Mineral water sales inventory management method based on deep learning

Abstract

The invention relates to the technical field of sales management, and discloses a mineral water sales inventory management method based on deep learning, and the method comprises the steps: 1, collecting a data source affecting the sales volume of mineral water, carrying out the format standardization, time alignment and feature extraction of the data source, and carrying out the feature extraction of the data source; generating a group of multi-dimensional input feature vectors for describing external sales driving factors; and 2, inputting the multi-dimensional input feature vector and historical sales volume data into a deep learning model, training to obtain a mineral water future sales volume predicted value, and carrying out inventory state analysis in combination with current inventory data, shelf life parameters and replenishment cycle parameters. According to the method, the technical scheme of feature fusion of fusing external sales driving factors and multi-source structured behavior data and unified input tensor construction is adopted, and the technical effects of dynamic modeling and real-time response of external factors such as weather, holidays and festivals, regional activities and sales promotion states in sales prediction are achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of sales management, and specifically to a mineral water sales inventory management method based on deep learning. Background Art

[0002] As a high-frequency, fast-moving commodity, mineral water is widely used in various sales channels, including retail outlets, supermarkets, and community delivery. It exhibits strong demand elasticity, significant sales volatility, and rapid inventory turnover. In actual operations, mineral water sales are often significantly affected by external factors such as weather changes, holiday promotions, and regional events. Sales fluctuations exhibit significant temporal nonlinearity and sudden disturbances.

[0003] For example, Chinese invention patent application publication number CN106447269A discloses a drug sales inventory management system, comprising an inventory management unit and a data storage unit; the inventory management unit includes a loss reporting management module, an inventory management module, a price adjustment management module, a price adjustment query module, an inventory query module, and a loss reporting query module. The present invention uses the data storage unit to store all data and information related to drug warehousing, delivery, sales, returns, and loss reporting. The inventory management unit manages drug loss reporting operations and inventory status, effectively managing all aspects of drug management and significantly improving management efficiency and accuracy.

[0004] For example, the Chinese invention application with publication number CN113159675B discloses a sales-based inventory management method and device. This application realizes the ordering and payment operations of goods on the client side through the interaction between the client side, the sales side and the warehouse side, so that the sales side and the warehouse side can pre-occupy and ship-out the goods based on the client side's ordering and payment operations; and realizes the pre-occupation operation of the goods ordered by the user and the ship-out operation of the goods paid by the user on the sales side, so that the strong consistency problem of the sales inventory quantity and the commodity warehouse quantity is solved by deducting the commodity image code and the quantity of the goods to be shipped out.

[0005] The shortcomings of the above patents are:

[0006] Existing mineral water sales forecasting methods generally rely on historical sales data to build deep learning models, failing to effectively identify and incorporate the impact of external factors such as weather, holidays, and regional events. In real-world retail scenarios, mineral water sales are often strongly impacted by external drivers, leading to short-term fluctuations. Traditional models, lacking the ability to model such unstructured, multi-source data, are unable to respond promptly to sudden sales fluctuations. This, in turn, results in delayed inventory warnings, delays in timely replenishment, and the potential for localized stockouts and backlogs.

[0007] Most existing deep learning inventory forecasting models infer inventory consumption trends based on sales data, failing to account for the complex behavioral factors inherent in actual inventory operations, such as returns, losses, transfers, and unrecorded shipments. These behavioral factors can lead to deviations between actual inventory and model predictions. However, existing models lack a mechanism for automatic feedback correction of these deviations, preventing timely correction of forecast errors. This can lead to systematic biases over long periods of operation, compromising the accuracy and effectiveness of inventory strategies.

[0008] To this end, the present invention proposes a mineral water sales inventory management method based on deep learning to solve the above-mentioned problems. Summary of the Invention

[0009] In view of the shortcomings of the existing technology, the present invention provides a mineral water sales inventory management method based on deep learning to solve the problems raised in the above background technology.

[0010] To achieve the above objectives, the present invention is implemented through the following technical solutions: A mineral water sales inventory management method based on deep learning, comprising:

[0011] Step 1: Collect data sources that influence mineral water sales, standardize the data sources, align their time, and extract features to generate a set of multi-dimensional input feature vectors that describe external sales drivers.

[0012] Step 2: The multi-dimensional input feature vector and historical sales data are fed into a deep learning model to train and obtain future sales forecasts for mineral water. The model then analyzes inventory status by combining current inventory data, shelf life parameters, and replenishment cycle parameters.

[0013] Step 3: Generate an inventory scheduling strategy based on the sales forecast and inventory status analysis results. The inventory scheduling strategy includes replenishment timing, replenishment quantity, and inventory risk warning notifications.

[0014] Step 4: After the inventory scheduling plan is output, inventory behavior data is collected synchronously during the actual inventory management process, and the inventory behavior data is structurally transformed to form an actual inventory change data stream that matches the sales forecast;

[0015] Step 5: Using the actual inventory change data stream, compare the sales forecast results with the inventory scheduling plan. Based on the comparison results, a feedback mechanism is established. The feedback mechanism constructs model parameter update rules based on the forecast deviation data, and dynamically corrects the parameters of the intermediate layers of the deep learning model by defining error reference values ​​for weight adjustment.

[0016] In step 6, the deep learning model with modified parameters receives the multi-source input data in step 1 and the inventory behavior data in step 4 again, and performs a new round of sales forecasting and inventory scheduling processing, forming an iterative optimization process that combines forecasting and feedback, and realizing continuous closed-loop management of mineral water sales and inventory scheduling.

[0017] Preferably, the step 1 further comprises:

[0018] Sub-step 1.1: Collect four types of external sales-driving data, including weather data, holiday and event schedules, regional popularity trends, and sales of neighboring stores, and perform preliminary cleaning and format standardization.

[0019] Weather data includes temperature, relative humidity, rainfall probability, and extreme weather warning levels;

[0020] Holiday and event arrangements are represented by a binary variable EventFlag. If there is an exhibition or holiday event on that day, EventFlag = 1, otherwise EventFlag = 0.

