Dynamic warehouse partitioning and inventory presetting prediction method and device based on live broadcast e-commerce traffic

Abstract

The application discloses a dynamic warehouse distribution and inventory presetting prediction method and device based on live broadcast e-commerce traffic, which comprises the following steps: collecting multi-dimensional live broadcast e-commerce related data and preprocessing to obtain a standardized feature set; inputting the standardized feature set into a fusion prediction model to output total order quantity, real-time order curve and order regional distribution prediction results of a target live broadcast; based on the prediction results, warehouse network data and commodity physical attributes, a preset inventory quantity matrix of each SKU in different warehouses is generated through a double-objective optimization algorithm to complete the pre-live broadcast inventory distribution and presetting; in the live broadcast process, the deviation between actual order data and the prediction results is monitored in real time, and when the deviation exceeds a preset threshold, emergency allocation calculation is started and the distribution strategy is updated; after the live broadcast ends, orders are distributed and fulfilled according to the real-time updated distribution strategy, and full-link actual data is collected and fed back to a data pool to iteratively optimize the fusion prediction model and the double-objective optimization algorithm parameters.

Description

Technical Field

[0001] This invention belongs to the fields of live-streaming e-commerce, intelligent logistics and warehouse management technology, specifically involving a method and device for dynamic warehouse allocation and inventory pre-positioning prediction based on live-streaming e-commerce traffic. Background Technology

[0002] With the rapid development of the live-streaming e-commerce industry, its strong interactivity and high sales efficiency have enabled peak order volumes in a single live stream to reach 3-10 times the daily sales volume, forming a unique sales model characterized by instantaneous traffic peaks, concentrated sales of best-selling products, and distinct regional consumption preferences. Under this model, order demand exhibits core characteristics such as short-term explosive growth (80% of orders in a single live stream are concentrated within 1-3 hours after the broadcast starts) and dynamic changes in geographical distribution (the geographical coverage of top streamers' fans exceeds 90% and fluctuates in real time), posing a severe challenge to the warehousing and inventory management systems of traditional supply chains.

[0003] In the warehousing and inventory management of the e-commerce industry, warehousing strategies and inventory pre-positioning directly impact logistics costs (accounting for 30%-40% of e-commerce operating costs), order fulfillment timeliness, and customer experience. Traditional e-commerce warehousing and inventory management primarily relies on historical sales data from all channels to formulate strategies, employing fixed warehousing ratios and experience-based inventory preparation. However, the emerging sales model of live-streaming e-commerce differs fundamentally from traditional models, and the existing supply chain system struggles to adapt to its demands: static warehousing cannot respond to real-time fluctuations in order geographic distribution, and experience-based inventory preparation easily leads to stockout rates exceeding 20% ​​or inventory backlog rates exceeding 30%. Therefore, an intelligent solution that can link live-streaming traffic with supply chain responsiveness is urgently needed.

[0004] In the existing technologies of the e-commerce industry, the commonly used static warehousing and inventory forecasting management solutions based on historical sales data have the following core logic: only collect annual / quarterly historical sales data from all channels, use the ARIMA model to perform linear extrapolation forecasts based on historical sales time series, determine a fixed warehousing ratio by calculating the proportion of historical sales in each region, allocate inventory to warehouses in each region according to a fixed ratio, and use a fixed rule of "nearest allocation" for order fulfillment without a dynamic adjustment mechanism.

[0005] The existing technology has the following significant drawbacks:

[0006] 1. Rigid warehouse allocation model: The static warehouse allocation strategy allocates warehouse resources based on the historical regional sales ratio, which cannot adapt to the real-time dynamic changes in the geographical distribution of live e-commerce orders, resulting in a mismatch between warehouse resources and order demand;

[0007] 2. Large inventory forecasting deviation: The ARIMA model relies only on linear historical data and cannot capture the short-term explosive and non-linear demand characteristics in the live streaming scenario. The prediction error rate is generally over 35%, which can easily lead to insufficient stock preparation or inventory backlog.

[0008] 3. Supply chain response lag: Inventory is concentrated in fixed warehouses. When live streaming drives sales, a large number of orders need to be shipped from long distances or transferred across warehouses, increasing transportation costs by 40%-60% and extending delivery time by 2-3 days, resulting in a customer complaint rate increase of more than 15%.

