A warehouse intelligent replenishment scheduling system

Abstract

The application relates to the technical field of warehouse management, and discloses a warehouse intelligent replenishment scheduling system, which comprises a heat analysis unit, a demand correction unit, an allocation optimization unit, an evaluation unit and a scheduling generation unit. Heat pulse sequences are generated through multi-platform real-time data feature extraction, time effectiveness weighting and trend abnormality decomposition, platform heat dynamics are accurately tracked, historical replenishment tasks and heat pulses are fused, virtual inventory allocation is adjusted in real time, replenishment urgency is accurately evaluated, scheduling priorities are constructed based on the urgency, position information and inventory health indexes, optimal cross-warehouse scheduling instruction sets are generated with the lowest cost and the fastest time effectiveness as targets, warehouse operation benefits are improved, real-time tracking of platform heat is achieved, replenishment quotas are adjusted in a timely manner, and real-time scheduling of the warehouse is adapted to new market conditions.

Description

Technical Field

[0001] This invention relates to the field of warehouse management technology, specifically to an intelligent warehouse replenishment scheduling system. Background Technology

[0002] In modern warehouse management, big data analytics and machine learning algorithms are used to accurately predict future demand for goods based on historical sales data, market trends, promotional activities, and other influencing factors. To maximize warehouse operational efficiency, multiple objectives such as transportation costs, warehousing costs, stockout costs, and time window constraints are considered. The optimal replenishment strategy is found through optimization algorithms. Once the optimal replenishment strategy is determined, the system can automatically generate replenishment instructions and schedule and allocate goods in each warehouse, thereby achieving intelligent replenishment.

[0003] However, as the sales model shifts towards unified sales across multiple platforms, existing warehouse replenishment scheduling systems have revealed significant shortcomings. Specifically, although these systems can predict future demand using big data and machine learning and formulate replenishment plans accordingly, their ability to track platform popularity in real time is severely inadequate in the face of dynamic changes in platform popularity across multiple sales scenarios. This results in the system being unable to adjust replenishment quotas in a timely manner when platform popularity changes, causing real-time warehouse scheduling to be unable to adapt to new market conditions. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an intelligent warehouse replenishment scheduling system that solves the aforementioned problems.

[0005] The above-mentioned technical objective of the present invention is achieved through the following technical solution:

[0006] A warehouse intelligent replenishment scheduling system includes:

[0007] The heat analysis unit is used to determine the heat pulse sequence based on real-time data from multiple platforms;

[0008] The demand correction unit is used to obtain the historical replenishment tasks of the target object, and fuse the historical replenishment tasks and the heat pulse sequence to obtain the real-time demand correction coefficient.

[0009] The allocation optimization unit is used to calculate the dynamic platform quota weight based on the real-time demand correction coefficient and real-time data, and adjust the virtual inventory allocation target of each target object to different platforms in real time according to the dynamic platform quota weight.

[0010] The evaluation unit is used to calculate the inventory turnover rate of the target object, analyze the inventory turnover rate and the virtual inventory allocation target, and determine the replenishment urgency.

[0011] The scheduling generation unit is used to obtain the location information of the target object and, based on the replenishment urgency and location information, obtain the dynamic optimal cross-warehouse scheduling instruction set.

[0012] Furthermore, the heat pulse sequence is determined based on real-time data from multiple platforms, including:

[0013] Extract features from real-time data to generate cross-platform raw feature datasets;

[0014] Based on the interval between the timestamps generated by the data in the cross-platform original feature dataset and the current time, the data in different time periods are weighted according to their timeliness to obtain a timeliness-corrected feature matrix.

[0015] The influence weight vector of different platforms is determined based on real-time data from multiple platforms.

[0016] The timeliness-corrected feature matrix is ​​decomposed to obtain the trend. Abnormal dual-track feature vector;

[0017] Influence weight vector and trend across different platforms The abnormal dual-track feature vectors are fused to generate a heat pulse sequence that reflects real-time heat fluctuations.

[0018] Furthermore, the timeliness-corrected feature matrix is ​​decomposed to obtain the trend. The abnormal dual-track feature vector includes:

[0019] The timeliness-corrected feature matrix is ​​initially decomposed to obtain the basic feature component matrix and the corresponding singular value vector.

[0020] Based on the basic feature component matrix and singular value vector, the data for each time dimension is fitted to generate a trend feature vector;

[0021] The abnormal fluctuation feature vector is obtained by calculating the timeliness correction feature matrix and the trend feature vector;

[0022] By combining the trend feature vector and the abnormal fluctuation feature vector, the trend can be obtained. Abnormal dual-track feature vector.

[0023] Furthermore, historical replenishment tasks and heatwave pulse sequences are fused to obtain real-time demand correction coefficients, including:

[0024] The historical replenishment tasks of the target object are decomposed in time series to obtain a multi-dimensional replenishment feature vector;

[0025] The heat pulse sequence is divided into grids according to the spatiotemporal dimension to generate a spatiotemporal heat matrix;

[0026] The time window width is adaptively adjusted based on the execution cycle of historical replenishment tasks, and the synchronization fluctuation coefficient between historical replenishment tasks and heat pulse sequences is calculated.

[0027] By analyzing the multidimensional replenishment feature vector and the spatiotemporal heat matrix, the historical influence weight vector and the heat correction weight vector are obtained. The historical influence weight vector, the heat correction weight vector and the synchronous fluctuation coefficient are then fused to obtain the real-time demand correction coefficient.

[0028] Furthermore, based on the real-time demand correction coefficient and real-time data, the dynamic platform quota weight is calculated, including:

[0029] The cost sensitivity matrix is ​​obtained by identifying real-time data from each platform.

[0030] Based on the real-time demand correction coefficient and historical replenishment tasks, the elastic demand threshold for each platform is calculated.

[0031] With the goals of minimizing total cost and maximizing demand satisfaction, the replenishment volume of each platform is dynamically optimized using the cost sensitivity matrix and elastic demand threshold as constraints, forming a Pareto optimal solution set.

[0032] The Pareto optimal solution set is solved based on the cost sensitivity matrix to generate the initial platform quota vector;

[0033] Obtain inventory data for each target object, analyze the inventory data, and derive the inventory health index;

[0034] The initial platform quota vector is adjusted based on the inventory health index to generate dynamic platform quota weights.

[0035] Furthermore, by solving the Pareto optimal solution set based on the cost sensitivity matrix, an initial platform quota vector is generated, including:

[0036] Based on the cost sensitivity matrix, a response surface between replenishment volume and total cost for each platform is constructed to generate a marginal cost gradient matrix.

[0037] Optimize the Pareto optimal solution set to generate a weighted Pareto front.

[0038] Calculate the overall utility value of each point on the weighted Pareto front, select the point that maximizes utility as the optimal solution, and use the replenishment volume ratio of each platform corresponding to that point as the initial platform quota vector.

[0039] Furthermore, the virtual inventory allocation targets for each target object across different platforms are adjusted in real time based on the dynamic platform quota weights, including:

[0040] Obtain the virtual inventory allocation target that should be allocated to each platform in each target object;

[0041] Determine the maximum capacity coefficient per unit time based on inventory data;

[0042] Based on dynamic platform quota weights and real-time data from multiple platforms, the urgency index of demand for each platform is analyzed.

[0043] The virtual inventory allocation adjustment coefficient is obtained by calculating the demand urgency index and maximum capacity coefficient of each platform.

[0044] The virtual inventory allocation target is adjusted based on the virtual inventory allocation adjustment coefficient to obtain the adjusted virtual inventory allocation target.

[0045] Furthermore, the inventory turnover rate of the target object is calculated, including:

[0046] Analyze the historical replenishment tasks and spatiotemporal heat matrix of the target object to form a spatiotemporal value decay matrix;

[0047] The dynamic inventory equivalent is obtained by weighted aggregation of the target object's inventory quantity and the spatiotemporal value decay matrix.

[0048] Analyze the delay response coefficients of historical replenishment tasks and heat pulse sequences, and combine them with real-time demand correction coefficients to generate demand response efficiency coefficients.

[0049] By integrating dynamic inventory equivalent, demand response efficiency coefficient, and sales cost, a corrected inventory turnover rate is obtained.

[0050] Furthermore, an analysis of inventory turnover rate and platform virtual inventory allocation targets yields replenishment urgency, including:

[0051] The revised inventory turnover rate is aligned with the platform's virtual inventory allocation target in both time and space to generate a multi-dimensional matching matrix.

[0052] A demand pressure index is constructed based on a multidimensional matching degree matrix, real-time demand correction coefficient, and demand urgency indicators of each platform.

[0053] The replenishment elasticity threshold is determined based on historical replenishment tasks, inventory health index, and maximum capacity coefficient.

[0054] The demand pressure index and replenishment elasticity threshold are analyzed to generate a replenishment urgency level.

