, and irrelevant values ​​from the multi-source data, extracting keywords, sentiment trends, and trend features using NLP, aligning the data using timestamps, and generating standardized multi-source data.

[0044] Based on standardized multi-source data, real-time sales data, user discussion data, timeliness data, and macroeconomic indicator data are extracted. Time series normalization is then performed on the four types of data to generate normalized multi-source data.

[0045] By systematically removing missing values, duplicate values, and irrelevant items during data preprocessing, redundant data interference can be effectively reduced, and sample quality can be improved. Introducing NLP for sentiment assessment and trend tag mining of user comments, tags, and social content transforms previously unquantifiable soft indicators such as user perception and market sentiment into computable structured data, thereby expanding the input dimensions and cognitive depth of the predictive model. Through a timestamp alignment mechanism, the offset problem of multi-source data at different time points is effectively avoided, ensuring the logical correlation and temporal consistency between variables, providing a solid temporal data foundation for subsequent time-series modeling. By performing time-series normalization on multiple variables such as sales volume, discussion level, logistics timeliness, and economic indicators, not only are the trend and periodic characteristics of each variable preserved, but the adaptability of data input under multiple algorithm frameworks and the training convergence speed are also improved.

[0046] Furthermore, decomposition is performed using discrete wavelet transform, and time alignment is achieved using eDTW distance, including:

[0047] Based on normalized multi-source data, discrete wavelet transform is used for decomposition, and the mean of detail coefficients is calculated. If the detail coefficients are greater than the mean of detail coefficients, the abrupt change point is marked as 1; otherwise, it is marked as 0. The formula is as follows:

[0048]

[0049] Where, μ Γ σ is the mean of the detail coefficients of the decomposition layer Γ. Γ dl represents the number of decomposition layers Γ. Γ Θ represents the detail coefficients of the decomposition layer Γ, obtained based on the decomposition results; Θ is the coefficient index.

[0050] Based on the mutation point, a three-dimensional feature vector is constructed, using the following formula:

[0051]

[0052] Among them, t ′ s t″ represents the start time of mutation at mutation point s. s The time at which the mutation ends at mutation point s. Time series The three-dimensional feature vector at time t, the Time series data of four types. Time series The normalized value at time t, Time series Information propagates during the rising event at time t. Time series Information propagates during the descent event at time t;

[0053] Based on the 3D feature vectors, the eDTW distance is calculated, and the normalized multi-source data is time-axis aligned using linear interpolation to generate an aligned sequence. The formula is as follows:

[0054]

[0055] Among them, eD(V 2 b V 2 c ) is a time series V 2 b and time series V 2 c The eDTW distance, where b and c are data types, W is the alignment path, is calculated through dynamic programming, and t b and t c For data types b and c, the time is...

[0056] By constructing a detailed coefficient mean judgment condition, local mutation behavior can be effectively identified while maintaining the overall picture of the data, greatly improving the sensitivity and accuracy of event detection and supporting more timely decision response. The three-dimensional features not only consider the normalized value of the time series, but also introduce the propagation indicators of rising and falling events, reflecting the linkage behavior and diffusion path of different data types in sudden events. eDTW introduces structural consistency constraints of data types to avoid overfitting problems in the alignment process of different types of data. Linear interpolation further smooths the time granularity, so that the aligned sequence maintains continuity in high-frequency data, with strong adaptability, stability and interpretability, providing a solid data foundation and analysis for sales forecasting, inventory scheduling and dynamic optimization.

[0057] Furthermore, the prediction requirements are initialized using the Poisson distribution and combined with the Gamma prior. Then, the requirements are quantized using the objective function and Shapley value to generate the final prediction requirements, including:

[0058] Based on the aligned sequence, it is organized into a high-dimensional feature matrix by stacking horizontal features;

[0059] Based on the high-dimensional feature matrix, a random projection matrix is ​​generated through Gaussian random projection, and an approximate kernel distance is calculated to generate a distance matrix. The formula is as follows:

[0060]

[0061] Where D is the approximate kernel distance, and is an element in the distance matrix. and These are the projections of the eigenvectors of region IDr, category IDp, and samples u and o into the random projection matrix at time t, respectively.

