A data-driven automatic replenishment method for oil products at gas stations

Abstract

The present invention belongs to the technical field of automatic replenishment scheduling, and provides a data-driven automatic replenishment method for oil products at gas stations. The method realizes data-driven automatic replenishment based on an automatic oil product replenishment system, and can take into account real-time changes in replenishment scenarios, both the real-time demand status of the consumer end and the real-time supply status of the supply end. Dynamic early warning analysis is performed based on monitored real-time data to trigger the generation of a replenishment plan in real time. When replenishment is triggered by an early warning, an oil product replenishment scheduling plan suitable for the current system state is generated quickly, efficiently and intelligently based on real-time data, so as to realize automatic replenishment of oil products at gas stations driven completely by data, which is conducive to on-demand, accurate, active and timely replenishment of oil products, and helps enterprises to cope with frequent fluctuations in oil product demand at a lower cost on the premise of ensuring timely and stable supply of oil products, so as to simplify and intelligentize the management of oil product replenishment.

Description

Technical Field

[0001] The invention relates to a data-driven automatic oil replenishment method for a gas station, belonging to the technical field of automatic replenishment scheduling. Background Art

[0002] Finished oil is a strategic resource related to the national economy and people's livelihood. Its timely, stable and sufficient supply is the key to ensuring the smooth operation of social production and people's lives. However, the inventory consumption status of gas stations driven by uncertain demand fluctuates frequently, and it is difficult to accurately predict the out-of-stock time of gas stations, which brings difficulties to the replenishment management of oil products. Data-driven automatic replenishment of oil products at gas stations can capture the state of replenishment in time and formulate replenishment scheduling plans in real time based on the real-time replenishment data of monitored gas stations, automatically respond to dynamic demand changes, and realize refined replenishment management. At the same time, the gross profit of finished oil retail enterprises has continued to decline in recent years, and the competition in the finished oil retail market has intensified. There is an urgent need to reduce costs, increase efficiency and increase competitiveness.

[0003] At present, the replenishment of oil products at gas stations is usually done by the staff of the dispatch center predicting the demand trend of oil products in the next working day at a fixed time, and judging whether the inventory level of the gas station will be lower than the inventory threshold based on the demand trend to drive replenishment. Automatic replenishment patents in other fields mainly focus on improving the accuracy of replenishment quantity by improving prediction accuracy, improving the accuracy of commodity inventory values, and dynamically calculating replenishment quantities. The patent "A smart replenishment system" with patent number CN 111626515 A of the State Intellectual Property Office of the People's Republic of China constructs, adjusts and optimizes the demand forecasting plan to improve the accuracy of the forecast, reasonably reduce inventory turnover, and avoid out-of-stock phenomena. The patent "Automatic inventory replenishment system based on daily sales coefficient and its use method" with patent number CN 111125140 A of the State Intellectual Property Office of the People's Republic of China is committed to improving the accuracy of commodity inventory values ​​and providing more accurate replenishment quantities to staff. The patent number CN108985691А issued by the State Intellectual Property Office of the People's Republic of China, "An automatic replenishment method and system based on dynamic inventory control", can dynamically calculate the safety stock, maximum stock and minimum stock of each material variety, and calculate the replenishment quantity on this basis, ensuring that the inventory quantity is the best balance between inventory and out-of-stock.

[0004] At present, the replenishment process of oil products at gas stations has the following shortcomings: the replenishment of oil products at gas stations is currently an extensive, threshold-driven static replenishment, which is prone to mismatch between supply and demand. For example, if the inventory of oil products at gas stations is too low, if customer demand suddenly increases in a short period of time, for example, gas stations around construction sites, gas stations next to highways during holidays, and gas stations in the city during the taxi shift period every day are facing a sharp increase in demand, which will cause the inventory to be unable to meet customer demand, resulting in oil shortage accidents, which will seriously affect production and life order, and even trigger a serious shortage chain reaction; if the inventory of oil products at gas stations is too high, if the customer demand trend is far lower than the historical demand, it will lead to excessive inventory backlogs, and the huge volume of nearly 120,000 gas stations in my country will generate high inventory holding costs, and will also increase the potential risk factor. In summary, under the traditional static replenishment method driven by thresholds, the misjudgment rate of replenishment triggering time is high and the accuracy is low, with the defects of high safety inventory, high holding costs, and poor decision-making scientificity.