[0009] 4. Lack of data linkage: The inventory management system is disconnected from the live streaming traffic data, and fails to utilize dynamic data such as the number of viewers, interaction rate, and conversion rate in the live streaming room in real time, resulting in a lack of foresight in inventory preparation decisions;

[0010] 5. Lack of closed-loop optimization mechanism: The actual data after order fulfillment is not fed back to the prediction and distribution model, the model parameters are fixed in the long term, and the accuracy of decision-making cannot be continuously improved. Summary of the Invention

[0011] To address the shortcomings of existing technologies, achieve precise matching of inventory and distribution strategies, and improve supply chain efficiency and customer experience, this invention adopts the following technical solution:

[0012] The dynamic warehousing and inventory pre-positioning method based on live-stream e-commerce traffic prediction includes the following steps:

[0013] Step S1: Collect multi-dimensional data related to the live stream and products;

[0014] Step S2: Construct a fusion model to obtain live streaming and product-related features from multi-dimensional data in order to predict order volume data and order distribution;

[0015] Step S3: Based on the inventory in the pre-positioned warehouse area, construct a linear programming model to generate dynamic warehouse allocation and inventory pre-positioning strategies, in order to minimize total cost and maximize the proportion of time-bound promised orders;

[0016] Step S4: Based on the live broadcast, perform corresponding inventory pre-positioning and dynamic warehouse allocation to trigger transfers;

[0017] Step S5: Obtain order volume data and order distribution for order fulfillment, and compare them with the prediction data of the fusion model to optimize the fusion model and the linear programming model in order to update the dynamic warehousing and inventory pre-positioning strategy.

[0018] Furthermore, in step S3, the total cost is constructed by allocating costs, warehousing costs, and delivery costs. The linear programming model is constructed by minimizing the total cost and maximizing the proportion of orders delivered within the promised delivery time using a bidirectional objective function. The weighted summation method is used to optimize the total cost and the order proportion in a dual-objective manner, which is used to generate dynamic warehousing and inventory pre-positioning strategies.

[0019] Furthermore, the formula for minimizing the total cost is as follows:

[0020]

[0021]

[0022]

[0023]

[0024] in, Represents the total cost. Indicates allocation cost, Indicates warehousing costs, This represents the delivery cost, M represents the number of minimum inventory units, N represents the number of warehouses, and K represents the number of regions. This represents the inventory quantity of the smallest inventory unit i in the corresponding area k pre-defined in warehouse j. Indicates central warehouse The cost of transferring goods between warehouse j and warehouse j This represents the storage cost of warehouse j. Indicates the pre-set inventory cycle. This represents the delivery cost from warehouse j to region k;

[0025] The formula for maximizing the proportion of orders with time-definite delivery commitments is as follows:

[0026]

[0027] in, This indicates that same-day or next-day delivery is possible from warehouse j to region k; otherwise... ;

[0028] A two-way objective function is constructed, and a weighted summation method is used to fuse the two objectives, as shown in the following formula:

[0029]

[0030] Where α represents the cost weight. This indicates the upper limit of the cost.

[0031] Furthermore, constraints are constructed, including inventory balance constraints, warehouse capacity constraints, and non-negativity constraints;

[0032] Inventory balance constraints: , , This indicates the order demand in each region;

[0033] Warehouse capacity constraints: , , Indicates inventory volume. Indicates warehouse capacity;

[0034] Nonnegativity constraint: , .

[0035] Furthermore, in step S5, the cost weights of the linear programming model are adjusted based on the actual delivery cost ratio and the expected ratio.

[0036] Furthermore, in step S2, static features related to live streaming and products and time-series features generated in historical live streams are obtained from multi-dimensional data. Nonlinear static patterns are extracted from the static features and time-series features to generate a first feature vector. Long-term dependencies are captured from the time-series features to generate a second feature vector. The first feature vector and the second feature vector are fused to obtain a fused feature vector, which is used to predict the category of order volume data and order distribution. A loss function is constructed based on the difference between the true value and the predicted value to train the fusion model.

[0037] Further, in step S1, feature engineering is used to construct derived features, including interaction conversion coefficient, traffic growth rate, and regional preference coefficient. Mutual information is used to filter features that are strongly correlated with the order quantity and order distribution. In step S2, the order volume data includes total order volume and real-time order curve, and the order distribution is the order geographical distribution. Based on the fused feature vector, the total order volume of the live broadcast and the real-time order curve corresponding to the live broadcast time slice are predicted respectively. The order geographical distribution is generated by using the predicted total order volume and the predicted regional preference coefficient in the live broadcast.

[0038] Furthermore, during the live broadcast in step S4, the actual order volume and the actual geographical distribution of the orders are collected based on the live broadcast time slice. Based on the deviation between the actual and predicted values, a transfer calculation is triggered, prioritizing the transfer of inventory from the warehouse closest to the target area, with sufficient stock and the lowest transfer cost. The transfer quantity is obtained based on the following formula:

[0039]

[0040] in, This represents the inventory quantity of the adjusted smallest inventory unit i in the corresponding area k pre-set in warehouse j. This represents the actual geographical distribution of region k within the live broadcast time slice t. This represents the occupied inventory of the smallest inventory unit i in the corresponding area k pre-set in warehouse j.

[0041] Furthermore, in step S5, the model parameters are updated using online gradient descent:

[0042]

[0043] in, , Indicates the number of iterations. It is a very small number. This represents the model gradient based on the actual order volume and actual geographical distribution.