[0021] The regional popularity trend is generated by combining the frequency index of hot words on social platforms and the LBS check-in density to generate the regional popularity score R:

[0022]

[0023] Among them, μ S and σ S is the mean and standard deviation of the hot word frequency, μ L and σ L is the mean and standard deviation of the check-in density, α and β are weight coefficients,

[0024] S is the frequency of regional social media hot words, and L is the user check-in density;

[0025] Sub-step 1.2: Perform outlier detection on the above-mentioned standardized data, using the triple standard deviation method:

[0026] is an outlier,

[0027] Among them, x t is the observed value at time t, μ and σ are the historical mean and standard deviation of the corresponding variables, and σ is the historical standard deviation;

[0028] For missing segments, a sliding window mean filling method based on local trends is used:

[0029]

[0030] Where k is the window length, is the estimated value of the variable used to complete the time point t, xt-i is the value of the variable at time point t forward i;

[0031] Sub-step 1.3: Combine the cleaned and completed feature variables to construct the feature vector of external sales drivers:

[0032] X t =[T t ,H t ,P t ,W t ,EventFlag t ,R t ,S t ,L t ,N t ],

[0033] Among them, X t is the external sales feature vector at time t, T t 、H t 、P t 、W t The temperature, humidity, rainfall probability and warning level at time t, EventFlag t R is the event existence flag, t Score the heat, S t is the hot word frequency, L t is the check-in density, N t is the sales index of neighboring stores, constructed using weighted average method:

[0034]

[0035] Among them, Sales i,t is the sales volume of the i-th neighboring store at time t, d i is the geographical distance between the i-th store and the target store, d j is the geographical distance between the jth store and the target store, n is the number of neighboring stores involved in the weighted calculation, and e is the base of the natural logarithm;

[0036] After the vector is constructed, t Perform time series alignment to ensure that X t and historical sales series Y t Completely consistent in timeline.

[0037] Preferably, the step 2 further comprises:

[0038] Sub-step 2.1: transform the external sales driving feature vector X constructed in step 1 into t and historical sales series Y t Concatenate and construct the input tensor Z t , defined as follows:

[0039] Z t =[X t-m ,X t-m+1 ,...,X t ; Y t-m ,Y t-m+1 ,...,Y t ],

[0040] Among them, m is the length of the historical lookback window, X t is the external sales feature vector at time t, Y t is the sales volume of mineral water at time t, Z t is the two-dimensional time series tensor of the input model;

[0041] If the historical sales data Z t If there are more than k consecutive missing data points in , the current data point will not participate in the training, and the missing segment number will be recorded and entered into the subsequent filling mark;

[0042] Sub-step 2.2, the tensor Z t Input is sent to the deep learning model M based on the long short-term memory network or transformer structure, and the model output is the future predicted sales volume in:

[0043]

[0044] in, is the sales forecast value at the τth time step from the current time t, τ is the length of the forecast period, M is the trained forecast model, which includes the trainable parameter set θ;

[0045] The model training goal is to minimize the prediction error loss function, and the mean square error function is used as the optimization target:

[0046]

[0047] Where L(·) is the loss function, c is the number of training samples, With Y (i) is the predicted value and true value of the i-th sample;

[0048] An early stopping strategy is added during the training process. When the validation set loss does not decrease for p consecutive cycles, the training is terminated.

[0049] Sub-step 2.3, the sales forecast results With the current inventory level I t , the product shelf life E and the replenishment cycle C are used together to analyze the inventory status and calculate the inventory adaptability index Ψ t :

[0050]

[0051] Among them, I t is the current inventory quantity, is the cumulative predicted sales volume in the next C time steps, E is the remaining shelf life of the product, Ψ t is the remaining inventory surplus rate under the unit shelf life;

[0052] If t <0.05, indicating that the inventory will be exhausted within the shelf life but there is a risk;

[0053] If t <-0.1, it is judged as insufficient inventory warning state, and enters the subsequent scheduling link and is marked as urgent replenishment;

[0054] If t >1, indicating excess inventory, entering the slow-moving reminder mark;

[0055] At the same time, output the inventory status label Label t ∈{urgent replenishment, normal inventory, slow-selling reminder}, which serves as the basic judgment basis for formulating the inventory scheduling strategy in the subsequent step 3.

[0056] Preferably, the step 3 further comprises:

[0057] Sub-step 3.1: Based on the sales forecast sequence output in step 2 Inventory adaptability index Ψ t and the inventory status label t , determine whether it is necessary to generate a replenishment window;

[0058] First, define the remaining available inventory:

[0059]

[0060] Among them, I t is the current inventory, is the sales forecast value at the τth time step from the current time t, and γ is the length of the inventory safety window;

[0061] If I 剩余 <θ s , where θ s is the inventory safety threshold, and marks the current time t as the candidate replenishment starting point;

[0062] Combined with the product replenishment cycle C, generate a set of available replenishment time windows:

[0063] Used for subsequent calculation of replenishment quantity;

[0064] Where W is the replenishment time window set, δ is the time offset, is the predicted inventory value at time t+δ, θ s is the inventory safety threshold;

[0065] Sub-step 3.2: Calculate the cumulative difference between the sales forecast and inventory expectations at each moment in the replenishment time window W to determine the recommended replenishment quantity Q 补 , the calculation formula is as follows:

[0066]

[0067] in, To predict the sales volume at time point τ, I t is the current inventory, ∈ is the conservative redundancy factor, Q 补 The number of bottles recommended for replenishment;

[0068] If the calculated Q 补 Exceeding the maximum storage capacity limit I max Subtract the current inventory I t , that is: Q 补 >I max -I t , Q 补 Corrected to I max -I t , to avoid exceeding the maximum inventory capacity that the store can hold;

[0069] Sub-step 3.3: Combine the replenishment time set W calculated in sub-step 3.1 and the replenishment quantity Q determined in sub-step 3.2 补 , build a complete inventory scheduling strategy object S t , defined as follows:

[0070] S t ={t 补 ,Q 补 ,Label t},

[0071] Among them, t 补 Q is the earliest time point that meets the replenishment trigger condition, 补 To recommend replenishment quantity, Label t It is the inventory status label;

[0072] If Label t = Emergency replenishment and Ψ t <-0.2, an additional risk mark high-risk warning is added, triggering a manual intervention prompt;

[0073] The final output strategy S tThis will serve as a reference for behavior tracking and difference comparison in the subsequent step 4, and will also be stored in the strategy history table for retrospective analysis.

