A perishable commodity procurement quantity optimization method based on a prediction model

Abstract

The application discloses a perishable commodity procurement quantity optimization method based on a prediction model, and comprises the following steps: obtaining historical demand data and related features of a commodity to be predicted; selecting k demand prediction models, respectively predicting the demand to obtain the predicted demand quantity, and taking the result as k random scenarios; substituting the random scenarios into a two-stage stochastic programming model to obtain the optimal pre-procurement quantity of each commodity, wherein the first stage determines the quantity of advance procurement, and the second stage determines the replenishment procurement quantity; and when the real demand is known, calculating the replenishment procurement quantity according to the real demand. The method combines multiple demand prediction models, can solve the defects that a single model is difficult to accurately predict and has stable good performance in cost saving, accurately predicts the demand of perishable commodities, can simultaneously predict and optimize procurement for a "basket of commodities", and has good practicability and feasibility.

Description

Technical Field

[0001] This invention relates to the field of perishable goods, and in particular to a method for optimizing the procurement quantity of perishable goods based on a predictive model. Background Technology

[0002] In the perishable goods sector, such as fast fashion apparel, fresh food, newspapers and magazines, and airline and shipping tickets, the products have a highly time-sensitive nature. Once the short time window is exceeded, the value of the goods will be significantly reduced. More worryingly, the order demand in these sectors is highly uncertain; for example, seasons, weather, weekends, and holidays can all cause drastic fluctuations in demand. These factors present a significant challenge for purchasing decision-makers in determining appropriate procurement quantities.

[0003] Due to the high degree of uncertainty in demand, multiple purchases to meet all order requirements are often unavoidable.

[0004] Even more challenging is that, because manufacturing requires a certain cycle, many practitioners in the field need to make production or procurement decisions before downstream demand is known. Inaccurate demand forecasts or incomplete consideration of procurement costs can lead to significant economic losses for decision-makers.

[0005] For example, in the fast fashion apparel industry, before each sales season, manufacturers need to develop new product production plans in advance, determine the production volume of each SKU, produce in advance, and then distribute them to stores for sale. If these garments are popular in the market, but the production quantity is small and supply cannot meet demand, stores will encounter stockouts, leading to lost orders and reduced profits; if the production quantity is too large, it will lead to unsold inventory and inventory costs.

[0006] Taking fresh food markets, which primarily supply vegetables, fresh meat, and dairy products, as an example, most downstream distributors (such as supermarkets and restaurants) typically need to prepare products in advance. For instance, supermarkets need to be ready before 5 a.m. Furthermore, due to complex processing operations, the sorting process for fresh food usually takes several hours. Because of these issues, downstream distributors must impose very strict time constraints on the delivery times of their upstream suppliers. Therefore, upstream distributors without known demand information need to purchase several hours in advance (known as advance purchasing) to ensure timely supply to their downstream customers. In this process, as time goes on, the actual needs of customers gradually become known.

[0007] like Figure 1 and Figure 2a -c is shown, where Figure 2aThe x-axis of the -c graph represents time, and the y-axis represents demand (in kilograms). Data from a wholesale company shows that demand is almost nonexistent from midnight to 6 AM, but increases after customers return to work. In particular, demand is highly concentrated between 8 PM and 11 PM. Furthermore, many factors can influence the demand for fresh food, such as season, weather, weekends, and holidays, inevitably leading to significant fluctuations in demand. Based on these facts, since demand cannot be accurately estimated, advance purchasing cannot meet all order demands for all goods. Therefore, additional purchases are typically made to satisfy remaining demand. We call this additional purchasing replenishment purchasing.

[0008] Uncertain demand presents significant challenges to procurement decisions. Inaccurate procurement can lead to substantial financial losses for distributors.

[0009] It should be noted that the information disclosed in the background section above is only for understanding the background of this application, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0010] The purpose of this invention is to solve the problem of accurately predicting the purchase volume of perishable goods and to provide a method for optimizing the purchase volume of perishable goods based on a prediction model.

[0011] To achieve the above objectives, the present invention adopts the following technical solution:

[0012] A method for optimizing the purchase quantity of perishable goods based on a predictive model includes the following steps:

[0013] S1: Obtain historical demand data and relevant characteristics of the product to be predicted;

[0014] S2: Select k demand forecasting models, forecast the demand for each model, and obtain the predicted demand amount. The results are used as k random scenarios.

