Material demand prediction method based on ARIMA model and LSTM model

Abstract

The invention discloses a material demand prediction method based on an ARIMA model and an LSTM model, and the method comprises the steps: constructing a time-serialized material demand data set, and carrying out the data preprocessing; time sequence analysis is carried out, linear trend modeling is carried out by using an ARIMA model, a preliminary prediction value is obtained, and a prediction residual error is calculated; constructing an LSTM neural network to perform nonlinear feature learning on the residual error, predicting a future residual error value, determining an optimal model by using an LSTM model evaluation index, and predicting future annual material consumption by using the optimal model; determining a calculation range of an annual purchase quantity, and establishing an annual recommendation purchase model; calculating an annual purchase quantity recommendation value of a plurality of years in the future, and forming an annual purchase schedule; and setting an early warning inventory value, and monitoring the inventory in real time. Compared with a traditional method depending on a single model, the method can more comprehensively capture tendency and volatility in material demands, and adapts to changes of material consumption modes in complex construction scenes.

Description

Technical Field

[0001] This invention belongs to the field of material procurement technology, and relates to the prediction of material demand in the construction process of LNG ships, specifically a material demand prediction method based on ARIMA model and LSTM model. Background Technology

[0002] In the construction of LNG (liquefied natural gas) vessels, the demand for materials is characterized by strong cyclicality, diverse types, large quantities, and complex fluctuations, encompassing various components such as pipelines, valves, welding materials, and insulation materials. Due to the multi-stage, long-cycle construction tasks and the lengthy material supply chain, the timeliness and accuracy of material supply directly impact the efficiency of construction plan execution and overall construction costs. Therefore, scientific and accurate material demand forecasting methods have become a crucial aspect of LNG vessel construction. Currently, the material demand forecasting methods commonly used in the industry are mainly based on manual experience or rely on traditional statistical models, such as linear regression, moving averages, or ARIMA time series analysis methods. While these methods have some practicality in scenarios with strong regularity, the forecast results are prone to deviation when facing the complex background of multiple parallel processes, frequent task changes, and unstable material usage rhythms in actual ship construction. Furthermore, a single model cannot fully characterize the multidimensional changes in material demand, especially its weak ability to identify sudden fluctuations and long-term trends. ARIMA models, as a classic time series forecasting method, are suitable for modeling linear trends and periodic changes in data and can handle stationary data well. However, their prediction accuracy is often limited when dealing with material demand containing significant nonlinear or abrupt characteristics. In contrast, LSTM (Long Short-Term Memory) models excel in handling nonlinear relationships in sequential data, capturing deep-seated features and long-term dependency structures in historical data. However, if the actual business logic and linear patterns of material usage are ignored, their prediction results are also prone to deviation. Summary of the Invention

[0003] Purpose of the invention: In order to overcome the shortcomings of the existing technology, a material demand forecasting method based on ARIMA model and LSTM model is provided.

[0004] Technical Solution: To achieve the above objectives, this invention provides a material demand forecasting method based on ARIMA and LSTM models, comprising the following steps:

[0005] S1: Based on the construction tasks and BOM data of LNG ship construction, collect historical procurement and consumption data of materials, construct a time-series material demand dataset, and perform data preprocessing.

[0006] S2: Perform time series analysis on the preprocessed material demand dataset to identify whether there are periodic patterns or trends in historical procurement and consumption data. Based on the material requisition records, use the ARIMA model to model linear trends, obtain preliminary predicted values ​​and calculate prediction residuals. Select the optimal order p, d, and q of the ARIMA model through the AIC criterion and complete the model fitting.

[0007] S3: Using the residual sequence calculated in step S2 as input, construct an LSTM neural network to learn the nonlinear features of the residuals, predict future residual values, make preliminary predictions of material consumption over the next few years, adjust the parameters based on the prediction results of material consumption over the next few years, use the LSTM model evaluation index to determine the optimal model, and use the optimal model to predict the material consumption of the future years.

[0008] S4: Based on historical procurement and consumption data, analyze changes in procurement demand, determine the calculation range for annual procurement quantities, construct procurement rules, and establish an annual recommended procurement model for procurement planning.

[0009] S5: Based on the annual material consumption predicted in step S3, and combined with the recommended procurement model in step S4, calculate the recommended annual procurement amount for the next few years and form an annual procurement plan.

[0010] S6: Set an early warning inventory value and monitor the inventory level in real time.

[0011] Furthermore, the historical procurement and consumption data in step S1 includes: material procurement records, construction requisition records, project construction schedule, material inventory records, BOM configuration data, construction team operation information, warehousing and circulation information, and other auxiliary attributes strongly related to material usage.

[0012] Furthermore, the method for constructing a time-series material demand dataset in step S1 includes: after completing the historical data collection, aggregating it into a continuous data sequence with a uniform time granularity (such as "week" or "ten-day period"), and generating the original time-series data of the target material by aggregating according to the uniform time granularity.

[0013]

[0014] In the formula: Indicates the first Requisition records within the period; This represents the set of dates covered by the period; This represents the periodic consumption; this sequence constitutes the direct input for ARIMA modeling and also serves as the target variable for LSTM training samples.