[0055] Furthermore, based on replenishment urgency and location information, a dynamically optimal cross-warehouse scheduling instruction set is obtained, including:

[0056] Construct a spatial distance weight matrix based on the location information of the target object;

[0057] Based on replenishment urgency, spatial distance weight matrix, and inventory health index, calculate the scheduling priority coefficient for each target object;

[0058] With the goal of minimizing scheduling costs and maximizing delivery time, and by combining scheduling priority coefficients, a dynamic optimal cross-warehouse scheduling instruction set is derived.

[0059] In summary, the present invention has the following main beneficial effects:

[0060] By extracting original features across platforms, combining data timeliness weighting, and then decomposing the matrix to obtain trend anomaly dual-track feature vectors, and integrating platform influence weights, a heat pulse sequence that can accurately reflect real-time heat fluctuations is generated. Compared with traditional systems, this solution can track heat changes on each platform in real time, avoiding the problem of delayed replenishment quotas caused by sudden changes in heat. For example, during e-commerce promotions, if a platform suddenly becomes popular due to its promotional strategy, the heat analysis unit can quickly sense and generate a corresponding heat pulse sequence, providing an accurate basis for subsequent replenishment scheduling and ensuring that warehouse replenishment can adapt to market dynamics in a timely manner.

[0061] By decomposing historical replenishment tasks into time series and combining spatiotemporal analysis of popularity pulse sequences, the system adaptively adjusts the time window width and calculates the synchronization fluctuation coefficient. Finally, it derives a correction coefficient that comprehensively considers historical patterns and real-time popularity. This allows the system to accurately adjust product demand in multi-platform sales scenarios by referencing past experience and combining it with current popularity changes. In complex scenarios such as different seasons and promotional activities, it can effectively avoid replenishment deviations caused by relying solely on historical data or failing to capture real-time popularity, achieving precise matching between demand and replenishment volume and reducing the risk of inventory backlog or stockouts.

[0062] By calculating dynamic platform quota weights based on real-time demand correction coefficients and real-time data, adjusting virtual inventory allocation targets in conjunction with factors such as inventory health index, determining replenishment urgency by assessing inventory turnover rate, and finally generating a dynamic optimal cross-warehouse scheduling instruction set based on location information, the entire process fully considers the complex factors of multi-platform sales and schedules with the goal of lowest cost and fastest delivery. This not only adapts to demand fluctuations caused by changes in platform popularity but also effectively balances transportation and warehousing costs, improves inventory turnover rate, enhances warehousing operation efficiency, and strengthens the company's competitiveness in the multi-platform sales market. Attached Figure Description

[0063] Figure 1 This is a schematic diagram of the intelligent warehouse replenishment scheduling system of the present invention. Detailed Implementation

[0064] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0065] refer to Figure 1 A warehouse intelligent replenishment scheduling system includes:

[0066] The heat analysis unit is used to determine the heat pulse sequence based on real-time data from multiple platforms;

[0067] The demand correction unit is used to obtain the historical replenishment tasks of the target object, fuse the historical replenishment tasks and the heat pulse sequence to obtain the real-time demand correction coefficient, where the target object is the warehouse;

[0068] The allocation optimization unit is used to calculate the dynamic platform quota weight based on the real-time demand correction coefficient and real-time data, and adjust the virtual inventory allocation target of each target object to different platforms in real time according to the dynamic platform quota weight.

[0069] The evaluation unit is used to calculate the inventory turnover rate of the target object, analyze the inventory turnover rate and the virtual inventory allocation target, and determine the replenishment urgency.

[0070] The scheduling generation unit is used to obtain the location information of the target object and, based on the replenishment urgency and location information, obtain the dynamic optimal cross-warehouse scheduling instruction set;

[0071] The real-time data includes: sales revenue, sales orders, sales time, sales location, platform traffic, user page views, add-to-cart volume, favorites volume, search popularity, discount level, event duration, and event scope.

[0072] Historical replenishment tasks include: replenishment time records, replenishment quantity records, replenishment source records, replenishment cost records, and replenishment delay records.

[0073] By combining real-time data from multiple platforms with the heat analysis unit to determine the heat pulse sequence, the bottleneck of real-time tracking in traditional systems is broken. The demand correction unit integrates historical replenishment tasks with the heat pulse sequence. The allocation optimization unit calculates the dynamic platform quota weight and adjusts the virtual inventory allocation target in real time. The evaluation unit analyzes the inventory turnover rate and the virtual inventory allocation target to determine the replenishment urgency. The scheduling generation unit combines location information to generate a dynamic optimal cross-warehouse scheduling instruction set, effectively balancing the needs of multiple platforms, maximizing operational efficiency, and reducing losses caused by untimely replenishment.

[0074] In one embodiment, determining the heat pulse sequence based on real-time data from multiple platforms includes:

[0075] Feature extraction is performed on real-time data to generate a cross-platform raw feature dataset. Specifically, this includes: preprocessing real-time data from each platform; extracting time-domain features using a 5-minute sliding window; decomposing the time series into frequency components using short-time Fourier transform to identify daily, weekly, and monthly periodic fluctuations; standardizing the feature vectors of each platform using Z-score; calculating the cross-platform feature correlation matrix using cosine similarity; clustering features based on the matrix; and weighted fusion of the clustered features to ultimately generate a cross-platform raw feature dataset containing multiple dimensions (time-domain traffic features, frequency-domain periodic features, cross-platform correlation features, inventory correlation features, and comprehensive scheduling features). The dataset includes the following features: time-domain traffic characteristics: peak order volume, average page views, add-to-cart conversion rate fluctuations, collection time, and platform traffic share within a time window; frequency-domain periodic characteristics: daily periodic fluctuation intensity, periodic traffic distribution, and seasonal periodic fluctuation components; cross-platform correlation characteristics: correlation of similar indicators across different platforms, cross-platform conversion rates, interest-based feature clustering, and abnormal traffic; inventory correlation characteristics: historical replenishment response delay rate, inventory turnover rate deviation, virtual inventory consumption rate, and platform delivery timeliness requirements; and comprehensive scheduling characteristics: geographical distance between the target warehouse and each platform, replenishment urgency, and probability of sudden demand surges.

[0076] Based on the interval between the timestamps generated by the centralized data in the cross-platform original feature dataset and the current time, the data in different time periods are weighted according to their timeliness to obtain a timeliness-corrected feature matrix. Specifically, the data is divided into multiple intervals according to the interval between the timestamp and the current time: the most recent 1 hour, 1-6 hours, 6-24 hours, 1-3 days, and more than 3 days. A fixed weight coefficient is assigned to each time interval: the most recent 1 hour has a weight of 0.9, 1-6 hours has a weight of 0.7, 6-24 hours has a weight of 0.5, 1-3 days has a weight of 0.3, and more than 3 days has a weight of 0.1. For each data point in the dataset, its time interval is determined based on its timestamp. All feature values ​​of the data are multiplied by the weight coefficient of the corresponding interval. After weighting all data, all weighted values ​​of the same feature dimension are standardized. First, the sum of all values ​​of that dimension is calculated, and then each value is divided by the sum to obtain the relative importance of the feature after timeliness correction. The above steps are repeated to process all feature dimensions, and finally a timeliness-corrected feature matrix is ​​formed. Each element in the matrix represents the feature value after timeliness weighting and standardization, reflecting the degree of influence of data from different time periods on the current decision.

[0077] The influence weight vector of different platforms is determined based on real-time data from multiple platforms. Specifically, this involves: selecting key indicators that reflect the influence of a platform from the real-time data, including sales indicators, traffic indicators, conversion indicators, search popularity indicators, and promotion indicators. For each key indicator, the data of each platform is divided by the sum of the data of all platforms for that indicator to calculate the proportion of each platform under that indicator. Then, the indicator is assigned a corresponding weight according to the proportion. The proportion of each platform under each key indicator is multiplied by the weight of the corresponding indicator, and the products are added together to obtain the influence score of each platform. The influence scores of all platforms are arranged and combined in order to form the influence weight vector of different platforms.