[0044] The dynamic warehousing and inventory pre-positioning prediction device based on live-stream e-commerce traffic includes a multi-dimensional data acquisition module, a multi-dimensional data feature fusion module, a dynamic warehousing and inventory pre-positioning generation module, an allocation module, and an update module. It employs the aforementioned dynamic warehousing and inventory pre-positioning prediction method based on live-stream e-commerce traffic to extract and fuse features from the collected multi-dimensional data to predict order volume and order distribution. A dynamic warehousing and inventory pre-positioning strategy is generated through linear programming, adjusted accordingly during the live stream, and updated after the live stream based on the feature fusion and linear programming approach.

[0045] The advantages and beneficial effects of this invention are as follows:

[0046] This invention employs multi-dimensional data fusion and a "GBDT-LSTM-MLP" fusion model to capture the nonlinear and temporal characteristics of live streaming scenarios. The prediction error rate is reduced from over 35% in existing technologies to below 15%, effectively reducing insufficient inventory and backlog, and significantly improving prediction accuracy. The dynamic warehouse allocation strategy of this invention is optimized in real-time based on the prediction results of a single live stream, adapting to the dynamic changes in the geographical distribution of orders. Warehouse resource utilization is increased by over 30%, avoiding the resource mismatch problem of static warehouse allocation and greatly improving warehouse allocation flexibility. The dynamic adjustment mechanism of this invention enables emergency allocation within 10 minutes during live streaming. The proportion of "same-day delivery" and "next-day delivery" orders has increased from below 60% in the past to over 85%, customer satisfaction has improved by 20%, and supply chain response speed has been accelerated. This invention reduces the overall logistics and warehousing costs by 15%-25% and increases inventory turnover by over 30% through localized delivery and optimized allocation strategies, while reducing capital occupation costs by 20%, significantly lowering overall costs. The closed-loop feedback learning system of this invention uses actual data to continuously iterate the model, and the prediction accuracy and warehouse optimization effect gradually improve with the number of uses, forming a virtuous cycle of "data-prediction-decision-feedback", which continuously optimizes decision-making capabilities. Attached Figure Description

[0047] Figure 1 This is a flowchart of the method in an embodiment of the present invention.

[0048] Figure 2 This is an architecture diagram of the fusion prediction model in an embodiment of the present invention.

[0049] Figure 3 This is a logic diagram for solving the linear programming model of the dynamic warehouse allocation strategy in this embodiment of the invention.

[0050] Figure 4 This is a flowchart illustrating the dynamic adjustment of inventory during the live streaming process according to an embodiment of the present invention.

[0051] Figure 5 This is a schematic diagram illustrating the working principle of the closed-loop feedback learning system in this embodiment of the invention.

[0052] Figure 6 This is a schematic diagram of the device in an embodiment of the present invention. Detailed Implementation

[0053] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0054] like Figure 1 As shown, this invention provides a method for dynamic warehouse allocation and inventory pre-positioning prediction based on live e-commerce traffic, including the following steps:

[0055] Step S1: Collect multi-dimensional live e-commerce related data and preprocess it to obtain a standardized feature set.

[0056] In this embodiment, data collection covers the following five core data categories:

[0057] Live streaming plan data: collected from the backend of the live streaming platform (such as Douyin e-commerce backend, Taobao Live control panel), including live streaming duration T=6 hours, number of products M=20, product category C=3 categories (beauty, apparel, home furnishing), promotion budget B=50,000 yuan, start time T0=2025 month X day 20:00 (accurate to the minute), and anchor level L=4 levels (levels 1-5, level 4 corresponds to top anchors);

[0058] Historical live streaming data: Time-series data for similar live streamers (level 4 and above) and similar products (beauty, apparel, home furnishings) over the past 180 days were extracted from the platform's historical database, with a time granularity of 5 minutes. This includes the number of viewers P(t) per 5 minutes, interaction rate I(t) = (number of likes + number of comments + number of shares) / number of viewers, conversion rate C(t) = number of orders / number of viewers, and total order volume Y. hist Order geographical distribution D hist (k) (k is the code for 34 provincial-level administrative regions);

[0059] Product characteristic data: collected from the OMS system, including SKU weight W (cosmetics 0.1-0.5kg, apparel 0.3-1.2kg, home furnishings 1.5-5kg), volume V (m³), unit price P (cosmetics 50-300 yuan, apparel 99-599 yuan, home furnishings 299-1999 yuan), best-selling attribute S (1 indicates best-selling, 0 indicates not best-selling, in this example, S=1 for 8 SKUs), and inventory turnover days Z (cosmetics 15-30 days, apparel 30-60 days, home furnishings 60-90 days).

[0060] External environmental data: Wea(k) of the weather in each region on the day of the live broadcast was obtained from a third-party weather platform (0=sunny, 1=cloudy, 2=rainy, 3=snow), holiday identifier Hol (1=holiday, 0=non-holiday) was obtained from the holiday database, and regional consumption index Con(k) (0-100, Con(k)≥60 for the eastern coastal region, and 30≤Con(k)<60 for the central and western regions) was obtained from the regional consumption index platform.

