A method and system for spare parts demand prediction and inventory ordering parameter generation based on spatio-temporal non-stationarity

Abstract

The application relates to the technical field of aviation material supply chain demand prediction, and discloses a method and system for aviation material demand prediction and inventory order parameter generation based on space-time non-stationarity, which comprises the following steps: extracting time series demand and space-time input vectors from aviation material use records, the number of running equipment, equipment monthly training hours, month serial numbers, aircraft type categories, supplier categories and equipment service life; performing local weighted regression seasonal trend decomposition on the time series demand, and determining a trend prediction value according to the number of running equipment and a long-term trend item; incorporating the trend prediction value, a check month pulse seasonal item and a random residual item into the space-time input vector, obtaining an aviation material demand prediction sequence through an extreme gradient boosting tree model and a long short-term memory network model; and generating safety stock, an order point and an order quantity according to the discrete degree of the prediction sequence, a service level coefficient and a procurement lead time. The application can adapt to non-stationary changes of aviation material demand, and improve the adaptability of demand prediction and inventory order parameter generation.

Description

Technical Field

[0001] This invention relates to the field of aviation material supply chain demand forecasting technology, specifically to a method and system for forecasting aviation material demand and generating inventory ordering parameters based on spatiotemporal nonstationarity. Background Technology

[0002] Aircraft material inventory management typically requires forecasting future demand based on historical requisition records, and determining safety stock, reorder point, and order quantity accordingly. Existing aircraft material demand forecasting methods mostly employ moving averages, exponential smoothing, time series regression, or static inventory models to smooth, extrapolate, or calculate requisition quantities with fixed parameters over several past periods.

[0003] In scenarios involving continuous changes in the number of training equipment, adjustments to equipment distribution bases, and changes in aircraft type structure, the demand sequence for aircraft materials often exhibits non-stationary characteristics. An increase in the number of operating equipment alters long-term demand levels; inspection tasks in different months create periodic peaks in requisition; and different aircraft types, suppliers, and equipment service ages also change the differences in aircraft material consumption. Traditional methods typically treat historical demand as a single time series, making it difficult to simultaneously express the combined impact of equipment scale, training hours, inspection months, aircraft type, supplier type, and equipment service age on aircraft material demand.

[0004] Furthermore, existing static inventory models typically employ fixed safety stock or fixed reorder points, making it difficult to incorporate forecast uncertainty, procurement lead time, and the importance level of aircraft materials into the same inventory calculation chain. When the forecast sequence exhibits discrete fluctuations in inspection months, equipment relocation months, or months of abnormal maintenance, fixed inventory parameters can easily become disconnected from the actual requisition rhythm, leading to insufficient inventory or excessive inventory holdings.

[0005] Therefore, this application provides a method and system for forecasting aircraft material demand and generating inventory ordering parameters based on spatiotemporal nonstationarity to solve the above-mentioned technical problems. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for forecasting aircraft material demand and generating inventory ordering parameters based on spatiotemporal nonstationarity, in order to solve the technical problems in the prior art where aircraft material demand forecasting is difficult to adapt to changes in equipment scale, inspection month pulses, aircraft model differences, supplier differences and equipment service age, and where static inventory parameters are difficult to dynamically adjust with forecast uncertainty.

[0007] To address the aforementioned technical problems, this invention provides a method for forecasting aircraft material demand and generating inventory ordering parameters based on spatiotemporal nonstationarity, comprising: Extract time-series demand and spatiotemporal input vectors from aircraft material requisition records, number of operating equipment, monthly training hours of equipment, month serial number, aircraft type, supplier type, and equipment service life; The time-series demand is subjected to local weighted regression seasonal trend decomposition to obtain a long-term trend term, an inspection month pulse seasonal term, and a random residual term. The number of operating equipment is logarithmically transformed and regressed with the long-term trend term to obtain a trend prediction. The trend prediction, the inspection month pulse seasonal term, and the random residual term are incorporated into the spatiotemporal input vector. The spatiotemporal input vector is input into the extreme gradient boosting tree model, and a static prediction output is obtained based on tree node splitting. The static prediction output and the spatiotemporal input vector arranged in time windows are input into the long short-term memory network model. Gating updates are performed based on the historical hidden state and cell state to obtain the aircraft material demand prediction sequence. The standard deviation of the forecast is generated based on the dispersion of the aircraft material demand forecast sequence. The safety stock is determined by combining the service level coefficient and the procurement lead time. The reorder point is determined based on the lead time forecast demand and the safety stock. The order quantity is determined based on the economic order quantity, the safety stock, the current inventory quantity, and the inventory in transit quantity.

[0008] In some specific embodiments, the temporal demand and spatiotemporal input vector are extracted from aircraft material requisition records, number of operating equipment, monthly training hours of equipment, month number, aircraft type, supplier type, and equipment service life, and further include: The time-series demand is formed by aggregating the aircraft material identification and the monthly aggregated requisition quantity; Write the number of operating devices in the same month into the time dimension data; Arrange the monthly training hours for each device into training intensity data by device and month; Convert the month serial number into quarterly inspection cycle code and annual cycle code; Convert the model category and supplier category into category values, and convert the equipment service life into service life values; The spatiotemporal input vector is obtained by concatenating the time dimension data, the training intensity data, the quarterly inspection cycle code, the annual cycle code, the category value, and the service age value.

[0009] In some specific embodiments, the time-series demand is subjected to a locally weighted regression seasonal trend decomposition to obtain a long-term trend term, a pulsed seasonal term for the month of inspection, and a random residual term, further including: The time-series demand is locally weighted smoothed using a preset seasonal window to separate candidate seasonal sequences that fluctuate repeatedly with the months. After deducting the candidate seasonal series from the time-series demand, the remaining series is locally weighted smoothed to obtain the long-term trend term; The random residual term is obtained by subtracting the long-term trend term and the candidate seasonal sequence from the time-series demand. Pulse amplitude preservation is performed on the candidate seasonal sequence for the period in which the month under inspection is located, and smoothing amplitude constraint is performed on the period in which the month outside the inspection is located, to obtain the pulse seasonal term for the month under inspection.

[0010] In some specific embodiments, the number of operating devices is logarithmically transformed and regressed with the long-term trend term to obtain a trend prediction. The trend prediction, the seasonal pulse term of the inspection month, and the random residual term are then incorporated into the spatiotemporal input vector, further including: Read the number of operating devices and the long-term trend item for each historical month; After performing a logarithmic transformation on the number of operating equipment, regression fitting was performed with the long-term trend term to obtain the equipment scale influence coefficient and constant term distinguished by aviation material category; The trend forecast quantity is determined based on the equipment scale influence coefficient, the logarithm of the number of operating equipment in the month to be predicted, and the constant term. The trend prediction, the seasonal pulse term for the month of inspection, and the random residual term are incorporated into the spatiotemporal input vector.

[0011] In some specific embodiments, the spatiotemporal input vector is input into an extreme gradient boosting tree model, and a static prediction output is obtained based on tree node splitting, further including: At each node of the regression tree, read the equipment size field, training hour field, month code field, model category field, supplier category field, and equipment service life field from the spatiotemporal input vector; Samples are assigned to leaf nodes based on node splitting thresholds; The outputs of each regression leaf node are summed into a tree-integrated output. The tree integration output is used as the static prediction output and input into the long short-term memory network model.

[0012] In some specific embodiments, the static prediction output and the spatiotemporal input vector arranged in time windows are input into a long short-term memory network model, and gating updates are performed based on historical hidden states and cell states to obtain the aircraft material demand prediction sequence, further including: The spatiotemporal input vectors are arranged in a preset time window, and the spatiotemporal input vectors in each window are concatenated with the static prediction output of the same time period to form a time-series input; Read the hidden state of the previous period, the cell state of the previous period, and the time sequence input for each prediction period; Perform weighted calculations using the forget gate, input gate, and output gate respectively, and update the hidden state and cell state for the current time period. The hidden state of the current time period and the static prediction output are input into a fully connected layer to obtain the aircraft material demand prediction sequence.

[0013] In some specific embodiments, after obtaining the aircraft material demand forecast sequence, the method further includes: For consumable aircraft materials, the number of operating equipment and the inspection month marker are read, and the theoretical demand intensity is determined based on the basic arrival rate, the number of operating equipment and the inspection month marker; A Poisson distribution is constructed based on the theoretical demand intensity, and the actual demand frequency in the requisition records is fitted to the Poisson distribution for a fitting test. When the fitting test results meet the preset matching conditions, the aircraft material demand forecast sequence is retained as the inventory calculation input. When the fitting test result does not meet the preset fit condition, the theoretical demand intensity is discretized by using a negative binomial distribution to obtain the corrected aircraft material demand forecast sequence, which is then used as the input for inventory calculation.

[0014] In some specific embodiments, a forecast standard deviation is generated based on the dispersion of the aircraft material demand forecast sequence, and a safety stock is determined by combining the service level coefficient and the procurement lead time, further including: In the Long Short-Term Memory network model, the random deactivation layer is kept in a sampling state, and multiple forward propagations are performed on the same month to be predicted to obtain multiple sampled predictions. Calculate the dispersion of the multiple sampled predictions within a single prediction period to obtain the prediction standard deviation; The importance level of aircraft materials is mapped to the service level coefficient, and the procurement channel category is mapped to the number of prediction cycles corresponding to the procurement lead time. The safety stock is determined based on the service level coefficient, the forecast standard deviation, and the forecast period number.

