A method for adaptive optimization of inventory demand fluctuations

Abstract

The application discloses a kind of inventory demand fluctuation adaptive optimization methods, belong to warehouse management and intelligent inventory optimization technical field, by obtaining the order of target inventory object, inventory, in-transit, location, warehousing execution, supply delivery date and abnormal event data, after washing, time alignment and windowing processing, extract demand fluctuation, inventory state, supply disturbance and warehousing execution disturbance feature, construct inventory demand fluctuation state vector and identify fluctuation state;With demand quantity in target period or demand quantile under specified confidence level, inventory supply capacity, location capacity and warehousing execution capacity calculate inventory-demand matching degree, generate inventory adjustment control instruction when meeting trigger condition, after executability verification, send to execution end, and update relevant parameters according to actual execution deviation, improve inventory fluctuation identification accuracy and control instruction executability.

Description

Technical Field

[0001] This invention belongs to the field of warehouse management and intelligent inventory optimization technology, specifically relating to an adaptive optimization method for inventory demand fluctuations. Background Technology

[0002] In warehouse management and supply chain management, inventory levels directly impact order fulfillment efficiency, warehousing efficiency, supply chain responsiveness, and inventory holding costs. Existing inventory management systems typically calculate future demand based on historical sales data, order data, inventory upper and lower limits, safety stock thresholds, or fixed replenishment cycles, and generate replenishment recommendations, procurement plans, or inventory alerts accordingly.

[0003] However, in actual warehousing and supply chain execution scenarios, fluctuations in inventory demand do not solely stem from changes in sales demand. Inventory status is also affected by a variety of factors, including in-transit inventory delays, supplier delivery deviations, order lock-in, inventory freezes, insufficient storage capacity, inbound delays, replenishment backlogs, picking congestion, transfer execution delays, logistics anomalies, promotional activities, holidays, and regional emergencies.

[0004] For example, a decrease in available inventory of a certain product in a warehouse could be due to increased actual market demand, or it could be caused by delayed arrival of in-transit inventory, inventory being locked by large orders, incomplete inbound tasks, insufficient storage capacity, or delays in synchronizing inventory records with physical inventory. If the existing system determines whether to replenish stock based solely on historical sales volume or current inventory levels, it may easily misinterpret supply disturbances, execution disturbances, or data synchronization anomalies as increased demand, triggering unnecessary procurement or allocation and causing inventory backlog. Conversely, when actual demand grows rapidly but the system fails to identify the causes of demand fluctuations in a timely manner, it may lead to untimely replenishment, allocation, or picking resource adjustments, resulting in stockouts and fulfillment delays.

[0005] Most existing inventory optimization methods focus on demand forecasting or safety stock calculations, typically outputting predicted demand, replenishment suggestions, or early warning information, lacking further processing of warehouse execution constraints. For example, existing methods often fail to adequately consider actual execution conditions such as storage capacity, replenishment routes, task queue load, inbound processing capacity, picking waves, supply lead times, transfer timeliness, and minimum packaging quantities. Therefore, even if the system can provide a theoretical replenishment quantity, it may be unable to be directly executed by the warehouse management system, warehouse control system, or transportation management system due to unmet execution conditions.

[0006] Furthermore, existing inventory optimization models largely rely on manual parameter tuning or periodic training, making it difficult to make timely corrections based on the actual execution results of control commands. When there are discrepancies between predicted and actual demand, estimated and actual delivery times, or estimated and actual picking completion times, existing systems typically cannot automatically adjust fluctuation identification thresholds, inventory matching model weights, and replenishment or transfer trigger conditions, leading to the recurrence of the same types of inventory fluctuation problems.

[0007] Therefore, there is an urgent need to propose an adaptive optimization method for inventory demand fluctuations in warehousing and supply chain execution systems. This method should be able to integrate multi-source inventory execution data, accurately identify the state of inventory demand fluctuations, calculate the dynamic matching degree between inventory and demand, generate control instructions that can be directly processed by the warehousing or supply chain execution system, and perform closed-loop adaptive updates based on actual execution deviations. Summary of the Invention

[0008] The purpose of this invention is to provide an adaptive optimization method for inventory demand fluctuations, which solves the problems of existing inventory optimization methods, such as single data sources for demand fluctuation identification, difficulty in distinguishing the causes of fluctuations, insufficient judgment of the matching relationship between inventory and demand, difficulty in directly executing optimization results, and lack of closed-loop adaptive updates in the model.

[0009] To achieve the above objectives, the present invention provides an adaptive optimization method for inventory demand fluctuations, comprising the following steps: Retrieve data on orders, inventory, in-transit, storage capacity, warehousing execution, supply delivery date, and abnormal events by inventory object, warehouse node, and target time window; A unified event timeline is established based on the event occurrence time, source system timestamp, and system reception time. Data is then deduplicated, time-corrected, anomaly-marked, missing data-completed, unit-converted, confidence-marked, and windowed to form a standardized inventory execution dataset. The dataset extracts features such as demand, inventory, supply, task queue, storage capacity, account-physical synchronization, and data confidence. A state vector is constructed based on the most recent N sliding windows to identify fluctuation states and confidence levels. Calculate the inventory-demand matching degree based on the demand gap, the time pressure caused by the expected stockout time and the available shelf time, the task queue load, and the storage capacity pressure. When the matching degree is lower than the matching threshold and the recognition confidence degree is higher than the confidence threshold, a machine-parseable inventory adjustment control instruction is generated. Before sending, the target warehouse capacity, processing capacity, completion time of shelving, task conflicts, available inventory in the source warehouse and transfer path are verified. If the verification fails, the following measures are taken: reducing quantity, splitting batches, changing source warehouse, adjusting warehouse location, postponing task or increasing priority, and then re-verifying. After verification, the data is sent to the execution end to generate or adjust the inventory execution task, and the fluctuation identification parameters, matching degree weights or control thresholds are updated in a single amplitude-limited manner according to the returned execution results.