[0074] Preferably, the step 4 further comprises:

[0075] Sub-step 4.1: Implement the inventory scheduling policy S generated in step 3 t After that, real-time collection of inventory-related behavior data, including the following types of behavior:

[0076] Warehousing operations, outbound sales, inventory loss reporting, and allocation and transfer;

[0077] Behavior records are converted into a unified structure in:

[0078] is the time when the i-th inventory behavior occurs;

[0079] a i ∈{R,S,D,T} is the behavior type;

[0080] ΔI i is the net impact of this action on inventory;

[0081] If a behavior record is missing a field, it will be considered an invalid record and deleted;

[0082] If the frequency of a certain type of behavioral event in a time period is higher than the threshold φ, an abnormal frequency warning is triggered;

[0083] Sub-step 4.2, the above behavior structure set {B i}Converted to actual inventory change sequence sorted by time series The definition is as follows:

[0084]

[0085] Based on this, we construct the actual inventory time series The recursive formula is:

[0086]

[0087] in, is the actual inventory level at time t;

[0088] The total amount of inventory increase or decrease at the current moment;

[0089] is the sales forecast sequence in step 2 Keep the time granularity consistent and resample and interpolate the actual inventory change series;

[0090] If there is no inventory behavior record in a certain period, then is 0;

[0091] Sub-step 4.3 is to evaluate the replenishment policy S in step 3 t The execution effect of each policy object S t Compare with the actual inventory behavior sequence, at t 补 to t 补 Check whether there is actual replenishment behavior that meets the following conditions within the time window of +C:

[0092] Make a r =R∧|ΔI r -Q 补 |≤∈ q ,

[0093] Among them, R is the behavior type of warehousing, ΔI r is the actual storage quantity, ∈ q is the allowable deviation tolerance, t r is the actual replenishment time, t 补 is the replenishment time, a r Identifies the behavior type;

[0094] If the above conditions are met, the policy is marked as executed, otherwise it is marked as unresponsive;

[0095] Then generate the response state label ρ t ∈{executed, delayed, unresponsive}:

[0096] If actual replenishment occurs in And the quantity condition is met, then ρ t =Executed;

[0097] If it occurs Then ρ t = Delayed execution; otherwise ρ t = No response;

[0098] Where, δ is the maximum tolerated response delay;

[0099] Finally, the response state label ρ generated at each moment t The behavior label stream R = {ρ t} t , serving as the basic input for the difference comparison and feedback correction mechanism in step 5.

[0100] Preferably, the step 5 further comprises:

[0101] Sub-step 5.1: Construct prediction bias indicators and perform difference comparisons

[0102] Take the sales forecast sequence output in step 2 Compared with the actual inventory change series constructed in step 4 Based on this, calculate the sales forecast deviation ε at each time point t , defined as follows:

[0103]

[0104] in, Forecast sales at time t, is the inventory reduction caused by actual sales at time t, ε t is the prediction deviation value;

[0105] Based on the deviation sequence {ε t}, calculate the mean absolute deviation indicator within the sliding window to characterize the overall prediction accuracy:

[0106]

[0107] Among them, w is the sliding window length, MAD t is the average prediction error in the current time window;

[0108] If MAD t >θ m , where θ m The deviation tolerance threshold enters the feedback correction process;

[0109] Sub-step 5.2, when MAD is satisfied t >θ m Under the condition of t :

[0110]

[0111] Among them, E t is the mean square error of the current window;

[0112] E t Based on this, an adaptive feedback factor α is introduced. t Perform weighted updates on the model's intermediate layer parameters, defined as follows:

[0113]

[0114] Where η is the error balance factor;

[0115] Further based on α t Construct the intermediate layer parameter adjustment rules, and set the set of trainable parameters of the model intermediate layer as θ (h) , the revised update formula is:

[0116]

[0117] in, are the updated model parameters, is the gradient recalculated based on the current deviation;

[0118] Sub-step 5.3, combined with the feedback factor α constructed in sub-step 5.2 t With the revised parameter set Integrate into the model structure to form an adaptive model M with feedback capability * , defined as follows:

[0119]

[0120] Among them, M * For deep models with integrated correction mechanisms, is the revised sales forecast result, are the updated model parameters;

[0121] If all k consecutive sliding windows satisfy MAD t ≤θ m , determine that the current model prediction has converged and freeze the feedback correction module;

[0122] The updated Model M * The output is used in step 6 to enter a new round of closed-loop iteration of sales forecasting and inventory scheduling.

[0123] Preferably, the step 6 further comprises:

[0124] Sub-step 6.1, based on the external sales driver feature vector Z constructed in step 1 t The actual inventory change sequence structured in step 4 Perform multi-dimensional feature fusion to build a unified input tensor The definition is as follows:

[0125]

[0126] Among them, Z t is the external sales driving feature vector at time t, is the actual inventory level, R t Label is the behavior response status label, t is the inventory risk level;

[0127] If any input dimension is missing, the nearest time series interpolation is performed to complete it. If it cannot be completed, the data at the current time t is marked as abnormal and the current round of training is skipped;

[0128] Sub-step 6.2, the unified input tensor constructed in the previous step Input to the modified model M generated in step 5.3 * , output the new round of sales forecast value At the same time, the inventory scheduling strategy reconstruction process is executed, which is defined as follows:

[0129]

[0130] Call the corresponding algorithms of sub-steps 3.1 to 3.3 in step 3 to regenerate the scheduling strategy:

[0131] based on and Compare to get the remaining inventory I 剩余 ;

[0132] Determine whether to enter the replenishment window to generate logic;

[0133] If the conditions are met, recalculate the recommended replenishment quantity based on the predicted sales and inventory capacity

[0134] Finally, a new strategy object is formed

[0135] If the new strategy Indicates that the strategy is stable, and subsequent feedback adjustments can be skipped, and only the stable state is recorded;

[0136] In sub-step 6.3, to prevent overfitting or frequent fluctuations in strategy, we introduce the strategy stability indicator Σ t , used to determine whether to terminate or continue the closed-loop optimization process, is defined as follows:

[0137]

[0138] Among them, [·] is the indicator function, Σ t is the stable proportion of the strategy in the past k iterations, k is the length of the lookback window;

[0139] If Σ t ≥θ 稳定 , where θ 稳定 =0.8, the strategy is determined to have converged and enters the maintenance phase, with the feedback mechanism frozen:

[0140]

[0141] Otherwise, continue to execute steps 1 to 5 to form a closed-loop system of continuous iteration.

[0142] Preferably, the weather information includes temperature, humidity, precipitation probability and extreme weather warning level, and the holiday information and activity arrangements are collected through a preset event calendar.

[0143] Preferably, the inventory scheduling strategy triggers a replenishment request in advance when the predicted sales volume exceeds the threshold by setting a dynamic safety stock threshold function;

[0144] The inventory forecast error uses the cumulative deviation and sliding average deviation within the cycle window as evaluation indicators to drive the model correction strategy.

[0145] Preferably, the mineral water sales inventory management method uses the prediction results of each round of iteration as the input for the next round of inventory scheduling, realizing a continuous optimization mechanism based on reinforcement learning.

[0146] The present invention provides a mineral water sales inventory management method based on deep learning. It has the following beneficial effects:

[0147] 1. The present invention adopts a technical solution of feature fusion and unified input tensor construction that integrates external sales drivers with multi-source structured behavioral data, achieving the technical effect of dynamically modeling and real-time response to external factors such as weather, holidays, regional activities, and promotional status in sales forecasts. Compared with the technical solution in the existing technology that only relies on historical sales data to build deep models, this solution solves the shortcomings of slow model response and delayed prediction results in scenarios with sudden changes in sales.