[0015] S3: Substitute the random scenario into a two-stage stochastic programming model to solve the problem and obtain the optimal pre-purchase quantity for each commodity;

[0016] S4: Once the actual demand is known, calculate the supplementary procurement quantity based on the actual demand;

[0017] In the two-stage stochastic programming model, the first stage determines the quantity to be purchased in advance, and the second stage determines the quantity to be purchased for replenishment.

[0018] In some embodiments, the objective function to be minimized by the two-stage stochastic programming model in step S3 consists of two parts: the expected cost of advance purchase of products in the first stage and the expected cost of replenishment purchase in the second stage, expressed by the following formula:

[0019]

[0020] Where, x i π refers to the pre-order quantity of product i, and π refers to the total cost incurred to fulfill all orders within a day. c refers to the collection of all perishable products. 1i The price of perishable products i purchased in advance. This refers to the expected cost during the replenishment procurement phase, where y represents the vector composed of the procurement quantities of each commodity during the replenishment procurement phase.

[0021] Furthermore, the expected cost of replenishment procurement in the second phase is expressed by the following formula:

[0022]

[0023] Where, x i This refers to the pre-order quantity of product i. c refers to the collection of all perishable products. 2i The price of product i in replenishment procurement, ξ s This refers to a possible scenario for all product demands, y i，s This refers to the quantity of product i purchased during replenishment. Ω refers to the predicted demand for product i on the current day in scenario s, and Ω refers to the set of all generated random scenarios.

[0024] Furthermore, the constraints of the objective function include:

[0025]

[0026] Where, x i This refers to the pre-order quantity of product i. It refers to the collection of all perishable products.

[0027] Furthermore, the constraints of the objective function include:

[0028]

[0029] Where, x i This refers to the pre-order quantity of product i. Refers to the collection of all perishable products. This refers to the predicted demand for product i on a given day, within scenario s.

[0030] Furthermore, the constraints on the expected cost of the replenishment procurement include:

[0031]

[0032] Among them, y i，sThis refers to the quantity of product i purchased during replenishment. Refers to the collection of all perishable products, ξ s Ω refers to a possible scenario for all product demands, while Ω refers to the set of all randomly generated scenarios.

[0033] Furthermore, the constraints on the expected cost of the replenishment procurement include:

[0034]

[0035] Among them, y i，s This refers to the quantity of product i purchased during replenishment. x refers to the collection of all perishable products. i The pre-order quantity of product i, ξ s This refers to a possible scenario for all product demands. Ω refers to the predicted demand for product i on the current day in scenario s, and Ω refers to the set of all generated random scenarios.

[0036] In some embodiments, the prediction model is a trained demand prediction model.

[0037] In some embodiments, the prediction model includes a 7-day median model and an ANN-MLP model.

[0038] The present invention also discloses a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, can implement the method described in any of the above-mentioned embodiments.

[0039] The present invention has the following beneficial effects:

[0040] This invention, based on a real-world problem scenario, innovatively models the problem as a two-stage stochastic programming model to optimize the decision-maker's procurement strategy, effectively modeling uncertainty. The first stage of the model aims to determine the quantity to be procured in advance, while the second stage aims to decide on the replenishment procurement quantity. Different prediction models have their own advantages, and a single model struggles to achieve optimal performance across all data. This invention combines multiple demand prediction models. Experimental results show that this method outperforms existing single models in terms of saving total cost across both stages. It overcomes the shortcomings of single models in accurately predicting and consistently achieving good cost savings, accurately predicting the procurement demand for perishable goods, and simultaneously predicting and optimizing the procurement of a "basket of goods," demonstrating good practicality and feasibility. Attached Figure Description

[0041] Figure 1 This is a schematic diagram showing the distribution of customer order quantities received during different time periods of the day in the prediction model-based method for optimizing the procurement volume of perishable goods in the background technology of this invention.

[0042] Figure 2a This is a diagram illustrating the demand for Grade A radish example SKU in the background technology of this invention over the past three months;

[0043] Figure 2b This is a schematic diagram illustrating the demand for Grade B tomatoes (SKU) in the background art example of this invention over the past three months.