[0010] Furthermore, the windowing establishes data units based on inventory objects, warehouse nodes, and target time windows, and writes data from different sources into the same data unit according to a unified event timeline; when there is a time conflict, the source system timestamp is calibrated with the event occurrence time, and the system reception time is used to determine the window to which the late event belongs.

[0011] Furthermore, when constructing the state vector, demand trends, demand surges, inventory coverage days, available inventory ratio, in-transit delay time, supply delivery deviation, replenishment task queue load, picking task queue load, inbound task queue load, storage capacity pressure, book-to-physical discrepancy rate, and data confidence are written in chronological order.

[0012] Furthermore, when identifying fluctuation states, candidate fluctuation states are first generated based on demand surges, inventory coverage days, in-transit delays, task queue loads, storage capacity pressures, and book-to-physical discrepancy rates. Then, the features of the most recent N sliding windows are input into the time series classification model to determine the output state from demand changes, supply delays, in-transit anomalies, warehouse execution congestion, limited storage capacity, book-to-physical synchronization anomalies, and compound fluctuation states.

[0013] Furthermore, the training sample labels of the time-series classification model are generated from the actual stockouts, actual arrival delays, task failure reasons, warehouse capacity alarms, and abnormal records of account-to-physical synchronization after the historical window; during training, the samples are divided in chronological order, and the classification threshold and confidence threshold are determined by verifying the samples.

[0014] Furthermore, the identification confidence level is calculated jointly by data integrity, timestamp consistency, and the probability of fluctuating state output. When the matching degree is lower than the matching threshold but the identification confidence level is not higher than the confidence threshold, the data is stopped from being sent to the execution end, and control candidates are generated for manual review.

[0015] Furthermore, when calculating the inventory-demand matching degree, the current available inventory and the on-time available inventory in transit are summed and then the locked inventory and frozen inventory are deducted to obtain the inventory supply. The demand gap is obtained based on the target cycle demand and the inventory supply. The mismatch risk value is calculated after normalizing the demand gap, time pressure, task queue load and storage capacity pressure.

[0016] Furthermore, the normalization is trimmed to [0,1] according to the historical percentile threshold or the physical execution upper limit; when the demand gap is the same but the time pressure, task queue load or storage capacity pressure is different, the control instruction generation logic for cross-warehouse transfer, wave adjustment, inbound priority adjustment or storage location adjustment is triggered respectively.

[0017] Furthermore, when generating inventory adjustment control instructions, the instruction identifier, instruction type, inventory object, target warehouse node, adjustment quantity, execution end identifier, idempotent key, version number, rollback condition, and status feedback address are written into the instruction message; duplicate instructions with the same idempotent key within the same fluctuation window are intercepted.

[0018] Furthermore, when updating parameters based on execution deviations, the fluctuation identification parameters are updated when the actual demand deviation exceeds the threshold but the supply and execution disturbances do not exceed the threshold; the demand gap weight and time pressure weight are updated when the expected stockout time deviation exceeds the threshold; and the executability verification threshold and instruction correction rules are updated when the task failure reason corresponds to the storage capacity, allocation path, source warehouse inventory, or queue load.

[0019] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. This invention addresses the inventory demand fluctuation control scenario in warehouse execution systems. When heterogeneous events such as orders, inventory, in-transit, storage locations, receiving, picking, and transfer arrive at different time granularities, existing systems struggle to distinguish the causes of fluctuations in a timely manner. By unifying the event timeline and performing time alignment and windowing processing on multi-source inventory execution data, events from different sources and at different granularities can participate in inventory demand fluctuation identification under the same time benchmark, thereby improving the timeliness and consistency of fluctuation identification results. 2. This invention constructs an inventory demand fluctuation state vector, transforming data such as order changes, inventory status, in-transit status, storage capacity, supply delivery date, warehousing tasks, picking tasks, transfer tasks, and abnormal events into feature information that can be used for status identification. This can distinguish different inventory demand fluctuation states such as real demand growth, supply delays, in-transit anomalies, limited storage capacity, warehouse execution congestion, and inventory synchronization anomalies, thereby reducing the probability of misjudging supply disturbances, execution disturbances, or data synchronization anomalies as real demand growth. 3. This invention calculates the inventory-demand matching degree based on the inventory demand fluctuation state vector, and introduces demand fulfillment gap, time pressure, task queue load and storage capacity pressure into the matching degree calculation. This makes the matching judgment between inventory and demand not only rely on the predicted demand or current inventory, but also reflect the impact of expected stockout time, available shelf time after replenishment or transfer, warehousing processing capacity and storage capacity constraints on inventory fulfillment, thereby improving the accuracy of inventory adjustment trigger judgment. 4. This invention generates inventory adjustment control instructions based on inventory-demand matching degree and warehousing execution constraints, so that control instructions such as replenishment, transfer, storage location adjustment, picking wave adjustment and inbound priority adjustment are matched with actual warehousing execution conditions. Before the instructions are issued, the feasibility of storage location capacity, processing capacity, shelving time, task conflict, available inventory and transfer path are verified, thereby reducing the probability of invalid control instructions, storage location conflict instructions and overloaded operation tasks. 5. The inventory adjustment control instructions generated by this invention can be directly parsed and executed by warehouse management systems, warehouse control systems, transportation management systems or other supply chain execution systems, so that the inventory optimization results are transformed from forecast results, replenishment suggestions or early warning information into executable system control instructions, thereby reducing manual judgment and secondary conversion links, and improving the warehouse execution response speed and inventory adjustment control efficiency. 6. This invention collects the actual execution results of control commands and adaptively updates the fluctuation identification parameters, matching degree model parameters, and control thresholds based on execution deviations such as actual completion time, actual execution quantity, actual shelving time, reasons for task failure, and actual stockout quantity. This enables the system to continuously correct subsequent control processes based on actual execution feedback from replenishment, allocation, storage location adjustment, picking wave adjustment, and inbound priority adjustment, thereby improving the accuracy of inventory demand fluctuation identification, the executability of inventory adjustment control commands, and the warehousing execution response speed. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating an adaptive optimization method for inventory demand fluctuations according to the present invention. Figure 2 This is a logical diagram illustrating the identification of inventory demand fluctuations and the calculation of inventory-demand matching degree in this invention. Figure 3 This is a schematic diagram illustrating the logic of inventory adjustment control instruction generation and execution deviation closed-loop update in this invention. Detailed Implementation

[0021] 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.