[0005] The present invention considers the real-time changes in replenishment scenarios, both the real-time demand status of the consumer side and the real-time supply status of the supply side, and proposes a real-time data-driven automatic replenishment method for oil products at gas stations. The present invention can drive the online intelligent generation of replenishment scheduling plans based on the monitored real-time oil consumption data and real-time inventory data, and realize on-demand, accurate, active, and timely replenishment of oil products, helping enterprises to cope with frequent fluctuations in oil demand at a lower cost while ensuring timely and stable supply of oil products. Summary of the invention

[0006] The purpose of the present invention is to provide a data-driven automatic replenishment method for oil products at gas stations. It performs dynamic early warning analysis based on real-time monitoring data of all links and all time periods of the gas station system, analyzes the key status of the gas station system, predicts the occurrence of oil outages in advance to automatically trigger replenishment, and when the early warning triggers replenishment, quickly, efficiently and intelligently generates an oil product replenishment scheduling plan suitable for the current system status based on real-time data, realizes the online intelligent generation of data-driven replenishment scheduling plans, ensures on-demand, accurate, active and timely replenishment of oil products, and can cope with dynamically changing replenishment scenarios at a lower cost.

[0007] Technical solution of the present invention

[0008] The data-driven automatic replenishment system for oil products at gas stations includes a prediction model training module, a value function training module, a data collection module, a real-time prediction module for oil product sales, a calculation module 1, a calculation module 2, a calculation module 3, a calculation module 4, a warning level judgment module, a replenishment vehicle decision module, and a replenishment quantity decision module. The prediction model training module uses the deep neural network prediction technology of big data to train prediction models with different prediction durations based on historical sales data; the value function training module uses the approximate dynamic programming algorithm to train the value function of the inventory level based on historical sales data and simulation data; the data collection module can collect the current time of the system, the remaining replenishment time of the station, the inventory level and real-time sales of oil products at the station, the location of the in-transit tanker, the number and replenishment quantity of the in-transit tanker station that has not been delivered, and the model and number of tankers specially used for rescue replenishment at the central warehouse; the real-time prediction module for oil product sales predicts the oil product sales value in the future period in real time; the calculation module 1 calculates whether the station definitely does not need replenishment before the next round of replenishment begins; the calculation module 2 determines whether the current inventory level of the station is higher than the predicted total sales of oil products within the remaining replenishment lead time Calculation module 3 determines whether the current time of the system reaches the beginning of the next period; calculation module 4 determines whether the current inventory level is greater than The difference between the accumulated oil inventory sold and the accumulated oil inventory sold; the warning level judgment module presents the oil inventory warning level of the station; the replenishment vehicle decision module selects the available replenishment vehicles; the replenishment quantity decision module determines for the warning station when and by which vehicle how much oil to transfer.

[0009] The data-driven automatic replenishment method for gas station oil products has the following steps:

[0010] Step 1: Offline training of sales forecast model and inventory level value function

[0011] In order to ensure that the predicted sales value is neither too low nor too high, a sales forecasting model considering the sales pattern category of each period is proposed. Firstly, an indicator that can measure the sales size and volatility of each period is proposed, and the sales pattern category of each period is fuzzily divided based on the indicator value; then, based on the sales pattern category of each period, a point prediction model based on the Long Short-Term Memory (LSTM) network or a quantile regression model based on the Quantile Long Short-Term Memory (QLSTM) network is selected to capture the long-term and short-term dependencies in the sales time series; finally, the sales forecast value based on the sales pattern category of each period is obtained.

[0012] In order to ensure that the replenishment scheduling decision can take into account the future random demand, a multi-stage random dynamic decision-making model with the goal of maximizing the overall profit is constructed. With the help of the concave adaptive value estimation algorithm (an approximate dynamic programming algorithm), simulation is carried out based on sales data, and it is continuously iterated until a convergent value function is determined to represent the future value of the inventory level at a certain moment.

[0013] Step 2: The oil product automatic replenishment system monitors the current time t and the remaining replenishment time ΔT of the station in real time t , the inventory level and real-time sales volume of oil products at the stations, the location of the in-transit tank trucks, the numbers and replenishment quantities of the stations that have not been delivered by the in-transit tank trucks, and the models and quantity of tank trucks specifically used for rescue replenishment at the central warehouse.