[0044] Figure 2c This is a diagram illustrating the demand for Grade A broccoli example SKU in the background art of this invention over the past three months;

[0045] Figure 3 This is a flowchart of the method for optimizing the purchase volume of perishable goods based on a prediction model in an embodiment of the present invention;

[0046] Figure 4 This is a schematic diagram comparing the performance of 14 methods under the average daily additional cost index in an embodiment of the present invention. Detailed Implementation

[0047] The embodiments of the present invention will be described in detail below. It should be emphasized that the following description is merely exemplary and not intended to limit the scope and application of the present invention.

[0048] To optimize decision-makers' procurement strategies, this invention provides a two-stage stochastic programming method (TS-SP). The first stage aims to determine the quantity to be procured in advance, while the second stage aims to determine the replenishment quantity to minimize the total cost throughout the decision-making process. However, traditional stochastic models assume that the parameters are unknown but follow some known distribution; for example, they assume demand follows a normal distribution and then use the Sampling Average Approximation (SAA) method to perform a series of samplings of this distribution. However, in the real-world market for perishable goods, demand rarely follows a normal distribution, and it is even difficult to fit a known distribution. Recent research on data-driven methods has also validated this view. Instead of assuming a demand distribution, they predict demand from historical data and explanatory variables. Since demand may not follow any known probability distribution, the traditional Sampling Average Approximation (SAA) method may fail to generate scenarios.

[0049] To address this issue, an intuitive approach is to accurately forecast demand and use the forecast as the purchase quantity, a method known as the "forecast-then-optimize" approach. However, due to the high volatility of demand and the asymmetric holding and stockout costs inherent in the classic newsboy model in inventory management, a single model struggles to accurately predict demand and consistently deliver good cost-saving performance.

[0050] See Figure 3 This invention proposes a method for optimizing the purchase quantity of perishable goods based on a predictive model, comprising the following steps:

[0051] The first step is to obtain historical demand data and relevant characteristics of the product to be predicted;

[0052] Relevant features include time features and price features. Time features include the month and week information corresponding to the daily sales volume. For example, to predict the demand for a certain product on June 1, 2021, the month information can be extracted as June and the week information as Tuesday.

[0053] The second step is to select k demand forecasting models, forecast the demand for each model, and obtain the predicted demand amount. The results are used as k random scenarios.

[0054] The third step is to substitute the random scenario into a two-stage stochastic programming model to solve the problem and obtain the optimal pre-purchase quantity for each commodity.

[0055] The fourth step is to make supplementary purchases for the demand that was not met by pre-procurement once the actual demand is known.

[0056] This invention also proposes a novel scenario generation method based on a prediction model, which can handle any type of demand uncertainty and combines multiple trained demand prediction models.

[0057] Extensive numerical experiments on real-world datasets demonstrate that the TS-SP method proposed in this invention outperforms 12 other "predict-then-optimize" methods in terms of overall purchasing cost and stability. Furthermore, the advantages of the TS-SP method increase with the number of SKUs (smallest stock units) involved and the number of prediction models considered.

[0058] Example

[0059] The following describes the solution to the two-stage procurement problem of perishable goods, including the first-stage decision (advance procurement) and the second-stage procurement decision (replenishment procurement).

[0060] Phase 1: Advance Procurement

[0061] Assuming the upstream and downstream partners cooperate through contracts, all downstream demands need to be met promptly. This represents the collection of all perishable products. All customers' views on the products... The total actual demand is d i Ideally, the best purchasing strategy is to wait until all orders are received before making another purchase, ensuring the purchased quantity is exactly equal to d. iHowever, in practice, peak order periods typically occur at night. Since the sorting process is time-consuming, it's impossible to deliver orders on time if procurement only occurs at the end of the customer's order period, which is not permitted in the contract. Therefore, for each product i, the buyer must first determine the price c for each perishable product based on demand forecasts. 1i Make pre-orders, with a total pre-order quantity of no less than μ for all perishable products, to ensure that workers in the sorting process have enough products to sort.

[0062] Phase Two: Replenishment Procurement

[0063] Pre-purchasing is unlikely to meet all demand. Therefore, after receiving all orders, the distributor needs to purchase an additional y for each product i. i To make up for the shortage, so that x i +y i =d i However, due to higher costs and reduced supply during nighttime transportation, the purchase price of perishable products during the replenishment phase is usually higher than that during the pre-order phase.