[0148] 2. The present invention adopts a dynamic correction mechanism for the model's intermediate layer parameters based on sliding window error feedback and error function drive, achieving the technical effects of traceable inventory behavior deviations, quantifiable model prediction errors, and sustainable adaptive optimization of prediction capabilities. Compared with the existing technical solutions that ignore inventory behavior factors such as returns, losses, and transfers and have no feedback correction mechanism for prediction errors, this solution solves the problem of inaccurate inventory strategies caused by the accumulation of systematic deviations after long-term prediction operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0149] Figure 1 Flowchart of the present invention. DETAILED DESCRIPTION

[0150] To help those skilled in the art understand the present invention, the following will provide a clear and complete description of the technical solutions in the embodiments of the present invention, in conjunction with the accompanying drawings. Obviously, the described embodiments are only partial embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0151] The present invention is described in detail below with reference to the accompanying drawings:

[0152] Example:

[0153] Please see the attached Figure 1, an embodiment of the present invention provides a mineral water sales inventory management method based on deep learning, comprising:

[0154] Step 1: Collect data sources that influence mineral water sales, standardize the data sources, align their time, and extract features to generate a set of multi-dimensional input feature vectors that describe external sales drivers.

[0155] Sub-step 1.1: Collect four types of external sales-driving data, including weather data, holiday and event schedules, regional popularity trends, and sales of neighboring stores, and perform preliminary cleaning and format standardization.

[0156] Weather data includes temperature, relative humidity, rainfall probability, and extreme weather warning levels;

[0157] Holiday and event arrangements are represented by a binary variable EventFlag. If there is an exhibition or holiday event on that day, EventFlag = 1, otherwise EventFlag = 0.

[0158] The regional popularity trend is generated by combining the frequency index of hot words on social platforms and the LBS check-in density to generate the regional popularity score R:

[0159]

[0160] Among them, μ S and σ S is the mean and standard deviation of the hot word frequency, μ L and σ L is the mean and standard deviation of the check-in density, α and β are weight coefficients,

[0161] S is the frequency of regional social media hot words, and L is the user check-in density;

[0162] Sub-step 1.2: Perform outlier detection on the above-mentioned standardized data, using the triple standard deviation method:

[0163] is an outlier,

[0164] Among them, x t is the observed value at time t, μ and σ are the historical mean and standard deviation of the corresponding variables, and σ is the historical standard deviation;

[0165] For missing segments, a sliding window mean filling method based on local trends is used:

[0166]

[0167] Where k is the window length, is the estimated value of the variable used to complete the time point t, x t-iis the value of the variable at time point t forward i;

[0168] Sub-step 1.3: Combine the cleaned and completed feature variables to construct the feature vector of external sales drivers:

[0169] X t =[T t ,H t ,P t ,W t ,EventFlag t ,R t ,S t ,L t ,N t ],

[0170] Among them, X t is the external sales feature vector at time t, T t 、H t 、P t 、W t The temperature, humidity, rainfall probability and warning level at time t, EventFlag t R is the event existence flag, t Score the heat, S t is the hot word frequency, L t is the check-in density, N t is the sales index of neighboring stores, constructed using weighted average method:

[0171]

[0172] Among them, Sales i,t is the sales volume of the i-th neighboring store at time t, d i is the geographical distance between the i-th store and the target store, d j is the geographical distance between the jth store and the target store, n is the number of neighboring stores involved in the weighted calculation, and e is the base of the natural logarithm;

[0173] After the vector is constructed, t Perform time series alignment to ensure that X t and historical sales series Y t Completely consistent in timeline;

[0174] Step 2: The multi-dimensional input feature vector and historical sales data are fed into a deep learning model to train and obtain future sales forecasts for mineral water. The model then analyzes inventory status by combining current inventory data, shelf life parameters, and replenishment cycle parameters.

[0175] Sub-step 2.1: transform the external sales driving feature vector X constructed in step 1 into tand historical sales series Y t Concatenate and construct the input tensor Z t , defined as follows:

[0176] Z t =[X t-m ,X t-m+1 ,...,X t ; Y t-m ,Y t-m+1 ,...,Y t ],

[0177] Among them, m is the length of the historical lookback window, X t is the external sales feature vector at time t, Y t is the sales volume of mineral water at time t, Z t is the two-dimensional time series tensor of the input model;

[0178] If the historical sales data Z t If there are more than k consecutive missing data points in , the current data point will not participate in the training, and the missing segment number will be recorded and entered into the subsequent filling mark;

[0179] Substep 2.2, transform the tensor Z t Input is sent to the deep learning model M based on the long short-term memory network or transformer structure, and the model output is the future predicted sales volume in:

[0180]

[0181] in, is the sales forecast value at the τth time step from the current time t, τ is the length of the forecast period, M is the trained forecast model, which includes the trainable parameter set θ;

[0182] The model training goal is to minimize the prediction error loss function, and the mean square error function is used as the optimization target:

[0183]

[0184] Where L(·) is the loss function, c is the number of training samples, With Y (i) is the predicted value and true value of the i-th sample;

[0185] An early stopping strategy is added during the training process. When the validation set loss does not decrease for p consecutive cycles, the training is terminated.

[0186] Sub-step 2.3, the sales forecast results With the current inventory level I t, the product shelf life E and the replenishment cycle C are used together to analyze the inventory status and calculate the inventory adaptability index Ψ t :

[0187]

[0188] Among them, I t is the current inventory quantity, is the cumulative predicted sales volume in the next C time steps, E is the remaining shelf life of the product, Ψ t is the remaining inventory surplus rate under the unit shelf life;

[0189] If t <0.05, indicating that the inventory will be exhausted within the shelf life but there is a risk;

[0190] If t <-0.1, it is judged as insufficient inventory warning state, and enters the subsequent scheduling link and is marked as urgent replenishment;

[0191] If t >1, indicating excess inventory, entering the slow-moving reminder mark;

[0192] At the same time, output the inventory status label Label t ∈{urgent replenishment, normal inventory, slow-selling reminder}, which serves as the basis for determining the inventory scheduling strategy in the subsequent step 3;

[0193] Step 3: Generate an inventory scheduling strategy based on the sales forecast and inventory status analysis results. The inventory scheduling strategy includes replenishment timing, replenishment quantity, and inventory risk warning notifications.

