CN-121563398-B - AI-based sales prediction and intelligent goods supplementing method and system

CN121563398BCN 121563398 BCN121563398 BCN 121563398BCN-121563398-B

Abstract

The invention provides an AI-based sales volume prediction and intelligent replenishment method and system, which relate to the field of sales management and comprise the steps of obtaining historical sales volume and commodity characteristic data, constructing and optimizing a sales volume prediction model, calculating a confidence interval boundary of a predicted value and a dynamic safety stock standard, determining replenishment demands by superposing compensation volumes when a basic gap is lower than the safety stock standard, and starting a multi-provider parallel replenishment mechanism according to provider delay coefficients. The invention realizes accurate sales volume prediction and efficient inventory management, reduces inventory risk and improves replenishment efficiency.

Inventors

BAO YILEI
WANG HANFU
Zhu Wufei
CHEN YAOHUI
DING XIANGLONG
LI LAN
XU XIAOPING

Assignees

杭州湖畔网络技术有限公司

Dates

Publication Date: 20260512
Application Date: 20260122

Claims (9)

1. AI-based sales prediction and intelligent replenishment method, characterized by comprising: Acquiring historical sales volume data, commodity characteristic data and inventory state data of a target commodity, performing time sequence decomposition on the historical sales volume data to obtain sales volume influence factors, and constructing an initial prediction model based on the sales volume influence factors and the commodity characteristic data; Optimizing a tree search space constructed by parameters of the initial prediction model, and backtracking to an optimal parameter combination when the continuous N times of prediction errors exceed a preset error threshold value to obtain a sales prediction model after training is completed; Inputting commodity characteristic data into a sales volume prediction model after training to obtain sales volume predicted values, and calculating confidence interval boundaries of the sales volume predicted values based on predicted deviation distribution of historical sales volume data; Calculating a dynamic safety stock standard based on a confidence interval boundary, taking the difference value between the sales volume predicted value and the current stock quantity in stock state data as a basic gap, superposing a safety compensation quantity when the basic gap is lower than the dynamic safety stock standard, and combining the historical sales volume data to obtain a final replenishment demand; Calculating historical supply delay coefficients of suppliers, starting a multi-supplier parallel replenishment mechanism when the historical supply delay coefficients exceed a preset delay threshold, performing proportional distribution on final replenishment demands, generating purchase order data of a plurality of staggered times, and outputting a purchase execution instruction; carrying out time sequence decomposition on the historical sales volume data to obtain sales volume influence factors, and constructing an initial prediction model based on the sales volume influence factors and commodity characteristic data comprises the following steps: Calculating fluctuation variance of the historical sales volume data under different time granularities, and determining the fluctuation variance ratio of adjacent time granularities as the optimal decomposition granularity when the fluctuation variance ratio of the adjacent time granularities is smaller than a preset ratio threshold; performing wavelet decomposition on the historical sales data under the optimal decomposition granularity to obtain a plurality of frequency band components, calculating the energy value of each frequency band component, sorting the frequency band components from high to low according to the energy value, calculating the energy ratio between adjacent frequency band components after sorting, and merging the frequency band components with the energy ratio smaller than a preset threshold into a new frequency band component; calculating the energy duty ratio of the new frequency band component as a reorganization weight, and carrying out weighted combination on the new frequency band component through the reorganization weight to obtain a sales volume influence factor; extracting attribute features from the commodity feature data, calculating a correlation coefficient of the attribute features and sales volume influence factors, carrying out weighted combination on the attribute features and the sales volume influence factors based on the correlation coefficient, constructing a feature sample matrix, calculating variance contribution rates of all features in the feature sample matrix, and selecting features with variance contribution rates higher than a preset contribution threshold to construct an optimized feature set; and mapping the optimized feature set to an input layer node, mapping a prediction time window to an output layer node, and constructing an initial prediction model.