[0064] Total cost

[0065] In summary, the total cost π of the entire procurement process consists of three parts: the total quantity of products procured in the advance procurement phase, c 1i x i The total number of products during the replenishment phase, c 2i y i ; and inventory or loss costs h i (x i -d i ), where h i This represents the holding cost of each excess product i. Therefore, the buyer needs to optimize the purchasing strategy to minimize the total cost.

[0066] Symbol definition

[0067] The definitions and parameter lists of the mathematical symbols used are shown in Table 1:

[0068] Table 1

[0069]

[0070]

[0071] Model Design

[0072] Unlike the traditional "predict first, then optimize" approach, this invention proposes a two-stage stochastic programming (TS-SP) method based on a prediction model. This method consists of two parts: a scenario production method based on the prediction model and the two-stage stochastic programming model.

[0073] Scene generation method based on prediction model

[0074] Sampling average (SAA) is an effective method for generating stochastic problem scenarios. The basic idea of ​​SAA is to approximate the true distribution using an empirical distribution of a random parameter. However, in reality, the demand for perishable foods is affected by historical sales, trends, seasons, weekdays, and unknown random factors, and rarely exhibits a stable statistical distribution.

[0075] In recent years, many data-driven inventory management models do not assume that demand follows a known distribution, but instead predict demand using forecasting models. Different models have their own advantages; for example, complex models can capture long-term dependencies, but they are not always better, especially on small datasets. Conversely, simpler models, such as regression models, may be more stable. Research also shows that hybrid approaches tend to outperform single models.

[0076] This work proposes a scenario generation method based on prediction models. As a novel approach to scenario generation, it does not assume any form of demand distribution, but instead uses the prediction results of k different prediction models as k scenarios to strive to leverage the strengths of various models. In each scenario ξ... s In this context, the prediction model provides its demand forecast.

[0077] For simplicity, we assume that the probability of each scenario is the same, that is... This means that the k prediction models involved have the same predictive power.

[0078] Here's an example to illustrate this scenario generation method: Assume there are only 3 SKUs for demand forecasting, k=3, meaning there are three forecasting models. On the 4th day of the actual test data, the actual demand for these 3 products is [400, 880, 720].

[0079] In this example, the first model's 7-day median demand forecast is [960, 820, 880], the second model's Exponential Smoothing (ETS) forecast is [1044, 900, 1133], and the third model's ANN-MLP (Neural Network-Multilayer Perceptron) forecast is [660, 564, 699]. Therefore, in subsequent stochastic programming, the three unknown parameters d can be used. i The scene.

[0080] -Scene ξ1:

[0081] -Scene ξ2:

[0082] -Scene ξ3:

[0083] In this scenario, assuming

[0084] Two-stage stochastic programming model

[0085] Given k scenarios generated from k prediction models, the mathematical formulation of the proposed two-stage stochastic programming model is as follows:

[0086] Phase 1: Advance Procurement

[0087]

[0088]

[0089]

[0090] Phase Two: Replenishment Procurement

[0091]

[0092]

[0093]

[0094] Equation (1) represents the objective function to be minimized, which consists of two parts: the expected cost of advance procurement in the advance procurement stage and the expected cost of replenishment procurement in the replenishment procurement stage. Where y represents the vector of purchase quantities for each commodity in the supplementary procurement phase. Constraint (2) states that the pre-order quantity i for each product should be non-negative. Constraint (3) specifies that the lower limit of the total purchase quantity in the lead-up phase is the minimum total demand forecast among all options, since the total demand forecast of each forecasting model is usually sufficient for ranking.

[0095] Formula (4) extends the expected cost formula for the replenishment procurement stage. In each scenario, the cost is determined by the different prices c at that stage. 2i The total amount consists of the holding costs of the excess purchase quantity compared to actual demand. Constraint (5) requires that the replenishment purchase quantity for each product i is non-negative. Constraint (6) guarantees that under any circumstances, the customer's actual demand for all products can be met through two purchase phases.

[0096] Experimental Example

[0097] The effectiveness of the proposed method was evaluated using a real dataset from a Chinese fresh food delivery company that procured dozens of products to serve local restaurants. The training set data spanned from December 1, 2020 to November 4, 2021, and the test set data spanned from November 4, 2021 to December 5, 2021 (30 days in total).