[0022] In the description of this invention, the terms "first," "second," etc., are used for distinguishing purposes only and should not be construed as indicating or implying relative importance. Unless expressly defined otherwise, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion.

[0023] The target inventory object in this invention can be a single SKU (Stock Keeping Unit), or it can be a combination of SKUs, a product category, a warehouse node, a regional warehouse network, or a supply chain node. The warehousing or supply chain execution system in this invention can be a WMS (Warehouse Management System), WCS (Warehouse Control System), TMS (Tracking Management System), ERP (Enterprise Resource Planning), OMS (Operating Management System), SRM (Supply Chain Relationship Management System), or other systems capable of receiving inventory adjustment control instructions and executing corresponding inventory adjustment operations.

[0024] Inventory supply capacity refers to the amount of inventory available to meet demand within a target time window H, calculated based at least on currently available inventory, in-transit inventory that can arrive on time and be put on the shelves, locked inventory, and frozen inventory. Warehouse execution capacity refers to the processing capacity of a target warehouse node to complete tasks such as receiving, shelving, replenishment, picking, transfer, or location adjustment within a target time window. It is determined at least based on the task queue length, personnel or equipment processing capacity, historical processing volume per unit time, and task failure rate. The inventory-demand matching degree M ranges from [0,1]. The lower the M, the higher the risk of mismatch. When M is lower than the first matching threshold and the identification confidence is higher than the first confidence threshold, the inventory adjustment control process is triggered. Similar inventory objects refer to inventory objects that meet at least three of the following preset similarity conditions: SKU category, volume, storage location type, historical demand variation coefficient, supply delivery stability, and storage node type. The unified event timeline includes mapping rules for event occurrence time, system reception time, and source system timestamp; Task queue load is the ratio of the number of tasks to be processed within the target time window to the processing capacity. Storage capacity pressure is the ratio of occupied capacity to total capacity; The expected stockout period is the point at which the current available inventory and the forecast demand consumption curves first intersect; The time when replenishment or allocation is completed and available for shelving is the estimated time when the goods arrive and are ready for shelving, not the time when they are transported to the warehouse. The triggering condition is that the inventory-demand matching degree M is lower than the first matching threshold and the identification confidence is higher than the first confidence threshold; the manual review condition is that the inventory-demand matching degree M is lower than the second matching threshold and the identification confidence is lower than the first confidence threshold.

[0025] Example 1: This embodiment provides an adaptive optimization method for inventory demand fluctuations, such as... Figure 1 As shown, it includes the following steps: Step 1: Obtain multi-source inventory execution data The system acquires multi-source inventory execution data for the target inventory unit SKU and target warehouse node; the multi-source inventory execution data includes order data, inventory snapshot data, inbound and outbound flow data, in-transit inventory data, locked inventory data, frozen inventory data, storage location capacity data, replenishment execution data, transfer execution data, picking execution data, warehousing execution data, supply delivery date data, and abnormal event data.

[0026] The data includes: order data (creation time, quantity, status, fulfillment time, region, and channel); inventory snapshot data (book inventory, available inventory, locked inventory, frozen inventory, safety stock, and location inventory); in-transit inventory data (quantity, estimated arrival time, actual arrival time, transportation milestones, and transportation anomaly status); location capacity data (location type, capacity, occupancy rate, picking area capacity, and storage area capacity); warehouse execution data (replenishment task status, transfer task status, inbound task status, picking task status, task creation time, start time, completion time, and reason for task failure); supply delivery time data (supplier promised delivery time, actual delivery time, historical delivery deviations, and supplier fulfillment stability); and anomaly event data (promotional activities, holidays, weather anomalies, logistics anomalies, supplier anomalies, and system synchronization anomalies).

[0027] When acquiring order data and abnormal event data, the system only collects fields necessary for inventory demand fluctuation identification and warehouse execution control. These fields include at least one of SKU, warehouse node, quantity, time, order status, channel category, region code, and abnormal event category. When the original order data contains natural person identification, contact information, precise address, or other personal information, the system deletes, desensitizes, aggregates, or anonymizes the corresponding fields before constructing the inventory demand fluctuation status vector. This ensures that status identification, matching degree calculation, and control instruction generation are processed at the granularity of SKU, warehouse node, storage area, or time window, rather than at the level of a single SKU. Natural persons are the controlled objects; the order region field is retained only up to the region code or delivery area required for inventory scheduling; the channel category is only used to identify the load of the fulfillment system and is not used for differentiated decisions on price, service level, or fulfillment quality for individual consumers; direct identity identifiers are removed from training samples, log samples, and execution feedback samples, and individual natural person orders are not used as model training labels; the system records audit logs for the data access, feature generation, model training, and control instruction generation processes, and generates control candidates for manual review when the data confidence level is lower than a preset threshold, the identification confidence level is lower than a preset threshold, or the control instruction is abnormal. The system sets access permissions and audit logs for data access, feature generation, model training, instruction generation, and manual review; when the identification confidence level is lower than the confidence threshold, the data confidence level is lower than the data threshold, or the control instruction involves abnormal high priority / abnormal large quantity / abnormal frequency, it is not directly sent to the execution end, but generates control candidates for manual review.

[0028] Step 2: Preprocess the multi-source inventory execution data The system maps data collected from different data sources to a unified timeline and divides the time window according to a preset time granularity.

[0029] For target SKU i Given the target warehouse node j and time window t, the system constructs the data unit: D(i,j,t)={order data, inventory data, in-transit data, storage location data, supply delivery date data, warehousing execution data, and abnormal event data}; During data preprocessing, the system removes duplicate data, corrects timestamps from different systems, marks or corrects obviously abnormal data, completes missing data, and generates data confidence scores based on the data completeness.