[0061] Warehouse network data: collected from WMS and TMS systems, including the number of warehouses N=5 (East China, North China, South China, Southwest China, and Northwest China warehouses), and the capacity Cap(j) of each warehouse (unit: m). 3 East China warehouse Cap=5000, other regional warehouses Cap=3000), warehouse coordinates (Lat(j), Lon(j)), and inter-warehouse transfer cost (Cost). trans (j1,j2) (Unit: Yuan / piece; intra-regional transfer cost: 0.5-1 Yuan / piece; inter-regional transfer cost: 2-5 Yuan / piece), Cost of delivery from warehouse to various regions. del (j,k) (unit: yuan / piece; local delivery cost: 3-8 yuan / piece; long-distance delivery cost: 8-15 yuan / piece), warehouse storage cost. stor (j) (Unit: RMB / piece / day, East China warehouse: RMB 0.8 / piece / day, other regional warehouses: RMB 0.5 / piece / day).

[0062] Data preprocessing includes the following steps:

[0063] 1) Data Cleaning: Outliers are removed using the Raida criterion. For time-series data such as P(t), I(t), and C(t), the mean μ and standard deviation σ of each indicator are calculated. If a data point x satisfies |x-μ|>3σ, it is identified as an outlier and replaced with the mean of the five data points before and after that moment. For example, if the number of viewers P(t) at a certain moment is 10000, and the mean of the five data points before and after is 3000, μ=2800, σ=800, then |10000-2800|=7200>3×800=2400, which is identified as an outlier and replaced with 3000.

[0064] 2) Data denoising: A moving average filter is applied to the time-series data. The formula is as follows:

[0065]

[0066] Where n=2 (window size is 5), x(t) is the original data, x smooth (t) represents the denoised data. For example, the original view count sequence is [2800, 3000, 3200, 2900, 3100], and the denoised data is x. smooth (t)=(2800+3000+3200+2900+3100) / 5=3000

[0067] 3) Data normalization: For numerical data, min-max normalization is used, with the following formula:

[0068]

[0069] Where, x min x max Let x be the minimum and maximum values ​​of the data, respectively. After normalization, the data range is [0,1]. For example, if the promotion budget B ranges from 10,000 to 100,000 yuan, and a certain live stream B = 50,000 yuan, then after normalization, x... norm =(50000-10000) / (100000-10000)=40000 / 90000≈0.444;

[0070] 4) Feature engineering:

[0071] Feature construction: Construct the interaction conversion coefficient IC=I(t)×C(t) and the traffic growth rate P. rate =(P(t)-P(t-1)) / P(t-1), Regional preference coefficient Pref(k)=D hist (k) / ∑ k =1 34 D hist (k) and other derived features. For example, at a certain moment, I(t)=0.3, C(t)=0.1, then IC=0.03; P(t)=3000, P(t-1)=2500, then P rate =(3000-2500) / 2500=0.2; A certain region D hist (k)=1000, ∑D hist If (k) = 10000, then Pref(k) = 0.1;

[0072] Feature selection: Mutual information (MI) is used to select features that are strongly correlated with the target variables (total order volume, geographical distribution). The mutual information formula is:

[0073]

[0074] Features with an MI > 0.3 are retained for the modeling stage. In this embodiment, 28 effective features are ultimately selected, including product weight W, volume V, unit price P, best-selling attribute S, promotion budget B, anchor level L, regional consumption index Con(k), holiday identifier Hol, number of viewers P(t), interaction rate I(t), conversion rate C(t), and traffic growth rate P. rate Interaction conversion coefficient (IC), regional preference coefficient (Pref(k)), etc.

[0075] Step S2: Live streaming traffic and order prediction based on the "GBDT-LSTM-MLP" fusion model; input the standardized feature set into the fusion prediction model, and output the total order volume of the target live stream, the real-time order curve, and the order geographical distribution prediction results.

[0076] Model architecture such as Figure 2 As shown, the fusion model adopts a three-level architecture of "GBDT-LSTM-MLP", which takes into account both the nonlinear mapping of static features and the long-term dependency capture of temporal features. The functions and connections of each module are as follows:

[0077] GBDT module: Receives static features and time-series features, extracts the nonlinear static patterns, and outputs the first feature vector. ;

[0078] LSTM module: Receives temporal features, captures long-term dependencies, and outputs a second feature vector. ;

[0079] MLP module: splicing and Obtain the fused feature vector (Total 164 dimensions), after processing by the hidden layer, three types of prediction results are output.

[0080] Model input:

[0081] Static features This includes product weight (W), volume (V), unit price (P), best-selling attributes (S), promotion budget (B), livestreamer level (L), and regional consumer index. Holiday signage (Hol), 8 dimensions in total;

[0082] Temporal characteristics This includes historical live stream metrics such as viewership per 5 minutes (P(t), interaction rate (I(t), conversion rate (C(t)), and traffic growth rate). Time step =5 minutes, sequence length =288 (corresponding to 24-hour time series data), totaling 4×288=1152 dimensions.