[0062] Based on the distance matrix, a low-rank decomposition is performed using LRI to obtain the low-rank coefficient matrix, as shown in the formula:

[0063]

[0064] Where Z is the low-rank coefficient matrix, E is the reconstruction error matrix, and S is the distance matrix;

[0065] Based on the low-rank coefficient matrix, iterative optimization using ADMM (based on a fixed iteration method) is performed to output a low-dimensional embedding, as shown in the formula:

[0066] Z (k+1) =(ρS Ω S+β 2 I) -1 (ρS Ω (SE 3(k) )+S Ω Λ (k) +β2 J (k) ),

[0067] Among them, Z (k+1) Let S be the low-rank coefficient matrix of iteration k+1, including low-dimensional embeddings, where ρ is the penalty parameter for the transpose of the distance matrix, and S... Ω Let β be the transpose of the distance matrix S. 2 For auxiliary matrix J (k) The penalty parameter, E 3(k) To reconstruct the error matrix, Λ(k) is the Lagrange multiplier matrix;

[0068] The auxiliary matrix, reconstruction error matrix, and Lagrange multiplier matrix include: calculating the difference between the distance matrix and the product of the low-rank coefficient matrix and the distance matrix; generating an initial reconstruction error matrix through a soft thresholding operation; generating an initial auxiliary matrix based on the difference between the distance matrix and the initial reconstruction error matrix through singular value decomposition and using a singular value thresholding function; initializing the Lagrange multiplier matrix; and obtaining the initial Lagrange multiplier matrix by calculating the difference between the distance matrix and the initial reconstruction error matrix and combining it with the product of the penalty parameter ρ.

[0069] The singular value threshold function is defined by the following formula:

[0070] J (0) =UCdiag(max(σ) i -τ,0))V 4Ω ,

[0071] Among them, J (0) Let σ be the initial auxiliary matrix, UC be the left singular vector matrix, and σ be the initial auxiliary matrix. i Let τ be the i-th singular value, and τ be the singular value threshold, based on the penalty parameter β. 2 The reciprocal gives V 4Ω V is a right singular vector matrix 4 The transpose of , the left and right singular vector matrices are both obtained by singular value decomposition, and diag is the operation to extract diagonal elements;

[0072] Based on low-dimensional embedding, the initial predicted demand follows a Poisson distribution, and a Gamma prior is set, as shown in the formula:

[0073] P(λ r,p,t )=Gam(λ r,p,t |α r,p,t ,β r,p,t ),

[0074] Where, λ r,p,t Let α be the Poisson parameter for region IDr, category IDp, and time t. r,p,t β represents the shape parameters of region IDr, category IDp, and time t. r,p,tLet IDr be the rate parameter for region, IDp be the category, and t be the time, Gam be the Gamma prior, and P be the prior probability density function.

[0075] Based on the Gamma prior, the posterior distribution is calculated using the following formula:

[0076]

[0077] Among them, P 1 Let Y be the posterior probability density function, Y be a low-dimensional embedding, and t be the posterior probability density function. 1 H represents the time within the historical time window, and x represents the actual demand, obtained based on e-commerce platform data.

[0078] Based on the posterior distribution and the Gamma prior, the predicted distribution is calculated using the following formula:

[0079]

[0080] Among them, P 2 For probability distribution, Let NB be the initial predicted demand for region IDr, category IDp, and time t, and let NB be a negative binomial distribution.

[0081] Based on the predicted distribution, the mean is calculated to obtain the predicted demand, using the following formula:

[0082]

[0083] in, Forecasting demand for region IDr, category IDp, and time t;

[0084] Based on the distance matrix, the predicted demand, and the predicted demand, the objective function is defined as follows:

[0085]

[0086] Where v(Cx) is the prediction contribution of the region subset Cx, Ο 2 It is a constant;

[0087] The Shapley value is calculated based on the objective function, using the following formula:

[0088]

[0089] Where, φ r The weight of region r is expressed as a Shapley value, and NT is the number of regions.

[0090] Based on the Shapley value and the predicted demand, the final predicted demand is obtained through multiplication.