2. The method of claim 1, wherein optimizing the parameter construction tree search space of the initial predictive model and backtracking to an optimal parameter combination when the continuous N number of prediction errors exceeds a preset error threshold, the obtaining the trained sales predictive model comprises: Dividing training parameters of an initial prediction model into a plurality of layers, calculating information entropy gains between adjacent layers to obtain layer dependency strength, and constructing layer nodes and connection relations in a tree-like search space according to the layer dependency strength; Searching training parameters layer by layer from a root node of a hierarchical node, calculating prediction errors, extracting N times of corresponding training parameter vectors when the continuous N times of prediction errors exceed a preset error threshold value, calculating a nonlinear coupling degree matrix among the training parameter vectors by utilizing core principal component analysis, and combining the training parameters with coupling degrees exceeding the preset coupling threshold value into a parameter coupling group based on the nonlinear coupling degree matrix; Creating a coupling sub-node under the current level node, constructing an adaptive search step size matrix based on the nonlinear coupling degree matrix, and performing joint optimization on the parameter coupling group by using the adaptive search step size matrix; mapping the training parameter vector into a failure point cloud, and extracting the principal component direction of the failure point cloud; And (3) backtracking the training parameter combination with the minimum prediction error along the opposite direction of the main component direction as an optimal parameter combination, and continuing to optimize until the prediction error is smaller than a preset error threshold value, so as to obtain the sales quantity prediction model after training is completed.
3. The method of claim 1, wherein inputting the merchandise feature data into the trained sales prediction model to obtain sales predictions, calculating confidence interval boundaries for the sales predictions based on predicted deviation distributions of historical sales data comprises: inputting commodity characteristic data into a sales volume prediction model after training is completed to obtain sales volume predicted values; Extracting predicted deviation distribution of historical sales data, constructing two-dimensional scattered point distribution by using predicted deviation values in the predicted deviation distribution and corresponding historical predicted values, performing nuclear density estimation on the two-dimensional scattered point distribution to obtain density field distribution, identifying boundary points with density gradients exceeding a preset gradient threshold value in the density field distribution, and connecting the boundary points to form a deviation boundary curve; extracting a predicted deviation value and a historical predicted value corresponding to the boundary point along the deviation boundary curve, dividing the deviation boundary curve into a plurality of curve segments according to the numerical range of the historical predicted value, and calculating the range of the predicted deviation value on each curve segment as the fluctuation amplitude of the curve segment; extracting a history predicted value range corresponding to each curve segment as a predicted value interval, and establishing a mapping relation between the predicted value interval and the curve segment fluctuation amplitude to obtain an interval fluctuation mapping table; searching a predicted value interval to which the sales predicted value belongs in an interval fluctuation mapping table, extracting curve segment fluctuation amplitude corresponding to the predicted value interval, and generating a confidence interval boundary by taking the sales predicted value as a center and the curve segment fluctuation amplitude as an offset.
4. The method of claim 1, wherein calculating a dynamic safety inventory benchmark based on confidence interval boundaries, taking a difference between the sales forecast and a current inventory in the inventory status data as a base gap, superimposing a safety compensation when the base gap is below the dynamic safety inventory benchmark, and combining the historical sales data to obtain the final replenishment demand comprises: Constructing a double-layer time window for the historical sales volume data, calculating a fluctuation characteristic value sequence and a trend inflection point through the double-layer time window, carrying out subsection correction on a confidence interval boundary of a sales volume predicted value, taking the ratio of the corrected confidence interval boundary to the sales volume predicted value as a dynamic risk factor, and determining a dynamic safety stock reference according to the product of the dynamic risk factor and a sales volume standard deviation; Calculating the difference value between the sales quantity predicted value and the current stock quantity to obtain a basic gap, combining the basic gap with a dynamic risk factor to construct an inventory state vector, constructing a dynamic state transition map based on a historical stock-out record, inputting the inventory state vector into the dynamic state transition map to obtain an evolution path of the inventory state, and predicting a risk level; When the basic gap is lower than the dynamic safety stock reference, mapping the stock state vector and the risk level to a multidimensional compensation space, constructing an optimal compensation path based on historical cost benefit data in the multidimensional compensation space, and calculating along the optimal compensation path to obtain a safety compensation amount; weighting the historical sales volume data by taking the dynamic risk factors as weight coefficients, and extracting seasonal features from the weighted historical sales volume data to obtain seasonal adjustment quantity; and adding the basic notch, the safety compensation quantity and the seasonal adjustment quantity to obtain the final replenishment demand.