[0083] Model training parameters:

[0084] GBDT module: Set the number of decision trees K=100, the tree depth d=6, the learning rate η=0.1, and use the gradient boosting algorithm to iteratively build decision trees. Each tree is trained based on the residual of the previous tree, and finally outputs a 100-dimensional feature vector.

[0085] LSTM module: Set the hidden layer dimension h=64, the number of layers l=2, and the dropout rate p=0.2 (to prevent overfitting). After inputting the temporal feature sequence, it captures long-term dependencies through a gating mechanism (input gate, forget gate, output gate) and outputs a 64-dimensional feature vector.

[0086] MLP module: The hidden layer is set to 2 layers, with the first layer having a dimension of 128 and the second layer having a dimension of 64. The activation function is ReLU (Rectified Linear Unit). The output layer uses linear activation, corresponding to the three prediction targets: total order volume, real-time order curve, and order geographical distribution.

[0087] The loss function uses mean squared error (MSE) combined with L1 regularization, and the formula is as follows:

[0088]

[0089] Where λ = 0.001 is the regularization coefficient, and θ is the model parameter. For the true value, The loss function, which is the predicted value, can balance prediction accuracy and model complexity and avoid overfitting. The optimizer uses the Adam optimizer with a learning rate of η=0.001, a decay rate of β1=0.9, and β2=0.999. The model converges quickly through adaptive learning rate adjustment.

[0090] Historical data from the past 180 days was used as the training set (80%), validation set (10%), and test set (10%). The training process was as follows: the preprocessed feature data was divided chronologically, with the first 144 days as the training set, the middle 18 days as the validation set, and the last 18 days as the test set. The number of iterations was set to 200 rounds. After each round of training, the loss value of the validation set was calculated. If the loss value of the validation set did not decrease for 10 consecutive rounds, training was stopped (early stopping mechanism). The prediction error rate on the test set was 12.8%, which meets the requirement of "prediction error rate controlled within 15%" in the claims.

[0091] Total order volume forecast:

[0092]

[0093] In this embodiment, the predicted total order quantity is 12,000 units;

[0094] Real-time order curve prediction:

[0095]

[0096] Where t is a 5-minute time slice within the live broadcast period, there are a total of 6×12=72 time slices, and the predicted order volume for each time slice is output. For example, the predicted order volume for the first time slice (0-5 minutes) after the broadcast starts is 500 items, and the predicted order volume for the 12th time slice (55-60 minutes) is 1800 items (peak period).

[0097] Order geographic distribution prediction:

[0098]

[0099] in, For the predicted regional preference coefficient, in this embodiment, the East China region... =0.35, predicted order quantity 12000 × 0.35 = 4200 pieces, South China region =0.25, predicted order volume 3000 pieces.

[0100] Step S3: Calculate the dynamic warehouse allocation strategy based on linear programming; based on the prediction results, warehouse network data and product physical attributes, generate the pre-set inventory quantity of each SKU in different warehouses through a bi-objective optimization algorithm, and complete the pre-set inventory allocation before the live broadcast.

[0101] like Figure 3 As shown, a linear programming model is constructed with the dual objectives of "minimizing total cost + maximizing the proportion of orders with promised delivery times" as the guiding principle. An optimization model is then built based on the constraints.

[0102] Decision variables: This indicates the inventory quantity of SKU i in the corresponding area k pre-set in warehouse j (i=1,2,…,20; j=1,2,…,5; k=1,2,…,34).

[0103] Objective function: Minimize total cost:

[0104]

[0105]

[0106]

[0107]

[0108] in, Indicates allocation cost, This is the central warehouse (in this example, the East China warehouse). Indicates warehousing costs, =3 days (pre-set inventory cycle). This indicates delivery costs and maximizes the proportion of orders with time-definite delivery commitments.

[0109]

[0110] in, This indicates that same-day / next-day delivery is possible from warehouse j to region k; otherwise... ;

[0111] A weighted summation method is used for dual-objective fusion:

[0112]

[0113] Where α = 0.6 is the cost weight. Given a cost ceiling (1 million yuan in this example), the dual-objective optimization is achieved by minimizing J;

[0114] Constraints:

[0115] Inventory balance constraints: , (Order demands from each region must be met).

[0116] Warehouse capacity constraints: , (Inventory volume shall not exceed warehouse capacity).

[0117] Nonnegativity constraint: , (Inventory quantity is non-negative).