[0078] The timeliness-corrected feature matrix is ​​decomposed to obtain the trend. Abnormal dual-track feature vector;

[0079] Influence weight vector and trend across different platforms The abnormal dual-track feature vectors are fused to generate a heat pulse sequence reflecting real-time heat fluctuations. Specifically, this includes: merging time... Time The trend component of each platform divided by the first The normalized trend term is obtained by finding the historical maximum value of the absolute value of the trend component of each platform. , time Time The abnormal components of the platform divided by the first The normalized anomaly term is obtained by finding the historical maximum value of the absolute value of the outlier components on each platform. , time Time The real-time volatility of each platform divided by the first The historical maximum volatility of each platform is used to obtain the normalized volatility term. ; platform The normalized trend term, normalized outlier term, and normalized fluctuation term are assigned weights respectively. , and Then, a weighted sum is performed to obtain the linear fusion factor. The linear fusion factor is passed through the hyperbolic tangent function. A nonlinear compression transformation is performed to obtain the platform heat compression term. ; platform The influence weight is multiplied by the platform's popularity compression factor to obtain the platform's corrected popularity contribution. ;Will The final result is obtained by summing the correction popularity contributions from each platform. Global heat pulse sequence In practical applications, the following calculation formula can be used, for example: ;

[0080] In the formula, Indicates time The thermal pulse sequence at that time Indicates the total number of platforms. express index, Indicates the first The influence weight of each platform, with a value of [value]. , Represents the hyperbolic tangent function. Indicates time Time Trend components of each platform Indicates the first The historical maximum value of the absolute value of the trend component of each platform. Indicates time Time Abnormal components of each platform Indicates the first The historical maximum value of the absolute value of outliers on each platform. Indicates time Time Real-time volatility of each platform Indicates the first The historical maximum volatility of each platform Indicates the trend weight (default 0.6). This represents the abnormal weight parameter (default 0.3). Indicates the volatility weight (default 0.1).

[0081] By employing multi-dimensional feature extraction and a time-sensitive weighting mechanism, the system's ability to track the popularity of multiple platforms in real time is significantly enhanced. Original feature datasets are constructed from dimensions such as time-domain traffic, frequency-domain period, and cross-platform correlation. Short-time Fourier transform and sliding window techniques are combined to capture dynamic data features. Simultaneously, cross-platform feature correlation analysis is achieved through Z-score standardization and cosine similarity. Based on this, time-sensitive weighting is applied to the data according to timestamps, enabling the system to perceive changes in platform popularity in real time. When a platform experiences a sudden surge in traffic or an abnormal conversion rate, the time-sensitive correction feature matrix can quickly reflect this change, providing data support for dynamically adjusting replenishment quotas. This solves the problem of delayed response to platform popularity in traditional systems, ensuring that warehouse scheduling can adapt to market fluctuations in a timely manner.

[0082] By fusing influence weight vectors with popularity pulse sequences, intelligent optimization allocation of replenishment quotas across multiple platforms is achieved. By calculating the influence weights of each platform through key indicators such as sales, traffic, and conversion, high-value platforms can be accurately identified. By combining the weighted fusion of trend anomaly dual-track feature vectors and fluctuation terms, a pulse sequence reflecting real-time popularity is generated. When a platform experiences a popularity peak due to promotional activities, the system can dynamically increase the replenishment quota based on its influence weight and popularity compression term. At the same time, it coordinates constraints such as inventory turnover rate and delivery timeliness to achieve a balance between transportation costs and stockout costs, thereby maximizing warehousing operation efficiency and optimizing multiple objectives.

[0083] In one embodiment, the timeliness correction feature matrix is ​​decomposed to obtain the trend. The abnormal dual-track feature vector includes:

[0084] The timeliness-corrected feature matrix is ​​initially decomposed to obtain the basic feature component matrix and the corresponding singular value vector. Specifically, the timeliness-corrected feature matrix is ​​regarded as a two-dimensional matrix, where the rows of the matrix represent different features and the columns represent different time points or platforms. The matrix is ​​decomposed into the product of three matrices using the SVD algorithm: the left singular matrix, the singular value matrix, and the transpose of the right singular matrix. The column vectors of the left singular matrix constitute the basic feature component matrix, which is used to reflect the main feature direction of the data. The non-zero elements on the diagonal of the singular value matrix are the corresponding singular value vectors. The singular value vectors are arranged in descending order, and the magnitude of the singular value vectors can represent the importance of each feature component.

[0085] Based on the basic feature component matrix and singular value vector, the data for each time dimension is fitted to generate a trend feature vector. Specifically, this involves: performing matrix multiplication on the basic feature component matrix and the singular value matrix to reconstruct the feature space, restoring the dimensionality-reduced matrix to the original feature matrix structure; then, for the column data corresponding to each time dimension in the reconstructed matrix (i.e., the feature set at each time point), seasonality and trend decomposition algorithms are used to decompose the time series data into three parts: long-term trend component, seasonal fluctuation component, and random noise component. Only the long-term trend component that can reflect the overall evolution direction of the data is retained, while seasonality and noise interference terms are removed. The retained long-term trend component is then smoothed using a moving average algorithm (taking the average of the most recent N time points). Finally, all trend features are dimensionless using the Z-score standardization method, and all standardized trend features are combined to generate a trend feature vector that can accurately reflect the long-term trend of the data.

[0086] The timeliness-corrected feature matrix and trend feature vector are calculated to obtain the abnormal fluctuation feature vector. Specifically, this involves: performing matrix multiplication on the trend feature vector and the basic feature component matrix to reconstruct a trend matrix reflecting the long-term trend of the data; subtracting this trend matrix from the original timeliness-corrected feature matrix to obtain a residual matrix containing data fluctuation information; and calculating the mean and standard deviation for each time dimension of the data in the residual matrix. Two standard deviations are used as the threshold for abnormal fluctuations. Fluctuation components exceeding this threshold are extracted from the residual matrix. Finally, the extracted abnormal fluctuation components are standardized by Z-score. The standardized abnormal fluctuation components are combined to generate an abnormal fluctuation feature vector that characterizes sudden abnormal changes in the data.

[0087] By combining the trend feature vector and the abnormal fluctuation feature vector, the trend can be obtained. The abnormal dual-track feature vector specifically includes: aligning the trend feature vector and the abnormal fluctuation feature vector along the time dimension to ensure a one-to-one correspondence between the trend and the outlier at each time point; using a vector concatenation algorithm to link the trend feature and the abnormal feature at the same time point end-to-end to form a composite vector with double the dimensions; and normalizing the composite vector by first calculating the mean and standard deviation of each dimension feature, then subtracting the mean from each element and dividing by the standard deviation to finally generate the trend. Abnormal dual-track feature vector, trend The first half of the abnormal dual-track feature vector represents the long-term evolution trend of the data, while the second half reflects short-term sudden fluctuations.

[0088] The timeliness matrix is ​​decomposed into basic feature components and singular values ​​by using the SVD algorithm. By reconstructing the feature space and applying the seasonal decomposition algorithm, seasonal noise in the long-term trend is removed. Then, the trend feature vector is generated by moving average smoothing and standardization. At the same time, the abnormal fluctuation threshold is calculated by using the residual matrix, and the components exceeding the threshold are extracted to generate abnormal feature vectors. This enables the system to simultaneously capture the long-term evolution trend and sudden fluctuations of the platform's popularity, solving the problem of the traditional system's lagging response to dynamic popularity.

[0089] By constructing a dual-track feature vector, dynamic optimization of replenishment quotas across multiple platforms is achieved. After trend and abnormal features are concatenated and normalized along the time dimension, the first half represents the trend direction (such as the quarterly traffic growth trend of a certain platform), and the second half reflects short-term anomalies (such as the instantaneous traffic peak of live-streaming e-commerce). The system combines platform influence weights to adjust replenishment strategies in real time. Specifically, when trend features show that a certain platform is continuously selling well, the long-term replenishment quota is increased. When abnormal features detect sudden traffic, a temporary replenishment mechanism is triggered. While balancing transportation costs and stockout risks, the system enables warehouse scheduling to accurately adapt to the dynamic changes in popularity across multiple platforms, thereby improving operational efficiency.

[0090] In one embodiment, historical replenishment tasks and heat pulse sequences are fused to obtain a real-time demand correction coefficient, including:

[0091] The historical replenishment tasks of the target object are decomposed into a time series to obtain a multi-dimensional replenishment feature vector. Specifically, the historical replenishment task data of the warehouse is arranged and cleaned in chronological order. When missing values ​​are processed using linear interpolation and outliers are removed, the data is converted into a suitable time granularity according to business needs. The STL algorithm is used to decompose the time series into three parts: long-term trend, seasonal cycle, and fluctuation. The long-term trend reflects the overall direction of change, the seasonal cycle reflects the fixed periodic fluctuation pattern, and the fluctuation includes random fluctuations and the impact of abnormal events. Then, the slope, inflection point, and stability of the trend line are calculated based on the long-term trend characteristics. The cycle length, fluctuation amplitude, and phase shift are identified based on the seasonal cycle characteristics. The residual standard deviation, frequency of abnormal events, and degree of impact are calculated based on the fluctuation characteristics. Finally, the extracted features are Z-score standardized to eliminate the influence of dimensions. The importance weight of each feature is determined through principal component analysis. The features and their corresponding importance weights are weighted and merged to form a multi-dimensional replenishment feature vector containing dimensions such as trend strength, cycle stability, and fluctuation amplitude.