[0194] Sub-step 3.1: Based on the sales forecast sequence output in step 2 Inventory adaptability index Ψ t and the inventory status label t , determine whether it is necessary to generate a replenishment window;

[0195] First, define the remaining available inventory:

[0196]

[0197] Among them, I t is the current inventory, is the sales forecast value at the τth time step from the current time t, and γ is the length of the inventory safety window;

[0198] If I 剩余 <θ s , where θ s is the inventory safety threshold, and marks the current time t as the candidate replenishment starting point;

[0199] Combined with the product replenishment cycle C, generate a set of available replenishment time windows:

[0200] Used for subsequent calculation of replenishment quantity;

[0201] Where W is the replenishment time window set, δ is the time offset, is the predicted inventory value at time t+δ, θ s is the inventory safety threshold;

[0202] Sub-step 3.2: Calculate the cumulative difference between the sales forecast and inventory expectations at each moment in the replenishment time window W to determine the recommended replenishment quantity Q 补 , the calculation formula is as follows:

[0203]

[0204] in, To predict the sales volume at time point τ, I t is the current inventory, ∈ is the conservative redundancy factor, Q 补 The number of bottles recommended for replenishment;

[0205] If the calculated Q 补 Exceeding the maximum storage capacity limit I max Subtract the current inventory I t , that is: Q 补 >I max -I t , Q 补 Corrected to I max -I t , to avoid exceeding the maximum inventory capacity that the store can hold;

[0206] Sub-step 3.3: Combine the replenishment time set W calculated in sub-step 3.1 and the replenishment quantity Q determined in sub-step 3.2 补 , build a complete inventory scheduling strategy object S t , defined as follows:

[0207] S t ={t 补 ,Q 补 ,Label t},

[0208] Among them, t 补 Q is the earliest time point that meets the replenishment trigger condition, 补 To recommend replenishment quantity, Label t It is the inventory status label;

[0209] If Label t = Emergency replenishment and Ψt <-0.2, an additional risk mark high-risk warning is added, triggering a manual intervention prompt;

[0210] The final output strategy S t This will serve as a reference for behavior tracking and difference comparison in step 4, and will also be stored in the strategy history table for retrospective analysis.

[0211] Step 4: After the inventory scheduling plan is output, inventory behavior data is collected synchronously during the actual inventory management process, and the inventory behavior data is structurally transformed to form an actual inventory change data stream that matches the sales forecast;

[0212] Sub-step 4.1: Implement the inventory scheduling policy S generated in step 3 t After that, real-time collection of inventory-related behavior data, including the following types of behavior:

[0213] Warehousing operations, outbound sales, inventory loss reporting, and allocation and transfer;

[0214] Behavior records are converted into a unified structure in:

[0215] is the time when the i-th inventory behavior occurs;

[0216] a i ∈{R,S,D,T} is the behavior type;

[0217] ΔI i is the net impact of this action on inventory;

[0218] If a behavior record is missing a field, it will be considered an invalid record and deleted;

[0219] If the frequency of a certain type of behavioral event in a time period is higher than the threshold φ, an abnormal frequency warning is triggered;

[0220] Sub-step 4.2, the above behavior structure set {B i}Converted to actual inventory change sequence sorted by time series The definition is as follows:

[0221]

[0222] Based on this, we construct the actual inventory time series The recursive formula is:

[0223]

[0224] in, is the actual inventory level at time t;

[0225] The total amount of inventory increase or decrease at the current moment;

[0226] is the sales forecast sequence in step 2 Keep the time granularity consistent and resample and interpolate the actual inventory change series;

[0227] If there is no inventory behavior record in a certain period, then is 0;

[0228] Sub-step 4.3 is to evaluate the replenishment policy S in step 3 t The execution effect of each policy object S t Compare with the actual inventory behavior sequence, at t 补 to t 补 Check whether there is actual replenishment behavior that meets the following conditions within the time window of +C:

[0229] Make a r =R∧|ΔI r -Q 补 |≤∈ q ,

[0230] Among them, R is the behavior type of warehousing, ΔI r is the actual storage quantity, ∈ q is the allowable deviation tolerance, t r is the actual replenishment time, t 补 is the replenishment time, a r Identifies the behavior type;

[0231] If the above conditions are met, the policy is marked as executed, otherwise it is marked as unresponsive;

[0232] Then generate the response state label ρ t ∈{executed, delayed, unresponsive}:

[0233] If actual replenishment occurs in And the quantity condition is met, then ρ t =Executed;

[0234] If it occurs Then ρ t = Delayed execution; otherwise ρ t = No response;

[0235] Where, δ is the maximum tolerated response delay;

[0236] Finally, the response state label ρ generated at each moment t The behavior label stream R = {ρ t} t , as the basic input for the difference comparison and feedback correction mechanism in step 5;

[0237] Step 5: Use the actual inventory change data stream to compare the sales forecast results with the inventory scheduling plan. Based on the comparison results, a feedback mechanism is established. The feedback mechanism constructs model parameter update rules based on the forecast deviation data. By defining the error reference value for weight adjustment, the parameters of the intermediate layer of the deep learning model are dynamically corrected.

[0238] Sub-step 5.1: Construct prediction bias indicators and perform difference comparisons

[0239] Take the sales forecast sequence output in step 2 Compared with the actual inventory change series constructed in step 4 Based on this, calculate the sales forecast deviation ε at each time point t , defined as follows:

[0240]

[0241] in, Forecast sales at time t, is the inventory reduction caused by actual sales at time t, ε t is the prediction deviation value;

[0242] Based on the deviation sequence {ε t}, calculate the mean absolute deviation indicator within the sliding window to characterize the overall prediction accuracy:

[0243]

[0244] Among them, w is the sliding window length, MAD t is the average prediction error in the current time window;

[0245] If MAD t >θ m , where θ m The deviation tolerance threshold enters the feedback correction process;

[0246] Sub-step 5.2, when MAD is satisfied t >θ m Under the condition of t :

[0247]

[0248] Among them, E t is the mean square error of the current window;

[0249] E t Based on this, an adaptive feedback factor α is introduced. tPerform weighted updates on the model's intermediate layer parameters, defined as follows:

[0250]

[0251] Where η is the error balance factor;

[0252] Further based on α t Construct the intermediate layer parameter adjustment rules, and set the set of trainable parameters of the model intermediate layer as θ (h) , the revised update formula is:

[0253]

[0254] in, are the updated model parameters, is the gradient recalculated based on the current deviation;

[0255] Sub-step 5.3, combined with the feedback factor α constructed in sub-step 5.2 t With the revised parameter set Integrate into the model structure to form an adaptive model M with feedback capability * , defined as follows:

[0256]

[0257] Among them, M * For deep models with integrated correction mechanisms, is the revised sales forecast result, are the updated model parameters;

[0258] If all k consecutive sliding windows satisfy MAD t ≤θ m , determine that the current model prediction has converged and freeze the feedback correction module;

[0259] The updated Model M * The output is used in step 6 to enter a new round of closed-loop iteration of sales forecasting and inventory scheduling;

[0260] In step 6, the parameter-corrected deep learning model receives the multi-source input data from step 1 and the inventory behavior data from step 4 again, and performs a new round of sales forecasting and inventory scheduling. This forms an iterative optimization process that combines forecasting and feedback, achieving continuous closed-loop management of mineral water sales and inventory scheduling.