[0118] The linear programming model is solved using the interior-point method, and a fast solution is achieved through the Gurobi optimizer (decision delay ≤ 30 seconds). The output is a matrix of pre-set inventory quantities for each SKU in different warehouses. And a cost estimation report. In this embodiment, the pre-positioned inventory in the East China warehouse accounts for 35% (4200 pieces), the South China warehouse accounts for 25% (3000 pieces), the North China warehouse accounts for 20% (2400 pieces), the Southwest warehouse accounts for 12% (1440 pieces), and the Northwest warehouse accounts for 8% (960 pieces). The total cost is estimated at 680,000 yuan, and the predicted proportion of time-committed orders is 88%.

[0119] Step S4: Inventory pre-positioning and dynamic adjustment; During the live broadcast, monitor the deviation between real-time order data and forecast results. When the deviation exceeds the preset threshold, initiate emergency allocation calculation and update the warehouse allocation strategy.

[0120] Before the live stream, inventory is pre-positioned. Based on the warehousing plan X, transfer instructions are issued through the TMS system. Inventory is transferred from the central warehouse (East China warehouse) to regional warehouses and put on shelves 24 hours before the live stream starts. The WMS system updates inventory data synchronously. For example, if 500 units of beauty SKU 1 need to be transferred to the South China warehouse, the TMS system plans the transfer route as "East China warehouse → South China warehouse," selecting road transportation (transfer cost 2 yuan / unit), with a transfer time of 12 hours, ensuring that the stock is completed before the live stream starts. The WMS system records the inventory of this SKU in the South China warehouse as 500 units.

[0121] like Figure 4 As shown, actual order data is collected every 5 minutes during the live stream. Actual geographical distribution Calculate the deviation from the predicted value:

[0122] Order quantity deviation: In this embodiment, the threshold .

[0123] Geographical distribution deviation: In this embodiment, the threshold .

[0124] In this embodiment, the actual order volume is recorded during the 12th time slot (55-60 minutes) after the broadcast begins. =2200 pieces, predicted value Y(t)=1800 pieces, ε γ (t) = (2200 - 1800) / 1800 ≈ 0.222 > 0.2, triggering dynamic adjustment; actual order volume in South China region. =800 items, predicted value D( ) = 600 pieces, South China, t) = (800-600) / 600 ≈ 0.333 > 0.3, further confirming the trigger adjustment.

[0125] Emergency allocation calculation and strategy update: A greedy algorithm is used to execute emergency allocations, prioritizing the allocation of inventory from warehouse j', which is closest to the target area k, has sufficient inventory, and has the lowest allocation cost. Allocation quantity:

[0126]

[0127] in, This is inventory that has already been used.

[0128] In this embodiment, the shortage in South China is 200 units. The nearest warehouse with sufficient inventory is the East China warehouse (1000 units remaining). The transfer cost is 2 yuan per unit. Therefore, 200 units will be transferred from the East China warehouse to the South China warehouse, with a transfer time of 2 hours (using expedited air transport). Adjusted inventory matrix. (ΔX is the allocation quantity matrix). The warehouse management system and the transportation system are updated in real time. The inventory of this SKU in the South China warehouse is updated to 700 units, and the inventory in the East China warehouse is updated to 800 units.

[0129] Step S5: Order Fulfillment and Feedback Learning; After the live broadcast ends, orders are allocated and fulfilled according to the real-time updated warehouse allocation strategy. Actual data from the entire chain is collected and fed back to the data pool. The parameters of the integrated medical model and the dual-objective optimization algorithm are iteratively optimized.

[0130] like Figure 5 As shown, the specific implementation process is as follows:

[0131] Order fulfillment: After receiving the order, the OMS system calculates the inventory based on the real-time inventory matrix. The fulfillment warehouse is allocated according to the principle of "nearest location + sufficient inventory". For example, SKU 1 ordered by a user in the South China region is allocated to the South China warehouse for fulfillment, a picking list is generated and sent to the WMS system to complete the picking, packing and shipping process, with a delivery cost of 5 yuan / piece and a fulfillment time of "next day delivery".

[0132] End-to-end data collection: gathering data from the entire order fulfillment process, including the actual total number of orders. =12,800 items, real-time order curve Actual geographical distribution (East China 32%, South China 28%, North China 19%, Southwest China 13%, Northwest China 8%), Allocation Costs =120,000 yuan, warehousing costs =80,000 yuan, delivery costs =450,000 yuan, performance period (The average delivery time is 1.2 days in East China, South China, and North China, and 2.5 days in Southwest and Northwest China).

[0133] Model iterative optimization, total order volume prediction error =12800-12000=800 items, regional distribution prediction error =12800×28%-3000=3584-3000=584 pieces; The model parameters are updated using the online gradient descent method:

[0134]

[0135] in, ( For the number of iterations, In this embodiment, t=100. =0.001 / 10=0.0001; The actual delivery cost accounts for 45 / (12+8+45)=45 / 65≈69.2%, which exceeds the expected proportion (60%). Therefore, the weight α of the linear programming model is adjusted to 0.5 to increase the weight of timeliness.