[0092] The heat pulse sequence is divided into grids according to the spatiotemporal dimension to generate a spatiotemporal heat matrix. Specifically, when generating the spatiotemporal heat matrix by dividing the heat pulse sequence into grids according to the spatiotemporal dimension, the time dimension of the heat pulse sequence is first discretized at fixed time intervals (1 hour), and the spatial dimension is divided according to geographical regions (latitude and longitude grids) to form spatiotemporal grid units. Then, each heat pulse data is mapped to the corresponding grid according to the timestamp and sales location. The data in the same grid are weighted and aggregated. Finally, the spatiotemporal heat matrix composed of the heat values ​​of each unit is constructed with the time grid as the rows and the spatial grid as the columns.

[0093] The time window width is adaptively adjusted based on the execution cycle of historical replenishment tasks, and the synchronization fluctuation coefficient between historical replenishment tasks and the heat pulse sequence is calculated. Specifically, this includes: performing autocorrelation analysis on the historical replenishment time series; calculating the autocorrelation coefficients of different time intervals (lag orders) in the historical replenishment time series; plotting autocorrelation diagrams to identify the maximum lag order corresponding to significantly non-zero correlation coefficients; determining the initial time window width based on the maximum lag order (if the maximum lag order corresponds to 7 days, then the initial window width is set to 7 days); and using a sliding window to calculate the standard deviation of the replenishment quantity. When the rate of change of the standard deviation of adjacent windows exceeds... When the preset threshold (20%) is reached, the window width adjustment mechanism is triggered. The exponential smoothing method is used to fit the historical window width data to obtain the optimal window width for the next period (if it is the peak sales season, the window width is automatically shortened to 3 days). When calculating the synchronous fluctuation coefficient, the heat pulse sequence within the adjusted optimal window width is first aligned with the historical replenishment volume data in time. The two sets of data are decomposed into high-frequency (sudden fluctuations) and low-frequency (long-term trend) components through wavelet transform. The Pearson correlation coefficient is calculated for each frequency component. Then, the weighted sum is performed according to the contribution of the frequency to the demand, and finally the synchronous fluctuation coefficient reflecting the consistency of the fluctuations of the two is obtained.

[0094] Specifically, by calculating the autocorrelation coefficients of different time intervals (lag orders) in the historical replenishment time series and plotting an autocorrelation graph, the maximum lag order corresponding to a significantly non-zero correlation coefficient is identified. The process involves arranging the historical replenishment data in chronological order and then calculating the correlation at different time intervals. For example, the correlation between today's replenishment volume and the replenishment volumes of different days ago, such as yesterday, the day before yesterday, and the day before yesterday, is calculated to obtain a series of correlation values. These correlation values ​​and their corresponding time intervals are plotted in a scatter plot, and two boundaries (confidence intervals) are drawn on the plot to determine whether the correlation is truly significant. If the correlation value corresponding to a certain time interval exceeds these two boundaries, it indicates that the correlation at this time interval is significantly non-zero, meaning that the replenishment volume is clearly associated at this time interval. Then, the plot is used to find the time interval at which the correlation value first changes from exceeding the boundaries to falling within the boundaries as the time interval increases. This time interval is the maximum time interval and can be used as the initial time window width.

[0095] Analyzing the multidimensional replenishment feature vector and spatiotemporal heat matrix yields historical influence weight vector and heat correction weight vector. These vectors, along with the synchronous fluctuation coefficient, are then fused to obtain the real-time demand correction coefficient. Specifically, this includes analyzing the historical replenishment feature vector... Normalization is performed to obtain the normalized historical replenishment feature vector. , attention key vector Perform the transpose operation to obtain Normalized historical replenishment feature vectors and Multiply to obtain the original similarity score. The final similarity score is obtained by dividing the original similarity score by the square root of the key dimension scaling factor. Application to the final similarity score The function transforms similarity into a probability distribution. The historical influence weight vector Perform the transpose operation to obtain ,Will Multiplying by the probability distribution yields the historical feature attention term. ; For heat feature vector Normalization is performed to obtain the normalized heat feature vector. Adjusting the weight vector for heat level Perform the transpose operation to obtain Normalized heat feature vector and Multiplying them together yields the heat correction term. The synchronization fluctuation coefficient based on the dynamic time warping function With synchronous fluctuation weight Multiplying them together yields the synchronous fluctuation term. Adjust the adaptive time window width Divide by the historical maximum window width of the adaptive time window ,get ,Will Multiply by window width weight Get the time window adjustment item The feature attention term, heat correction term, synchronization fluctuation term, and time window adjustment term are summed to obtain a summation result. This summation result is then mapped to a sigmoid function. The obtained value is the real-time demand correction factor. In practical applications, this can be achieved using the following calculation formula, for example: ;

[0096] In the formula, This represents the real-time demand adjustment factor, with a value of [value missing]. , This represents the sigmoid function. Represents the historical influence weight vector. This indicates the transpose operation. Represents the normalized exponential function, Represents the normalization function. This represents the historical replenishment feature vector. Represents the attention key vector. Indicates the key dimension scaling factor. This represents the heat adjustment weight vector. Represents the heat feature vector, Indicates synchronous fluctuation weights. The value range is 0.3. 0.9, This represents the synchronization fluctuation coefficient based on the dynamic time warping function. The range of values ​​is 0. 1, Indicates window width weight. The value range is 0.3. 0.6, Indicates the width of the adaptive time window. express The widest window in history;

[0097] Among them, when When the value is 0.3, it indicates that the demand for the product is not high, and replenishment should be reduced. A value of 0.7 indicates relatively stable demand, necessitating the continuation of the existing replenishment plan. A value of 0.7 indicates increased demand, requiring additional replenishment. The closer the value is to 1, the more urgent the need for real-time demand adjustments.

[0098] The following is about Explanation of the range of values:

[0099] When sales fluctuations across multiple platforms are highly synchronized, and the deviation of the sales fluctuation amplitude from the average level of sales fluctuation on each platform is less than 5%, then... The value range is 0.3-0.4. Because the fluctuations across multiple platforms are highly consistent, it is necessary to suppress the weight of real-time correction in order to enhance the reliability of historical data.

[0100] When sales fluctuations across multiple platforms are moderately synchronized, and the deviation of the sales fluctuation amplitude from the average level of sales fluctuation on each platform is less than 10%, then... The value range is 0.41-0.7. Because the fluctuations across multiple platforms are at a moderate level, it is necessary to increase the weight of the real-time fluctuation coefficient in order to capture the different needs between platforms.

[0101] When sales fluctuations across multiple platforms are highly asynchronous, and the deviation of the sales fluctuation amplitude from the average level of sales fluctuation on each platform is less than 15%, then... The value range is 0.7-0.9. Because the popularity varies significantly across multiple platforms, it is necessary to actively strengthen real-time correction in order to respond to the real-time fluctuations of each platform to the greatest extent.

[0102] The following is about Explanation of the range of values:

[0103] When the turnover rate of goods sold is slow, at this time The value range is 0.3-0.35, thereby avoiding short-term fluctuations from interfering with long-term warehousing planning and reducing the window width weight to a minimum;

[0104] When the turnover rate of goods sold is moderate, at this time The value range is 0.36-0.45, which can take into account both trend stability and seasonal fluctuations, thereby maintaining the response of the baseline window width;

[0105] When the turnover rate of goods sold is relatively fast, at this time The value range is 0.46-0.6. At this point, the sales of goods fluctuate greatly in a short period of time, so it is necessary to increase the sensitivity of the window width to facilitate order tracking.

[0106] The following is about Explanation of the range of values:

[0107] When multiple platforms are experiencing stable sales, the level of interest is relatively stable. The value range is 0-0.3, because the real-time sales volume is relatively close to the historical trend at this time, so the original replenishment plan is maintained.

[0108] When sales status fluctuates across a single platform, and popularity also experiences localized fluctuations, at this time... The value range is 0.31-0.6. Because the platform is experiencing regional fluctuations at this time, in order to ensure the rationality of replenishment, the replenishment quota is increased on the original replenishment plan.

[0109] When sales status fluctuates significantly across multiple platforms, while popularity remains consistently high, at this time... The value range is 0.61-1, because the popularity is very high at this time, and the surface sales volume may increase significantly, requiring the initiation of cross-warehouse transfer and incremental replenishment.

[0110] By decomposing historical replenishment tasks over time, key features such as trend strength and periodic stability are extracted. Principal component analysis is used to determine weights, allowing for a precise understanding of historical replenishment patterns. Furthermore, the heat pulse sequences are aggregated in a grid to construct a spatiotemporal heat matrix, enabling real-time capture of sales dynamics across multiple platforms. Autocorrelation analysis and exponential smoothing are used to adaptively adjust the time window width, and wavelet transform is combined to calculate the synchronization fluctuation coefficient. This allows for flexible matching of analysis granularity based on market fluctuation characteristics. Finally, historical influence weights, heat correction weights, and synchronization fluctuation coefficients are integrated to generate a real-time demand correction coefficient. This ensures that the system can leverage historical experience to guarantee replenishment stability while rapidly responding to sudden changes in platform heat, effectively addressing the shortcomings of traditional systems in real-time heat tracking.