[0030] For example, when order data and inventory data are complete, but in-transit data are missing within a certain time window, the system can estimate the missing data based on historical in-transit arrival data for the same period, in-transit patterns of similar SKUs, or feedback data from transportation nodes, and reduce the confidence level of the in-transit data within that time window.

[0031] Through the above preprocessing, the system obtains a standardized inventory execution dataset.

[0032] Step 3: Extracting Multiple Fluctuation Features The system extracts demand fluctuation characteristics, inventory status characteristics, supply disturbance characteristics, and warehouse execution disturbance characteristics based on a standardized inventory execution dataset.

[0033] Demand fluctuation characteristics include short-cycle average demand, long-cycle average demand, demand coefficient of variation, month-on-month growth rate, year-on-year deviation rate, percentage of zero demand window, demand surge index, trend slope, and degree of order fulfillment time compression; among which, demand trend characteristics can be expressed as: Among them, MA short MA represents the short-term moving average of demand. long This represents the long-term moving average demand, and ε is a smoothing parameter to prevent the denominator from being zero.

[0034] The characteristics of a sudden surge in demand can be represented as follows: ; where Q current Median(Q) represents the demand within the current time window. history MED(Q) represents the historical median demand. history () represents the absolute median difference in historical demand.

[0035] Inventory status characteristics include available inventory, locked inventory, frozen inventory, inventory in transit, inventory coverage days, safety stock deviation, inventory turnover rate, inventory discrepancy between book and physical inventory, and storage space occupancy rate.

[0036] Supply disruption characteristics include supplier delivery time deviations, in-transit arrival delays, in-transit inventory anomaly rates, purchase order delay rates, and the interval between expected arrival time and expected stockout time.

[0037] Warehouse execution disturbance characteristics include replenishment task queue length, picking task queue length, inbound task completion delay, transfer task completion delay, storage space capacity occupancy rate, equipment load, task failure rate, and picking wave completion rate.

[0038] Step 4: Construct an inventory demand fluctuation state vector and identify fluctuation states. like Figure 2 As shown, the system constructs an inventory demand fluctuation state vector based on various extracted features: V(i,j,t)=[F demand F inventory F supply F execution F event ,Conf]; where F demand F represents the set of demand fluctuation characteristics. inventory F represents the set of inventory status features.supply F represents the set of characteristics of supply disturbances. execution F represents the set of warehouse execution disturbance characteristics. event This represents the set of features of abnormal events, and Conf represents the data confidence or identification confidence.

[0039] The system identifies the inventory demand fluctuation state based on the inventory demand fluctuation state vector; the identification method can be rule-based threshold identification, machine learning classification identification, or a combination of rule-based threshold and classification model.

[0040] In one specific implementation, the inventory demand fluctuation state identification module includes a candidate state generation submodule, a feature encoding submodule, a state classification submodule, and a confidence calibration submodule. The candidate state generation submodule generates a set of candidate fluctuation states based on demand surge index, inventory coverage days, in-transit arrival delay, replenishment task queue load, picking task queue load, storage capacity pressure, and inventory discrepancy rate. The feature encoding submodule normalizes continuous features, encodes abnormal event types, warehouse node types, and SKU categories, and concatenates the features of the most recent N sliding time windows in chronological order into an input matrix. The state classification submodule uses a gradient boosting tree model or a one-dimensional temporal convolutional network to output the probability of each inventory demand fluctuation state. The confidence calibration submodule determines the identification confidence based on data completeness, timestamp consistency, and model output probability, and determines the state with the highest probability and confidence meeting the threshold as the final inventory demand fluctuation state.

[0041] When using a classification model to identify inventory demand fluctuations, the target inventory object, warehouse node, and sliding time window are used as sample units. The demand fluctuation characteristics, inventory status characteristics, supply disturbance characteristics, warehousing execution disturbance characteristics, abnormal event characteristics, and data confidence characteristics within the most recent N sliding time windows are concatenated in chronological order to form an input matrix, where N is an integer greater than 1.

[0042] The classification model comprises an input normalization layer, a temporal feature extraction layer, a temporal pooling layer, and an output layer. The temporal feature extraction layer is formed by connecting at least two one-dimensional convolutional layers or recurrent units. The output layer outputs the probability of each inventory demand fluctuation state. Inventory demand fluctuation states include demand change states, supply delay states, in-transit anomaly states, warehouse execution congestion states, warehouse capacity limited states, inventory-to-physical synchronization anomaly states, and composite fluctuation states.

[0043] Training sample labels are generated based on actual stockout records, actual arrival delay records, task failure reasons, warehouse capacity alarm records, and records of abnormal synchronization between accounts and physical inventory after the historical window. Training samples are divided into training, validation, and test sets in chronological order, and weighted cross-entropy loss is used for training. The classification threshold and confidence threshold are determined through the validation set. Identification confidence is determined based on data completeness, timestamp consistency, and the output probability of the classification model.

[0044] When using a one-dimensional temporal convolutional network for state classification, the input is the feature matrix X(i,j,t)∈R of the N most recent sliding time windows. N×d , where d is the total dimension of demand fluctuation characteristics, inventory status characteristics, supply disturbance characteristics, warehousing execution disturbance characteristics, abnormal event characteristics, and confidence characteristics; the one-dimensional temporal convolutional network includes an input normalization layer, at least two one-dimensional convolutional layers, a nonlinear activation layer, a temporal pooling layer, and a fully connected output layer; the output layer uses Softmax to output the probability of K types of inventory demand fluctuation states; the training labels are generated from actual stockouts, actual arrival delays, task failure reasons, warehouse capacity alarms, and abnormal records of account-physical synchronization within the historical window; the loss function is a weighted cross-entropy loss, and the class weights are determined according to the frequency of occurrence of each fluctuation state in the training samples.