[0015] In some specific embodiments, determining the safety stock based on the service level coefficient, the forecast standard deviation, and the number of forecast periods further includes: The number of forecast periods is used in the determination of the safety stock in the form of a square root. The safety stock is determined by the service level coefficient, the forecast standard deviation, and the number of forecast periods. The lead time forecast demand includes the forecast quantity of aircraft material demand during the lead time, and the reorder point is determined based on the forecast quantity of aircraft material demand during the lead time and the safety stock. The target inventory level is determined based on the economic order quantity and the safety stock, and the available inventory level is determined based on the current inventory level and the inventory in transit. When the target inventory level is higher than the available inventory level, the order quantity is generated based on the difference between the target inventory level and the available inventory level; When the order quantity is less than the preset minimum purchase quantity, the preset minimum purchase quantity is output. When the order quantity reaches the preset minimum purchase quantity, the order quantity is output.

[0016] Based on the same concept, the present invention also provides a system for predicting aircraft material demand and generating inventory ordering parameters based on spatiotemporal nonstationarity, comprising: The input vector formation module is configured to extract time-series demand quantities and spatiotemporal input vectors from aircraft material requisition records, number of operating equipment, monthly training hours of equipment, month serial number, aircraft type, supplier type, and equipment service age. The non-stationary trend processing module is configured to perform local weighted regression seasonal trend decomposition on the time-series demand to obtain a long-term trend term, a check month pulse seasonal term, and a random residual term; perform logarithmic transformation on the number of operating equipment and regress and fit it with the long-term trend term to obtain a trend prediction; and incorporate the trend prediction, the check month pulse seasonal term, and the random residual term into the spatiotemporal input vector. The tree prediction module is configured to input the spatiotemporal input vector into an extreme gradient boosting tree model and obtain a static prediction output based on tree node splitting; The time-series prediction module is configured to input the static prediction output and the spatiotemporal input vector arranged in time windows into the long short-term memory network model, and perform gating updates based on historical hidden states and cell states to obtain the aircraft material demand prediction sequence. The inventory parameter generation module is configured to generate a forecast standard deviation based on the dispersion of the aircraft material demand forecast sequence, determine the safety stock by combining the service level coefficient and the procurement lead time, determine the reorder point based on the lead time forecast demand and the safety stock, and determine the order quantity based on the economic order quantity, the safety stock, the current inventory quantity, and the in-transit inventory quantity.

[0017] Compared with existing technologies, its advantages are as follows: This invention discloses a method and system for forecasting aircraft material demand and generating inventory ordering parameters based on spatiotemporal nonstationarity. It generates time-series demand quantities and spatiotemporal input vectors from aircraft material requisition records, number of operating equipment, monthly training hours of equipment, month serial number, aircraft type, supplier type, and equipment service age. It can incorporate changes in equipment scale, training intensity, inspection month, aircraft type differences, supplier differences, and equipment aging degree into the same forecast input, avoiding insufficient input information caused by relying solely on extrapolation of a single historical demand sequence.

[0018] By decomposing time-series demand into a long-term trend term, a seasonally impulsive term for inspection months, and a random residual term, and performing a regression fit between the logarithmic transformation of the number of operating equipment and the long-term trend term, we can model the long-term demand changes caused by the increase in the number of equipment separately, while retaining the impulsive seasonal fluctuations caused by concentrated maintenance in inspection months. This ensures that non-stationary demand sequences have a clear data composition before entering the forecasting model.

[0019] By inputting the spatiotemporal input vector into the extreme gradient boosting tree model and obtaining the static prediction output based on tree node splitting, the nonlinear interaction relationship between different static features can be expressed, forming a static prediction output that provides input containing spatial, device, and class differences for subsequent time series prediction.

[0020] By inputting the static prediction output and the spatiotemporal input vector arranged by time window into the long short-term memory network model, and using the historical hidden state and cell state to perform gating updates on the time period input, it is possible to introduce cross-month demand dependence while retaining the static nonlinear prediction results, so that the aircraft material demand prediction sequence can simultaneously reflect the characteristics of the current month and the changes of historical months.

[0021] The theoretical demand intensity of consumable aircraft materials is constructed and a Poisson distribution fitting test is performed. When the fitting test result does not meet the preset fit conditions, a negative binomial distribution is used to discretize and correct the theoretical demand intensity. This can verify the distribution pattern of the demand frequency of consumable aircraft materials and transform the excessive discretization case into a corrected aircraft material demand prediction sequence.

[0022] The forecast standard deviation is generated based on the dispersion of the aircraft material demand forecast sequence. Combined with the service level coefficient, procurement lead time, lead time forecast demand, current inventory quantity and in-transit inventory quantity, safety stock, reorder point and order quantity are determined. This can transfer the forecast uncertainty to the inventory parameter generation stage, so that aircraft materials of different importance levels and different procurement channels can obtain matching inventory ordering parameters. Attached Figure Description

[0023] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a flowchart illustrating some specific embodiments of the method for predicting aircraft material demand and generating inventory ordering parameters based on spatiotemporal nonstationarity of the present invention. Figure 2 This is a schematic diagram of the structure of a system for predicting demand for aviation materials and generating inventory ordering parameters based on spatiotemporal nonstationarity, according to some specific embodiments of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0025] The terminology used in the embodiments of this application is for the purpose of describing particular embodiments only and is not intended to limit the application. The singular forms "a," "the," and "the" as used in the embodiments of this application are also intended to include the plural forms, unless the context clearly indicates otherwise, and "multiple" generally includes at least two.

[0026] It should be understood that the term "and / or" used in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0027] It should be understood that although the terms first, second, third, etc., may be used in the embodiments of this application, these descriptions should not be limited to these terms. These terms are only used to distinguish the descriptions. For example, first may also be referred to as second without departing from the scope of the embodiments of this application, and similarly, second may also be referred to as first.

[0028] Depending on the context, the words “if” or “suppose” as used here can be interpreted as “when” or “in response to determination” or “in response to detection.” Similarly, depending on the context, the phrases “if determination” or “if detection (of the stated condition or event)” can be interpreted as “when determination” or “in response to determination” or “when detection (of the stated condition or event)” or “in response to detection (of the stated condition or event).”

[0029] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that an article or device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such an article or device. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the article or device that includes said element.

[0030] It should be noted that any symbols and / or numbers present in the specification that are not marked in the accompanying drawings are not reference numerals.

[0031] Reference Figure 1 A method for forecasting aircraft material demand and generating inventory ordering parameters based on spatiotemporal nonstationarity includes: S101 extracts the time-series demand and spatiotemporal input vector from the aircraft material requisition records, the number of operating equipment, the monthly training hours of the equipment, the month number, the aircraft type, the supplier type, and the service life of the equipment; S102, perform local weighted regression seasonal trend decomposition on the time-series demand to obtain long-term trend term, inspection month pulse seasonal term and random residual term, perform logarithmic transformation on the number of operating equipment and regress and fit with the long-term trend term to obtain trend prediction, and incorporate the trend prediction, the inspection month pulse seasonal term and random residual term into the spatiotemporal input vector. S103, input the spatiotemporal input vector into the extreme gradient boosting tree model, and obtain the static prediction output based on tree node splitting; S104, The static prediction output and the spatiotemporal input vector arranged in time windows are input into the long short-term memory network model, and the gating update is performed based on the historical hidden state and cell state to obtain the aircraft material demand prediction sequence. S105, generate the forecast standard deviation based on the dispersion of the aircraft material demand forecast sequence, and determine the safety stock by combining the service level coefficient and the procurement lead time. Determine the reorder point based on the lead time forecast demand and the safety stock, and determine the order quantity based on the economic order quantity, the safety stock, the current inventory quantity, and the in-transit inventory quantity.

[0032] To provide a clearer explanation, the steps in the embodiments of the present invention are described in detail below: S101 extracts the time-series demand and spatiotemporal input vector from the aircraft material requisition records, the number of operating equipment, the monthly training hours of the equipment, the month number, the aircraft type, the supplier type, and the service life of the equipment; In this embodiment, the aircraft material requisition records can be derived from the training center's aircraft material management ledger, inbound and outbound records, maintenance work orders, requisition application forms, or historical records in the inventory management system. Each aircraft material requisition record includes at least the aircraft material identifier, requisition date, requisition quantity, equipment or equipment group used, aircraft type, base, supplier type, and maintenance or training related information.

[0033] The number of operating equipment can be counted monthly to determine the number of simulators put into operation and participating in training tasks; the monthly training hours of the equipment can be counted by equipment number and month to indicate the usage intensity of each equipment in each month; the month serial number is used to form a seasonal code; the aircraft type is used to distinguish the differences in the consumption of aviation materials of different aircraft types of simulators; the supplier type is used to distinguish the differences in the consumption of equipment from different suppliers in terms of structure, maintenance habits or spare parts life; the service life of the equipment is used to characterize the service life of the equipment after it has been put into operation.

[0034] Furthermore, the time-series demand is formed by aggregating the aircraft material identification and the monthly requisition quantity.

[0035] Specifically, when using the natural month as the time granularity, the quantities of multiple requisitions of the same aviation material identifier within the same month are added together to obtain the time-series demand of the aviation material in that month.