[0014] Step 3: Based on the real-time oil product sales and offline trained sales forecasting model, real-time forecast of the remaining replenishment lead time ΔT t Oil sales in

[0015] Step 4: Perform dynamic early warning of oil inventory based on real-time monitoring data. Divide a work cycle into N time periods, each of which is Δt long, and perform early warning analysis on a rolling basis: First, at the beginning of each time period, perform early warning analysis based on the predicted value of oil sales; then, within the time period, perform early warning analysis based on real demand events and the predicted value of oil sales at the beginning of the time period. The specific steps are as follows.

[0016] Step 4.1: Inventory warning analysis at the beginning of the period, first proceed to step 4.1.1.

[0017] Step 4.1.1: Based on the maximum inventory consumption rate v max Determine whether the site will definitely not be out of stock: If the inventory level is I t The site has been maintained max , ΔT can still be satisfied t The demand within, that is, I t ≥ΔT t *v max , it means that the site definitely does not need to replenish before the next round of replenishment begins, and automatic replenishment ends; otherwise, go to step 4.1.2.

[0018] Step 4.1.2: Based on the predictions from step 3 Judgment I t ≤D t,ΔT Whether it occurs. If it occurs, it means that the current moment is the replenishment triggering moment, triggering the warning, and proceeding to step 5; otherwise, proceeding to step 4.2.

[0019] Step 4.2: Inventory warning analysis within the time period, go to step 4.2.1.

[0020] Step 4.2.1: Determine whether the current time has reached the beginning of the next time period. If so, return to step 2; otherwise, proceed to step 4.2.2;

[0021] Step 4.2.2: Real-time monitoring of whether there is real demand during the period. If real demand occurs at a certain time tt, let the real inventory level at this time be I tt , and calculate the cumulative oil inventory TD sold from the beginning of the period to time tt [t,tt] If TD [t,tt] ≥Forecast value of oil sales volume in this period D t,Δt , then further judge Whether it occurs, to determine whether time tt is the replenishment triggering time. If it occurs, it means that the current time is the replenishment time, triggering the warning and entering step 5; otherwise, return to step 4.2.1.

[0022] The technical key points of step 4 are: based on the early warning analysis driven by both time and real demand events, comprehensively and in real time analyze the dynamic changes of replenishment scenario factors such as inventory levels, remaining replenishment lead time, and future consumption trends of oil products, analyze the key states that can trigger early warnings, fully consider the impact of real-time scenario states on replenishment time, and improve the scenario adaptability of replenishment management.

[0023] Step 5: Determine the warning level. According to the resources currently available for rescue, the current distribution plan, the degree of urgency of inventory, and the difficulty of subsequent emergency measures, the warning levels are divided into Level 1 (high), Level 2 (low), Level 3 (medium), and Level 4 (low). Level 1 warning requires the dispatch decision maker to find a way to expand the available resources, which mainly includes two situations: (1) the distribution system currently has no redundant resources available for rescue; (2) although there are vehicles on the road that deliver this type of oil to other gas stations, these vehicles cannot reach the warning station in time. Level 2 warning requires the dispatch decision maker to have rich experience in re-dispatching or dispatching in an emergency mode, which mainly includes three situations: (1) There are no available vehicles for rescue and replenishment at the distribution center, but there are vehicles on the road that deliver this type of oil to other gas stations, and some of them can arrive at the warning station in time for rescue; (2) There are vehicles that can be used for rescue and replenishment, but if the previous dispatch method is used, it will not be able to arrive at the warning station in time. The dispatcher needs to dispatch vehicles in an emergency mode according to the remaining time (for example, in order to save fuel filling time, a small amount of fuel can be loaded first instead of filling the car all at once); (3) There are vehicles that meet both situations (1) and (2), but the vehicle that meets situation (2) needs to be called in an emergency mode. Therefore, in situation (3), the vehicle that meets situation (1) is preferred. Level 3 warning means that there are only vehicles in the system that can arrive at the warning station in time for rescue and replenishment. The dispatch decision maker only needs to dispatch emergency rescue vehicles from the distribution center in the conventional way. Level 4 warning means that the rescue resources available in the system are abundant, and there are vehicles dedicated to rescue and replenishment and vehicles for other distribution tasks that can arrive at the warning site in time. The scheduling decision maker can decide what method to use for emergency rescue based on the actual situation and the difficulty of scheduling. If the site meets the situation (1) and situation (3) in the level 2 warning and the level 4 warning, go to step 6 to generate a scientific replenishment scheduling plan in real time.