[0045] In one embodiment, N is a sliding time window of 6 to 48, and d includes at least demand trend characteristics, demand surge characteristics, inventory coverage days, available inventory ratio, in-transit delay duration, supply delivery date deviation, replenishment task queue load, picking task queue load, inbound task queue load, storage capacity pressure, inventory discrepancy rate, abnormal event coding, and data confidence. Continuous features are normalized using median absolute deviation (MAD) or z-score, and categorical features are normalized using one-hot encoding or low-dimensional embedding encoding. The one-dimensional temporal convolutional network includes an input normalization layer, a first one-dimensional convolutional layer, a first activation layer, a second one-dimensional convolutional layer, a second activation layer, a temporal pooling layer, and a fully connected output layer connected in sequence; the convolutional kernel width is 3 or 5, the number of channels is 16 to 128, and the activation function is ReLU, LeakyReLU, or GELU; the number of nodes K in the output layer corresponds to demand changes, supply delays, in-transit anomalies, warehouse execution congestion, limited storage capacity, inventory synchronization anomalies, and compound fluctuation states.

[0046] Training samples are divided into training, validation, and test sets in chronological order to avoid the same anomalous event window appearing in both the training and test sets simultaneously; class weights w k The frequency of samples in class k in the training set is determined by the reciprocal of their occurrence frequency, and is limited to preset upper and lower limits; the loss function is weighted cross-entropy loss; the recognition confidence is jointly determined by data completeness, timestamp consistency, and classification output probability, Conf=q data ×qtime ×max(Softmax(X)), where q data q is determined based on the number of missing fields and the degree of missing key fields. time The determination is based on the consistency of the event occurrence time, the source system timestamp, and the system reception time.

[0047] In one implementation, for each SKU-warehouse node-time window sample, if the actual available inventory falls below zero or below the safety stock threshold within the forecast period H after window t, and the supply disturbance, in-transit anomaly, and warehouse execution disturbance are all below the corresponding thresholds, it is marked as a demand change state; if the actual arrival time is later than the promised arrival time by more than a first delay threshold, and the replenishment task queue load is lower than the second load threshold, it is marked as a supply delay state; if the transportation node is abnormal or the in-transit status is stagnant for more than a third time threshold, it is marked as an in-transit anomaly state; if the replenishment, warehousing, or picking task queue load exceeds the second load threshold and causes the actual completion of shelving to be later than the planned completion of shelving, it is marked as a warehouse execution congestion state; if the ratio of the target warehouse area's occupied capacity to the total capacity exceeds the capacity threshold and causes instruction verification failure or shelving task delay, it is marked as a warehouse location capacity limited state; if the difference rate between book inventory and physical inventory exceeds the book-physical difference threshold, it is marked as an inventory book-physical synchronization anomaly state; when two or more states are simultaneously satisfied and all affect the expected stockout time or instruction executability, it is marked as a compound fluctuation state.

[0048] In one implementation, the system first determines candidate fluctuation states based on preset thresholds, and then outputs the final fluctuation state and identification confidence level through a classification model. For example, when the demand surge feature B is greater than the preset demand surge threshold, the inventory coverage days are lower than the preset coverage threshold, and the arrival time in transit is later than the expected stockout time, the system identifies the target SKU as a composite fluctuation state of "sudden demand growth plus in-transit inventory delay".

[0049] When the ratio of replenishment task queue length to processing capacity exceeds the preset execution load threshold, the system further identifies a warehouse execution congestion state; when the storage location occupancy rate exceeds the preset storage location capacity threshold, the system further identifies a storage location capacity limitation state.

[0050] Step 5: Calculate the inventory-demand match. After identifying the fluctuations in inventory demand, the system further calculates the inventory-demand matching degree.

[0051] The demand gap within the target period H can be expressed as: Gap = Demand H -AvailableSupply H Among them, Demand HIndicates the demand within the target period H or the demand quantile at a specified confidence level. AvailableSupply H This represents the amount of inventory available to meet demand within the target period H; where AvailableSupply H This can be represented as: AvailableSupply H =AvailableStock+OnTimeInTransitStock-LockedStock-FrozenStock; where AvailableStock represents the currently available inventory, OnTimeInTransitStock represents the inventory in transit that is expected to arrive and be available for sale within the target period H, LockedStock represents the inventory that has been locked by orders but has not yet been shipped, and FrozenStock represents the frozen or unavailable inventory.

[0052] The system further calculates the time pressure between the estimated stockout time and the estimated replenishment arrival time: P time =max(0,ArrivalTime replenishment -StockOutTime estimated ArrivalTime replenishment StockOutTime indicates the estimated time before stock can be put back on the shelves after replenishment or allocation is completed. estimated This indicates the estimated stockout duration based on current inventory demand fluctuations; P represents the stockout period when replenishment or allocation is completed and available for shelving no later than the estimated stockout duration. time It is 0.

[0053] The system also calculates warehouse execution load: Where TaskQueueLength represents the number of tasks to be processed within the target time window, and ProcessingCapacity represents the processing capacity of the corresponding job resources.

[0054] The system also calculates the storage space capacity pressure: Where OccupiedCapacity represents the occupied capacity of the target storage area or storage location, and TotalCapacity represents the total capacity.

[0055] In one implementation, the system calculates the inventory-demand mismatch risk value R as follows: R = α1 × Norm(Gap) + α2 × Norm(P) time )+α3×Norm(P exec )+α4×Norm(P loc)+α5×Norm(L)+α6×Norm(E); where Norm represents the normalization process, L represents the supply disturbance characteristics, determined by the supply delivery time deviation and the in-transit anomaly rate; E represents the execution disturbance characteristics, determined by the timeout rate and failure rate of warehousing, replenishment, picking, or transfer tasks; Gap represents the demand fulfillment gap, P time P represents time pressure. exec P represents the task queue load. loc α1 to α6 represent the weights of each factor, and satisfy α1+α2+α3+α4+α5+α6=1; when any normalized feature exceeds [0,1] due to abnormal data, extreme fluctuations or missing value completion, the normalized feature is pruned to [0,1].