[0036] If a piece of aviation material has records of being returned, cancelled, or duplicated within a month, the record status field, document number, and inbound / outbound direction field can be verified. Records that have been issued or actually consumed can be retained, while cancelled and duplicate records can be removed, and the net issued quantity can be summed.

[0037] If there are no usage records for a certain month, the time series demand for that month can be recorded as 0, and the number of operating devices, training hours and seasonal code for that month can be retained so that subsequent trend decomposition can identify low demand months and maintain the continuity of the time axis of the demand sequence.

[0038] The time-series demand can be expressed as ,in, Indicates the month index. Indicates the first The monthly demand for aircraft materials.

[0039] Write the number of operating devices in the same month into the time dimension data.

[0040] Specifically, the number of operating devices can be expressed as ,in, For the first The number of devices that are operational and participating in training schedules each month. The number of operational devices reflects the impact of changes in equipment scale on the demand for aviation materials.

[0041] When the training center adds new equipment, relocates equipment to different bases, or experiences equipment downtime or resumption of operation, the operational status is recorded monthly. .

[0042] For the same model of equipment that changes from a centralized base to a multi-base distribution, the equipment can be confirmed in the equipment ledger at the [number]th [location]. The operating location and status for the past month, including equipment that meets the operating conditions. .

[0043] If the operating status of equipment changes mid-month, the equivalent number of operating equipment for the month can be calculated based on the actual number of operating days of the corresponding equipment in that month.

[0044] The monthly training hours for each device are arranged into training intensity data by device and month.

[0045] Specifically, the monthly training hours for the equipment can be expressed as: It can also be expanded as ,in, Indicates the equipment number. Indicates the month.

[0046] For an embodiment with 34 devices excluding flight trainers, the training intensity data can be constructed as a matrix of training hours for the 34 devices and 12 months, where each matrix element represents the training hour of the 34 devices over 12 months. The first piece of equipment Monthly training hours.

[0047] If a device is out of service in a certain month, the matrix elements can be 0. If a device is missing training hour records in a certain month, the missing records can be filled in by reading the scheduling records, flight simulation training records, or equipment operation logs. If the above records still cannot provide the corresponding data, the median of training hours for the same aircraft type, the same base, and adjacent months can be used as an example filling value.

[0048] Training intensity data and aircraft material requisition records are aligned by month, allowing the impact of different aircraft types and training intensities on aircraft material demand to be incorporated into subsequent forecasting processes.

[0049] Convert the month serial number into quarterly inspection cycle code and annual cycle code.

[0050] Specifically, month number Can be converted to seasonal codes : ; in, and Together they are used to represent the characteristics of quarterly inspections over a 3-month period. and Together they are used to represent the annual variation characteristics of a 12-month cycle. Indicates the month index. It represents pi (π).

[0051] For the inspection peaks in March, June, September, and December, the quarterly inspection cycle code is used to characterize the position of the corresponding month in the 3-month cycle, and the annual cycle code is used to characterize the position of the corresponding month in the annual cycle, so that subsequent trend decomposition and tree prediction processing can read the periodic time characteristics.

[0052] Convert the model category and supplier category into category values, and convert the equipment service life into service life values.

[0053] Specifically, the aircraft type category can include Airbus series, Boeing series and other training equipment aircraft type categories, and the supplier category can include different simulator manufacturers or self-developed equipment categories.

[0054] Model category and supplier category can be formed using unique hot coding, target coding or embedded coding to generate category values.

[0055] Equipment service life It can be by the first The time difference between the month and the equipment commissioning date is determined, and the unit can be either a year or a month.

[0056] If a certain aviation material serves multiple aircraft types, the aircraft type category value can be weighted according to the monthly training hours of the equipment to obtain the corresponding aircraft type category input for that month; if multiple supplier equipment are running at the same time in a certain month, the supplier category input can be formed according to the corresponding training hours of the supplier equipment or the proportion of the number of running equipment.

[0057] The spatiotemporal input vector is obtained by splicing time dimension data, training intensity data, quarterly inspection cycle encoding, annual cycle encoding, category value and service age value.

[0058] Specifically, the time dimension data corresponding to the number of running devices is aligned with the monthly index, the training intensity data corresponding to the training hours of the devices, the quarterly inspection cycle code, the annual cycle code, the category values ​​corresponding to the machine type and supplier type, and the service life value corresponding to the service life of the devices, forming a time-series demand. Spatiotemporal input vectors on the same time axis.

[0059] If a field is missing, perform a field integrity check on the corresponding data source. If the number of operating devices is missing, retrieve the device operation log to supplement it; if training hour data is missing, retrieve the training schedule or device operation log to supplement it; if category values ​​are missing, retrieve the device master data to supplement it.

[0060] In this embodiment, the base location can also be written as a grouping attribute of the training intensity data into the extended field of the spatiotemporal input vector.

[0061] The extended fields can be used to differentiate spare parts requirements for the same model operating at different bases such as Zhuhai, Xinjiang, North China, and Chongqing. When using extended fields, monthly training hours for equipment can be summarized by base and combined with model category and supplier category to form grouped training intensity inputs.

[0062] Even when the extended fields are missing, the spatiotemporal input vector is still formed using the device's monthly training hours, device type, and supplier type, and the subsequent prediction data stream remains unchanged.

[0063] S102, perform local weighted regression seasonal trend decomposition on the time-series demand to obtain long-term trend term, inspection month pulse seasonal term and random residual term, perform logarithmic transformation on the number of operating equipment and regress and fit with the long-term trend term to obtain trend prediction, and incorporate the trend prediction, the inspection month pulse seasonal term and random residual term into the spatiotemporal input vector. In this embodiment, the timing requirement... Performing local weighted regression seasonal trend decomposition transforms the long-term changes, periodic inspection fluctuations, and random fluctuations in the aircraft material demand sequence into long-term trend terms, inspection month pulse seasonal terms, and random residual terms, respectively.

[0064] The long-term trend term reflects long-term demand changes caused by the increase in the number of operating equipment, equipment migration, and model expansion; the inspection month pulse seasonal term reflects periodic demand changes caused by peak equipment inspections in March, June, September, and December; and the random residual term reflects random changes caused by sudden failures, temporary maintenance, and non-periodic requisition.

[0065] Furthermore, a local weighted smoothing is performed on the time-series demand using a preset seasonal window to separate candidate seasonal sequences that fluctuate repeatedly with the months.

[0066] Specifically, the preset seasonal window can be a 12-month annual cycle window, or a 3-month cycle window can be set simultaneously for aircraft materials with obvious quarterly inspection characteristics.

[0067] When performing local weighted smoothing, the demand of adjacent months is selected around each month to be decomposed, and the local curve is fitted in such a way that the closer to the month to be decomposed, the greater the weight, to obtain the candidate seasonal sequence.

[0068] For months under inspection, the candidate seasonal sequence retains the time position of the demand peak; for months not under inspection, the candidate seasonal sequence records the periodic changes of ordinary months.

[0069] If the time-series demand for a certain month is 0, the local weighted smoothing still retains the sample for that month to maintain the continuity of the time axis.

[0070] If records for consecutive months are missing, the corresponding missing month markers can be written into the residual processing data, enabling the local weighted smoothing process to distinguish between missing records and months with actual low demand.

[0071] After deducting the candidate seasonal series from the time-series demand, the remaining series is locally weighted smoothed to obtain the long-term trend term.

[0072] Specifically, the residual sequence represents the change in demand for aircraft materials after deducting seasonal variations.

[0073] When performing local weighted smoothing on the remaining sequence, the long-term demand level resulting from the increase in the number of operating equipment, base relocation, and changes in aircraft type structure is retained in the long-term trend term. middle.

[0074] For scenarios where the number of simulators gradually increases from a few to 34, the long-term trend item records the increase in demand as the scale of equipment expands.

[0075] If a certain aviation material has economies of scale, the long-term trend term may change at a rate lower than linear as the number of operating equipment increases; if a certain aviation material is greatly affected by equipment interaction, operation frequency, or high-intensity training, the long-term trend term may change at a rate higher than linear.

[0076] The random residual term is obtained by subtracting the long-term trend term and the candidate seasonal sequence from the time-series demand.

[0077] Specifically, the stochastic residual term is used to record demand changes in time-series demand that are not explained by the long-term trend term and candidate seasonal series.

[0078] For events such as sudden failures, temporary maintenance, and abnormal monthly requisitions, the random residual term can represent the corresponding demand deviation.

[0079] When the random residual term exceeds the outlier threshold set based on the historical residual distribution, the corresponding month can be marked as an outlier. Outliers can still be retained in the model training data, and the impact of extreme biases on model parameter updates can be reduced during model training using Hubell loss.

[0080] When data verification determines that the random residual discrepancy is due to an error in the requisition record, the corresponding requisition record and inventory documents are reviewed, the time-series demand is updated based on the verified requisition quantity, and the trend decomposition for the corresponding month is re-executed.

[0081] Pulse amplitude preservation is performed on the candidate seasonal sequence of the period in which the month under inspection is located, and smoothing amplitude constraint is performed on the period in which the month outside the inspection is located, to obtain the pulse seasonal term of the month under inspection.

[0082] Specifically, the inspection months can include months that are part of quarterly or annual inspections. In this embodiment, March, June, September, and December can be used as examples of inspection months.