[0136] like Figure 6 As shown, the present invention also provides a dynamic warehousing and inventory pre-positioning prediction device based on live e-commerce traffic, including a multi-dimensional data collection and preprocessing module, a live order multi-dimensional fusion prediction module, a pre-live SKU warehousing pre-positioning dual-objective optimization module, a live order deviation monitoring and emergency allocation optimization module, and a post-live order fulfillment and model algorithm iteration optimization module.

[0137] A multi-dimensional data acquisition and preprocessing module is used to collect and preprocess multi-dimensional live-stream e-commerce related data to obtain a standardized feature set. In this embodiment of the invention, the multi-dimensional data acquisition and preprocessing module for live-stream e-commerce is specifically used to collect multi-source data from live-streaming platforms (Douyin, Taobao Live, etc.), OMS, WMS, TMS, and third-party weather and consumer index platforms. Through steps such as removing outliers using the Laida criterion, moving average filtering for noise reduction, min-max normalization, and mutual information feature filtering, the data preprocessing is completed, and a standardized static feature set and time-series feature set are output.

[0138] The live streaming order multi-dimensional fusion prediction module is used to input the standardized feature set into the fusion prediction model and output the total order volume, real-time order curve, and order regional distribution prediction results for the target live stream. In this embodiment of the invention, the live streaming order multi-dimensional fusion prediction module is specifically used to load the trained "GBDT-LSTM-MLP" fusion model, receive the preprocessed feature data, extract nonlinear static laws through the GBDT module, capture long-term temporal dependencies through the LSTM module, and realize feature fusion and prediction output through the MLP module to obtain the total order volume, real-time order curve, and order regional distribution prediction results and store them in the MySQL database.

[0139] The pre-livestream SKU warehouse pre-setting dual-objective optimization module is used to generate a pre-set inventory quantity matrix for each SKU in different warehouses based on the prediction results, warehouse network data, and product physical attributes, through a dual-objective optimization algorithm, thus completing the pre-livestream inventory warehouse pre-setting. In this embodiment of the invention, the pre-livestream SKU warehouse pre-setting dual-objective optimization module is specifically used to construct a dual-objective function of "minimizing total cost + maximizing the proportion of time-committed orders", setting inventory balance, warehouse capacity, and non-negativity constraints, solving the linear programming model using the interior-point method, and outputting the pre-set inventory quantity matrix and cost estimation report for each SKU in each region of multiple warehouses, which is then sent to the inventory pre-setting and scheduling execution module through a message queue.

[0140] The live-stream order deviation monitoring and emergency allocation optimization module is used to monitor the deviation between actual order data and the predicted results in real time during the live stream. When the deviation exceeds a preset threshold, emergency allocation calculation is initiated and the warehouse allocation strategy is updated. In this embodiment of the invention, the live-stream order deviation monitoring and emergency allocation optimization module is specifically used to collect actual order data every 5 minutes, calculate the order volume deviation (threshold 0.2) and the regional distribution deviation (threshold 0.3). When the deviation exceeds the threshold, a greedy algorithm is used to allocate inventory from the warehouse closest to the target area, with sufficient inventory and the lowest allocation cost, update the inventory matrix and synchronize it to the WMS and TMS systems.

[0141] The post-livestream order fulfillment and model algorithm iteration optimization module is used to allocate and fulfill orders according to the real-time updated warehouse allocation strategy after the livestream ends, collect actual data from the entire chain and feed it back to the data pool, and iteratively optimize the parameters of the fusion prediction model and the dual-objective optimization algorithm. In this embodiment of the invention, the post-livestream order fulfillment and model algorithm iteration optimization module is specifically used to connect to the OMS system to allocate fulfillment warehouses according to the principle of "nearest + sufficient inventory", collect actual order data, cost data and fulfillment timeliness data, calculate prediction error, update model parameters through online gradient descent method, and adjust the weight α of the linear programming model according to the actual cost data to form a closed-loop iterative optimization.

[0142] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A dynamic warehouse allocation and inventory pre-positioning method based on live-stream e-commerce traffic prediction, characterized in that... Includes the following steps: Step S1: Collect multi-dimensional data related to the live stream and products; Step S2: Construct a fusion model to obtain live streaming and product-related features from multi-dimensional data in order to predict order volume data and order distribution; Step S3: Based on the inventory in the pre-positioned warehouse area, construct a linear programming model to generate dynamic warehouse allocation and inventory pre-positioning strategies, in order to minimize total cost and maximize the proportion of time-bound promised orders; Step S4: Based on the live broadcast, perform corresponding inventory pre-positioning and dynamic warehouse allocation to trigger transfers; Step S5: Obtain order volume data and order distribution for order fulfillment, and compare them with the prediction data of the fusion model to optimize the fusion model and the linear programming model in order to update the dynamic warehousing and inventory pre-positioning strategy.