[0056] The inventory-demand matching degree M can be expressed as: M = 1 - R; M ∈ [0, 1], the lower the M, the higher the mismatch risk; when the inventory-demand matching degree M is lower than the preset matching threshold, it indicates that there is a risk of inventory and demand mismatch within the target inventory object in the target time window, and the inventory adjustment control process needs to be triggered; the inventory-demand matching degree M is used to control whether the warehousing execution system generates and what kind of inventory adjustment control instruction is generated; unlike replenishment judgment based solely on predicted demand, the matching degree calculation of this invention simultaneously introduces the time pressure between the expected stockout time and the expected replenishment or transfer completion available shelf time, the task queue load of the target warehouse area or picking area, the warehouse capacity pressure, and supply disturbances and execution disturbances; when the demand gap is the same but the time pressure, Different types of control instructions are generated by the system depending on the load of the task queue or the capacity pressure of the storage location. For example, when the purchase replenishment cannot arrive before the expected stockout time, the system prioritizes generating cross-warehouse transfer instructions. When the load of the picking task queue exceeds the threshold, the system generates picking wave adjustment instructions. When the capacity pressure of the target storage area exceeds the threshold, the system generates storage location adjustment instructions or inbound priority adjustment instructions. Each weight α1 to α6 is initialized based on historical execution samples and normalized again after being updated and limited to a preset range. When M is lower than the first matching threshold and the recognition confidence is higher than the first confidence threshold, the control instruction generation process begins. When M is lower than the second matching threshold but the confidence is insufficient, only control candidates awaiting manual review are generated and not directly issued to the execution end.

[0057] Step Six: Generate Inventory Adjustment Control Instructions like Figure 3 As shown, when the inventory-demand matching degree meets the preset triggering conditions, the system generates an inventory adjustment control instruction based on the inventory-demand fluctuation status, the inventory-demand matching degree, and the warehouse execution constraints.

[0058] When the system identifies an upward trend in demand, a sudden increase in demand, or insufficient inventory coverage, and the supplier's delivery time can meet the expected stockout time requirement, the system generates a replenishment instruction. The replenishment instruction includes SKU number, warehouse node, replenishment quantity, target arrival time, supplier or upstream warehouse, task priority, minimum packaging quantity constraint, and safety stock adjustment parameters.

[0059] When the target warehouse has insufficient inventory, other warehouses have available inventory that can be transferred, and the transfer arrival time is earlier than the purchase replenishment arrival time, the system generates a transfer instruction. The transfer instruction includes the warehouse to be transferred out, the warehouse to be transferred in, the SKU number, the quantity to be transferred, the estimated shipping time, the estimated arrival time, the transportation route, and the transfer priority.

[0060] When the storage capacity of the target SKU is limited, the inventory in the picking area is insufficient, or the storage layout affects replenishment efficiency, the system generates a storage location adjustment instruction. The storage location adjustment instruction includes the target SKU, the original storage location, the target storage location, the adjustment quantity, the recommended adjustment time, the storage location type, and whether to enter the fast picking area.

[0061] When order fulfillment time is shortened, picking tasks are congested, or high-priority orders occur in a concentrated manner, the system generates a picking wave adjustment instruction. The picking wave adjustment instruction includes the target order set, target SKU, picking wave number, wave priority, picking path adjustment parameters, and equipment or personnel resource allocation parameters.

[0062] When in-transit inventory is about to arrive and the target SKU has a high risk of stockout, the system generates an inbound priority adjustment instruction. The inbound priority adjustment instruction includes the inbound order number, SKU number, target shelving area, shelving priority, expected completion time, and whether to prioritize entry into the picking area.

[0063] When changes in demand trends, supply delivery dates, or execution delays occur within multiple consecutive time windows, the system generates a safety stock threshold adjustment instruction. The safety stock threshold adjustment instruction includes the SKU number, warehouse node, original safety stock threshold, adjusted safety stock threshold, threshold validity period, triggering reason, and rollback conditions.

[0064] Each inventory adjustment control instruction includes an instruction identifier, instruction type, target SKU, target warehouse node, source location or warehouse, target location or warehouse, adjustment quantity, planned start time, planned completion time, priority, constraint fields, target execution system identifier, idempotent key, version number, rollback conditions, and status feedback address. The target execution system identifier indicates that the instruction is executed by one or more systems, including WMS, WCS, TMS, ERP, OMS, SRM, or automated replenishment systems. The idempotent key is used to prevent the same instruction from being issued repeatedly within the same fluctuation window. The version number is used to cancel old instructions and generate new instructions when inventory snapshots, task queue status, or warehouse location status changes. The rollback conditions are used to cancel, split, or regenerate control instructions when the target warehouse capacity is insufficient, the transfer path fails, the source warehouse inventory is below the safety threshold, or the task queue exceeds the load threshold.

[0065] In one specific implementation, the instruction generation rules are as follows: the idempotent key is generated by hashing the target inventory object, the target warehouse node, the sliding time window number, the fluctuation status, the instruction type, and the target execution terminal identifier; the version number increases with the changes in the inventory snapshot version, the task queue version, or the storage location status version; when the status feedback shows that the old version instruction has not been executed and the new version instruction has been generated, the old version instruction enters the cancellation or rollback process; the rollback conditions include the source warehouse inventory being lower than the safety threshold, the transfer path failing, the target storage location capacity being insufficient, and the task queue load exceeding the second load threshold.

[0066] Step 7: Verify and issue control command executability Before issuing inventory adjustment control instructions, the system performs an executability check on the instructions. The executability check includes: checking whether the target storage location has sufficient capacity; checking whether there is sufficient inbound, replenishment, picking, or transfer processing capacity within the target time window; checking whether the supplier or upstream warehouse has sufficient available inventory; checking whether the transfer route is available; checking whether the estimated arrival time is earlier than the estimated stockout time; checking whether the minimum order quantity, minimum packaging quantity, or full pallet constraint is met; checking whether there is a conflict with existing tasks; and checking whether it will cause the inventory in other warehouses to fall below the safety threshold.

[0067] If a control command fails the executability check, the system corrects the control command; the correction includes reducing the replenishment quantity, splitting task batches, changing the warehouse from which the command is transferred, adjusting the target storage location, postponing non-urgent tasks, or increasing the priority of the current task.