[0083] When pulse amplitude retention is applied to the time period of the inspection month, the high demand value formed by concentrated maintenance in the candidate seasonal sequence is retained as the pulse amplitude of the corresponding inspection month; when smoothing amplitude constraint is applied to the time period of non-inspection months, the seasonal items of ordinary months maintain the continuous change between adjacent months.

[0084] If a piece of aircraft material does not experience a significant peak in demand during the inspection month, the material can maintain normal seasonal variations for the corresponding month.

[0085] If the peak inspection month for a certain aircraft material at a certain base is earlier or later than expected, the pulse month can be shifted by one month to the adjacent month within the preset seasonal window, and the shifted month will be used as the corresponding inspection month position for that aircraft material.

[0086] In this embodiment, the demand for aviation materials can also be categorized before seasonal decomposition.

[0087] Consumable aircraft materials, repairable aircraft materials, and low-frequency, high-value aircraft materials can each use different seasonal windows. Consumable aircraft materials can retain monthly issuance quantities, while low-frequency, high-value aircraft materials can use quarterly cumulative demand trends for decomposition.

[0088] The above classification changes the aggregation granularity of the corresponding aviation material input sequence. After trend decomposition, it still forms a long-term trend term, an inspection month pulse seasonal term, and a random residual term.

[0089] After completing the trend decomposition, the number of operating devices is logarithmically transformed and regressed with the long-term trend term.

[0090] Read the number of operating devices and the long-term trend item for each historical month.

[0091] Specifically, the number of operating devices in historical months Long-term trend items can be obtained from equipment operation logs. It is obtained from the aforementioned local weighted regression seasonal trend decomposition.

[0092] Reading is aligned according to the month index. and This forms the corresponding historical regression samples.

[0093] If there is equipment relocation or equipment downtime in a certain month, it will be determined based on the actual operating status. Furthermore, the migration base can be retained as an extended field.

[0094] For historical data examples from 2003 to 2024, the number of operating devices can be gradually increased from a small number of devices in the early stages to 34 devices. The long-term trend item records the long-term demand changes for different aviation materials such as projector aviation materials, control display unit aviation materials, and oxygen mask aviation materials.

[0095] If the operating status of equipment changes mid-month, the number of operating days can be converted into the equivalent number of operating equipment for the month.

[0096] After performing a logarithmic transformation on the number of operating equipment, regression fitting is performed with the long-term trend term to obtain the equipment scale influence coefficient and constant term distinguished by aviation material category.

[0097] The long-term trend term and the number of operating devices can be correlated using the following logarithmic-linear relationship: ; in, Indicates the first Long-term trend items over several months; Indicates the first Number of operating devices per month; This represents the equipment scale influence coefficient, used to indicate the degree of influence of changes in the number of operating equipment on the long-term trend term. This represents the constant term, obtained by regression from historical data; This represents the random error term, used to indicate the trend deviation after regression fitting that is not explained by the number of operating devices; Represents the natural logarithm operation.

[0098] Equipment size influence coefficient Estimates can be made separately according to the category of aircraft materials.

[0099] In one embodiment, the projector's aviation materials 0.85 can be taken; control display unit aviation materials 1.12 can be taken; oxygen mask aviation materials You can take 1.05, depending on the type of aircraft material. It is used to characterize the degree of influence of the logarithmic change in the number of operating devices on the long-term trend term.

[0100] The above values ​​are parameter values ​​used in the embodiment. In actual implementation, they can be re-determined based on historical data.

[0101] If the number of historical months for a certain aircraft material is insufficient to complete a separate regression, the corresponding historical samples can be merged according to the aircraft material function category, aircraft type category, or supplier category to determine the equipment scale impact coefficient and constant term.

[0102] The trend forecast quantity is determined based on the equipment scale influence coefficient, the logarithm of the number of operating equipment in the month to be predicted, and the constant term.

[0103] The number of operating devices for the month to be predicted is expressed as At that time, the trend forecast can be expressed as: ; in, This indicates the trend forecast amount corresponding to the month to be predicted; This represents the equipment scale impact coefficient for the corresponding aviation material category; This indicates the number of operating devices for the month to be predicted; This represents the corresponding constant term.

[0104] The trend prediction, the seasonal pulse term for the month of inspection, and the random residual term are incorporated into the spatiotemporal input vector.

[0105] Specifically, the trend prediction, the seasonal term of the monthly pulse, and the random residual term are input into the spatiotemporal vector according to the monthly index, so that the subsequent extreme gradient boosting tree model can jointly read the equipment scale change, the monthly pulse, and the random fluctuation information.

[0106] If the month to be predicted is the month to be checked, the corresponding seasonal term of the checked month participates in the subsequent tree node splitting; if the month to be predicted is not the month to be checked, the corresponding seasonal term provides seasonal variation information for ordinary months; if the random residual term has an anomalous sample label, the anomalous sample label can be used as sample weight adjustment information in model training.

[0107] For low-frequency, high-value aviation materials, the equipment scale impact coefficient can be estimated based on the cumulative demand over a longer period; for high-frequency consumable aviation materials, it can be estimated based on the monthly demand.

[0108] If there is a lack of historical usage records for new aircraft materials, the equipment scale impact coefficient corresponding to aircraft materials of the same model, supplier, or functional category can be used as the initial value.

[0109] After new usage records are generated in subsequent months, the new months can be added to the regression sample set. The equipment scale influence coefficient and constant term can be updated in subsequent prediction periods to keep the model parameters used in the current prediction period stable.

[0110] S103, input the spatiotemporal input vector into the extreme gradient boosting tree model, and obtain the static prediction output based on tree node splitting; In this embodiment, the extreme gradient boosting tree model is used to capture the nonlinear relationship between static fields in the spatiotemporal input vector and aircraft material requirements.

[0111] Each regression tree divides samples into corresponding leaf nodes layer by layer according to the input fields corresponding to the tree nodes and the node splitting threshold. The leaf nodes form local predictions of aviation material demand, and the outputs of the leaf nodes of multiple regression trees together form static prediction outputs.

[0112] Furthermore, at each node of the regression tree, the equipment size field, training hour field, month code field, model category field, supplier category field, and equipment service life field are read from the spatiotemporal input vector.

[0113] Specifically, the equipment scale field may include the number of operating equipment and the trend forecast; the training hours field may include the total monthly training hours of the equipment, the training hours component of the same aircraft type, and the training hours component of the same supplier; the month code field may include the quarterly inspection cycle code and the annual cycle code; the aircraft type field may include categories such as Airbus series and Boeing series; the supplier type field may include the category value corresponding to the equipment of each supplier; and the equipment service age field may include the service age of a single piece of equipment and the weighted average service age.

[0114] Missing values ​​can be marked before each field enters a tree node. Extreme gradient boosting tree models can determine the default branch for missing values ​​during node splitting.

[0115] If the training hour field has abnormally high values, it can be truncated according to the quantile interval corresponding to the training hours of the same model in history, and the truncated mark can be added to the model input as an auxiliary input field.

[0116] Samples are assigned to leaf nodes based on node splitting thresholds.

[0117] Specifically, each tree node stores the selection results of the corresponding field and the split threshold.

[0118] During the model training phase, the corresponding fields and their splitting thresholds are determined based on the training objective. During the model inference phase, the spatiotemporal input vector of the month to be predicted is read, the field values ​​are compared with the node splitting thresholds, and the corresponding branches are entered based on the comparison results until the leaf node is reached.

[0119] For the category field, the category code can be converted into the corresponding split value, and the split can be performed according to the node configuration.

[0120] For the month code field, the tree node can distinguish the corresponding month samples based on the quarterly inspection cycle code and the annual cycle code.

[0121] For the training hour field, tree nodes can assign samples to the corresponding training intensity branches based on differences in training intensity.

[0122] When a field is missing, the sample enters the corresponding default branch determined during the model training phase.

[0123] The outputs of each regression leaf node are summed into a tree-integrated output.

[0124] Specifically, each regression tree generates a corresponding leaf node prediction based on the leaf node that the sample finally reaches. The leaf node predictions of all regression trees are accumulated to form the tree ensemble output.

[0125] During the model training phase, the demand for aircraft materials or the corresponding demand data after trend decomposition and trend regression can be used as the training target, and the prediction residual of the previous round can be fitted by an iterative method.

[0126] The model training objective can include a loss term and a tree complexity constraint term. The loss term can use squared error or Hubel loss, and the tree complexity constraint term is used to limit the number of leaf nodes and the magnitude of the leaf node weights.

[0127] If the training samples include abnormal usage caused by sudden maintenance, the Hubel loss is treated according to the squared error relationship when the prediction error is small, and according to the linear error relationship when the prediction error is large, thereby reducing the impact of extreme samples on the tree structure and model parameter updates.

[0128] The tree integration output is used as the static prediction output and input into the long short-term memory network model.

[0129] Specifically, the static prediction output is used to characterize the nonlinear demand estimate formed by equipment size, training hours, month code, model type, supplier type, and equipment age in the same month.

[0130] The static prediction output, together with the spatiotemporal input vector of the corresponding month, constitutes the temporal input of the Long Short-Term Memory network model.

[0131] If the static prediction output of a certain aviation material in a certain month differs from the long-term trend term by more than the preset deviation range, the corresponding deviation value can be written into the model running record and used as auxiliary input information in the time window.

[0132] If the tree model lacks effective historical samples for new aircraft materials, the average output of tree ensemble of similar aircraft materials can be used to form the initial static prediction output, and the corresponding tree model can be retrained after the accumulation of new aircraft material requisition records.