5. The method of claim 1, wherein constructing a dynamic state transition map based on the historical backorder records, inputting the inventory state vector into the dynamic state transition map to obtain an evolution path of the inventory state, and predicting the risk level comprises: Extracting inventory state data from the historical stock-out record, performing time sequence grouping on the inventory state data, and connecting the grouped inventory state data in time sequence to form a state evolution sequence; Performing density clustering on the state evolution sequence to obtain a state transition mode, marking different inventory states as state nodes in the state transition mode, calculating transition probability and transition time length between the state nodes, and constructing a dynamic state transition map by taking the state nodes as map nodes and the transition probability and the transition time length as connection attributes between the nodes; inputting an inventory state vector into the dynamic state transition map, searching a target node with highest similarity with the inventory state vector in state nodes of the dynamic state transition map, taking the target node as a starting point, extracting a multi-state evolution path along the inter-node connection attribute, and calculating path weight according to transition probability in the multi-state evolution path; and calculating the historical backorder probability corresponding to each state node in the state evolution path, combining the historical backorder probability with the path weight to obtain a risk evolution trend, and predicting the risk level according to the risk evolution trend.
6. The method of claim 1, wherein calculating a historical supply delay factor for the suppliers, activating a multi-supplier parallel replenishment mechanism when the historical supply delay factor exceeds a preset delay threshold, proportioning final replenishment demands, generating a plurality of staggered time purchase order data, and outputting a purchase execution instruction comprises: Acquiring a historical supply record of a supplier, carrying out weighting treatment on the historical supply record according to time sequence, extracting the actual arrival time of each order, subtracting the appointed time to obtain time deviation, and carrying out exponential weighting calculation on the time deviation to obtain a historical supply delay coefficient; when the historical supply delay coefficient exceeds a preset delay threshold, supply data and available capacity of alternative suppliers are extracted from a qualified supplier library; Calculating historical delay coefficients of alternative suppliers, carrying out normalization processing to obtain a supply timeliness score, calculating a supply capacity score based on the ratio of the supply timeliness score to the available capacity of each supplier, and selecting a plurality of suppliers meeting the final replenishment demand as a replenishment supplier combination according to the order of the supply capacity score from high to low; Distributing the final replenishment demands according to the supply capacity scores of all suppliers in the replenishment supplier combination to obtain order distribution quantity, and verifying whether the order distribution quantity exceeds the upper limit of the available capacity of all suppliers; and acquiring target arrival time of the final replenishment demand, calculating differentiated order advance periods according to historical supply delay coefficients of all suppliers, generating purchase order data comprising order allocation amount and order time, and outputting a purchase execution instruction.
7. AI-based sales prediction and intelligent restocking system for implementing the method of any of the preceding claims 1-6, comprising: the first unit is used for acquiring historical sales volume data, commodity characteristic data and inventory state data of the target commodity, carrying out time sequence decomposition on the historical sales volume data to obtain sales volume influence factors, and constructing an initial prediction model based on the sales volume influence factors and the commodity characteristic data; The second unit is used for optimizing the tree-shaped search space of parameter construction of the initial prediction model, and backtracking to an optimal parameter combination when the continuous N times of prediction errors exceed a preset error threshold value to obtain a sales prediction model after training is completed; The third unit is used for inputting the commodity characteristic data into the sales volume prediction model after training to obtain sales volume predicted values, and calculating confidence interval boundaries of the sales volume predicted values based on predicted deviation distribution of the historical sales volume data; A fourth unit, configured to calculate a dynamic safety inventory benchmark based on the confidence interval boundary, take a difference value between the sales volume predicted value and the current inventory volume in the inventory state data as a basic gap, superimpose a safety compensation volume when the basic gap is lower than the dynamic safety inventory benchmark, and obtain a final replenishment demand by combining the historical sales volume data; and a fifth unit for calculating the historical supply delay coefficient of the suppliers, starting the multi-supplier parallel replenishment mechanism when the historical supply delay coefficient exceeds a preset delay threshold value, performing proportional distribution on the final replenishment demands, generating purchasing order data with a plurality of staggered times and outputting purchasing execution instructions.
8. An electronic device, comprising: A processor; A memory for storing processor-executable instructions; Wherein the processor is configured to invoke the instructions stored in the memory to perform the method of any of claims 1 to 6.
9. A computer readable storage medium having stored thereon computer program instructions, which when executed by a processor, implement the method of any of claims 1 to 6.