[0091] After generating a high-dimensional feature matrix by aligning time series, dimensionality reduction is achieved using Gaussian random projection, and an approximate kernel distance is constructed. This not only reduces computational complexity without significantly losing feature information but also preserves the geometric structure of the feature space, providing effective input for subsequent distance metrics and low-rank decomposition. Low-rank decomposition maps the complex high-dimensional feature structure to a lower-dimensional latent space, and iterative solutions using the alternating direction multiplier method significantly reduce redundant noise and the influence of local anomalies, enhancing the model's representational power and sensitivity to changing trends. The low-dimensional embedding results are used as input to the Poisson distribution parameters, and Bayesian modeling is performed using Gamma priors. This approach effectively avoids the destruction of expectation estimates by extreme data while maintaining the flexibility of the prediction model, ensuring the stability and robustness of the distribution fit. By using Gamma prior and historical data to perform posterior inference of the Poisson parameter, a negative binomial distribution form of the prediction distribution is generated, giving the model a stronger ability to handle excessive discretization and improving robustness in scenarios with drastic fluctuations or unpredictability. An objective function with prediction accuracy and regional contribution as the core is defined, and the marginal contribution of each regional subset in the overall prediction is decomposed by combining Shapley value theory, avoiding the uninterpretability of the regional dimension, and providing a decision-making basis for resource allocation and strategy formulation.

[0092] S2. Calculate the weighted average satisfaction based on multi-source data, obtain the final allocation matrix through the PSG objective function and stochastic gradient, define a multi-objective function based on the weighted average satisfaction and set constraints in combination with the final prediction requirements, initialize the particle swarm, determine the individual and global optimal positions through position and velocity updates, perform local guided search, output the best solution through domain operations and update the particle swarm, perform seeding on the updated particle swarm, perform SA perturbation and DE-GA mutation respectively, generate the optimal logistics path and scheduling scheme, and distribute it through the API interface.

[0093] Specifically, the weighted average satisfaction rate is calculated based on multi-source data, including:

[0094] The delivery waiting time is obtained based on logistics platform data and is the difference between the actual delivery time and the order placement time.

[0095] Based on user reviews, the BERT model is used for classification to obtain positive and negative reviews, and the star ratings are extracted. These ratings are then normalized in conjunction with the delivery waiting time.

[0096] Based on normalized delivery waiting time, normalized rating stars, and positive and negative evaluations (where positive evaluation values ​​are 1 and negative values ​​are 0), a satisfaction model is defined with the following formula:

[0097]

[0098] in, as well as The weights for normalized delivery waiting time, normalized rating stars, and positive and negative evaluations are respectively set based on expert experience. TR, SW, and ER are normalized delivery waiting time, normalized rating stars, and positive and negative evaluations, respectively. SW is the initial satisfaction score, which is represented as the satisfaction model.

[0099] Based on the satisfaction model, the weighted average satisfaction level is calculated using the following formula:

[0100]

[0101]

[0102] Where w is the order freshness weight, For order i 1 The order freshness weighting is defined by RT (current time), RU (order placement time), SV (weighted average satisfaction), N (number of orders), and Ξ′ (decay period). For order i 1 The initial satisfaction score.

[0103] By introducing the sentiment classification results of the BERT model as a key feature and combining it with the quantitative data of star ratings, the satisfaction model can simultaneously take into account users' emotional tendencies and rational ratings, improving the model's expressiveness and interpretability. By using an exponential decay function to weight order freshness, the system pays more attention to the satisfaction ratings of recent orders, thereby quickly reflecting changes in service levels and avoiding the lagging impact of historical data on the assessment of current service status, effectively improving the system's responsiveness. Finally, by aggregating the satisfaction scores of all orders, a single weighted average satisfaction index is output, enhancing the system's closed-loop control capability.

[0104] Furthermore, the final allocation matrix is ​​obtained through the PSG objective function and stochastic gradient, including:

[0105] The constraint set is defined based on the current inventory level in the warehouse, including the following: the inventory allocation variable is less than or equal to the current inventory level, the inventory allocation variable is greater than or equal to the product of the final predicted demand and the weighted average satisfaction level, and the inventory allocation variable is greater than or equal to 0.

[0106] Based on the constraint set, the PSG objective function is defined as follows:

[0107] K(q)=∑ r,z Cy·sf r,z +∑ z CQ·(I z -∑ r,z sf r,z ),

[0108] Where K(q) is the objective function of the inventory allocation matrix q, Cy is the transportation cost, and sf r,z I allocates the inventory quantity of warehouse z to region r. z Let Z be the current inventory level in warehouse z, and CQ be the inventory holding cost.