[0111] By understanding the synchronicity of sales fluctuations across multiple platforms, inventory turnover efficiency, and platform popularity, different threshold ranges are set for the correction coefficient. When sales fluctuations across multiple platforms are highly synchronized, the weight of real-time correction is reduced to rely on historical data, ensuring the stability of the replenishment strategy. When there are significant differences in popularity, real-time correction is proactively strengthened to quickly respond to the personalized needs of each platform. For different inventory turnover efficiencies, the sensitivity of the time window is dynamically adjusted to avoid short-term fluctuations or to promptly capture high-frequency demand changes. This refined, tiered response mechanism enables the warehousing system to dynamically balance historical experience and real-time data when facing complex multi-platform sales scenarios, optimize replenishment quota allocation, and significantly improve the market adaptability and operational efficiency of warehousing scheduling.

[0112] In one embodiment, the dynamic platform quota weight is calculated based on the real-time demand correction coefficient and real-time data, including:

[0113] The real-time data of each platform is identified to obtain a cost sensitivity matrix. Specifically, this includes: extracting cost-related features such as discount intensity, activity cost, delivery timeliness requirements, and inventory turnover deviation from the pre-processed real-time data of each platform; using principal component analysis to reduce dimensionality; calculating the contribution weight of each feature to cost; and weighting and summing the feature values ​​with their corresponding weights according to the platform to obtain the cost sensitivity value of each platform. The cost sensitivity values ​​of all platforms are then arranged by platform as rows and feature as columns to form a cost sensitivity matrix.

[0114] Based on the real-time demand correction coefficient and historical replenishment tasks, the elastic demand threshold for each platform is calculated. Specifically, this involves: arranging the warehouse's historical replenishment tasks for each platform in chronological order to form a replenishment volume time series for each platform; performing kernel density estimation on the replenishment volume time series for each platform; using the historical replenishment volume of each platform as samples to calculate the probability density of different replenishment volume intervals, generating a platform-specific historical replenishment probability density distribution; based on Bayesian inference principles, using the historical probability density distribution of each platform as prior probabilities; and for each platform, transforming the real-time demand correction coefficient into a likelihood function. Specifically, if the real-time demand correction coefficient for a platform is 0.8, the probability density of the platform's real-time demand fluctuation interval is strengthened, making the probability value of the likelihood function in that interval higher than the historical normal interval. A Markov chain Monte Carlo algorithm is used to iteratively sample each platform: from each platform's... Initial samples are generated from the historical prior distribution, and the posterior probability of new samples on each platform is calculated (i.e., the probability combining historical patterns and current real-time correction). After more than 500 iterations, the sample distribution of each platform converges to the posterior distribution (for example, the posterior distribution of a certain platform shows that the probability of replenishment quantity of 150-200 units is the highest). The 95th percentile is calculated for the converged posterior sample set of each platform: all posterior samples of a certain platform are sorted by value, and the value at the 95th position is taken as the basic threshold of the platform. At the same time, the time decay weight is calculated according to the historical replenishment timestamp of each platform: the historical replenishment data within 3 days of a certain platform is assigned a weight of 1.0, the data from 3 to 7 days is assigned a weight of 0.7, and the data from 7 days ago is assigned a weight of 0.3. The posterior samples of each platform are re-sorted after being weighted by time weight, and the 95th percentile is taken as the elastic demand threshold of the platform that is dynamically adjusted according to real-time demand.

[0115] With the objectives of minimizing total cost and maximizing demand satisfaction rate, the replenishment quantity of each platform is dynamically optimized using the cost sensitivity matrix and elastic demand threshold as constraints to form a Pareto optimal solution set. Specifically, the process involves: using the sum of the products of cost sensitivity values ​​and replenishment quantity for each platform in the cost sensitivity matrix as the total cost, and the average ratio of replenishment quantity to elastic demand threshold for each platform as the demand satisfaction rate. The replenishment quantity of each platform is constrained to be no less than the elastic demand threshold, and the product of cost sensitivity value and replenishment quantity does not exceed the total budget. Particle swarm optimization algorithm is used to initialize the replenishment quantity particle swarm, calculate the total cost and demand satisfaction rate of each particle, and iteratively update the particle positions, retaining solutions that cannot deteriorate one objective without worsening the other, thus forming a Pareto optimal solution set.

[0116] The Pareto optimal solution set is solved based on the cost sensitivity matrix to generate the initial platform quota vector;

[0117] The process involves acquiring inventory data for each target object, analyzing the inventory data, and deriving an inventory health index. Specifically, this includes extracting current inventory, total inventory, maximum inventory, turnover rate, batch and distribution data from each warehouse, calculating turnover rate deviation rate, inventory saturation (current / maximum inventory), batch health, and distribution balance (variance of inventory proportion in each warehouse), standardizing each indicator, and summing them using a weighted average method to obtain the inventory health index. The inventory data includes: current inventory quantity, total inventory quantity, maximum inventory quantity, inventory turnover rate, inventory batch information, and inventory distribution data.

[0118] The initial platform quota vector is adjusted based on the inventory health index to generate dynamic platform quota weights. Specifically, the health index is mapped to a correction factor through a linear transformation, so that platforms with better health status receive stronger correction gains. Then, the initial platform quota vector is multiplied by the corresponding correction factor according to the platform dimension to obtain the initially adjusted quota vector. After that, it is verified whether the adjusted quota meets the constraints that the sum of the product of the cost sensitivity value and the adjusted replenishment quantity of each platform does not exceed the total budget and the replenishment quantity is not lower than the elastic demand threshold. If not, the correction factor is fine-tuned using the gradient descent method. Finally, the adjusted quota vector is normalized to ensure that the sum of the weights of each platform is 1, thereby generating dynamic platform quota weights.

[0119] By extracting cost features such as discount intensity and delivery time, a sensitivity matrix is ​​constructed. Combined with Bayesian inference and time decay weighting, an elastic demand threshold is generated. This allows the system to adjust the coefficients based on real-time demand (dynamically calibrating the replenishment benchmark). When a platform experiences a sudden surge in promotional activity, the elastic threshold is quickly adjusted using high-weight historical data from the past 3 days. The total cost and demand fulfillment rate are optimized using a particle swarm optimization algorithm to form a Pareto optimal solution. This solves the problem of traditional systems lagging in responding to platform activity and ensures that replenishment quotas are adapted to market changes in real time.

[0120] The system calculates a health index by measuring indicators such as turnover rate deviation and inventory saturation. The quota adjustment factor for platforms with high health (such as those with high turnover rate and balanced inventory) is increased by 20%. Combined with cost-sensitive value constraints (such as total budget not exceeding the threshold), the quota vector is finely adjusted. When the inventory health of a platform declines, the system automatically reduces its quota weight and triggers cross-warehouse transfers. While balancing transportation costs and stockout risks, the system enables multi-platform replenishment strategies to respond to real-time heat fluctuations and ensure inventory operation efficiency.

[0121] In one embodiment, the initial platform quota vector is generated by solving the Pareto optimal solution set based on the cost sensitivity matrix, including:

[0122] Based on the cost sensitivity matrix, a response surface is constructed between the replenishment quantity and total cost of each platform, generating a marginal cost gradient matrix. Specifically, this includes: extracting the replenishment quantity and cost sensitivity value of each platform based on the cost sensitivity matrix; fitting the nonlinear relationship between replenishment quantity and total cost through a multiple regression algorithm to generate a response surface; then calculating the change in total cost, i.e., marginal cost, for each unit increase or decrease in replenishment quantity on each platform using the finite difference method; and finally arranging the marginal costs of each platform in order to generate a marginal cost gradient matrix that characterizes the rate of cost change.

[0123] The Pareto optimal solution set is optimized to generate a weighted Pareto front. Specifically, this involves: setting an importance ratio (cost 60%, demand 40%) for minimizing total cost and maximizing demand satisfaction rate based on the Pareto optimal solution set; adding the total cost and demand satisfaction rate of each solution according to this ratio to obtain a comprehensive evaluation value; repeatedly calculating and adjusting the evaluation value of each solution using the particle swarm optimization algorithm; selecting the solution with the best comprehensive performance under this ratio; and sorting the solutions according to the evaluation value to form a weighted Pareto front that reflects the preference for the importance of different objectives.