[0261] Sub-step 6.1, based on the external sales driver feature vector Z constructed in step 1 t The actual inventory change sequence structured in step 4 Perform multi-dimensional feature fusion to build a unified input tensor The definition is as follows:

[0262]

[0263] Among them, Z t is the external sales driving feature vector at time t, is the actual inventory level, R t Label is the behavior response status label, t is the inventory risk level;

[0264] If any input dimension is missing, the nearest time series interpolation is performed to complete it. If it cannot be completed, the data at the current time t is marked as abnormal and the current round of training is skipped;

[0265] Sub-step 6.2, the unified input tensor constructed in the previous step Input to the modified model M generated in step 5.3 * , output the new round of sales forecast value At the same time, the inventory scheduling strategy reconstruction process is executed, which is defined as follows:

[0266]

[0267] Call the corresponding algorithms of sub-steps 3.1 to 3.3 in step 3 to regenerate the scheduling strategy:

[0268] based on and Compare to get the remaining inventory I 剩余 ;

[0269] Determine whether to enter the replenishment window to generate logic;

[0270] If the conditions are met, recalculate the recommended replenishment quantity based on the predicted sales and inventory capacity

[0271] Finally, a new strategy object is formed

[0272] If the new strategy Indicates that the strategy is stable, and subsequent feedback adjustments can be skipped, and only the stable state is recorded;

[0273] In sub-step 6.3, to prevent overfitting or frequent fluctuations in strategy, we introduce the strategy stability indicator Σ t , used to determine whether to terminate or continue the closed-loop optimization process, is defined as follows:

[0274]

[0275] Among them, [·] is the indicator function, Σ t is the stable proportion of the strategy in the past k iterations, k is the length of the lookback window;

[0276] If Σ t ≥θ 稳定 , where θ 稳定 =0.8, the strategy is determined to have converged and enters the maintenance phase, with the feedback mechanism frozen:

[0277]

[0278] Otherwise, continue to execute steps 1 to 5 to form a closed-loop system of continuous iteration.

[0279] The benefits of Step 1 are that by comprehensively collecting and standardizing multi-source external sales-driving data, such as weather, holidays, regional popularity, and neighboring store sales, we achieve structured modeling of external disturbance factors, enabling the model to proactively perceive market changes. Through outlier detection and intelligent filling of missing values, data quality and integrity are improved, significantly enhancing the input accuracy and generalization capabilities of subsequent deep learning models. This provides a stable and reliable input foundation for sales forecasting, improving overall forecast accuracy, and is particularly adaptable to high-frequency fluctuations, such as unusual changes in mineral water sales during hot weather or major promotions.

[0280] The benefits of step 2 are that multidimensional external features are integrated with historical sales data and fed into a deep neural network model, enabling multi-factor dynamic modeling of mineral water sales and accurate short-term trend forecasting. Furthermore, by combining parameters such as current inventory status, shelf life, and replenishment cycle, inventory adaptability indicators are calculated, enabling detailed classification of different inventory scenarios. This enables the system to assess replenishment risks in real time and provide forward-looking inventory scheduling capabilities, significantly reducing out-of-stock and overstock situations and improving inventory structure rationality and product turnover efficiency.

[0281] The benefits of step 3 include automatically generating precise inventory scheduling strategies based on sales forecasts and inventory adaptability, including replenishment timing, quantity, and warning levels. This improves the intelligence and responsiveness of inventory management. By defining inventory safety thresholds and forecast windows, potential inventory risks are proactively translated into scheduling actions, effectively shortening replenishment response cycles and avoiding sales losses caused by delayed decisions. Furthermore, the strategies consider replenishment capacity constraints and the rationality of time windows to ensure their feasibility and adaptability to practical constraints.

[0282] The benefits of Step 4 are the establishment of a structured management system for inventory behavior data. Through standardized collection and conversion of actions such as warehousing, outbound shipments, transfers, and loss reports, a sequence of actual inventory changes consistent with sales forecast granularity is generated, enabling the system to fully recover inventory change paths. When comparing policy execution, it can accurately identify whether the policy was executed, the execution time, and any deviations from the execution, providing fundamental data support for subsequent feedback mechanisms. This also enhances the visualization and traceability of inventory operations, helping operators promptly identify management loopholes and behavioral anomalies.

[0283] The benefit of Step 5 is that by systematically comparing sales forecasts with actual inventory behavior, an error feedback mechanism is established, enabling quantitative perception and dynamic correction of model forecast errors. A sliding window calculates error metrics and, combined with adaptive factors, adjusts model intermediate parameters, enabling the model to learn and self-calibrate over time. This effectively suppresses the accumulation of systematic biases caused by inventory fluctuations and enhances model stability and adaptability. Through this closed feedback loop, the model gradually approaches actual operational performance, providing the core technical foundation for precise inventory management.

[0284] The benefit of step 6 is that the deep model, modified through feedback, is re-entered into the forecasting and scheduling process, achieving an iterative optimization loop between forecasts and actual behavior. By integrating the latest external features with inventory behavior data for unified modeling, the system consistently adjusts based on the latest business dynamics, enabling continuous learning and policy adaptation. The introduction of a policy stability assessment mechanism effectively mitigates frequent policy fluctuations or model overfitting, ensuring stable scheduling policy output and long-term system maintainability.

[0285] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A mineral water sales inventory management method based on deep learning, characterized in that: include: Step 1: Collect data sources that influence mineral water sales, standardize the data sources, align their time, and extract features to generate a set of multi-dimensional input feature vectors that describe external sales drivers. Step 2: The multi-dimensional input feature vector and historical sales data are fed into a deep learning model to train and obtain future sales forecasts for mineral water. The model then analyzes inventory status by combining current inventory data, shelf life parameters, and replenishment cycle parameters. Step 3: Generate an inventory scheduling strategy based on the sales forecast and inventory status analysis results. The inventory scheduling strategy includes replenishment timing, replenishment quantity, and inventory risk warning notifications. Step 4: After the inventory scheduling plan is output, inventory behavior data is collected synchronously during the actual inventory management process, and the inventory behavior data is structurally transformed to form an actual inventory change data stream that matches the sales forecast; Step 5: Using the actual inventory change data stream, compare the sales forecast results with the inventory scheduling plan. Based on the comparison results, a feedback mechanism is established. The feedback mechanism constructs model parameter update rules based on the forecast deviation data, and dynamically corrects the parameters of the intermediate layers of the deep learning model by defining error reference values ​​for weight adjustment. In step 6, the deep learning model with modified parameters receives the multi-source input data in step 1 and the inventory behavior data in step 4 again, and performs a new round of sales forecasting and inventory scheduling processing, forming an iterative optimization process that combines forecasting and feedback, and realizing continuous closed-loop management of mineral water sales and inventory scheduling.