[0098] Demand forecasting models and forecast accuracy

[0099] Ten different forecasting models were trained and applied to predict demand for the next 30 days on a daily basis. To predict demand in as many ways as possible, these ten models were selected from historical methods, regression models, decision tree-based models, and neural networks.

[0100] Model (1): 7-day Median

[0101] Model (2): 7-day Mean

[0102] Model (3): Exponential Smoothing (ETS)

[0103] Model (4): Support Vector Regressor (SVR)

[0104] Model (5): Ridge Regression (LASSO Regreesion)

[0105] Model (6): Autoregressive Model with Differential Integrated Moving Average (ARIMA)

[0106] Model (7): Extreme Gradient Boosting Algorithm (XGBoost)

[0107] Model (8): Efficient Gradient Boosting Algorithm (LightGBM)

[0108] Model (9): Artificial Neural Network-Multilayer Perceptron (ANN-MLP)

[0109] Model (10): Long Short-Term Memory Network (LSTM)

[0110] Table 2 shows the prediction accuracy results of different prediction models under different indicators. The best-performing results for each indicator are bolded and underlined. Using the training set data, the optimal parameters for each model were obtained, and then predictions were made daily on the test set. Table 2 shows the performance of these 10 models on the test set, with SKU numbers of 10, 20, 30, 40, and 50 respectively. We used three indicators to measure accuracy: Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Mean Percentage Error (MAPE). Due to the high instability of demand, the two models that performed best in terms of prediction accuracy were the 7-day median and ANN-MLP.

[0111] Table 2

[0112]

[0113] Continued from Table 2

[0114]

[0115] Scene generation method based on prediction model

[0116] We used 14 methods to make purchasing decisions and evaluated cost-related metrics on real-world test data.

[0117] • Methods (1-10): Ten prediction models were applied using the traditional “predict first, then optimize” approach.

[0118] • Method (11) - AvgPred (10 models): which is an abbreviation for AveragePrediction. The arithmetic mean of the demand forecast results of all 10 models is used as the final forecast result, and then the forecast-optimization method is applied.

[0119] • Method (12) - AvgPred (2 models): The arithmetic mean of the demand forecasts of the two models that perform best in terms of prediction accuracy;

[0120] • Method (13) - TS-SP (2 scenarios): The 7-day median and the demand forecast results of ANN-MLP are used as two scenarios (k=2), and then the TS-SP method based on the forecast model is solved;

[0121] • Method (14) - TS-SP (10 scenarios): The demand prediction results of all 10 models are used as 10 scenarios (k=10), and the TS-SP method based on the prediction model is solved.

[0122] The cost-related evaluation metrics are as follows:

[0123] We define the following three metrics to evaluate the total cost performance of these 14 methods on real-world data.

[0124] (1) Average Additional Cost

[0125] This metric measures the total cost incurred by price increases during the replenishment phase, and the holding or waste costs resulting from over-purchasing during the lead-up phase. Assume c 2i =1.2c 1i and h i =0.4c 1i ,So:

[0126]

[0127] (2) Number of winning days

[0128] This metric records the number of days each method achieved the lowest additional cost among all methods within a 30-day testing period. A higher metric indicates a better method.

[0129] (3) Average ranking

[0130] This metric is the average daily performance ranking of each model over a 30-day testing period. The lower this metric, the better the method.

[0131] Experimental results

[0132] Table 3 shows the performance comparison of different methods on the test set, with the best-performing results for each metric highlighted in bold and underlined. Table 3 lists the evaluation results of these 14 methods on the test set. For different numbers of SKUs, the table reports the number of winning days, average ranking, average additional cost, and Diff / % (relative difference percentage), which represents the cost savings of the TS-SP (10 scenarios) proposed in this embodiment compared to other methods.

[0133] Table 3

[0134]

[0135] Continued from Table 3

[0136]

[0137] As can be seen from the results table, for any number of SKUs, the TS-SP (10 scenarios) method proposed in this embodiment of the invention consistently outperforms other methods in terms of average additional cost and average ranking metrics. In other words, the method proposed in this embodiment of the invention has the lowest average cost and the highest average ranking.