[0109] Based on the PSG objective function, the stochastic gradient is calculated using the following formula:

[0110] g(q,ξ)=Cy-CQ,

[0111] Where g is the stochastic gradient and ξ is a random variable, representing a randomly selected region-warehouse;

[0112] Initialize the allocation matrix, update it using stochastic gradient descent until the maximum number of updates is reached, then stop. Based on experimental experience, output the final allocation matrix using the following formula:

[0113] y (A) =y (A-1) -τ·g(y (A-1) ,ξ)

[0114] Among them, y (A) Let τ be the allocation matrix for iteration A, and τ be the step size.

[0115] By replacing the full gradient with stochastic gradients, computational overhead is significantly reduced, enabling the algorithm to handle complex relationships between multiple warehouses and distribution areas in real time. The constraint set dynamically correlates and predicts demand and satisfaction, ensuring that resource allocation not only meets quantitative requirements but also takes service quality into account, thereby improving the overall supply chain responsiveness. The PSG objective function achieves an optimal balance between transportation costs and inventory costs, effectively reducing logistics costs and increasing turnover. Through projection and step size control mechanisms, the iterative process is stable and has strong global search capabilities. The maximum number of iterations set based on experimental experience ensures that an executable allocation scheme is output within a reasonable time.

[0116] Furthermore, the particle swarm is initialized, the optimal solution is output and updated through domain operations, SA perturbation and DE-GA mutation are performed, and the optimal logistics path and scheduling scheme are generated, including:

[0117] Based on the comprehensive transportation cost (obtained from logistics platform data) and the weighted average satisfaction, a multi-objective function is defined, with the following formula:

[0118] OB = ψ 1 ·CV+ψ 2 ·(1-SV),

[0119] Where OB is the multi-objective function value, CV is the overall transportation cost, and ψ 1 and ψ 2The weights for comprehensive transportation cost and weighted average satisfaction are respectively, and their sum is 1;

[0120] The delivery time window (based on the final predicted demand), delivery distance, load (based on logistics platform data), and allocation matrix are set as constraints to initialize the particle swarm. Each particle is defined as an initial scheduling scheme, including vehicle allocation and delivery route. A random number generator is used for random initialization to obtain a randomly initialized particle swarm.

[0121] Based on a randomly initialized particle swarm, the initial multi-objective function value is calculated. Through position and velocity updates, and using the early stopping method, the optimal individual position and the global optimal position are obtained, as shown in the following formula:

[0122] V i (t+1)=Φ·V i (t)+Υ 1 ·rand()·(PO best -X i (t))+Υ 2 rand() (GO best -X i (t)),

[0123] X i (t+1)=X i (t)+V i (t+1),

[0124] Among them, V i (t+1) is the velocity vector of particle i at time t+1, PO best For the optimal position of an individual, X i (t) is the position vector of particle i at time t, and rand() is a random number between 0 and 1. best The global optimal position is given by Φ, where Φ is the inertia weight and Υ is the weight. 1 and Υ 2 These are individual learning factors and global learning factors, respectively.

[0125] The initial multi-objective function values ​​are sorted in descending order, and the particles corresponding to the top K multi-objective function values ​​are selected. Based on the Top-K selection method, a local guided search is performed, including swapping or redistribution. Through domain operations and guidance strategies, the maximum number of iterations is set based on empirical rules, and the optimal domain solution is iteratively output.

[0126] Replace the optimal neighborhood scheme with the initial scheduling scheme of the corresponding particles, and output the updated particle swarm;

[0127] Based on the goal orientation, the updated particle swarm is initially divided into a satisfaction seed group, a cost seed group, and a reward seed group.

[0128] The satisfaction seed group refers to the order time sorted in descending order and the top C selected; the cost seed group refers to the path cost sorted in ascending order and the top E selected; and the reward seed group refers to the initial multi-objective function value sorted in ascending order and the top N selected. Here, C, E, and N are the updated particle numbers, set by a fixed ratio method.