[0124] The comprehensive utility value of each point on the weighted Pareto front is calculated, and the point that maximizes utility is selected as the optimal solution. The replenishment volume ratio of each platform corresponding to this point is used as the initial platform quota vector. Specifically, based on the weighted Pareto front, the total cost and demand satisfaction rate of each point are standardized. The standardized indicators are weighted and summed with a cost weight of 0.6 and a demand weight of 0.4 to obtain the utility value of each point. The utility values ​​of all points are compared, and the point corresponding to the maximum value is selected as the optimal solution. The proportion of replenishment volume of each platform to the total replenishment volume at this point is extracted to form the initial platform quota vector.

[0125] By constructing a cost response surface and a marginal cost matrix, dynamic cost optimization of replenishment quotas across multiple platforms is achieved. Based on the cost sensitivity matrix, the nonlinear relationship between replenishment quantity and total cost is fitted to generate a response surface that intuitively reflects the cost change pattern. Then, the marginal cost gradient matrix is ​​used to quantify the impact of unit changes in replenishment quantity on total cost for each platform. When a platform experiences a surge in delivery costs due to promotional activities, the system can quickly identify high-cost-sensitive platforms based on the marginal cost gradient and prioritize adjusting their replenishment quotas in Pareto optimization. This achieves a dynamic balance between total cost control and demand fulfillment rate, solving the problem of traditional system cost control lagging behind changes in platform popularity.

[0126] By setting a weighting ratio of 60% for cost and 40% for demand, the Pareto optimal solution is converted into a comprehensive evaluation value. Then, the particle swarm optimization algorithm is used to select the optimal solution that balances cost minimization and demand fulfillment rate maximization. For example, when the popularity of a platform suddenly increases, causing the demand fulfillment rate weight to increase, the system automatically increases the proportion of demand target, while recalculating the utility value and adjusting the quota, so as to increase the replenishment volume of the platform and control the increase in total cost. This dynamic weighting mechanism enables warehouse scheduling to respond to the real-time popularity of the platform while maintaining the overall benefits of multi-objective optimization.

[0127] In one embodiment, adjusting the virtual inventory allocation targets for each target object across different platforms in real time based on the dynamic platform quota weight includes:

[0128] To obtain the virtual inventory allocation target that each platform should allocate to each target object, the following steps are taken: First, extract the dynamic quota weight of each platform, the real-time inventory data of each warehouse, and the geographical distance between the warehouse and the platform to quantify the delivery timeliness (where the timeliness weight of the nearby warehouse is set to 0.7 and that of the distant warehouse is set to 0.3). Then, according to the quota weight of each platform, allocate the total virtual inventory target to each platform proportionally. Next, for each platform, calculate the allocation weight of each warehouse: divide the available capacity of the warehouse by the total available capacity of all warehouses and multiply by the delivery timeliness weight. Then, normalize these allocation weights to ensure that the sum is 1. Finally, split the total virtual inventory of the platform into each warehouse according to the warehouse allocation weight, and check whether the allocation exceeds the available capacity of the warehouse. If it exceeds, reduce it proportionally. Finally, determine the virtual inventory allocation target of each platform in each warehouse.

[0129] The maximum capacity coefficient per unit time is determined based on inventory data. Specifically, after standardizing the deviation rate, available capacity ratio, batch turnover efficiency, and distribution balance, the values ​​are weighted and summed with a weight of 40% for turnover rate deviation, 30% for available capacity, 20% for batch efficiency, and 10% for distribution balance. Then, the sum is multiplied by a time decay factor (data from the last 3 days is weighted at 1.0, data from 3-7 days is weighted at 0.8, and data from 7 days ago is weighted at 0.5) to obtain the maximum capacity coefficient per unit time.

[0130] Based on dynamic platform quota weights and real-time data from multiple platforms, we analyze the urgency indicators of demand for each platform. Specifically, this includes: extracting real-time data from each platform, including indicators that reflect demand such as sales volume, sales order volume, website traffic, search popularity, and discount intensity; standardizing these indicators with different dimensions; extracting key principal components from numerous indicators through principal component analysis; determining the weight of each principal component based on its variance contribution rate; obtaining the influence weight of each indicator on demand urgency; and combining the dynamic platform quota weights with the standardized real-time indicators to perform a weighted summation (influence weight multiplied by dynamic platform quota weight multiplied by real-time indicator) to form an initial base value for demand urgency. Using inventory health index and delivery timeliness as constraints, we use particle swarm optimization algorithm for iterative optimization, adjusting various parameters to make the urgency indicators more closely match actual demand, and generating a comprehensive urgency indicator that reflects the urgency of demand on each platform.

[0131] The virtual inventory allocation adjustment coefficient is calculated by analyzing the demand urgency index and maximum capacity coefficient of each platform. Specifically, the demand urgency index and maximum capacity coefficient of each platform are standardized and converted to a value between 0 and 1. The standardized demand urgency index and maximum capacity coefficient are weighted and summed with a weight of 0.6 for demand urgency and 0.4 for maximum capacity coefficient to obtain the initial adjustment coefficient. By comparing the deviation data between the historical adjustment coefficient and the actual inventory turnover rate, the weight ratio is dynamically adjusted using an iterative optimization method (5% weight adjustment each time) until the matching degree between the adjustment coefficient and inventory fluctuation reaches a preset threshold (90%), and finally the virtual inventory allocation adjustment coefficient is generated.

[0132] The virtual inventory allocation target is adjusted according to the virtual inventory allocation adjustment coefficient to obtain the adjusted virtual inventory allocation target. Specifically, the initial virtual inventory allocation target value of each platform in each warehouse is multiplied by the virtual inventory allocation adjustment coefficient of the corresponding platform to obtain the preliminary adjusted allocation amount. The adjustment amount is checked for each warehouse to see if it exceeds the available capacity of the warehouse (the maximum inventory minus the current inventory is the available capacity of the warehouse). If it exceeds the limit, the allocation amount is reduced by the ratio of "available capacity divided by adjusted allocation amount" to finally determine the adjusted virtual inventory allocation target of each platform in each warehouse.

[0133] When allocating initial targets, the system comprehensively considers dynamic platform quota weights, real-time warehouse inventory, and geographical distance to divide virtual inventory for each platform in different warehouses. At the same time, it calculates the maximum capacity coefficient, which includes indicators such as turnover rate deviation and available capacity, and uses a time decay factor to ensure that inventory allocation fully matches the real-time operating status of the warehouse. This effectively avoids the inventory backlog or shortage problems caused by traditional systems neglecting the dynamic capacity of the warehouse, and ensures that inventory resource allocation always matches the actual warehousing capacity.

[0134] Principal component analysis of multiple indicators such as sales volume and search popularity is used to calculate the urgency of demand by combining dynamic platform quota weights. With inventory health index and delivery timeliness as constraints, particle swarm optimization is used for iterative optimization to accurately capture real-time demand changes on each platform. In the adjustment stage, adjustment coefficients are generated with a weight of 0.6 for demand urgency and 0.4 for maximum capacity coefficient. The weights are iteratively corrected based on historical data to make the adjustment strategy highly consistent with actual inventory fluctuations. For example, if the urgency of demand on a certain platform surges due to promotional activities, the system will prioritize increasing its virtual inventory allocation ratio. At the same time, it will combine the warehouse maximum capacity coefficient to avoid over-allocation, breaking the static allocation limitations of traditional systems and realizing intelligent replenishment scheduling of inventory resources in scenarios where the popularity of multiple platforms changes rapidly.

[0135] In one embodiment of this invention, calculating the inventory turnover rate of the target object includes:

[0136] Analyzing the historical replenishment tasks and spatiotemporal heat matrix of the target object, a spatiotemporal value decay matrix is ​​formed. Specifically, this includes: preprocessing historical replenishment tasks and sorting them by time; extracting features from the preprocessed heat pulse sequences; dividing the time into hourly granularities and the space into spatiotemporal grids based on latitude and longitude; calculating the trend slope, inflection points, and stability characteristics of the long-term trend, seasonal cycle, and fluctuation components decomposed from the historical replenishment data; assigning decay weights to data within the time dimension: 1.0 weight for data within 3 days, 0.7 weight for data between 3 and 7 days, and 0.3 weight for data more than 7 days ago; constructing a multidimensional feature vector based on the trend slope, inflection point density, and fluctuation stability; and then combining this vector with the data within the segmented time periods. The weights are used to perform tensor product operations to form a time decay factor; at the same time, from the spatial dimension, the geographical distance between each warehouse and the platform is calculated based on the latitude and longitude coordinates of the replenishment location, and the decay coefficient is set according to the distance (higher weight for closer distances and lower weight for farther distances). By multiplying the geographical distance with the decay coefficient, a spatial decay factor is formed; finally, the time decay factor and the spatial decay factor are concatenated by dimension to construct a spatiotemporal decay feature that includes time decay and spatial decay. The spatiotemporal decay feature of historical replenishment is matched with the spatiotemporal heat matrix by grid, and the value decay coefficient of each grid is calculated. The decay coefficients of all grids are arranged by the coordinates of the time grid (row) and the spatial grid (column) to form a spatiotemporal value decay matrix.