[0068] After the executability verification is passed, the system sends the inventory adjustment control instruction to the corresponding warehousing or supply chain execution system; the warehousing or supply chain execution system performs replenishment, transfer, storage location adjustment, picking wave adjustment, inbound priority adjustment or safety stock threshold adjustment operations according to the received control instruction.

[0069] Step 8: Perform deviation closed-loop adaptive update After the inventory adjustment control command is executed, the system collects the actual execution results. The actual execution results include whether the control command was executed, the actual execution start time, the actual execution completion time, the actual execution quantity, the actual replenishment arrival time, the actual transfer arrival time, the actual shelving time, the actual outbound quantity, the actual stockout quantity, the actual inventory coverage days, the actual task failure reason, the actual warehouse location occupancy change, and the actual picking completion rate. The system calculates the execution deviation based on the actual execution results and the target execution results.

[0070] Demand deviation can be expressed as: The execution time deviation can be expressed as: When the execution deviation exceeds the preset deviation threshold, the system adaptively updates the inventory demand fluctuation identification parameters, inventory-demand matching model parameters, and inventory adjustment control parameters. Execution deviations include demand deviation, arrival time deviation, task completion time deviation, actual execution quantity deviation, actual stockout quantity deviation, and task failure type. The system selects different update objects based on the deviation type: when the actual demand deviation exceeds the threshold but supply disturbances, in-transit anomalies, and warehouse execution disturbances do not exceed the threshold, the demand fluctuation identification parameters are updated; when the deviation between the expected stockout time and the actual stockout time exceeds the threshold, the demand gap weight and time pressure weight in the inventory-demand matching model are updated; when the task failure is due to insufficient storage capacity, unavailable transfer paths, insufficient source warehouse inventory, or queue overload, the control instruction generation parameters and executability verification threshold are updated. The above updates are preferentially applied to the target SKU, target warehouse node, or SKU group with similar characteristics, and a maximum single update amplitude is set to avoid global model parameter oscillations caused by local abnormal events.

[0071] For example, the system can update the weights of the inventory-demand matching model as follows: α(k+1)=Normalize[α(k)+η×Err×X]; where α(k) represents the weight vector before the k-th update, η represents the learning rate, Err represents the execution bias, X represents the feature vector within the corresponding time window, and Normalize represents the normalization process.

[0072] The replenishment trigger threshold can be updated as follows: θ(k+1)=Clip[θ(k)+β×(Err)] stockout -Err target ]; where θ(k) represents the replenishment trigger threshold before the k-th update, β represents the threshold update step size, and Err stockout Indicates stockout deviation, Err target This indicates the target deviation, and Clip means that the updated threshold will be limited to a preset range.

[0073] When the execution deviation of a target SKU or warehouse node exceeds the preset deviation threshold in multiple consecutive time windows, the system triggers a partial update, updating only the model parameters corresponding to the target SKU, the warehouse node, or the same group of SKUs.

[0074] Through the aforementioned closed-loop adaptive update, this invention can continuously correct the fluctuation identification rules, inventory-demand matching degree model, and control command generation parameters based on actual execution results, thereby improving the accuracy of subsequent inventory adjustment control.

[0075] Example 2 This embodiment illustrates the specific application process of the present invention using a multi-warehouse inventory transfer scenario. An e-commerce company has set up a first warehouse W1 and a second warehouse W2, with the target SKU being SKU-A. The system collects order data, inventory data, in-transit inventory data, storage capacity data, replenishment task data, transfer task data, picking task data, and inbound task data for SKU-A in W1 and W2 every hour.

[0076] Within a certain time window, the system detected that the current order volume of SKU-A in W1 increased by 80% compared to the historical median for the same period. The current available inventory in W1 can only cover the demand for the next 10 hours, and a batch of inventory in W1 is expected to arrive in 24 hours. At the same time, the picking task queue length for SKU-A in W1 increased by 60% compared to the normal level, and the occupancy rate of the fast picking area reached 92%. The system further detected that there was sufficient inventory of SKU-A in W2, and the estimated arrival time for transfers from W2 to W1 was 8 hours.

[0077] The system first aligns the order data, inventory data, in-transit inventory data, storage capacity data, and warehousing execution data with time to form a standardized inventory execution dataset for SKU-A under the current time window of W1.

[0078] Subsequently, the system extracts features such as sudden demand surge characteristics, inventory coverage days, in-transit arrival delays, picking task load, and storage capacity pressure, and constructs an inventory demand fluctuation state vector. Based on this inventory demand fluctuation state vector, the system identifies the inventory demand fluctuation state of SKU-A in W1 as a composite fluctuation state of sudden demand growth + in-transit replenishment lag + picking execution congestion + limited capacity of the fast picking area.

[0079] The system further calculates the inventory-demand matching degree; since the expected stockout time is earlier than the arrival time of W1 in-transit inventory, and the picking task queue length and storage capacity pressure are both high, the inventory-demand matching degree is lower than the preset matching threshold. The system triggers the inventory adjustment control process.

[0080] After verifying the executability of the optional control strategies, the system generates the following inventory adjustment control instructions: First, transfer a predetermined quantity of SKU-A inventory from W2 to W1 and set this transfer task as a high priority; Second, increase the priority of the inbound task corresponding to SKU-A in W1, so that the inventory that is about to arrive in transit can be inbound and put on the shelf first; Third, move part of the inventory of SKU-A from the storage area to the fast picking area, while removing low-priority SKUs from the fast picking area to free up storage space; Fourth, adjust the picking wave of orders related to SKU-A, and process high-time-sensitive orders in advance; Fifth, temporarily increase the safety stock threshold of SKU-A in W1 within the preset validity period and set rollback conditions.

[0081] The aforementioned inventory adjustment control instructions are sent to the warehouse management system, transportation management system, and warehouse control system, respectively. The warehouse management system adjusts the warehouse operation tasks according to the location adjustment instructions and picking wave adjustment instructions. The transportation management system generates transfer and transportation tasks according to the transfer instructions. The warehouse control system adjusts the processing order of automated equipment or operation queues according to the inbound priority adjustment instructions.