[0129] Based on the satisfaction seed group, the corresponding initial multi-objective function values ​​are sorted in ascending order, the top M values ​​are selected, and new particles are generated by SA perturbation based on the threshold truncation method.

[0130] The SA perturbation includes initializing the initial temperature and cooling rate, exchanging or redistributing two orders on the same path through random perturbation, stopping after reaching the number of iterations, outputting perturbation particles based on the fixed budget method, and calculating their corresponding multi-objective function values;

[0131] Based on the multi-objective function value, the acceptance probability is calculated using the following formula:

[0132] P = exp(-Δ / TU),

[0133] Where P is the acceptance probability, Δ is the difference between the multi-objective function value of the perturbed particle and the initial multi-objective function value, and TU is the initial temperature;

[0134] If the acceptance probability is greater than the acceptance probability threshold, then the perturbation particle is accepted; otherwise, the original particle is retained and the perturbation particle is output.

[0135] Based on the cost seed group, the corresponding multi-objective function values ​​are sorted in ascending order, and the top O values ​​are selected. Adaptive differential weights are calculated based on the adaptive scaling method, using the following formula:

[0136]

[0137] Where F is the adaptive differential weight, DW is the initial multi-objective function value of the first O particles, AR is the mean of the initial multi-objective function values ​​of the first O particles, and Λ 1 Based on the weights, Λ 2 This is the scaling factor;

[0138] Based on adaptive differential weights, DE-GA mutation is performed on the cost seed group to obtain the positions of the mutated particles, as shown in the formula:

[0139] X′ i =X″ i +F·(X″ Π1 -X″ Π2 )+0.1·(X″ best -X″ i ),

[0140] Where, X′ i Let X″ be the position of particle i after mutation. i Let X″ be the current position of particle i. Π2 and X″ Π1 X″ represents the current positions of particles Π2 and Π1 in the first O particles, respectively. best The globally optimal particle;

[0141] If the position of the mutated particle is greater than the threshold for the position of the mutated particle, then accept the mutated particle; otherwise, retain the original particle and output "accept mutated particle".

[0142] Replace the corresponding screening particles with the perturbation particles and the mutation particles to generate a new particle swarm.

[0143] Based on the new particle swarm optimization, multi-objective function values ​​are calculated. Through velocity and position updates, the optimal individual position and the new global optimal position are output, generating the optimal logistics path and scheduling scheme, which are then distributed via API interface.

[0144] The particle swarm is initialized using a random number generator, and each particle is modeled as an initial scheduling scheme including vehicle allocation and delivery routes. A multi-objective function is introduced, comprehensively considering logistics costs and weighted average satisfaction, to achieve scheduling goals more aligned with actual business needs. Velocity and position vector update formulas guide particles towards individual and global optima, gradually narrowing the solution space while maintaining global search efficiency and improving convergence. Early stopping is introduced to avoid overfitting, and fine-grained optimization through neighborhood redistribution and exchange operations significantly improves the local adaptability and feasibility of the solution, providing a high-quality initial structure for the global solution. A three-class classification logic—satisfaction seed group, cost seed group, and reward subgroup—combined with a fixed ratio method and a threshold truncation method, enhances the algorithm's hierarchical response capability in multi-objective scenarios, enabling the system to adapt to actual priority requirements. To achieve more targeted optimization, an SA perturbation mechanism is introduced in the satisfaction-oriented subgroup, effectively overcoming the shortcomings of standard PSO which is prone to getting trapped in local optima. At the same time, it ensures that the perturbation particles have interpretability and stability under the control of acceptance probability, and improves the diversity of the search process. In the cost-oriented subgroup, adaptive differential weights are used for mutation operations, introducing local differences and global optimum guidance to improve the speed of approaching the global minimum of the multi-objective function. Through a mutation particle position threshold screening mechanism, the accuracy and practicality of mutation operations are ensured. In each iteration, the multi-objective function value is recalculated through the updated new particle swarm, and the final individual optimum and global optimum are iteratively optimized to ensure that the logistics path and scheduling scheme can dynamically respond to changes in demand and environmental disturbances. Finally, the scheduling results are issued through the API interface, which has good system integration and real-time deployment capabilities.