[0137] The inventory quantity of the target object is weighted and aggregated with the spatiotemporal value decay matrix to obtain the dynamic inventory equivalent. Specifically, this includes: mapping the inventory quantity of each warehouse to the corresponding grid according to the spatiotemporal coordinates, obtaining the value decay coefficient within the grid, multiplying the inventory of each warehouse with the decay coefficient of the corresponding grid, and weighting and aggregating according to the time grid and spatial grid dimensions, that is, summing the inventory decay values ​​of each grid to obtain the dynamic inventory equivalent of comprehensive spatiotemporal value decay.

[0138] The process involves analyzing the delay response coefficients of historical replenishment tasks and heat pulse sequences, and combining them with real-time demand correction coefficients to generate a demand response efficiency coefficient. Specifically, this includes: preprocessing the delay records of historical replenishment tasks and aligning them with the heat pulse sequence time; calculating the time difference between heat changes and replenishment responses at each time point; calculating the delay response coefficient using the Pearson correlation coefficient; and then weighting and fusing the delay response coefficient and the real-time demand correction coefficient according to preset weights (0.6 for the delay response coefficient and 0.4 for the real-time demand correction coefficient). After normalization, the demand response efficiency coefficient is generated.

[0139] The corrected inventory turnover rate is obtained by integrating dynamic inventory equivalent, demand response efficiency coefficient, and sales cost. Specifically, the following steps are taken: using dynamic inventory equivalent as the adjusted inventory base, dividing sales cost by dynamic inventory equivalent to obtain the basic turnover rate, multiplying the demand response efficiency coefficient as a correction factor with the basic turnover rate, and normalizing the product to obtain the corrected inventory turnover rate that integrates spatiotemporal value, response efficiency, and cost factors.

[0140] By assigning decay factors to historical replenishment data according to time and space, a spatiotemporal decay feature is constructed and integrated with the popularity matrix, so that the inventory equivalent calculation can accurately reflect the spatiotemporal differences in popularity across multiple platforms, and the inventory base accounting can be dynamically matched with the real-time popularity across multiple platforms.

[0141] By quantifying historical replenishment delays and real-time demand correction coefficients using the Pearson correlation coefficient, a response efficiency factor is generated and embedded into turnover rate calculations. This allows the turnover rate metric to simultaneously reflect inventory turnover speed and market responsiveness. When a platform experiences a sudden surge in traffic, the system adjusts its expected inventory turnover rate based on the demand response efficiency coefficient, while simultaneously reducing the weight of distant warehouse inventory through a spatiotemporal decay matrix to prioritize replenishment of nearby warehouses. This not only solves the problem of traditional systems' turnover rate calculations lagging behind changes in demand but also achieves dynamic optimization of warehousing resources across multiple platforms through spatiotemporal and efficiency-based corrections.

[0142] In one embodiment, the inventory turnover rate and the platform's virtual inventory allocation target are analyzed to determine the replenishment urgency, including:

[0143] The corrected inventory turnover rate is aligned with the platform's virtual inventory allocation target in both time and space to generate a multidimensional matching matrix. Specifically, the corrected inventory turnover rate is divided into one-hour time units, and spatially into grid regions based on latitude and longitude. Similarly, the platform's virtual inventory allocation target is divided according to the same time and space standards, ensuring a one-to-one correspondence between their time and space coordinates. In the time dimension, since the length and rhythm of data from different time periods may differ, a dynamic time warping algorithm is used to find the optimal matching path between the two time series, eliminating time misalignment. In the spatial dimension, the actual distance between different warehouses and the platform's spatial location is calculated, and then the distance is converted into a similarity value; the closer the distance, the higher the similarity. Finally, using the time grid as rows and the spatial grid as columns, the inventory turnover rate data and virtual inventory allocation target data in each spatiotemporal unit, after alignment with the optimal matching path in the time dimension and similarity processing in the spatial dimension, are placed together. By calculating the difference index between the two values, these difference indices are filled into the matrix according to the corresponding spatiotemporal units, ultimately constructing the multidimensional matching matrix.

[0144] Based on a multidimensional matching degree matrix, real-time demand correction coefficients, and demand urgency indicators for each platform, a demand pressure index is constructed. Specifically, this involves: extracting the difference between each spatiotemporal unit based on the multidimensional matching degree matrix; weighting and summing the results according to the time decay factor and spatial distance similarity to obtain a spatiotemporal pressure baseline value; and standardizing the real-time demand correction coefficients and demand urgency indicators to 0. Within a given interval, a demand pressure index is generated by weighting and integrating the time-space pressure benchmark value (40%), the real-time correction coefficient (30%), and the urgency index (30%).

[0145] Based on historical replenishment tasks, inventory health index, and maximum capacity coefficient, the replenishment elasticity threshold is determined. Specifically, this includes: performing kernel density estimation on historical replenishment task data to generate a historical replenishment probability distribution, taking the 95th percentile as the base threshold, standardizing the inventory health index and maximum capacity coefficient, and obtaining a correction factor by weighting inventory health at 40% and maximum capacity at 60%, and using the correction factor to linearly adjust the base threshold to generate a replenishment elasticity threshold that takes into account both historical patterns and the current inventory status.

[0146] Analyzing the demand pressure index and replenishment elasticity threshold to generate replenishment urgency involves: uniformly converting the demand pressure index and replenishment elasticity threshold to a 0-1 range; subtracting the replenishment elasticity threshold from the demand pressure index to obtain the difference; and if the difference range is... The urgency level is mapped according to the ratio of "(difference + 1) / 2". For example, when the difference is 1, the urgency level is 1, and when the difference is -1, the urgency level is 0. The numerical difference is converted into an urgency level of 0-1 through a linear ratio to generate the replenishment urgency level. The larger the difference, the higher the urgency of replenishment.

[0147] By aligning the data in the spatiotemporal dimensions, the dynamic time warping algorithm solves the problem of time series misalignment between inventory turnover rate and virtual inventory allocation target. Spatial similarity calculation eliminates the evaluation bias caused by geographical distance, ensuring the accuracy of data comparison. When constructing the demand pressure index, the algorithm integrates a multi-dimensional matching degree matrix, real-time demand correction coefficient, and demand urgency index, organically combining platform popularity changes, historical demand patterns, and real-time demand dynamics. Compared with traditional solutions that rely on only a single indicator, this approach can more comprehensively and timely reflect the actual replenishment pressure.

[0148] By estimating historical replenishment data using kernel density, a basic threshold is determined. This threshold is then dynamically adjusted in conjunction with the inventory health index and maximum capacity coefficient. This ensures that the threshold reflects both historical replenishment patterns and the current state of warehouse operations. When comparing the demand pressure index with the replenishment elasticity threshold, the difference is linearly mapped to an urgency level between 0 and 1, accurately distinguishing replenishment priorities. For example, if a warehouse's inventory health declines, causing its elasticity threshold to increase, while a platform's demand pressure index rises simultaneously due to a surge in traffic, the system can promptly determine that the platform's replenishment urgency is close to 1 through difference calculation. This prioritizes triggering cross-warehouse transfers or expedited replenishment processes, avoiding replenishment delays or over-replenishment caused by traditional static threshold settings, and effectively improving the efficiency of warehouse resource scheduling and operational effectiveness.

[0149] In one embodiment, a dynamically optimal cross-warehouse scheduling instruction set is obtained based on replenishment urgency and location information, including:

[0150] Constructing a spatial distance weight matrix based on the location information of the target object involves: first, obtaining the latitude and longitude coordinates of each warehouse and the platform; calculating the actual distance between each pair using geographic information; normalizing the actual distance to obtain a distance coefficient in the 0-1 range; then converting the distance into a similarity weight by subtracting the distance coefficient from 1, with higher weights for closer distances; and filling the weight values ​​into the corresponding positions with warehouses as rows and platforms as columns to generate the spatial distance weight matrix.

[0151] Based on replenishment urgency, spatial distance weight matrix, and inventory health index, the scheduling priority coefficient of each target object is calculated. Specifically, the inventory health index is normalized, the weight values ​​in the spatial distance weight matrix corresponding to the target object are extracted, and the replenishment urgency, spatial distance weight matrix, and inventory health index are weighted and summed according to the weights of replenishment urgency (40%), spatial distance weight (30%), and inventory health index (30%) to obtain the scheduling priority coefficient. The higher the coefficient, the higher the scheduling priority.