[0024] Step 6: Generate replenishment scheduling plan in real time. After the early warning triggers replenishment, take the following steps to generate a replenishment scheduling plan in real time to effectively intervene in or reschedule the current distribution plan.

[0025] Step 6.1: Based on the warning level and the real-time predicted oil sales volume, select available replenishment vehicles. If the site meets the conditions (1) and (3) in the Level 2 warning, then after selecting the rescue warning site, there will still be no vehicles in transit at the site originally planned for delivery. If the warning level of the site is Level 4, after selecting the rescue warning site, there will still be no vehicles in transit at the site originally planned for delivery, and there will be idle vehicles at the central warehouse dedicated to rescue replenishment, and then proceed to Step 6.2. The steps for selecting the vehicles in transit are as follows: Let the current time be t, and the replenishment cycle end time be t e , the replenishment quantity of station i is C i The time it takes for the in-transit tanker m, which is originally scheduled to deliver oil to station i, to arrive at the warning station j from the current location is Δt(V m,t ,j), the time from the current location to site j and then back to site i is Δt(V m,t ,j)+Δt(i,j), the current inventory levels of site i and site j are I i,t ,I j,t . if and and That is, if the originally planned delivery station i will not be out of stock, and station j will not be out of stock before the rescue vehicle arrives, then the on-the-way vehicle m can rescue the warning station; otherwise, vehicle m is unavailable.

[0026] Step 6.2: Based on the real-time predicted oil sales and the value function trained in the offline stage that can represent the future value of inventory levels, an operations research model with the goal of maximizing overall revenue is constructed, and the Gurobi solver is used to decide when and by which vehicle to transfer how much oil.

[0027] The technical key points of step 6 are: based on the real-time replenishment status and sales forecast value, an online decision rule for replenishment vehicles that takes into account the impact of out-of-stock is proposed to expand the offline approximate dynamic programming algorithm, to ensure that stations affected by transshipment do not run out of stock, and to adopt a replenishment scheduling plan that is conducive to maximizing overall revenue.

[0028] The effects and benefits of the present invention are:

[0029] 1. The present invention proposes a data-driven automatic replenishment system and method for oil products at gas stations, which has the ability to dynamically respond to real-time data. First, dynamic early warning analysis is performed based on the monitored real-time data to trigger the generation of a replenishment scheduling plan in real time; and when the early warning triggers replenishment, an oil product replenishment scheduling plan suitable for the current system status is generated quickly, efficiently and intelligently based on the real-time data, thereby realizing fully data-driven automatic replenishment of oil products at gas stations, which is conducive to on-demand, accurate, proactive and timely replenishment of oil products, and helps enterprises to cope with frequent fluctuations in oil product demand at a lower cost while ensuring timely and stable supply of oil products, thereby simplifying and intelligentizing oil product replenishment management.

[0030] 2. The dynamic early warning analysis stage of the automatic replenishment system and method combines the deep neural network prediction technology based on big data to analyze the key status of the replenishment system, predict the occurrence of oil outage in advance to automatically trigger the generation of the next replenishment scheduling plan. In this way, the present invention can cope with the dynamically changing demand environment, comprehensively analyze the replenishment scenario to avoid potential out-of-stock risks, and ensure the real-time response of replenishment, which has the advantages of strong pertinence and real-time performance.

[0031] 3. In the generation stage of the replenishment scheduling plan, the automatic replenishment system and method propose online decision rules for replenishment vehicles that take into account the impact of out-of-stock, expand the approximate dynamic programming solution algorithm, and generate a real-time oil product replenishment scheduling plan that is suitable for the current system state and aims to maximize overall benefits by combining online data monitoring and offline data training results. In this way, the present invention can generate a replenishment scheduling plan that ensures that there is no out-of-stock and takes into account future random demand in real time based on the real-time system state, which can improve the real-time, scientific and intelligent nature of the replenishment scheduling plan.