[0133] In this embodiment, field importance can also be calculated during the training phase.

[0134] Field importance can be accumulated based on the amount of loss reduction caused by node splitting.

[0135] The number of devices, monthly training hours, seasonal coding, model type, supplier type, and equipment service life can all be used to calculate the importance of a field.

[0136] Field importance results are not used directly as the final predicted value, but are used to verify whether the prediction model effectively reads the corresponding input fields.

[0137] When a field has not participated in tree node splitting for a long time, you can check whether the field has coding errors, too much missing data, or is misaligned with the timeline of aircraft material demand, and save the field importance results as model maintenance records.

[0138] S104, The static prediction output and the spatiotemporal input vector arranged in time windows are input into the long short-term memory network model, and the gating update is performed based on the historical hidden state and cell state to obtain the aircraft material demand prediction sequence. In this embodiment, a long short-term memory network model is used to handle the time dependency of aircraft material demand.

[0139] The model input consists of a spatiotemporal input vector arranged in time windows and the corresponding static prediction output.

[0140] The time window can be 12 months, using input data from the past 12 months to predict demand for the next month.

[0141] The hidden layer dimension of the model can be 64, and the random deactivation coefficient can be 0.2.

[0142] Furthermore, the spatiotemporal input vectors are arranged in a preset time window, and the spatiotemporal input vectors in each window are concatenated with the static prediction output of the same time period to form a time-series input.

[0143] Specifically, for the month to be predicted You can select Moonlit The spatiotemporal input vector for each month is concatenated with the static prediction output corresponding to the same month to form a time series input with a length of 12 months.

[0144] The time-series input for each month may include equipment size, training hours, month code, model type, supplier type, equipment age, trend forecast quantity, check month pulse seasonal item, random residual item, and static forecast output.

[0145] If the historical months are less than 12 months, the historical sequence of similar aircraft materials can be used to supplement them, or the data corresponding to the existing months can be used to fill the front of the time window, and the corresponding supplementary mark can be added to the input.

[0146] If there are abnormal requisition records in a certain month, the corresponding abnormal marker can be retained so that the Long Short-Term Memory network model can read the abnormal state during the gating update process.

[0147] Read the hidden state of the previous period, the cell state of the previous period, and the time sequence input for each prediction period.

[0148] Specifically, the model initializes the hidden state and cell state at the beginning of the time window.

[0149] The initialization method can be a zero vector, or it can be the hidden state and cell state saved in the previous prediction cycle of the same type of aircraft material.

[0150] For each prediction period within the time window, read the time sequence input of the current period, the hidden state of the previous period, and the cell state of the previous period.

[0151] The hidden state of the previous period is used to retain short-term demand changes corresponding to the previous period, and the cell state of the previous period is used to retain long-term state information spanning months.

[0152] If the state of the previous period is empty, the initial state is used; if the state cache is abnormal, the corresponding hidden state and cell state are regenerated based on the historical input in the time window, and the state cache is updated.

[0153] Perform weighted calculations for the forget gate, input gate, and output gate respectively, and update the hidden state and cell state for the current time period.

[0154] Specifically, the forget gate reads the current time sequence input and the hidden state of the previous time period to determine the degree to which historical demand information in the cell state of the previous time period is retained.

[0155] The input gate reads the current time-series input and the hidden state of the previous time period to determine the extent to which new information for the current month has been written into the cell state.

[0156] The candidate cell state is formed by nonlinear mapping based on the current time-series input and the hidden state of the previous time period. The cell state of the current time period is updated based on the historical information corresponding to the forget gate and the new information corresponding to the input gate.

[0157] The output gate determines the current hidden state based on the current time sequence input, the hidden state of the previous time sequence, and the cell state of the current time sequence.

[0158] Changes in the number of operating devices, training hours, inspection month codes, model type, or supplier type can be adjusted by changing the corresponding gating weights through time-series input, so that the model state is updated monthly along the time window.

[0159] The hidden state of the current time period and the static prediction output are input into a fully connected layer to obtain the aircraft material demand prediction sequence.

[0160] Specifically, the fully connected layer receives the hidden state and static prediction output for the current time period, and fuses the static nonlinear relationship of the same month captured by the extreme gradient boosting tree model with the cross-month time series relationship captured by the long short-term memory network model to output the predicted value of aviation material demand for the month to be predicted.

[0161] When multiple months' forecasts need to be output, a rolling forecast method can be used. After obtaining the forecast value corresponding to the current month to be forecasted, the planned number of operating equipment, planned training hours, month code, model type, supplier type and equipment service life of the subsequent months are used as new inputs to recursively generate the forecast values ​​corresponding to the subsequent months.

[0162] If the predicted value for a certain category of aviation materials is less than 0, the corresponding demand prediction value can be adjusted to 0; if the predicted value exceeds the upper limit of the percentile determined based on historical demand data, the original prediction value can be retained and an anomaly review mark can be generated, so that the corresponding abnormal prediction status can be identified during the inventory parameter generation process.

[0163] In this embodiment, the model training, inference, and evaluation processes may also be disclosed.

[0164] Training data may include historical aircraft material demand and characteristics data from 2003 to 2020, validation data may include data from 2021 to 2022, and test data may include data from 2023.

[0165] During training, the spatiotemporal input vectors of historical months and the static prediction output are used as inputs, and the actual demand for aircraft materials in the corresponding month is used as the training objective. Hubel loss is used to update the parameters.

[0166] Huber loss uses a squared error relationship when the absolute value of the prediction error is less than a preset threshold, and a linear error relationship when the absolute value of the prediction error reaches the preset threshold.

[0167] Model parameters can be updated via backpropagation.

[0168] The validation data is used to select the number of training epochs, number of trees, learning rate, hidden layer dimension, time window, and random deactivation coefficient.

[0169] During inference, the saved extreme gradient boosting tree model parameters, long short-term memory network model parameters, field encoding rules, and standardized parameters are loaded, and a prediction sequence of aircraft material demand is formed based on the corresponding inputs.

[0170] During evaluation, mean absolute error, root mean square error, and mean absolute percentage error can be used as evaluation indicators. The evaluation results are used to adjust model parameters and input field processing rules.

[0171] After obtaining the aircraft material demand forecast sequence, it also includes: For consumable aviation materials, the number of operating equipment and the inspection month marker are read, and the theoretical demand intensity is determined based on the basic arrival rate, the number of operating equipment, and the inspection month marker.

[0172] Specifically, for consumable aircraft materials, the demand forecast sequence can be validated using a Poisson distribution.

[0173] The basic arrival rate is used to characterize the basic demand arrival level of a unit of operating equipment within a unit forecast period.

[0174] The theoretical demand intensity for a normal month is determined based on the basic arrival rate and the number of operating equipment in the corresponding month. When the corresponding month is an inspection month, the theoretical demand intensity is adjusted according to the inspection month marker and seasonal coefficient so that the concentrated maintenance demand corresponding to the inspection month is included in the theoretical demand intensity.

[0175] The months for inspection can include March, June, September, and December.

[0176] For one embodiment of oxygen mask-type aviation materials, the basic arrival rate can be taken as 0.085 pieces / unit / month, and the seasonal coefficient can be taken as 2.8.

[0177] A Poisson distribution is constructed based on the theoretical demand intensity, and the actual demand frequency in the requisition records is fitted to the Poisson distribution for a fit test.

[0178] Specifically, the theoretical frequency corresponding to the Poisson distribution is determined based on the theoretical demand intensity. The actual demand frequency in the historical requisition records is grouped according to the demand frequency to form the corresponding observation frequency. The chi-square test or likelihood test is performed on the observation frequency and the theoretical frequency.

[0179] The results of the fit test can include the fit test statistic, probability value, and the degree of dispersion corresponding to the number of demands.

[0180] The results of this fitting test are used to determine whether the aircraft material demand forecast sequence needs to be discretized.

[0181] When the fitting test results meet the preset matching conditions, the aircraft material demand forecast sequence is retained as the inventory calculation input.

[0182] Specifically, the preset matching conditions may include a chi-square test probability value greater than 0.05 and the ratio of the variance to the mean of the number of demands being within a preset discrete range.

[0183] In one embodiment, when the chi-square test statistic is 7.32 and the probability value is 0.292, it can be determined that the fitting test result corresponding to the Poisson distribution meets the preset fitting conditions.

[0184] When the fitting test result does not meet the preset fit condition, the theoretical demand intensity is discretized by using a negative binomial distribution to obtain the corrected aircraft material demand forecast sequence, which is then used as the input for inventory calculation.

[0185] Specifically, when the actual demand frequency exhibits a degree of dispersion exceeding a preset range relative to the Poisson distribution, a negative binomial distribution is used to characterize the corresponding excessively discrete demand.

[0186] In one embodiment, when the ratio of the variance to the mean of the number of demands is approximately 1.2 and exceeds a preset discrete range, a negative binomial distribution can be used for correction.

[0187] During the discrete correction process, the deviation between the observed frequency of the actual required number of times in the fit test and the theoretical frequency corresponding to the Poisson distribution can be read.

[0188] The deviation can be expressed as the difference between the observed frequency and the theoretical frequency, or as a standardized residual.

[0189] For each demand frequency, the number of months corresponding to the demand frequency in the historical records is counted and compared with the theoretical number of months corresponding to the Poisson distribution to form the corresponding deviation.