[0152] With the goal of minimizing scheduling costs and maximizing delivery time, and incorporating scheduling priority coefficients, a dynamically optimal cross-warehouse scheduling instruction set is derived. Specifically, this involves: compiling the scheduling costs and delivery times from each supply warehouse to the target warehouse; sorting the target warehouses by their scheduling priority coefficients from highest to lowest; for the highest priority target warehouse, iterating through all supply warehouses, dividing each supply warehouse's scheduling cost by its highest scheduling cost and normalizing it, dividing its delivery time by the fastest delivery time from that supply warehouse to the target warehouse and normalizing it; calculating the weighted sum of costs and delivery times for each supply warehouse according to the cost and delivery time weights set by the business; selecting the supply warehouse with the smallest weighted sum as the scheduling source; processing all target warehouses sequentially; and generating the dynamically optimal scheduling instruction set.

[0153] By constructing a spatial distance weight matrix, the geographical distance between the warehouse and the platform is transformed into a similarity weight, enabling scheduling decisions to prioritize allocation to nearby warehouses, effectively shortening delivery routes. When calculating the scheduling priority coefficient, the urgency of replenishment, spatial distance weight, and inventory health index are comprehensively considered to achieve an organic combination of demand urgency, geographical advantages, and warehouse status. This allows for a comprehensive and dynamic assessment of scheduling needs, avoiding scheduling delays or resource mismatches caused by incomplete information.

[0154] By prioritizing the lowest scheduling costs and fastest delivery times, and ranking target warehouses based on scheduling priority coefficients, the system prioritizes urgent needs. When selecting a supply warehouse, scheduling costs and delivery times are normalized, and a weighted sum is calculated based on business weights to accurately select the optimal scheduling source. For example, if a target warehouse has a high priority but the scheduling cost of a nearby warehouse is too high, the system will comprehensively weigh the cost and timeliness and select the next nearest warehouse. This dynamic trade-off strategy can ensure a rapid response to urgent needs while controlling operating costs, avoiding cost surges due to blindly pursuing timeliness or delivery delays caused by excessive cost reduction, and effectively improving the overall replenishment scheduling efficiency of the warehousing system in multi-platform sales scenarios.

[0155] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A warehouse intelligent replenishment scheduling system, characterized in that, include: The heat analysis unit is used to determine the heat pulse sequence based on real-time data from multiple platforms, including: extracting features from real-time data and generating a cross-platform raw feature dataset; Based on the interval between the timestamps generated by the data in the cross-platform original feature dataset and the current time, the data in different time periods are weighted according to their timeliness to obtain a timeliness-corrected feature matrix. The influence weight vector of different platforms is determined based on real-time data from multiple platforms. The timeliness-corrected feature matrix is ​​decomposed to obtain a trend-anomaly dual-track feature vector, including: The timeliness-corrected feature matrix is ​​initially decomposed to obtain the basic feature component matrix and the corresponding singular value vector. Based on the basic feature component matrix and singular value vector, the data for each time dimension is fitted to generate a trend feature vector; The abnormal fluctuation feature vector is obtained by calculating the timeliness correction feature matrix and the trend feature vector; By combining the trend feature vector and the abnormal fluctuation feature vector, we obtain the trend-abnormal dual-track feature vector; The influence weight vectors and trend-anomaly dual-track feature vectors of different platforms are fused to generate a heat pulse sequence that reflects real-time heat fluctuations. The demand correction unit is used to obtain the historical replenishment tasks of the target object, and fuse the historical replenishment tasks and the heat pulse sequence to obtain the real-time demand correction coefficient. The allocation optimization unit is used to calculate the dynamic platform quota weight based on the real-time demand correction coefficient and real-time data, and adjust the virtual inventory allocation target of each target object to different platforms in real time according to the dynamic platform quota weight. The evaluation unit is used to calculate the inventory turnover rate of the target object, analyze the inventory turnover rate and the virtual inventory allocation target, and determine the replenishment urgency. The scheduling generation unit is used to obtain the location information of the target object and, based on the replenishment urgency and location information, obtain the dynamic optimal cross-warehouse scheduling instruction set.

2. The intelligent warehouse replenishment scheduling system according to claim 1, characterized in that, By fusing historical replenishment tasks and heatwave pulse sequences, a real-time demand correction coefficient is obtained, including: The historical replenishment tasks of the target object are decomposed in time series to obtain a multi-dimensional replenishment feature vector; The heat pulse sequence is divided into grids according to the spatiotemporal dimension to generate a spatiotemporal heat matrix; The time window width is adaptively adjusted based on the execution cycle of historical replenishment tasks, and the synchronization fluctuation coefficient between historical replenishment tasks and heat pulse sequences is calculated. By analyzing the multidimensional replenishment feature vector and the spatiotemporal heat matrix, the historical influence weight vector and the heat correction weight vector are obtained. The historical influence weight vector, the heat correction weight vector and the synchronous fluctuation coefficient are then fused to obtain the real-time demand correction coefficient.

3. The intelligent warehouse replenishment scheduling system according to claim 2, characterized in that, Based on the real-time demand correction factor and real-time data, the dynamic platform quota weight is calculated, including: The cost sensitivity matrix is ​​obtained by identifying real-time data from each platform. Based on the real-time demand correction coefficient and historical replenishment tasks, the elastic demand threshold for each platform is calculated. With the goals of minimizing total cost and maximizing demand satisfaction, the replenishment volume of each platform is dynamically optimized using the cost sensitivity matrix and elastic demand threshold as constraints, forming a Pareto optimal solution set. The Pareto optimal solution set is solved based on the cost sensitivity matrix to generate the initial platform quota vector; Obtain inventory data for each target object, analyze the inventory data, and derive the inventory health index; The initial platform quota vector is adjusted based on the inventory health index to generate dynamic platform quota weights.

4. The intelligent warehouse replenishment scheduling system according to claim 3, characterized in that, Based on the cost sensitivity matrix, the Pareto optimal solution set is solved to generate an initial platform quota vector, including: Based on the cost sensitivity matrix, a response surface between replenishment volume and total cost for each platform is constructed to generate a marginal cost gradient matrix. Optimize the Pareto optimal solution set to generate a weighted Pareto front. Calculate the overall utility value of each point on the weighted Pareto front, select the point that maximizes utility as the optimal solution, and use the replenishment volume ratio of each platform corresponding to that point as the initial platform quota vector.

5. The intelligent warehouse replenishment scheduling system according to claim 4, characterized in that, The virtual inventory allocation targets for each target object across different platforms are adjusted in real time based on the dynamic platform quota weights, including: Obtain the virtual inventory allocation target that should be allocated to each platform in each target object; Determine the maximum capacity coefficient per unit time based on inventory data; Based on dynamic platform quota weights and real-time data from multiple platforms, the urgency index of demand for each platform is analyzed. The virtual inventory allocation adjustment coefficient is obtained by calculating the demand urgency index and maximum capacity coefficient of each platform. The virtual inventory allocation target is adjusted based on the virtual inventory allocation adjustment coefficient to obtain the adjusted virtual inventory allocation target.

6. The intelligent warehouse replenishment scheduling system according to claim 5, characterized in that, Calculate the inventory turnover rate of the target object, including: Analyze the historical replenishment tasks and spatiotemporal heat matrix of the target object to form a spatiotemporal value decay matrix; The dynamic inventory equivalent is obtained by weighted aggregation of the target object's inventory quantity and the spatiotemporal value decay matrix. Analyze the delay response coefficients of historical replenishment tasks and heat pulse sequences, and combine them with real-time demand correction coefficients to generate demand response efficiency coefficients. By integrating dynamic inventory equivalent, demand response efficiency coefficient, and sales cost, a corrected inventory turnover rate is obtained.

7. The intelligent warehouse replenishment scheduling system according to claim 6, characterized in that, Analyzing inventory turnover rate and platform virtual inventory allocation targets yields replenishment urgency, including: The revised inventory turnover rate is aligned with the platform's virtual inventory allocation target in both time and space to generate a multi-dimensional matching matrix. A demand pressure index is constructed based on a multidimensional matching degree matrix, real-time demand correction coefficient, and demand urgency indicators of each platform. The replenishment elasticity threshold is determined based on historical replenishment tasks, inventory health index, and maximum capacity coefficient. The demand pressure index and replenishment elasticity threshold are analyzed to generate a replenishment urgency level.

8. The intelligent warehouse replenishment scheduling system according to claim 7, characterized in that, Based on replenishment urgency and location information, a dynamically optimal cross-warehouse scheduling instruction set is obtained, including: Construct a spatial distance weight matrix based on the location information of the target object; Based on replenishment urgency, spatial distance weight matrix, and inventory health index, calculate the scheduling priority coefficient for each target object; With the goal of minimizing scheduling costs and maximizing delivery time, and by combining scheduling priority coefficients, a dynamic optimal cross-warehouse scheduling instruction set is derived.

Images