[0032] 4. The present invention can provide decision support and reference for the automatic replenishment of oil products at gas stations, and has certain inspirational significance for data-driven dynamic decision-making problems.

[0033] 5. The present invention can be applied to automatic replenishment in fields where replenishment-related factors are diverse, uncertain, and dynamically changing. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a flowchart of the data-driven automatic replenishment process of oil products at a gas station according to the present invention.

[0035] Figure 2 It is a schematic diagram of the structural mode of the data-driven automatic oil replenishment system for gas stations of the present invention and the connection and transmission relationship between the internal modules.

[0036] Figure 3It is a flowchart of the oil product inventory dynamic early warning process of the present invention.

[0037] Figure 4 It is a flowchart of the replenishment scheduling plan generation process of the present invention. DETAILED DESCRIPTION

[0038] The specific implementation of the present invention is further described below in conjunction with the accompanying drawings and technical solutions.

[0039] like Figure 1 As shown, the data-driven automatic replenishment system for gas stations of the present invention is composed of a prediction model training module, a value function training module, a data acquisition module, a real-time prediction module for oil sales, a calculation module 1, a calculation module 2, a calculation module 3, a calculation module 4, a warning level judgment module, a replenishment vehicle decision module, and a replenishment quantity decision module. The functions of each module include:

[0040] (1) The prediction model training module uses deep neural network prediction technology based on historical sales data to train prediction models with different prediction durations.

[0041] (2) The value function training module constructs a multi-stage stochastic dynamic decision-making model with the goal of maximizing overall revenue. With the help of an approximate dynamic programming algorithm, simulation is performed based on sales data. The model is iterated continuously until a convergent value function is determined to represent the future value of the inventory level at a certain moment.

[0042] (3) Data collection module, which collects data on elements related to the replenishment scenario in real time: current time data, remaining replenishment time data of the station, inventory level data and real-time sales data of the station's oil products, location data of in-transit tankers, number data and replenishment quantity data of the stations that have not been delivered by in-transit tankers, and model data and quantity data of tankers specifically used for rescue replenishment at the central warehouse.

[0043] (4) The oil product sales real-time prediction module predicts the oil product sales value in the future period based on the real-time oil product sales data and the sales prediction model trained offline.

[0044] (5) Calculation module 1 calculates the time when the gas station is within the remaining replenishment lead time ΔT t Always maintain the maximum inventory consumption rate v max Total sales volume ΔT t *v max , compare this value with the site’s inventory level to determine whether the site does not need replenishment before the next round of replenishment begins.

[0045] (6) Calculation module 2, based on the oil product sales real-time prediction module, predicts the remaining replenishment lead time ΔT of the gas station t Oil sales Determine the current inventory level of the site I t Is it higher than the sales volume?

[0046] (7) Calculation module 3 determines whether the current time of the system has reached the beginning of the next time period.

[0047] (8) Calculation module 4 predicts the oil sales volume D of the gas station in this period Δt according to the real-time prediction module of oil sales volume t,Δt , when there is real demand (assuming that the system time is tt and the inventory level is I tt ), calculate the cumulative oil inventory TD sold from the beginning of the period t to time tt [t,tt] . Determine TD [t,tt] ≥D t,Δt Whether it happens, if it happens, further judge Whether it happens.

[0048] (9) The warning level judgment module determines the specific warning level based on the status of the tank trucks used for rescue replenishment and the tank trucks in transit at the system central warehouse obtained by the data collection module and the inventory level of the tank trucks themselves.

[0049] (10) Replenishment vehicle decision module, which selects available replenishment vehicles. The selection of in-transit replenishment vehicles is based on the prediction of the oil sales real-time prediction module. And the current inventory level I of site i and j collected by the data collection module i,t ,I j,t ,judge and Whether they occur simultaneously.

[0050] (11) The replenishment quantity decision module builds an operations research model with the goal of maximizing overall revenue based on the sales volume forecast value of the oil product sales real-time forecast module and the value function of the inventory level output by the value function training module, and uses the solver to decide when and by which vehicle to transport how much oil.