2. The dynamic warehousing and inventory pre-positioning method based on live-stream e-commerce traffic prediction according to claim 1, characterized in that: In step S3, the total cost is constructed by allocating costs, warehousing costs, and delivery costs. The linear programming model is constructed by minimizing the total cost and maximizing the proportion of orders delivered within the promised delivery time using a two-way objective function. The weighted summation method is used to optimize the total cost and the proportion of orders in a dual-objective manner, which is used to generate dynamic warehousing and inventory pre-positioning strategies.

3. The dynamic warehousing and inventory pre-positioning method based on live-stream e-commerce traffic prediction according to claim 2, characterized in that: The formula for minimizing the total cost is as follows: in, Represents the total cost. Indicates allocation cost, Indicates warehousing costs, This represents the delivery cost, M represents the number of minimum inventory units, N represents the number of warehouses, and K represents the number of regions. This represents the inventory quantity of the smallest inventory unit i in the corresponding area k pre-defined in warehouse j. Indicates central warehouse The cost of transferring goods between warehouse j and warehouse j This represents the storage cost of warehouse j. Indicates the pre-set inventory cycle. This represents the delivery cost from warehouse j to region k; The formula for maximizing the proportion of orders with time-definite delivery commitments is as follows: in, This indicates that same-day or next-day delivery is possible from warehouse j to region k; otherwise... ; A two-way objective function is constructed, and a weighted summation method is used to fuse the two objectives, as shown in the following formula: Where α represents the cost weight. This indicates the upper limit of the cost.

4. The dynamic warehousing and inventory pre-positioning method based on live-stream e-commerce traffic prediction according to claim 3, characterized in that: Construct constraints, including inventory balance constraints, warehouse capacity constraints, and non-negativity constraints; Inventory balance constraints: , , This indicates the order demand in each region; Warehouse capacity constraints: , , Indicates inventory volume. Indicates warehouse capacity; Nonnegativity constraint: , .

5. The dynamic warehousing and inventory pre-positioning method based on live-stream e-commerce traffic prediction according to claim 3, characterized in that: In step S5, the cost weights of the linear programming model are adjusted based on the actual delivery cost ratio and the expected ratio.

6. The dynamic warehousing and inventory pre-positioning method based on live-stream e-commerce traffic prediction according to claim 1, characterized in that: In step S2, static features related to live streaming and products and time-series features generated in historical live streams are obtained from multi-dimensional data. Nonlinear static patterns are extracted from the static features and time-series features to generate a first feature vector. Long-term dependencies are captured from the time-series features to generate a second feature vector. The first feature vector and the second feature vector are fused to obtain a fused feature vector, which is used to predict the category of order volume data and order distribution. A loss function is constructed based on the difference between the true value and the predicted value to train the fusion model.

7. The dynamic warehousing and inventory pre-positioning method based on live-stream e-commerce traffic prediction according to claim 6, characterized in that: In step S1, feature construction is performed through feature engineering to build derived features, including regional preference coefficients. In step S2, the order volume data includes total order volume and real-time order curves, and the order distribution is the order geographical distribution. Based on the fused feature vector, the total order volume of the live broadcast and the real-time order curves corresponding to the live broadcast time slices are predicted respectively. The order geographical distribution is generated by using the predicted total order volume and the predicted regional preference coefficients in the live broadcast.

8. The dynamic warehousing and inventory pre-positioning method based on live-stream e-commerce traffic prediction according to claim 7, characterized in that: During the live broadcast in step S4, the actual order volume and the actual geographical distribution of the orders are collected based on the live broadcast time slice. Based on the deviation between the actual and predicted values, a transfer calculation is triggered. Inventory is prioritized for transfer from the warehouse closest to the target area, with sufficient stock and the lowest transfer cost. The transfer quantity is obtained based on the following formula: in, This represents the inventory quantity of the adjusted smallest inventory unit i in the corresponding area k pre-set in warehouse j. This represents the actual geographical distribution of region k within the live broadcast time slice t. This represents the occupied inventory of the smallest inventory unit i in the corresponding area k pre-set in warehouse j.

9. The dynamic warehousing and inventory pre-positioning method based on live-stream e-commerce traffic prediction according to claim 8, characterized in that: In step S5, the model parameters are updated using the online gradient descent method: in, , Indicates the number of iterations. It is a very small number. This represents the model gradient based on the actual order volume and actual geographical distribution.

10. A dynamic warehousing and inventory pre-positioning prediction device based on live-stream e-commerce traffic, comprising a multi-dimensional data acquisition module, a multi-dimensional data feature fusion module, a dynamic warehousing and inventory pre-positioning generation module, an allocation module, and an update module, characterized in that: The method for dynamic warehouse allocation and inventory pre-positioning prediction based on live e-commerce traffic, as described in any one of claims 1 to 9, extracts and fuses features from the collected multidimensional data to predict order volume data and order distribution. It generates dynamic warehouse allocation and inventory pre-positioning strategies through linear programming, makes corresponding adjustments during the live broadcast, and updates the feature fusion and linear programming methods after the live broadcast.

Images