[0138] As the number of SKUs increases, i.e. on a larger dataset, the method proposed in this embodiment of the invention performs better and more stably. For example, when SKU=10, it is optimal for only 7 out of 30 days on the test set, with an average ranking of 3.5, while when SKU=50, it outperforms other methods over 12 days, with an average ranking of 1.9. Figure 4Table 3 details the average daily additional cost of these 14 methods under different SKU numbers. Compared with the other 12 prediction-then-optimization methods, when SKU=10, the TS-SP method (10 scenarios) saves an average additional cost of 14.6% to 65.9% (average 40.2%); when SKU=20, it saves 20.3% to 63.2% (average 41.5%); when SKU=20, it saves 19.8% to 66.2% (average 45.2%); when SKU=40, it saves 26.8% to 75.0% (average 55.9%); and when SKU=50, it saves 9.9% to 73.5% (average 47.6%). This demonstrates that the prediction model-based method for optimizing the purchase quantity of perishable goods proposed in this embodiment of the invention significantly reduces costs.

[0139] When comparing AvgPred (10 models) and TS-SP (10 scenarios), it can be found that although they both utilize a combination of 10 prediction models, the two-stage stochastic optimization model for demand forecasting based on different scenarios proposed in this embodiment of the invention significantly outperforms simple average prediction results. The underlying implication is that optimizing procurement decisions based on various practical factors such as price, stockout costs, and holding costs is crucial, while making decisions solely based on averaging prediction results is not advisable.

[0140] In addition, increasing the number of scenarios can improve performance. Both TS-SP (2 scenarios) and TS-SP (10 scenarios) use two-stage stochastic programming to optimize decisions, but the latter is significantly better across all metrics. The reason behind this is that even two models with strong predictive power struggle to cover enough cases where demand is uncertain, while more stochastic scenarios allow the model to optimize under more possibilities.

[0141] The problem solved by this embodiment is as follows:

[0142] The demand for perishable goods, such as fresh food, newspapers and magazines, fast fashion apparel, and airline and ship tickets, is highly uncertain. Factors such as seasons, weather, weekends, and holidays can cause significant fluctuations in demand. This invention considers these factors and combines multiple machine learning models for demand forecasting. Different forecasting models have their own advantages, and a single model is unlikely to achieve optimal results across all data. This invention combines multiple demand forecasting models, and experimental results show that it outperforms existing single models in terms of saving total two-stage costs. The perishable goods sector involves a wide variety of products, and procurement involves optimizing multiple product purchases. This invention can simultaneously forecast and optimize the procurement of a "basket of goods," demonstrating good practicality and feasibility. Addressing a new problem in real life, the perishable goods market cannot tolerate stockouts, presenting a unique two-stage procurement problem. This invention proposes a data-driven method based on a forecasting model, using a two-stage stochastic programming model from operations research to model this two-stage procurement problem. In supply chain inventory management, stockout costs and replenishment costs are often different, and inaccurate procurement can lead to significant economic losses for distributors. This method not only forecasts demand but also optimizes procurement decisions based on different cost scenarios, minimizing total procurement costs.

[0143] The advantages of this embodiment are as follows:

[0144] This invention relates to the supply chain field, specifically addressing the characteristics of perishable goods (such as fast-moving consumer goods and fresh produce) where demand fluctuates significantly and multiple purchases may be necessary to meet all needs. It proposes a data-driven "demand forecasting + cost optimization" method, a novel approach combining machine learning and operations research. Based on a real-world problem scenario, this invention innovatively models the problem as a two-stage stochastic programming model (TS-SP) to optimize the decision-maker's purchasing strategy, effectively modeling uncertainty. The first stage of the model aims to determine the quantity to be purchased in advance, while the second stage aims to decide on the replenishment purchase quantity. Traditional stochastic programming models typically require estimation of the probability distribution of parameters using the sampled average approximation (SAA) method. However, this method requires parameters to follow a known probability distribution, but in reality, the demand for goods is often not consistently distributed and exhibits temporal dependencies, and may fluctuate drastically due to external factors such as seasons and weather. This invention proposes a novel scenario generation method based on a prediction model, combining multiple pre-trained demand forecasting models to handle decision-making problems under any type of demand uncertainty. Extensive numerical experiments on real-world datasets demonstrate that the TS-SP method proposed in this invention outperforms 12 other "predict-then-optimize" methods in terms of overall purchase cost and stability. The advantages of using the method proposed in this invention further increase as the number of product SKUs involved and the number of considered options grows.