2. The method for mineral water sales inventory management based on deep learning according to claim 1, characterized in that: The step 1 further comprises: Sub-step 1.1: Collect four types of external sales-driving data, including weather data, holiday and event schedules, regional popularity trends, and sales of neighboring stores, and perform preliminary cleaning and format standardization. Weather data includes temperature, relative humidity, rainfall probability, and extreme weather warning levels; Holiday and event arrangements are represented by a binary variable EventFlag. If there is an exhibition or holiday event on that day, EventFlag = 1, otherwise EventFlag = 0. The regional popularity trend is generated by combining the frequency index of hot words on social platforms and the LBS check-in density to generate the regional popularity score R: Among them, μ S and σ S is the mean and standard deviation of the hot word frequency, μ L and σ L is the mean and standard deviation of the check-in density, α and β are weight coefficients, S is the frequency of regional social media hot words, and L is the user check-in density; Sub-step 1.2: Perform outlier detection on the above-mentioned standardized data, using the triple standard deviation method: is an outlier, Among them, x t is the observed value at time t, μ and σ are the historical mean and standard deviation of the corresponding variables, and σ is the historical standard deviation; For missing segments, a sliding window mean filling method based on local trends is used: Where k is the window length, is the estimated value of the variable used to complete the time point t, x t-i is the value of the variable at time point t forward i; Sub-step 1.3: Combine the cleaned and completed feature variables to construct the feature vector of external sales drivers: X t =[T t ,H t ,P t ,W t ,EventFlag t ,R t ,S t ,L t ,N t ], Among them, X t is the external sales feature vector at time t, T t 、H t 、P t 、W t The temperature, humidity, rainfall probability and warning level at time t, EventFlag t R is the event existence flag, t Score the heat, S t is the hot word frequency, L t is the check-in density, N t is the sales index of neighboring stores, constructed using weighted average method: Among them, Sales i,t is the sales volume of the i-th neighboring store at time t, d i is the geographical distance between the i-th store and the target store, d j is the geographical distance between the jth store and the target store, n is the number of neighboring stores involved in the weighted calculation, and e is the base of the natural logarithm; After the vector is constructed, t Perform time series alignment to ensure that X t and historical sales series Y t Completely consistent in timeline.

3. The method for mineral water sales inventory management based on deep learning according to claim 1, characterized in that: The step 2 further comprises: Sub-step 2.1: transform the external sales driving feature vector X constructed in step 1 into t and historical sales series Y t Concatenate and construct the input tensor Z t , defined as follows: Z t =[X t-m ,X t-m+1 ,...,X t ;AND t-m ,AND t-m+1 ,...,AND t ], Among them, m is the length of the historical lookback window, X t is the external sales feature vector at time t, Y t is the sales volume of mineral water at time t, Z t is the two-dimensional time series tensor of the input model; If the historical sales data Z t If there are more than k consecutive missing data points in , the current data point will not participate in the training, and the missing segment number will be recorded and entered into the subsequent filling mark; Sub-step 2.2, the tensor Z t Input is sent to the deep learning model M based on the long short-term memory network or transformer structure, and the model output is the future predicted sales volume in: in, is the sales forecast value at the τth time step from the current time t, τ is the length of the forecast period, M is the trained forecast model, which includes the trainable parameter set θ; The model training goal is to minimize the prediction error loss function, and the mean square error function is used as the optimization target: Where L(·) is the loss function, c is the number of training samples, With Y (i) is the predicted value and true value of the i-th sample; An early stopping strategy is added during the training process. When the validation set loss does not decrease for p consecutive cycles, the training is terminated. Sub-step 2.3, the sales forecast results With the current inventory level I t , the product shelf life E and the replenishment cycle C are used together to analyze the inventory status and calculate the inventory adaptability index Ψ t : Among them, I t is the current inventory quantity, is the cumulative predicted sales volume in the next C time steps, E is the remaining shelf life of the product, Ψ t is the remaining inventory surplus rate under the unit shelf life; If t <0.05, indicating that the inventory will be exhausted within the shelf life but there is a risk; If t <-0.1, it is judged as insufficient inventory warning state, and enters the subsequent scheduling link and is marked as urgent replenishment; If t >1, indicating excess inventory, entering the slow-moving reminder mark; At the same time, output the inventory status label Label t ∈{urgent replenishment, normal inventory, slow-selling reminder}, which serves as the basic judgment basis for formulating the inventory scheduling strategy in the subsequent step 3.

4. The method for mineral water sales inventory management based on deep learning according to claim 1, characterized in that: The step 3 further comprises: Sub-step 3.1: Based on the sales forecast sequence output in step 2 Inventory adaptability index Ψ t and the inventory status label t , determine whether it is necessary to generate a replenishment window; First, define the remaining available inventory: Among them, I t is the current inventory, is the sales forecast value at the τth time step from the current time t, and γ is the length of the inventory safety window; If I 剩余 <θ s , where θ s is the inventory safety threshold, and marks the current time t as the candidate replenishment starting point; Combined with the product replenishment cycle C, generate a set of available replenishment time windows: Used for subsequent calculation of replenishment quantity; Where W is the replenishment time window set, δ is the time offset, is the predicted inventory value at time t+δ, θ s is the inventory safety threshold; Sub-step 3.2: Calculate the cumulative difference between the sales forecast and inventory expectations at each moment in the replenishment time window W to determine the recommended replenishment quantity Q 补 , the calculation formula is as follows: in, To predict the sales volume at time point τ, I t is the current inventory, ∈ is the conservative redundancy factor, Q 补 The number of bottles recommended for replenishment; If the calculated Q 补 Exceeding the maximum storage capacity limit I max Subtract the current inventory I t , that is: Q 补 >I max -I t , Q 补 Corrected to I max -I t , to avoid exceeding the maximum inventory capacity that the store can hold; Sub-step 3.3: Combine the replenishment time set W calculated in sub-step 3.1 and the replenishment quantity Q determined in sub-step 3.2 补 , build a complete inventory scheduling strategy object S t , defined as follows: S t ={t 补 ,Q 补 ,Label t }, Among them, t 补 Q is the earliest time point that meets the replenishment trigger condition, 补 To recommend replenishment quantity, Label t It is the inventory status label; If Label t = Emergency replenishment and Ψ t <-0.2, an additional risk mark high-risk warning is added, triggering a manual intervention prompt; The final output strategy S t This will serve as a reference for behavior tracking and difference comparison in the subsequent step 4, and will also be stored in the strategy history table for retrospective analysis.