[0082] After the control command is executed, the system continues to collect the actual execution results. If the system detects that the actual transfer arrival time is 10 hours, which is 2 hours later than the expected transfer arrival time, and the picking task is still delayed, the system calculates the transfer completion deviation and the picking execution deviation, and increases the weight of transportation timeliness and execution congestion in the inventory-demand matching model. When the same or similar fluctuation state occurs again, the system can trigger the transfer command earlier and adjust the picking wave and warehouse layout earlier.

[0083] Therefore, this invention can not only generate inventory adjustment plans based on changes in demand, but also identify the supply, in-transit, storage location and warehousing execution factors behind demand fluctuations, convert the optimization results into executable control instructions, and then perform closed-loop adaptive updates based on the actual execution results.

[0084] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An adaptive optimization method for inventory demand fluctuations, applied to an inventory control system communicating with a warehousing or transportation execution terminal, characterized in that, Perform the following steps: Retrieve data on orders, inventory, in-transit, storage capacity, warehousing execution, supply delivery date, and abnormal events by inventory object, warehouse node, and target time window; A unified event timeline is established based on the event occurrence time, source system timestamp, and system reception time. Data is then deduplicated, time-corrected, anomaly-marked, missing data-completed, unit-converted, confidence-marked, and windowed to form a standardized inventory execution dataset. The dataset extracts features such as demand, inventory, supply, task queue, storage capacity, account-physical synchronization, and data confidence. A state vector is constructed based on the most recent N sliding windows to identify fluctuation states and confidence levels. Calculate the inventory-demand matching degree based on the demand gap, the time pressure caused by the expected stockout time and the available shelf time, the task queue load, and the storage capacity pressure. When the matching degree is lower than the matching threshold and the recognition confidence degree is higher than the confidence threshold, a machine-parseable inventory adjustment control instruction is generated. Before sending, the target warehouse capacity, processing capacity, completion time of shelving, task conflicts, available inventory in the source warehouse and transfer path are verified. If the verification fails, the following measures are taken: reducing quantity, splitting batches, changing source warehouse, adjusting warehouse location, postponing task or increasing priority, and then re-verifying. After verification, the data is sent to the execution end to generate or adjust the inventory execution task, and the fluctuation identification parameters, matching degree weights or control thresholds are updated in a single amplitude-limited manner according to the returned execution results.

2. The adaptive optimization method for inventory demand fluctuations according to claim 1, characterized in that, The windowing is established by inventory object, warehouse node and target time window to create data unit, and writes data from different sources into the same data unit according to a unified event timeline; when there is a time conflict, the source system timestamp is calibrated by the event occurrence time, and the system reception time is used to determine the window to which the late event belongs.

3. The adaptive optimization method for inventory demand fluctuations according to claim 1, characterized in that, When constructing the state vector, demand trends, demand surges, inventory coverage days, available inventory ratio, in-transit delay time, supply delivery deviation, replenishment task queue load, picking task queue load, inbound task queue load, storage capacity pressure, book-to-physical discrepancy rate, and data confidence are written in chronological order.

4. The adaptive optimization method for inventory demand fluctuations according to claim 1, characterized in that, When identifying fluctuation states, candidate fluctuation states are first generated based on demand surges, inventory coverage days, in-transit delays, task queue loads, storage capacity pressures, and book-to-physical discrepancy rates. Then, the features of the most recent N sliding windows are input into the time series classification model to determine the output state from demand changes, supply delays, in-transit anomalies, warehouse execution congestion, limited storage capacity, book-to-physical synchronization anomalies, and compound fluctuation states.

5. The adaptive optimization method for inventory demand fluctuations according to claim 4, characterized in that, The training sample labels of the time-series classification model are generated from the actual stockout, actual arrival delay, task failure reasons, warehouse capacity alarm, and abnormal records of account-to-physical synchronization after the historical window; during training, the samples are divided in chronological order, and the classification threshold and confidence threshold are determined by verifying the samples.

6. The adaptive optimization method for inventory demand fluctuations according to claim 1, characterized in that, The confidence level is calculated by combining data completeness, timestamp consistency, and the probability of fluctuating output states. When the matching degree is lower than the matching threshold but the recognition confidence degree is not higher than the confidence threshold, the transmission to the execution end is stopped, and control candidates are generated for manual review.

7. The adaptive optimization method for inventory demand fluctuations according to claim 1, characterized in that, When calculating the inventory-demand matching degree, the current available inventory and the on-time available inventory in transit are summed and then the locked inventory and frozen inventory are deducted to obtain the inventory supply. The demand gap is obtained by comparing the target cycle demand and the inventory supply. The mismatch risk value is then calculated after normalizing the demand gap, time pressure, task queue load, and storage capacity pressure.

8. The adaptive optimization method for inventory demand fluctuations according to claim 7, characterized in that, The normalization is trimmed to [0,1] based on the historical percentile threshold or the physical execution limit; when the demand gap is the same but the time pressure, task queue load or storage capacity pressure is different, the control instruction generation logic for cross-warehouse transfer, wave adjustment, inbound priority adjustment or storage location adjustment is triggered respectively.

9. The adaptive optimization method for inventory demand fluctuations according to claim 1, characterized in that, When generating inventory adjustment control instructions, the instruction identifier, instruction type, inventory object, target warehouse node, adjustment quantity, execution end identifier, idempotent key, version number, rollback condition, and status feedback address are written into the instruction message. Duplicate instructions with the same idempotent key within the same fluctuation window are intercepted.

10. The adaptive optimization method for inventory demand fluctuations according to claim 1, characterized in that, When updating parameters based on execution deviation, the fluctuation identification parameters are updated when the actual demand deviation exceeds the threshold but the supply and execution disturbances do not exceed the threshold; the demand gap weight and time pressure weight are updated when the expected stockout time deviation exceeds the threshold; and the executability verification threshold and instruction correction rules are updated when the task failure reason corresponds to the storage capacity, allocation path, source warehouse inventory, or queue load.

Images