[0145] Comparative example

[0146] The Monte Carlo method using the Sampling Average Approximation (SAA) is employed to generate random scenarios: random sampling is performed on historical demand, and the resulting possible values ​​are used to simulate uncertainty. However, this sampling strategy assumes that the daily demand distribution for each type of blood is identical and that demand on different days is independent and does not change over time. This assumption does not hold true in the supply chain of perishable goods, or in other areas where demand is time-dependent and fluctuates significantly over time.

[0147] In contrast, this invention proposes a "scenario generation method based on prediction models." This method utilizes multiple prediction models to forecast unknown daily demand from different perspectives, treating the prediction result of each model as a possible scenario, thus eliminating the need to assume that demand follows any distribution. Furthermore, the method of this invention allows for time-series dependencies in demand, and the prediction models partially consider the impact of historical demand trends, seasons, holidays, and commodity prices on demand, making it a more data-driven approach.

[0148] The differences between the embodiments of the present invention and the prior art are as follows:

[0149] Combining advanced machine learning and operations research techniques, this invention models and solves the two-stage procurement problem in the perishable goods market. No known method currently exists to solve this problem. The method in this invention is a fully data-driven approach, utilizing historical demand data for scientific prediction. It does not require assumptions about demand following a specific distribution. Unlike the traditional "predict first, then optimize" method, a two-stage stochastic programming (TS-SP) method based on a prediction model is proposed, which can minimize costs for different stockout and replenishment costs. Compared to the "predict first, then optimize" method, this model can save approximately 40% of the total cost. The method proposed in this invention is highly scalable; numerical experimental results show that the cost-saving advantage of this method gradually increases as the number of procured goods and the number of prediction models considered increases.

[0150] The field used in this embodiment is as follows:

[0151] It can be used in any field where purchasing or production decisions need to be made under uncertain demand.

[0152] 1. In the field of perishable goods, such as fresh produce wholesalers in various regions, this method can be used to minimize total procurement costs;

[0153] 2. In the fast fashion sector, such as clothing brands, the method described in this invention can guide the production quantity of new apparel products;

[0154] 3. In the newspaper and magazine industry, the print run is determined when reader demand is uncertain;

[0155] 4. In the high-end manufacturing sector, the pre-production quantity of equipment will be determined given the uncertainty of demand from partner manufacturers in the next quarter;

[0156] In the agricultural sector, this method can be used to predict the demand and prices of various crops in the following year, thereby guiding the input of crops to be planted in the following year and improving returns.

[0157] The application scenarios for this embodiment are as follows:

[0158] It can be applied to various procurement optimization problems involving multiple stages of procurement, saving total procurement costs.

[0159] For example, fresh food is a perishable commodity, and its supply is a core issue of people's livelihood in any country in the world, involving a massive scale. Currently, the Chinese fresh food market has reached 5 trillion RMB, but despite its size, the distribution channels are still relatively traditional agricultural markets, relying mainly on purchasing staff to estimate order volumes based on their experience. This results in significant losses, approaching 20%. Numerical experiments show that this method can save more than 20% of costs compared to traditional methods, with a potential cost saving of approximately 20 million RMB per year.

[0160] In the fast fashion sector, the current scale of China's fast fashion industry has exceeded 250 billion yuan. Fast fashion apparel is also a sector characterized by strong seasonality, short fashion cycles, and highly uncertain demand. Apparel brands can use the methods of this invention to guide the production quantity of new apparel. If demand forecasting is performed using models and production decisions are guided by the models of this invention, inventory costs caused by overproduction and order losses due to underproduction can be reduced.

[0161] The same applies to other industries; specific profit forecasts mainly depend on the effectiveness of the original decisions made in that specific industry.

[0162] In the perishable goods sector, this approach can be applied to fresh produce wholesalers across various regions, helping companies make more effective purchasing decisions.

[0163] This can be used by wholesalers or distribution companies of perishable goods to forecast demand and develop cost-minimizing inventory strategies for these items. It is estimated to save an average of 10% in total additional costs (i.e., costs beyond the purchase of the goods themselves, primarily inventory costs and replenishment markups).

[0164] If patent fees are charged to companies that use this method as a source of profit, substantial revenue can be generated.