[0190] If the deviation is concentrated in the high-demand range of the inspection month, the corresponding discrete correction parameter can be determined based on the demand fluctuation of the inspection month; if the deviation is concentrated in the demand range of non-inspection months, the corresponding discrete correction parameter can be determined based on the demand fluctuation of ordinary months.

[0191] The deviations are grouped according to the inspection month label and the non-inspection month label, and the mean, variance and variance-to-mean ratio are calculated respectively to form the corresponding discrete correction parameters.

[0192] If the number of historical samples for a certain group is insufficient, historical samples corresponding to the same month in adjacent years can be merged, or the grouping parameters corresponding to the same type of aircraft materials can be used as the initial value.

[0193] When correcting the negative binomial distribution, the theoretical demand intensity is used as the mean parameter of the negative binomial distribution, and the degree of dispersion of the negative binomial distribution is determined according to the discrete correction parameter.

[0194] The discrete correction parameter corresponding to the month of inspection is used, and the discrete correction parameter corresponding to the month of non-inspection is used for non-inspection months.

[0195] The corrected demand distribution can be used to form a corrected aircraft material demand forecast sequence based on the corresponding distribution mean, target service level quantile, or upper bound of the forecast interval.

[0196] When the dispersion is within the preset range corresponding to the Poisson distribution, the original aircraft material demand forecast sequence can continue to be used; when the dispersion exceeds the preset abnormal range, an abnormality verification mark can be generated, and the corresponding forecast uncertainty information can be used in the subsequent safety stock processing.

[0197] S105, generate the forecast standard deviation based on the dispersion of the aircraft material demand forecast sequence, and determine the safety stock by combining the service level coefficient and the procurement lead time. Determine the reorder point based on the lead time forecast demand and the safety stock, and determine the order quantity based on the economic order quantity, the safety stock, the current inventory quantity, and the in-transit inventory quantity.

[0198] In this embodiment, the forecast standard deviation is formed based on the degree of uncertainty corresponding to the aircraft material demand forecast results, and the safety stock is determined by combining the importance level of the aircraft materials and the procurement lead time corresponding to the procurement channels.

[0199] Safety stock can be represented as: ; in, Indicates the first Safety stock for one month; Indicates the service level coefficient; This represents the standard deviation of the forecast within a single forecast period; This indicates the number of forecast cycles calculated by adjusting the procurement lead time.

[0200] Number of prediction periods Using the square root form in the determination of safety stock allows the forecast uncertainty corresponding to the increase in procurement lead time to be reflected in the safety stock.

[0201] Furthermore, in the Long Short-Term Memory network model, the randomly deactivated layer is kept in a sampling state, and multiple forward propagations are performed on the same month to be predicted to obtain multiple sampled predictions.

[0202] Specifically, when performing predictions after the model training is complete, the randomly deactivated layer is kept in a sampling state.

[0203] For the same month to be predicted, the same spatiotemporal input vector, static prediction output, historical hidden state, and cell state are input, and forward propagation is performed multiple times.

[0204] In one embodiment, 100 forward propagations can be performed to obtain 100 sampled predictions.

[0205] During each forward propagation, the random deactivation layer randomly masks the output of some neurons in the hidden layer, and the different sampled predictions thus reflect the prediction uncertainty corresponding to the relationship between the model parameters and the input time sequence.

[0206] If the number of operating devices, training hours, inspection month code, or supplier category changes significantly in the month to be predicted, the dispersion of multiple sampled predictions can change with the corresponding input.

[0207] The dispersion of the multiple sampled predictions within a single prediction period is calculated to obtain the prediction standard deviation.

[0208] Specifically, the standard deviation of multiple sampled forecasts for the same month to be predicted is calculated, and the obtained standard deviation is used as the forecast standard deviation for the corresponding single forecast period.

[0209] In a further embodiment, the sampling prediction quantity can be divided according to the model category and the supplier category.

[0210] When a single piece of aviation material serves multiple aircraft models or multiple suppliers' equipment, grouping labels corresponding to aircraft model categories and supplier categories can be retained across multiple sampled prediction quantities.

[0211] Group labels are derived from category values.

[0212] For each sampled prediction, you can read the model category weight and supplier category weight in the corresponding input window and assign the sampled prediction to the category group with the highest proportion. Alternatively, you can divide the corresponding sampled prediction into multiple category groups according to the category weight.

[0213] Calculate the standard deviation, range, and quantile interval for each model category group and supplier category group.

[0214] Standard deviation is used to form the standard deviation of category prediction, range can be used to identify outlier sampling, and quantile intervals can be used to determine the distribution range of the uncertainty of the corresponding category prediction.

[0215] If the number of samples in a certain category group is insufficient, the sampled forecasts of the same category in adjacent months can be combined for calculation; if the number of samples after merging is still insufficient, the dispersion of all sampled forecasts can be used as the initial value of the dispersion of the corresponding category.

[0216] The preset discrete threshold can be determined based on the historical prediction error distribution, for example, by using the 75th percentile of the standard deviation of historical sampling of similar aircraft materials.

[0217] When the dispersion of the sampled prediction for any category exceeds a preset dispersion threshold, the category prediction standard deviation is determined based on the sampled prediction for the corresponding category, and the category prediction standard deviation is used to determine the safety stock of the corresponding category of aviation materials.

[0218] When the dispersion of the sampled predictions for each category is within the preset dispersion range, the prediction standard deviation corresponding to all sampled predictions is used to determine the subsequent safety stock.

[0219] The importance level of aircraft materials is mapped to the service level coefficient, and the procurement channel category is mapped to the number of prediction cycles corresponding to the procurement lead time.

[0220] Specifically, the importance level of aircraft materials can be classified according to their safety impact, training disruption impact, and difficulty of replacement.

[0221] In one embodiment, first-importance aircraft materials can be mapped to a service level coefficient of 2.33, corresponding to a service level of 99%; second-importance aircraft materials can be mapped to a service level coefficient of 1.65, corresponding to a service level of 95%; and third-importance aircraft materials can be mapped to a service level coefficient of 1.28, corresponding to a service level of 90%.

[0222] The procurement channels can include imported parts, domestically produced parts, and locally processed parts.

[0223] In the example using a monthly forecasting period, when the lead time for imported parts is 90 days, it corresponds to 3 forecasting periods; when the lead time for domestically produced parts is 30 days, it corresponds to 1 forecasting period; and when the lead time for locally processed parts is 7 days, it can correspond to... One prediction cycle.

[0224] If the importance level data for aircraft materials is missing, a temporary importance level for aircraft materials can be generated based on the safety impact field and the training suspension impact field, and the updated importance level for aircraft materials will be adopted after the master data for aircraft materials is updated.

[0225] If procurement channel category data is missing, a temporary procurement lead time can be generated based on the average delivery cycle of historical purchase orders, and the updated procurement lead time can be adopted after the procurement master data is updated.

[0226] The safety stock is determined based on the service level coefficient, the forecast standard deviation, and the forecast period number.

[0227] Specifically, the safety stock is obtained by substituting the service level coefficient, the forecast standard deviation corresponding to a single forecast period, and the number of forecast periods corresponding to the procurement lead time into the safety stock calculation relationship.

[0228] The predicted period number is used in the determination of the safety stock in the form of a square root.

[0229] When using the category forecast standard deviation, the safety stock of the corresponding category of aircraft materials is determined based on the corresponding category forecast standard deviation; when using the forecast standard deviation corresponding to all sampled forecast quantities, the corresponding safety stock is determined based on the aircraft material identification.

[0230] The safety stock calculation result can be rounded to a decimal, and the result will be processed according to the minimum quantity unit corresponding to the aircraft material when the inventory ordering parameters are output.

[0231] The lead time forecast demand includes the forecast quantity of aircraft materials demand during the lead time period, and the reorder point is determined based on the forecast quantity of aircraft materials demand during the lead time period and the safety stock.

[0232] Specifically, the forecast of aircraft material demand during the lead time period can be formed by summing up the forecasts of aircraft material demand corresponding to each forecast period covered by the procurement lead time period.

[0233] If the aircraft material demand forecast sequence adopts the negative binomial distribution correction result, the aircraft material demand forecast during the lead time period is formed based on the corrected aircraft material demand forecast sequence.

[0234] The corresponding reorder point is determined based on the forecast of aircraft material demand and safety stock during the lead time period.

[0235] During inventory management, the current inventory level and the inventory in transit can be compared with the reorder point. When the corresponding inventory level is lower than the reorder point, the order quantity determination process begins; when the inventory level meets the inventory requirements corresponding to the reorder point, the current inventory level can be maintained and the corresponding inventory processing result can be output.

[0236] The target inventory level is determined based on the economic order quantity and the safety stock, and the available inventory level is determined based on the current inventory level and the inventory in transit.

[0237] Specifically, the current inventory quantity comes from the inventory ledger, while the inventory in transit quantity comes from purchase orders or pending receipt records.

[0238] Economic order quantity (EOQ) can be determined by existing rules of the inventory management system, or it can be determined based on ordering costs, holding costs, and annual demand.

[0239] Economic order quantity and safety stock together form the target inventory level, while current inventory and in-transit inventory together form the available inventory level.

[0240] When the target inventory level is higher than the available inventory level, the order quantity is generated based on the difference between the target inventory level and the available inventory level.