5. The method for mineral water sales inventory management based on deep learning according to claim 1, characterized in that: The step 4 further comprises: Sub-step 4.1: Implement the inventory scheduling policy S generated in step 3 t After that, real-time collection of inventory-related behavior data, including the following types of behavior: Warehousing operations, outbound sales, inventory loss reporting, and allocation and transfer; Behavior records are converted into a unified structure in: is the time when the i-th inventory behavior occurs; a i ∈{R,S,D,T} is the behavior type; ΔI i is the net impact of this action on inventory; If a behavior record is missing a field, it will be considered an invalid record and deleted; If the frequency of a certain type of behavioral event in a time period is higher than the threshold φ, an abnormal frequency warning is triggered; Sub-step 4.2, the above behavior structure set {B i }Converted to actual inventory change sequence sorted by time series The definition is as follows: Based on this, we construct the actual inventory time series The recursive formula is: in, is the actual inventory level at time t; The total amount of inventory increase or decrease at the current moment; is the sales forecast sequence in step 2 Keep the time granularity consistent and resample and interpolate the actual inventory change series; If there is no inventory behavior record in a certain period, then is 0; Sub-step 4.3 is to evaluate the replenishment policy S in step 3 t The execution effect of each policy object S t Compare with the actual inventory behavior sequence, at t 补 to t 补 Check whether there is actual replenishment behavior that meets the following conditions within the time window of +C: Make a r =R∧|ΔI r -Q 补 |≤∈ q , Among them, R is the behavior type of warehousing, ΔI r is the actual storage quantity, ∈ q is the allowable deviation tolerance, t r is the actual replenishment time, t 补 is the replenishment time, a r Identifies the behavior type; If the above conditions are met, the policy is marked as executed, otherwise it is marked as unresponsive; Then generate the response state label ρ t ∈{executed, delayed, unresponsive}: If actual replenishment occurs in And the quantity condition is met, then ρ t =Executed; If it occurs Then ρ t = Delayed execution; otherwise ρ t = No response; Where, δ is the maximum tolerated response delay; Finally, the response state label ρ generated at each moment t The behavior label stream R = {ρ t } t , serving as the basic input for the difference comparison and feedback correction mechanism in step 5.

6. The method for mineral water sales inventory management based on deep learning according to claim 1, characterized in that: The step 5 further includes: Sub-step 5.1: Construct prediction bias indicators and perform difference comparisons Take the sales forecast sequence output in step 2 Compared with the actual inventory change series constructed in step 4 Based on this, calculate the sales forecast deviation ε at each time point t , defined as follows: in, Forecast sales at time t, is the inventory reduction caused by actual sales at time t, ε t is the prediction deviation value; Based on the deviation sequence {ε t }, calculate the mean absolute deviation indicator within the sliding window to characterize the overall prediction accuracy: Among them, w is the sliding window length, MAD t is the average prediction error in the current time window; If MAD t >θ m , where θ m The deviation tolerance threshold enters the feedback correction process; Sub-step 5.2, when MAD is satisfied t >θ m Under the condition of t : Among them, E t is the mean square error of the current window; E t Based on this, an adaptive feedback factor α is introduced. t Perform weighted updates on the model's intermediate layer parameters, defined as follows: Where η is the error balance factor; Further based on α t Construct the intermediate layer parameter adjustment rules, and set the set of trainable parameters of the model intermediate layer as θ (h) , the revised update formula is: in, are the updated model parameters, is the gradient recalculated based on the current deviation; Sub-step 5.3, combined with the feedback factor α constructed in sub-step 5.2 t With the revised parameter set Integrate into the model structure to form an adaptive model M with feedback capability * , defined as follows: Among them, M * For deep models with integrated correction mechanisms, is the revised sales forecast result, are the updated model parameters; If all k consecutive sliding windows satisfy MAD t ≤θ m , determine that the current model prediction has converged and freeze the feedback correction module; The updated Model M * The output is used in step 6 to enter a new round of closed-loop iteration of sales forecasting and inventory scheduling.

7. The method for mineral water sales inventory management based on deep learning according to claim 1, characterized in that: The step 6 further includes: Sub-step 6.1, based on the external sales driver feature vector Z constructed in step 1 t The actual inventory change sequence structured in step 4 Perform multi-dimensional feature fusion to build a unified input tensor The definition is as follows: Among them, Z t is the external sales driving feature vector at time t, is the actual inventory level, R t Label is the behavior response status label, t is the inventory risk level; If any input dimension is missing, the nearest time series interpolation is performed to complete it. If it cannot be completed, the data at the current time t is marked as abnormal and the current round of training is skipped; Sub-step 6.2, the unified input tensor constructed in the previous step Input to the modified model M generated in step 5.3 * , output the new round of sales forecast value At the same time, the inventory scheduling strategy reconstruction process is executed, which is defined as follows: Call the corresponding algorithms of sub-steps 3.1 to 3.3 in step 3 to regenerate the scheduling strategy: based on and Compare to get the remaining inventory I 剩余 ; Determine whether to enter the replenishment window to generate logic; If the conditions are met, recalculate the recommended replenishment quantity based on the predicted sales and inventory capacity Finally, a new strategy object is formed If the new strategy Indicates that the strategy is stable, and subsequent feedback adjustments can be skipped, and only the stable state is recorded; In sub-step 6.3, to prevent overfitting or frequent fluctuations in strategy, we introduce the strategy stability indicator Σ t , used to determine whether to terminate or continue the closed-loop optimization process, is defined as follows: Among them, [·] is the indicator function, Σ t is the stable proportion of the strategy in the past k iterations, k is the length of the lookback window; If Σ t ≥θ 稳定 , where θ 稳定 =0.8, the strategy is determined to have converged and enters the maintenance phase, with the feedback mechanism frozen: Otherwise, continue to execute steps 1 to 5 to form a closed-loop system of continuous iteration.

8. The method for mineral water sales inventory management based on deep learning according to claim 1, characterized in that: The weather information includes temperature, humidity, precipitation probability and extreme weather warning level, and the holiday information and activity schedule are collected through a preset event calendar.

9. The method for mineral water sales and inventory management based on deep learning according to claim 1, characterized in that: The inventory scheduling strategy sets a dynamic safety stock threshold function to trigger replenishment requests in advance when the predicted sales volume exceeds the threshold; The inventory forecast error uses the cumulative deviation and sliding average deviation within the cycle window as evaluation indicators to drive the model correction strategy.

10. The method for mineral water sales and inventory management based on deep learning according to claim 1, characterized in that: The mineral water sales inventory management method uses the prediction results of each round of iteration as the input for the next round of inventory scheduling, realizing a continuous optimization mechanism based on reinforcement learning.

Images