Description

AI-based sales prediction and intelligent goods supplementing method and system Technical Field The invention relates to a sales management technology, in particular to an AI-based sales volume prediction and intelligent goods supplementing method and system. Background In the field of retail and supply chain management, accurate sales volume prediction and efficient inventory replenishment are key links to ensure smooth operation of business. With the rapid development of artificial intelligence technology, AI-based sales prediction and intelligent replenishment methods are becoming industry research hotspots. Traditional sales predictions rely primarily on statistical methods and rules of thumb, such as moving average, exponential smoothing, etc., while restocking decisions are often based on fixed safety stock levels and empirical decisions. The existing AI-based sales volume prediction and intelligent replenishment methods still have some obvious defects and shortcomings, most of the existing models lack an effective parameter self-optimization mechanism, and when facing market fluctuation or abnormal conditions, the prediction models are difficult to adjust parameters in time to adapt to changes, so that prediction accuracy is reduced, and the replenishment decision accuracy is affected. The existing sales volume prediction method generally only provides point estimation results, lacks quantitative analysis on prediction uncertainty, cannot provide confidence interval information, makes a replenishment decision difficult to cope with sales fluctuation risks and easily causes overstock or shortage situations, and the traditional replenishment system often adopts a single supplier mode, fails to fully consider supply delay risks of suppliers, and cannot timely start alternative schemes when the suppliers have delay delivery situations, increases the shortage risks and influences normal sales and customer satisfaction. Disclosure of Invention The embodiment of the invention provides an AI-based sales prediction and intelligent goods supplementing method and system, which can solve the problems in the prior art. In a first aspect of the embodiments of the present invention, there is provided an AI-based sales prediction and intelligent replenishment method, including: Acquiring historical sales volume data, commodity characteristic data and inventory state data of a target commodity, performing time sequence decomposition on the historical sales volume data to obtain sales volume influence factors, and constructing an initial prediction model based on the sales volume influence factors and the commodity characteristic data; Optimizing a tree search space constructed by parameters of the initial prediction model, and backtracking to an optimal parameter combination when the continuous N times of prediction errors exceed a preset error threshold value to obtain a sales prediction model after training is completed; Inputting commodity characteristic data into a sales volume prediction model after training to obtain sales volume predicted values, and calculating confidence interval boundaries of the sales volume predicted values based on predicted deviation distribution of historical sales volume data; Calculating a dynamic safety stock standard based on a confidence interval boundary, taking the difference value between the sales volume predicted value and the current stock quantity in stock state data as a basic gap, superposing a safety compensation quantity when the basic gap is lower than the dynamic safety stock standard, and combining the historical sales volume data to obtain a final replenishment demand; And calculating historical supply delay coefficients of suppliers, starting a multi-supplier parallel replenishment mechanism when the historical supply delay coefficients exceed a preset delay threshold, performing proportional distribution on the final replenishment demands, generating purchase order data of a plurality of staggered times, and outputting a purchase execution instruction. Carrying out time sequence decomposition on the historical sales volume data to obtain sales volume influence factors, and constructing an initial prediction model based on the sales volume influence factors and commodity characteristic data comprises the following steps: Calculating fluctuation variance of the historical sales volume data under different time granularities, and determining the fluctuation variance ratio of adjacent time granularities as the optimal decomposition granularity when the fluctuation variance ratio of the adjacent time granularities is smaller than a preset ratio threshold; performing wavelet decomposition on the historical sales data under the optimal decomposition granularity to obtain a plurality of frequency band components, calculating the energy value of each frequency band component, sorting the frequency band components from high to low according to the energy value, calculating