[0051] See also Figure 1 As shown in the figure, the flowchart shows the data-driven automatic replenishment process of gas station oil products. The specific implementation steps are as follows:

[0052] Step 1: Offline training of sales prediction model and value function

[0053] Using the prediction model training module, prediction models with different prediction durations are trained and called by the system's oil sales real-time prediction module; using the value function training module, the converged value function is trained to represent the future value of the inventory level at a certain moment, and is used in the replenishment quantity decision module to generate a replenishment scheduling plan that takes into account future random demand.

[0054] Step 2: Data Collection

[0055] Run the oil product automatic replenishment system and use the data acquisition module of the system to collect the current time data t and the remaining replenishment time data ΔT of the station in real time t , Site oil inventory level data I t , real-time sales data of oil products at the stations, location data of in-transit tankers, number data and replenishment quantity data of stations that have not been delivered by in-transit tankers, model data and quantity data of tankers specially used for rescue replenishment at the central warehouse, and transmit these data to the system's oil product sales real-time prediction module, calculation module 1, calculation module 2, calculation module 3, calculation module 4, warning level judgment module, replenishment vehicle decision module and replenishment quantity decision module.

[0056] Step 3: Real-time forecast of sales volume during the remaining replenishment lead time

[0057] Use the oil sales real-time forecasting module of the oil automatic replenishment system to predict the remaining replenishment lead time ΔT t Sales volume within Mainly based on ΔT t The prediction model trained by the prediction model training module is called according to the length of time.

[0058] Step 4: Oil inventory dynamic warning (see Figure 3 , the flowchart shows the oil inventory dynamic warning process)

[0059] Step 4.1: Use calculation module 1 to determine whether the site definitely does not need replenishment. If the calculation result of calculation module 1 is that the site definitely does not need replenishment before the next round of replenishment begins, then stop the subsequent steps of automatic replenishment; otherwise, go to step 4.2 and enter the calculation module 2 of the system.

[0060] Step 4.2: Use calculation module 2 to determine the current inventory level of the site I t Is it higher than the sales value within the remaining replenishment lead time? If it is higher, go to step 4.3 and enter the calculation module 3 of the system; otherwise, it means that the current moment is the replenishment triggering moment, triggering the warning, and go to step 5 and enter the warning level judgment module of the system.

[0061] Step 4.3: Use calculation module 3 to determine whether the current time has reached the next time period. If not, go to step 4.4 and enter the calculation module 4 of the system; otherwise, go to step 2 and enter the data acquisition module of the system;

[0062] Step 4.4: Use calculation module 4 to determine whether there is real demand. If real demand occurs at a certain time tt, let the real inventory level at this time be I tt , and calculate the cumulative oil inventory TD sold from the beginning of the period t to time tt [t,tt] Once TD [t,tt] ≥ Forecasted oil sales volume for this period D t,Δt , further judge If it happens, it means that the current moment is the replenishment time, triggering the warning, and going to step 5 to enter the system's warning level judgment module; otherwise, returning to step 4.3 and entering the system's calculation module 3.

[0063] Step 5: Use the warning level judgment module to judge the warning level, and then go to step 6 to enter the system's replenishment vehicle decision module. According to the status of the tank trucks dedicated to rescue replenishment and the tank trucks on the way at the system's central warehouse obtained by the data acquisition module and its own inventory level, the specific warning level is determined.

[0064] Step 6: Generate replenishment scheduling plan (see Figure 4 , the flowchart shows the replenishment scheduling plan generation process)

[0065] Step 6.1: Use the replenishment vehicle decision module to determine the optional replenishment vehicles, and then go to step 6.2 to enter the replenishment quantity decision module of the system. The steps for selecting the available in-transit tanker are as follows:

[0066] (1) Use the data acquisition module to collect the current time t of the system and the end time of the cycle t e , the inventory level I at site i i,t , the replenishment quantity of station i is C i , the inventory level I of the warning site j j,t , the time Δt (V m,t ,j), the time Δt (V m,t ,j)+Δt(i,j).

[0067] (2) Use the real-time forecasting module for oil product sales to predict the following sales values:

[0068] (3)Judgment and Are they both true (i.e., the originally planned delivery site i will not be out of stock and site j will not be out of stock before the rescue vehicle arrives)? If they are both true, vehicle m can rescue the warning site j; otherwise, the in-transit tanker m is unavailable.