[0145] This embodiment also provides a sales supply chain management system based on Internet data, including:

[0146] The prediction collection module is used to collect and preprocess multi-source data, generate normalized multi-source data, decompose it through discrete wavelet transform, extract detail coefficients and mark abrupt change points, construct a three-dimensional feature vector, perform time alignment using eDTW distance, calculate the approximate kernel distance through Gaussian random projection, generate low-dimensional embeddings through low-rank decomposition and ADMM iterative optimization, initialize the prediction requirements of the Poisson distribution and combine them with the Gamma prior, calculate the prediction requirements through the posterior distribution and the prediction distribution, quantize it using the objective function and Shapley value, and generate the final prediction requirements.

[0147] The scheduling optimization module is used to calculate the weighted average satisfaction based on multi-source data. It obtains the final allocation matrix through the PSG objective function and stochastic gradient, defines a multi-objective function based on the weighted average satisfaction, sets constraints in combination with the final prediction requirements, initializes the particle swarm, determines the individual and global optimal positions through position and velocity updates, performs local guided search, outputs the best solution through neighborhood operations and updates the particle swarm, further seeds the updated particle swarm, performs SA perturbation and DE-GA mutation respectively, generates the optimal logistics path and scheduling scheme, and distributes them through the API interface.

[0148] This embodiment also provides a computer device applicable to the sales supply chain management method based on Internet data, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the sales supply chain management method based on Internet data as proposed in the above embodiment.

[0149] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0150] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the sales supply chain management method based on Internet data as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0151] In summary, this invention decomposes the data using discrete wavelet transform, calculates the approximate kernel distance using Gaussian random projection, generates low-dimensional embeddings through low-rank decomposition and ADMM iterative optimization, initializes the prediction requirements based on the Poisson distribution, quantizes the data using Shapley values, initializes the particle swarm optimization, obtains the individual and global optimal positions, performs local guided search, updates the data by outputting the optimal solution through neighborhood operations, and generates the optimal logistics path and scheduling scheme through SA perturbation and DE-GA mutation. This improves prediction accuracy and enhances the adaptability and operational efficiency of the sales supply chain in dynamic market environments.

[0152] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A sales supply chain management method based on internet data, characterized by: include, Multi-source data is collected and preprocessed to generate normalized multi-source data. It is then decomposed using discrete wavelet transform, detail coefficients are extracted and abrupt change points are marked, a three-dimensional feature vector is constructed, time alignment is performed using eDTW distance, approximate kernel distance is calculated using Gaussian random projection, low-dimensional embedding is generated through low-rank decomposition and ADMM iterative optimization, the prediction requirement is initialized with Poisson distribution and combined with Gamma prior, the prediction requirement is calculated using posterior distribution and prediction distribution, quantization is performed using objective function and Shapley value, and the final prediction requirement is generated. The weighted average satisfaction is calculated based on multi-source data. The final allocation matrix is ​​obtained through the PSG objective function and stochastic gradient. A multi-objective function is defined based on the weighted average satisfaction and constraints are set in combination with the final prediction requirements. The particle swarm is initialized. The optimal positions of individuals and the global optimal positions are determined through position and velocity updates. Local guided search is performed. The best solution is output through neighborhood operation and the particle swarm is updated. The updated particle swarm is seeded and subjected to SA perturbation and DE-GA mutation respectively to generate the optimal logistics path and scheduling scheme. The scheme is then distributed through the API interface. The multi-source data includes real-time sales data, user discussion data, timeliness data, and macroeconomic indicator data; The initial particle swarm optimization (PSO) process outputs the optimal solution and updates the PSO through neighborhood operations, performs SA perturbation and DE-GA mutation, and generates the optimal logistics path and scheduling scheme, including: Based on the final allocation matrix and weighted average satisfaction, a multi-objective function is defined and constraints are set. The particle swarm is initialized, the initial multi-objective function value is calculated, sorted in descending order, and the particles corresponding to the top K multi-objective function values ​​are selected. Local guided search is performed, and the best neighborhood scheme is iteratively output through neighborhood operations and guidance strategies. The initial scheduling scheme of the corresponding particles is replaced, the updated particle swarm is output, and initial partitioning is performed, including satisfaction seed group, cost seed group and reward seed group. Based on the satisfaction seed group, the corresponding initial multi-objective function values ​​are sorted in ascending order, the top M values ​​are selected, and perturbation particles are generated through SA perturbation. Based on the cost seed group, adaptive differential weights are calculated, and DE-GA mutation is performed to generate mutated particles. The corresponding selected particles are replaced by perturbation particles and mutation particles to generate a new particle swarm. The multi-objective function value is calculated to generate the optimal logistics path and scheduling scheme, which is then distributed through the API interface.