[0241] Specifically, when the target inventory level is higher than the available inventory level, the difference between the target inventory level and the available inventory level is used as the basis for determining the corresponding quantity of aviation materials to be procured.

[0242] When the target inventory level is not higher than the available inventory level, the current purchasing status can be maintained and a no-order flag can be output.

[0243] When the order quantity is less than the preset minimum purchase quantity, the preset minimum purchase quantity is output; when the order quantity reaches the preset minimum purchase quantity, the order quantity is output.

[0244] Specifically, the preset minimum purchase quantity can be set based on the supplier's minimum order requirement or the minimum packaging quantity for aircraft materials.

[0245] When the generated order quantity is lower than the corresponding preset minimum purchase quantity, the preset minimum purchase quantity is used as the actual purchase quantity; when the generated order quantity reaches the preset minimum purchase quantity, the generated order quantity is used as the purchase quantity.

[0246] The final order quantity can also be rounded up according to the smallest procurement packaging unit corresponding to aviation materials.

[0247] The final output of inventory ordering parameters may include aircraft material identification, forecast month, lead time forecast demand, safety stock, reorder point, current inventory quantity, in-transit inventory quantity, order quantity, and calculation time.

[0248] In this embodiment, a detectable implementation method can also be set.

[0249] Each time reorder points and order quantities are generated, the spatiotemporal input vector version, long-term trend term, monthly pulse seasonal term, random residual term, static prediction output, hidden state summary, cell state summary, aircraft material demand prediction sequence, sampled prediction standard deviation, Poisson distribution or negative binomial distribution fitting result, safety stock, reorder points and order quantities can be recorded.

[0250] The above operation records can be saved in the operation log of the inventory management system, or they can be read through the result verification interface.

[0251] When data conflicts occur between the aircraft material demand forecast sequence, distribution correction results, and inventory parameters, the result verification interface can output a conflict flag and retain the original forecast value, corrected forecast value, and inventory calculation parameters for maintenance personnel to review.

[0252] The following description, using specific examples, elaborates on this embodiment.

[0253] In one example, for projector components, enter the monthly requisition records, equipment operation logs, and training hour records from 2003 to 2024.

[0254] The time-series demand is obtained based on the monthly quantity of aviation materials requisitioned for composite projectors. Read the number of running devices The monthly training hours of the equipment are arranged into a training intensity matrix corresponding to 34 equipment for 12 months. Quarterly inspection cycle codes and annual cycle codes are constructed, and the model category, supplier category and equipment service age are converted into corresponding category values ​​and service age values ​​to form a spatiotemporal input vector.

[0255] right Perform local weighted regression seasonal trend decomposition.

[0256] In the example of projector aircraft parts, the long-term trend term can change from 4 pieces / year in 2003 to 28 pieces / year in 2024; the inspection month pulse seasonal term forms demand peaks in March, June, September and December; the random residual term records deviations caused by sudden failures or temporary maintenance, and the standard deviation of historical random residuals can be used as reference information to identify the degree of anomalies during model training.

[0257] After logarithmically transforming the number of operating equipment, regression fitting is performed with the long-term trend term, and the equipment scale influence coefficient corresponding to projector aviation materials can be taken as 0.85.

[0258] The trend forecast is obtained by taking the logarithmic value of the number of operating devices in the month to be predicted and the corresponding constant term, and then the trend forecast, the seasonal pulse term of the month to be checked and the random residual term are incorporated into the spatiotemporal input vector.

[0259] Input the spatiotemporal input vector into the extreme gradient boosting tree model.

[0260] If March 2024 is the inspection month, the tree node can read fields such as the inspection month code, the number of running devices, training hours, and device type. Based on the node splitting threshold, the samples for that month are assigned to the corresponding leaf nodes, and the outputs of each regression leaf node are accumulated to obtain the static prediction output.

[0261] The spatiotemporal input vectors of the past 12 months and the corresponding static prediction outputs are combined to form the temporal input. The long short-term memory network model recursively updates the corresponding cell states with 64-dimensional hidden states, and outputs the predicted sequence of aircraft material demand through a fully connected layer.

[0262] In an example record, the predicted value for January 2024 could be 22 items, and the actual number of items used was 23; the predicted value for March 2024 could be 28 items, and the actual number of items used was 26; the predicted value for the whole year could be 285 items, and the actual number of items used was 278.

[0263] The above data is used to illustrate the data processing procedure and prediction result output format of this embodiment.

[0264] In another example, demand distribution validation and inventory parameter generation are performed for oxygen mask materials.

[0265] After obtaining the aircraft material demand forecast sequence, the number of operating equipment and the inspection month marker are read, and the theoretical demand intensity is determined based on the basic arrival rate, the number of operating equipment and the inspection month marker.

[0266] Among them, the basic arrival rate of oxygen mask aviation materials can be taken as 0.085 pieces / unit / month, and the seasonal coefficient can be taken as 2.8.

[0267] A Poisson distribution is constructed based on the theoretical demand intensity, and a fit test is performed on the actual demand frequency.

[0268] When the chi-square test statistic is 7.32 and the probability value is 0.292, it can be determined that the corresponding fit test result meets the preset fit condition.

[0269] When the ratio of the variance to the mean of the number of demands is approximately 1.2 and exceeds the preset dispersion range, a negative binomial distribution can be used to correct the corresponding excessive dispersion.

[0270] Oxygen masks are classified as the most important type of aircraft material, with a service level coefficient of [missing information]. 2.33 is acceptable.

[0271] The lead time for imported parts procurement is 90 days. When the forecast period is in months, the corresponding forecast period is... Take 3.

[0272] After performing multiple forward propagations on the same month to be predicted using the Monte Carlo random inactivation method, the prediction standard deviation for a single prediction period is obtained. Item.

[0273] The safety stock is determined based on a service level coefficient of 2.33, a forecast standard deviation of 3.2 units, and a forecast period of 3. The safety stock is approximately 12.9 units.

[0274] When the lead time forecast demand is 45 units, the reorder point determined based on the lead time forecast demand and safety stock is approximately 57.9 units, which can be output as 58 units after processing according to the minimum quantity unit in inventory management.

[0275] When the economic order quantity is 30 units, the current inventory is 15 units, and the in-transit inventory is 10 units, the target inventory determined based on the economic order quantity and safety stock is approximately 42.9 units, the available inventory determined based on the current inventory and in-transit inventory is 25 units, and the corresponding order quantity generated is approximately 17.9 units.

[0276] After processing according to the minimum purchasing unit, the order quantity can be output as 18 pieces.

[0277] The above values ​​are used to illustrate the process of determining the forecast standard deviation, safety stock, reorder point, and order quantity. In actual application, the corresponding parameters are determined based on historical data of aviation materials, the importance level of aviation materials, procurement channels, and inventory status.

[0278] For the purpose of simplicity, the method steps disclosed in the above embodiments are described as a series of actions. However, those skilled in the art should understand that the embodiments of the present invention are not limited to the described order of actions, because according to the embodiments of the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all preferred embodiments, and the actions involved are not necessarily essential to the embodiments of the present invention.

[0279] like Figure 2 As shown, the present invention also provides a system for predicting aircraft material demand and generating inventory ordering parameters based on spatiotemporal nonstationarity, comprising: The input vector forming module 201 is configured to extract the time-series demand quantity and spatiotemporal input vector from the aircraft material requisition record, the number of operating equipment, the monthly training hours of the equipment, the month number, the aircraft type, the supplier type, and the service life of the equipment; The non-stationary trend processing module 202 is configured to perform local weighted regression seasonal trend decomposition on the time-series demand to obtain a long-term trend term, a check month pulse seasonal term and a random residual term, perform logarithmic transformation on the number of operating equipment and regress and fit it with the long-term trend term to obtain a trend prediction, and incorporate the trend prediction, the check month pulse seasonal term and the random residual term into the spatiotemporal input vector. Tree prediction module 203 is configured to input the spatiotemporal input vector into an extreme gradient boosting tree model and obtain a static prediction output based on tree node splitting; The time series prediction module 204 is configured to input the static prediction output and the spatiotemporal input vector arranged in time windows into the long short-term memory network model, and perform gating updates based on the historical hidden state and cell state to obtain the aircraft material demand prediction sequence. The inventory parameter generation module 205 is configured to generate a forecast standard deviation based on the dispersion of the aircraft material demand forecast sequence, determine the safety stock in combination with the service level coefficient and the procurement lead time, determine the reorder point based on the lead time forecast demand and the safety stock, and determine the order quantity based on the economic order quantity, the safety stock, the current inventory quantity and the in-transit inventory quantity.

[0280] It is worth noting that although only some basic functional modules are disclosed in the embodiments of this invention, it does not mean that the composition of this system is limited to the above-mentioned basic functional modules. On the contrary, based on the above-mentioned basic functional modules, those skilled in the art can arbitrarily add one or more functional modules in combination with existing technology to form an infinite number of embodiments or technical solutions. That is to say, this system is open rather than closed. The fact that this embodiment only discloses a few basic functional modules should not be considered as the scope of protection of this invention being limited to the disclosed basic functional modules. At the same time, for the convenience of description, the above devices are described separately according to their functions as various units and modules. Of course, in implementing this invention, the functions of each unit and module can be implemented in one or more software and / or hardware.