[0069] Step 6.2: Use the replenishment decision module to decide when and by which vehicle to transfer how much oil. Using the sales forecast output by the oil sales real-time prediction module and the value function of the inventory level output by the value function training module, an operations research model with the goal of maximizing overall revenue is constructed and the solver is used to decide when and by which vehicle to transfer how much oil.

Claims

1. A data-driven automatic oil replenishment method for a gas station, characterized in that: Here are the steps: Step 1: Offline training of sales forecast model and inventory level value function Firstly, an index that can measure the sales volume and volatility of a period is proposed, and the sales pattern categories of the period are fuzzily divided based on the index values; Afterwards, based on the sales pattern category of each period, a point prediction model based on the long short-term memory network or a quantile regression model based on the quantile long short-term memory network is selected to capture the long-term and short-term dependencies in the sales time series; A multi-stage stochastic dynamic decision-making model with the goal of maximizing overall revenue is constructed. With the help of an approximate dynamic programming algorithm, simulation is performed based on sales data, and the model is iterated continuously until a convergent value function is determined to represent the future value of the inventory level at a certain moment. Step 2: The oil product automatic replenishment system monitors the current time t and the remaining replenishment time ΔT of the station in real time t , the inventory level and real-time sales volume of oil products at the stations, the location of the in-transit tank trucks, the numbers and replenishment quantities of the stations that the in-transit tank trucks have not delivered, and the models and numbers of the tank trucks used for rescue replenishment at the central warehouse; Step 3: Based on the real-time oil product sales and offline trained sales forecasting model, real-time forecast of the remaining replenishment lead time ΔT t Oil sales in Step 4: Conduct dynamic early warning of oil inventory based on real-time monitoring data; Step 4.1: Early inventory analysis at the beginning of the period, first proceed to step 4.1.1; Step 4.1.1: Based on the maximum inventory consumption rate v max Determine whether the site will definitely not be out of stock: If the inventory level is I t The site has been maintained max , ΔT can still be satisfied t The demand within, that is, I t ≥ΔT t *v max , it means that the site definitely does not need to replenish before the next round of replenishment, and automatic replenishment ends; otherwise, go to step 4.1.2; Step 4.1.2: Based on the predictions from step 3 Judgment I t ≤D t，ΔT Whether it occurs; if it occurs, it means that the current moment is the replenishment triggering moment, triggering an early warning, and proceeding to step 5; Otherwise, go to step 4.2; Step 4.2: Inventory warning analysis within the time period, go to step 4.2.1; Step 4.2.1: Determine whether the current time has reached the beginning of the next time period. If so, return to step 2; otherwise, proceed to step 4.2.2; Step 4.2.2: Real-time monitoring of whether there is real demand during the period. If real demand occurs at a certain time tt, let the real inventory level at this time be I tt , and calculate the cumulative oil inventory TD sold from the beginning of the period to time tt [t，tt] ; If TD [t，tt] ≥Forecast value of oil sales volume in this period D t，Δt , then further judge Whether it occurs, to determine whether time tt is the replenishment triggering time; if it occurs, it means that the current time is the replenishment time, triggering an early warning and entering step 5; otherwise, return to step 4.2.1; Step 5: Determine the warning level; based on the resources currently available for rescue, the current distribution plan, the degree of urgency of inventory, and the difficulty of subsequent emergency measures, the warning levels are divided into Level 1 (high), Level 2 (low), Level 3 (medium), and Level 4 (low); Level 1 warning requires the dispatch decision maker to find a way to expand the available resources, including two situations: (1) the distribution system currently has no redundant resources available for rescue; (2) although there are vehicles on the road that deliver this type of oil to other gas stations, these vehicles cannot reach the warning station in time; Level 2 warning requires the dispatch decision maker to have rich experience in re-dispatching or dispatching in emergency mode, including three situations: (1) the distribution center has no available separate rescue and replenishment vehicles, but there are vehicles on the road that deliver this type of oil to other gas stations, and some of them can arrive at the warning station in time for rescue; (2) there are vehicles that can be used specifically for rescue and replenishment, but if the previous dispatch method is used, it will not be able to arrive at the warning station in time, and the dispatcher needs to dispatch vehicles in emergency mode according to the remaining time; (3) there are vehicles that meet both situations (1) and (2), but the vehicle that meets situation (2) needs to be called in emergency mode, so situation ( 3) In the next step, vehicles that meet condition (1) are given priority; in level 3 warning, the system only has vehicles that can arrive at the warning site in time for rescue and replenishment, and the dispatch decision maker only needs to dispatch emergency rescue vehicles from the distribution center in the conventional way; in level 4 warning, the system has abundant rescue resources available, and has vehicles that can arrive at the warning site in time for rescue and replenishment and vehicles that perform other distribution tasks, and the dispatch decision maker can decide which method to use for emergency rescue based on the actual situation and the difficulty of dispatch; if the site meets condition (1), condition (3) and level 4 warning in level 2 warning, then go to step 6 to generate a scientific replenishment dispatch plan in real time; Step 6: Generate replenishment scheduling plan in real time. After the early warning triggers replenishment, take the following steps to generate replenishment scheduling plan in real time, and effectively intervene or reschedule the current distribution plan; Step 6.1: Based on the warning level and the real-time predicted oil sales volume, select available replenishment vehicles. If the site meets the conditions (1) and (3) in the Level 2 warning, then after selecting the rescue warning site, there will still be no vehicles in transit at the site originally planned for delivery. If the warning level of the site is Level 4, then after selecting the rescue warning site, there will still be no vehicles in transit at the site originally planned for delivery, and there will be idle vehicles at the central warehouse specifically used for rescue replenishment, and then proceed to Step 6.