2. The sales supply chain management method based on Internet data as described in claim 1, characterized in that: The decomposition via discrete wavelet transform and time alignment using eDTW distance include: Based on normalized multi-source data, discrete wavelet transform is used for decomposition, and the mean of detail coefficients is calculated. If the detail coefficients are greater than the mean of detail coefficients, abrupt change points are marked, a three-dimensional feature vector is constructed, the eDTW distance is calculated, and the normalized multi-source data is aligned on the time axis through linear interpolation to generate an aligned sequence.

3. The sales supply chain management method based on Internet data as described in claim 2, characterized in that: The initialization of the Poisson distribution prediction requirement, combined with the Gamma prior, is quantized using the objective function and Shapley value to generate the final prediction requirement, including: Based on the aligned sequence, it is concatenated into a high-dimensional feature matrix. A random projection matrix is ​​generated by Gaussian random projection, and the approximate kernel distance is calculated to generate a distance matrix. Low-rank decomposition is performed by LRI to obtain a low-rank coefficient matrix. ADMM is used for iterative optimization to output a low-dimensional embedding. The initial prediction demand follows a Poisson distribution, and Gamma is set as a prior. The posterior distribution is then calculated to obtain the prediction distribution. Based on the predicted distribution, the mean is calculated to obtain the predicted demand. An objective function is defined, the Shapley value is calculated, and the final predicted demand is obtained through multiplication.

4. The sales supply chain management method based on Internet data as described in claim 3, characterized in that: The calculation of the weighted average satisfaction rate based on multi-source data includes: A satisfaction model is defined based on normalized delivery waiting time, normalized rating stars, and positive and negative evaluations, and a weighted average satisfaction rate is calculated.

5. The sales supply chain management method based on Internet data as described in claim 4, characterized in that: The process of obtaining the final allocation matrix through the PSG objective function and stochastic gradient includes: Define a set of constraints based on the final predicted demand, define the PSG objective function, and calculate the stochastic gradient; Initialize the allocation matrix, update the allocation matrix using stochastic gradient descent, and output the final allocation matrix.

6. The sales supply chain management method based on Internet data as described in claim 5, characterized in that: The process of collecting and preprocessing multi-source data to generate normalized multi-source data includes: Multi-source data is collected through API interfaces and preprocessed to generate standardized multi-source data. Normalized multi-source data is then generated through time series normalization.

7. A sales supply chain management system based on internet data, used to implement the sales supply chain management method based on internet data as described in any one of claims 1 to 6, characterized in that: include, The prediction collection module is used to collect and preprocess multi-source data, generate normalized multi-source data, decompose it through discrete wavelet transform, extract detail coefficients and mark abrupt change points, construct a three-dimensional feature vector, perform time alignment using eDTW distance, calculate the approximate kernel distance through Gaussian random projection, generate low-dimensional embeddings through low-rank decomposition and ADMM iterative optimization, initialize the prediction requirements of the Poisson distribution and combine them with the Gamma prior, calculate the prediction requirements through the posterior distribution and the prediction distribution, quantize it using the objective function and Shapley value, and generate the final prediction requirements. The scheduling optimization module is used to calculate the weighted average satisfaction based on multi-source data. It obtains the final allocation matrix through the PSG objective function and stochastic gradient, defines a multi-objective function based on the weighted average satisfaction, sets constraints in combination with the final prediction requirements, initializes the particle swarm, determines the individual and global optimal positions through position and velocity updates, performs local guided search, outputs the best solution through neighborhood operations and updates the particle swarm, seeds the updated particle swarm, performs SA perturbation and DE-GA mutation respectively, generates the optimal logistics path and scheduling scheme, and distributes them through the API interface.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the sales supply chain management method based on Internet data as described in any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the sales supply chain management method based on Internet data as described in any one of claims 1 to 6.

Images