[0165] The plan and outlook are as follows:

[0166] 1. In the future, it can be applied to optimize procurement decisions in the circulation of perishable goods, reduce losses in the circulation process of fast fashion apparel manufacturing, fresh food and other commodities, and save costs throughout the entire supply chain.

[0167] 2. By leveraging the consulting services and product recommendations provided by industry SaaS software service companies, it is possible to quickly capture market share in the perishable goods delivery industry.

[0168] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0169] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0170] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0171] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0172] The above description provides a further detailed explanation of the present invention in conjunction with specific / preferred embodiments, and it should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various substitutions or modifications can be made to these described embodiments without departing from the concept of the present invention, and all such substitutions or modifications should be considered within the scope of protection of the present invention. In the description of this specification, the reference to terms such as "an embodiment," "some embodiments," "preferred embodiment," "example," "specific example," or "some examples," etc., indicates that the specific features, structures, materials, or characteristics described in connection with that embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in any suitable manner in one or more embodiments or examples. Without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification and the features of different embodiments or examples. Although the embodiments of the present invention and their advantages have been described in detail, it should be understood that various changes, substitutions, and modifications can be made herein without departing from the scope of protection of the patent application.

Claims

1. A method for optimizing the purchase quantity of perishable goods based on a predictive model, characterized in that, Includes the following steps: S1: Obtain historical demand data and relevant characteristics of the product to be predicted; S2: Select k demand forecasting models and forecast the demand respectively to obtain the predicted demand amount. The results are used as k random scenarios. No demand distribution is assumed. Instead, the prediction results of k different forecasting models are used as k scenarios, and each scenario has the same probability. S3: Substitute the random scenario into a two-stage stochastic programming model to solve the problem and obtain the optimal pre-purchase quantity for each commodity; the objective function of the two-stage stochastic programming model consists of three parts: advance purchase cost, replenishment purchase cost, and inventory or loss cost, so as to minimize the total cost; S4: Once the actual demand is known, calculate the supplementary purchase quantity based on the actual demand; In the two-stage stochastic programming model, the first stage determines the quantity to be purchased in advance, and the second stage determines the quantity to be purchased for replenishment.

2. The method as described in claim 1, characterized in that, In step S3, the objective function to be minimized in the first stage of the two-stage stochastic programming model consists of two parts: the expected cost of advance procurement in the first stage and the expected cost of replenishment procurement in the second stage, expressed by the following formula: ， in, Refers to products The number of pre-orders, This refers to the total cost incurred in fulfilling all orders within a day. Refers to the collection of all perishable products. Perishable products purchased in advance The price This refers to the expected costs during the replenishment procurement phase, among which The meaning is a vector composed of the purchase quantities of each commodity in the supplementary procurement stage.

3. The method as described in claim 2, characterized in that, The expected cost of replenishment procurement in the second phase is expressed by the following formula: ， in, Refers to products The number of pre-orders, Refers to the collection of all perishable products. Products in replenishment procurement The price This refers to a possible scenario for all product demands. Products during replenishment Purchase volume, Products on the same day Predicted demand in the scenario The following products Forecast demand, It refers to the set of all randomly generated scenes.

4. The method as described in claim 2, characterized in that, The constraints of the objective function include: ， in, Refers to products The number of pre-orders, It refers to the collection of all perishable products.

5. The method as described in claim 2, characterized in that, The constraints of the objective function include: ， in, Refers to products The number of pre-orders, Refers to the collection of all perishable products. This refers to the predicted demand for product i on that day in the scenario. The following products Forecast demand.

6. The method as described in claim 2, characterized in that, The constraints on the expected cost of the replenishment procurement include: ， in, Products during replenishment Purchase volume, Refers to the collection of all perishable products. This refers to a possible scenario for all product demands. It refers to the set of all randomly generated scenes.

7. The method as described in claim 5, characterized in that, The constraints on the expected cost of the replenishment procurement include: ， in, Products during replenishment Purchase volume, Refers to the collection of all perishable products. Refers to products The number of pre-orders, This refers to a possible scenario for all product demands. Products on the same day Predicted demand in the scenario The following products Forecast demand, It refers to the set of all randomly generated scenes.

8. The method as described in claim 1, characterized in that, The prediction model is a pre-trained demand prediction model.

9. The method as described in claim 1, characterized in that, The prediction models include a 7-day median model and an ANN-MLP model.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program can implement the method described in any one of claims 1-9.

Images