2. Step 6.2: Based on the real-time predicted oil sales and the value function trained in the offline stage that can represent the future value of inventory levels, an operations research model with the goal of maximizing overall revenue is constructed, and the Gurobi solver is used to decide when and by which vehicle to transfer how much oil.

2. A data-driven automatic oil replenishment method for a gas station as claimed in claim 1, characterized in that: In step 6.1, the selection steps of the vehicles in transit are as follows: Let the current time be t, and the end time of the replenishment cycle be t e , the replenishment quantity of station i is C i The time it takes for the in-transit tanker m, which is originally scheduled to deliver oil to station i, to arrive at the warning station j from the current location is Δt(V m，t , j), the time from the current location to site j and then back to site i is Δt(V m，t , j)+Δt(i, j), the current inventory levels of site i and site j are I i，t ,I j，t ;if and and That is, if the originally planned delivery station i will not be out of stock, and station j will not be out of stock before the rescue vehicle arrives, then the on-the-way vehicle m can rescue the warning station; otherwise, vehicle m is unavailable.

3. A data-driven automatic oil replenishment method for a gas station as claimed in claim 1, characterized in that: In the first step, the approximate dynamic programming algorithm adopts a concave adaptive value estimation algorithm.

4. The replenishment system used in the data-driven automatic replenishment method for oil products at a gas station according to any one of claims 1 to 3, characterized in that: The data-driven automatic replenishment system for oil products at gas stations includes a prediction model training module, a value function training module, a data collection module, a real-time prediction module for oil product sales, a calculation module 1, a calculation module 2, a calculation module 3, a calculation module 4, a warning level judgment module, a replenishment vehicle decision module, and a replenishment quantity decision module; the prediction model training module is based on historical sales data and uses the deep neural network prediction technology of big data to train prediction models with different prediction durations; the value function training module is based on historical sales data and simulation data and uses the approximate dynamic programming algorithm to train the value function of the inventory level; The data acquisition module can collect in real time the current time of the system, the remaining replenishment time of the station, the inventory level and real-time sales of oil products at the station, the location of the in-transit tanker, the number and replenishment quantity of the station that the in-transit tanker has not delivered, and the model and number of tankers specially used for rescue replenishment at the central warehouse; the oil product sales real-time prediction module predicts the oil product sales value in the future period in real time; the calculation module 1 calculates whether the station definitely does not need to replenish before the next round of replenishment starts; Calculation module 2 determines whether the current inventory level of the site is higher than the total sales volume of oil products within the predicted remaining replenishment lead time Calculation module 3, determines whether the current time of the system reaches the beginning of the next time period; Calculation module 4: determines whether the current inventory level is greater than The difference between the accumulated oil inventory and the accumulated oil inventory sold; the warning level judgment module presents the oil inventory warning level of the station; the replenishment vehicle decision module selects the available replenishment vehicles; The replenishment quantity decision module determines for the early warning station how much oil should be transferred by which vehicle and when.

Images