[0281] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for forecasting aircraft material demand and generating inventory ordering parameters based on spatiotemporal nonstationarity, characterized in that, include: Extract time-series demand and spatiotemporal input vectors from aircraft material requisition records, number of operating equipment, monthly training hours of equipment, month serial number, aircraft type, supplier type, and equipment service life; The time-series demand is subjected to local weighted regression seasonal trend decomposition to obtain a long-term trend term, an inspection month pulse seasonal term, and a random residual term. The number of operating equipment is logarithmically transformed and regressed with the long-term trend term to obtain a trend prediction. The trend prediction, the inspection month pulse seasonal term, and the random residual term are incorporated into the spatiotemporal input vector. The spatiotemporal input vector is input into the extreme gradient boosting tree model, and a static prediction output is obtained based on tree node splitting. The static prediction output and the spatiotemporal input vector arranged in time windows are input into the long short-term memory network model. Gating updates are performed based on the historical hidden state and cell state to obtain the aircraft material demand prediction sequence. The standard deviation of the forecast is generated based on the dispersion of the aircraft material demand forecast sequence. The safety stock is determined by combining the service level coefficient and the procurement lead time. The reorder point is determined based on the lead time forecast demand and the safety stock. The order quantity is determined based on the economic order quantity, the safety stock, the current inventory quantity, and the inventory in transit quantity.

2. The method for forecasting aircraft material demand and generating inventory ordering parameters based on spatiotemporal nonstationarity as described in claim 1, characterized in that, The time-series demand and spatiotemporal input vector are extracted from aircraft material requisition records, number of operating equipment, monthly training hours of equipment, month serial number, aircraft type, supplier type, and equipment service age. Further, this includes: The time-series demand is formed by aggregating the aircraft material identification and the monthly aggregated requisition quantity; Write the number of operating devices in the same month into the time dimension data; Arrange the monthly training hours for each device into training intensity data by device and month; Convert the month serial number into quarterly inspection cycle code and annual cycle code; Convert the model category and supplier category into category values, and convert the equipment service life into service life values; The spatiotemporal input vector is obtained by concatenating the time dimension data, the training intensity data, the quarterly inspection cycle code, the annual cycle code, the category value, and the service age value.

3. The method for forecasting aircraft material demand and generating inventory ordering parameters based on spatiotemporal nonstationarity as described in claim 1, characterized in that, The time-series demand is subjected to a locally weighted regression seasonal trend decomposition to obtain a long-term trend term, a seasonal pulse term for the month of inspection, and a random residual term, further including: The time-series demand is locally weighted smoothed using a preset seasonal window to separate candidate seasonal sequences that fluctuate repeatedly with the months. After deducting the candidate seasonal series from the time-series demand, the remaining series is locally weighted smoothed to obtain the long-term trend term; The random residual term is obtained by subtracting the long-term trend term and the candidate seasonal sequence from the time-series demand. Pulse amplitude preservation is performed on the candidate seasonal sequence for the period in which the month under inspection is located, and smoothing amplitude constraint is performed on the period in which the month outside the inspection is located, to obtain the pulse seasonal term for the month under inspection.

4. The method for forecasting aircraft material demand and generating inventory ordering parameters based on spatiotemporal nonstationarity as described in claim 1, characterized in that, Logarithmically transform the number of operating devices and regress it with the long-term trend term to obtain a trend prediction. The trend prediction, the pulse seasonal term of the inspection month, and the random residual term are then incorporated into the spatiotemporal input vector, further including: Read the number of operating devices and the long-term trend item for each historical month; After performing a logarithmic transformation on the number of operating equipment, regression fitting was performed with the long-term trend term to obtain the equipment scale influence coefficient and constant term distinguished by aviation material category; The trend forecast quantity is determined based on the equipment scale influence coefficient, the logarithm of the number of operating equipment in the month to be predicted, and the constant term. The trend prediction, the seasonal pulse term for the month of inspection, and the random residual term are incorporated into the spatiotemporal input vector.

5. The method for forecasting aircraft material demand and generating inventory ordering parameters based on spatiotemporal nonstationarity as described in claim 1, characterized in that, The spatiotemporal input vector is input into the extreme gradient boosting tree model, and a static prediction output is obtained based on tree node splitting, further including: At each node of the regression tree, read the equipment size field, training hour field, month code field, model category field, supplier category field, and equipment service life field from the spatiotemporal input vector; Samples are assigned to leaf nodes based on node splitting thresholds; The outputs of each regression leaf node are summed into a tree-integrated output. The tree integration output is used as the static prediction output and input into the long short-term memory network model.

6. The method for forecasting aircraft material demand and generating inventory ordering parameters based on spatiotemporal nonstationarity as described in claim 1, characterized in that, The static prediction output and the spatiotemporal input vector arranged in time windows are input into a long short-term memory network model. Gating updates are performed based on historical hidden states and cell states to obtain the aircraft material demand prediction sequence, which further includes: The spatiotemporal input vectors are arranged in a preset time window, and the spatiotemporal input vectors in each window are concatenated with the static prediction output of the same time period to form a time-series input; Read the hidden state of the previous period, the cell state of the previous period, and the time sequence input for each prediction period; Perform weighted calculations using the forget gate, input gate, and output gate respectively, and update the hidden state and cell state for the current time period. The hidden state of the current time period and the static prediction output are input into a fully connected layer to obtain the aircraft material demand prediction sequence.

7. The method for forecasting aircraft material demand and generating inventory ordering parameters based on spatiotemporal nonstationarity as described in claim 1, characterized in that, After obtaining the aircraft material demand forecast sequence, it also includes: For consumable aircraft materials, the number of operating equipment and the inspection month marker are read, and the theoretical demand intensity is determined based on the basic arrival rate, the number of operating equipment and the inspection month marker; A Poisson distribution is constructed based on the theoretical demand intensity, and the actual demand frequency in the requisition records is fitted to the Poisson distribution for a fitting test. When the fitting test results meet the preset matching conditions, the aircraft material demand forecast sequence is retained as the inventory calculation input. When the fitting test result does not meet the preset fit condition, the theoretical demand intensity is discretized by using a negative binomial distribution to obtain the corrected aircraft material demand forecast sequence, which is then used as the input for inventory calculation.

8. The method for forecasting aircraft material demand and generating inventory ordering parameters based on spatiotemporal nonstationarity as described in claim 1, characterized in that, The forecast standard deviation is generated based on the dispersion of the aforementioned aircraft material demand forecast sequence, and safety stock is determined by combining the service level coefficient and procurement lead time, further including: In the Long Short-Term Memory network model, the random deactivation layer is kept in a sampling state, and multiple forward propagations are performed on the same month to be predicted to obtain multiple sampled predictions. Calculate the dispersion of the multiple sampled predictions within a single prediction period to obtain the prediction standard deviation; The importance level of aircraft materials is mapped to the service level coefficient, and the procurement channel category is mapped to the number of prediction cycles corresponding to the procurement lead time. The safety stock is determined based on the service level coefficient, the forecast standard deviation, and the forecast period number.

9. The method for forecasting aircraft material demand and generating inventory ordering parameters based on spatiotemporal nonstationarity as described in claim 8, characterized in that, Determining the safety stock based on the service level coefficient, the forecast standard deviation, and the number of forecast periods further includes: The number of forecast periods is used in the determination of the safety stock in the form of a square root. The safety stock is determined by the service level coefficient, the forecast standard deviation, and the number of forecast periods. The lead time forecast demand includes the forecast quantity of aircraft material demand during the lead time, and the reorder point is determined based on the forecast quantity of aircraft material demand during the lead time and the safety stock. The target inventory level is determined based on the economic order quantity and the safety stock, and the available inventory level is determined based on the current inventory level and the inventory in transit. When the target inventory level is higher than the available inventory level, the order quantity is generated based on the difference between the target inventory level and the available inventory level; When the order quantity is less than the preset minimum purchase quantity, the preset minimum purchase quantity is output. When the order quantity reaches the preset minimum purchase quantity, the order quantity is output.

10. A system for forecasting demand for aviation materials and generating inventory ordering parameters based on spatiotemporal nonstationarity, characterized in that, include: The input vector formation module is configured to extract time-series demand quantities and spatiotemporal input vectors from aircraft material requisition records, number of operating equipment, monthly training hours of equipment, month serial number, aircraft type, supplier type, and equipment service age. The non-stationary trend processing module is configured to perform local weighted regression seasonal trend decomposition on the time-series demand to obtain a long-term trend term, a check month pulse seasonal term, and a random residual term; perform logarithmic transformation on the number of operating equipment and regress and fit it with the long-term trend term to obtain a trend prediction; and incorporate the trend prediction, the check month pulse seasonal term, and the random residual term into the spatiotemporal input vector. The tree prediction module is configured to input the spatiotemporal input vector into an extreme gradient boosting tree model and obtain a static prediction output based on tree node splitting; The time-series prediction module is configured to input the static prediction output and the spatiotemporal input vector arranged in time windows into the long short-term memory network model, and perform gating updates based on historical hidden states and cell states to obtain the aircraft material demand prediction sequence. The inventory parameter generation module is configured to generate a forecast standard deviation based on the dispersion of the aircraft material demand forecast sequence, determine the safety stock by combining the service level coefficient and the procurement lead time, determine the reorder point based on the lead time forecast demand and the safety stock, and determine the order quantity based on the economic order quantity, the safety stock, the current inventory quantity, and the in-transit inventory quantity.

Images