[0015] Data preprocessing includes data cleaning, missing data imputation, standardization, and smoothing to reduce the impact of short-term fluctuations on modeling, particularly for sequences. The following preprocessing procedure is performed:

[0016] A1: Missing Value Imputation: Blank segments in the construction records are filled using linear interpolation or periodic mean method. The missing points are estimated as follows:

[0017] If time points are missing If two adjacent points are known, then:

[0018]

[0019] If the missing time period is long, it is supplemented using the average value of the same historical period:

[0020]

[0021] In the formula: The period length (e.g., 7 indicates a period of one week) is suitable for periodic construction tasks.

[0022] The imputation result replaces the missing positions in the original sequence and updates. ;

[0023] A2: Outlier Correction and Sequence Smoothing: Performing moving average smoothing on detected abrupt changes:

[0024]

[0025] In the formula: The width is the window width, typically 3 to 5; for points with abrupt changes, if:

[0026]

[0027] If it is identified as an outlier, it is corrected using the neighboring mean:

[0028]

[0029] Smoothed sequence More suitable for stationarity testing in ARIMA fitting;

[0030] A3: Standardization Processing: To eliminate differences in numerical scales among different material categories or different engineering stages, and to unify training scales, [the following steps are taken]. Sequence Z-score normalization:

[0031]

[0032] In the formula: This is the historical average over a period of time. Standard deviation;

[0033] Received It can be used for both ARIMA modeling and LSTM training. The standardized parameters should be recorded for future reconstruction of the predicted values.

[0034] A4: Training Sample Construction: Based on the standardized sequences processed above. Construct input-output sliding window sample pairs for the prediction model:

[0035]

[0036] In the formula: For input windows of ARIMA or LSTM models; Predict the target value for the next cycle.

[0037] Furthermore, in step S2, based on several years of construction and consumption data, the periodicity and trend of historical procurement and consumption data are judged through autocorrelation plots and partial autocorrelation plots to initially determine the correlation order p, d, and q of the ARIMA model, i.e., ARIMA(p,d,q); where p is the order of the autoregressive term, q is the order of the moving average term, and d is the order of the difference.

[0038] Furthermore, step S3 involves adjusting the parameters based on the predicted annual material requisition amounts for the next few years, including the following steps:

[0039] B1: Divide the dataset into a training set and a test set. Train the model on the training set and predict the values ​​on the test set. Then, use the mean squared error or root mean square error to compare the actual values ​​with the predicted values. Compare the difference between the predicted values ​​and the actual values, calculate the prediction error, and evaluate the accuracy of the prediction.

[0040] B2: Based on the prediction error results, adjust the order of the autoregressive term, the order of the differencing, and the parameters of the moving average term in the ARIMA model.

[0041] Furthermore, step S3 utilizes the AIC model evaluation index to determine the optimal model, including the following steps:

[0042] C1: Calculate the AIC value of the ARIMA model at each order;

[0043] C2: Compare the AIC values ​​of the models at different orders, and select the order that minimizes the AIC value as the optimal order of the model.

[0044] Further, step C1 specifically includes:

[0045] C1-1: Preliminary Order Identification: Based on the material consumption time series data constructed in step S2 The autocorrelation function (ACF) and partial autocorrelation function (PACF) are used to plot graphs, analyze the lag characteristics of the series, and preliminarily identify the possible order of the autoregressive term. and the order of the moving average term And determine whether it is necessary to perform a differential test. Order differencing is used to make the sequence stationary, forming a set of candidate model combinations:

[0046]

[0047] Supplement: Time series data of material consumption constructed based on step S2 The autocorrelation function (ACF) and partial autocorrelation function (PACF) are used to plot graphs, analyze the lag characteristics of the series, and preliminarily identify the possible order of the autoregressive term. and the order of the moving average term And determine whether a differential test is needed. The order difference is used to make the sequence stationary, forming a set of candidate model combinations. Where: This is the maximum order search boundary, usually set to 3~5.

[0048] C1-2: Fitting candidate models one by one: For each group in the model set... The ARIMA model was fitted to each combination separately to obtain the log-likelihood function value for each group. and residual sum of squares ;

[0049] C1-3: AIC Value Calculation and Sorting:

[0050] Based on the Akaike Information Criterion (AIC), the fitting effect of each model is comprehensively evaluated, and its AIC value is calculated:

[0051]

[0052] Or, the equivalent form is:

[0053]

[0054] In the formula: The length of the time series; This represents the number of free parameters in the model (including constant terms). It is the sum of squared residuals; Likelihood function value.

[0055] Further, step C2 specifically includes:

[0056] C2-1: Optimal Model Selection: Sort all candidate models by their AIC values ​​and select the combination with the lowest AIC value as the optimal model structure.

[0057]

[0058] As an information criterion, AIC considers not only the model fitting accuracy (the smaller the residual, the better) but also the model complexity (the more parameters, the greater the penalty). Therefore, its minimum value corresponds to the optimal structure that balances accuracy and generalization ability.

[0059] C2-2: Optimal Model Fitting and Prediction

[0060] Based on the selected The structure is remodeled completely on the training set, and the prediction residuals are passed to the LSTM residual compensation module to complete the subsequent fusion prediction.

[0061] Furthermore, in step S4, it is assumed that the quantity purchased each time is not less than M0. Based on the current annual material quantity P, the future annual material consumption Pi (i = 1, 2, ... i ... n), the average procurement cycle Ta, and the shelf life T, a formula for calculating the annual procurement quantity M is established. The procurement quantity M satisfies the following formula:

[0062]

[0063] in, Minimum order quantity for a single purchase (contract or supply chain constraints). The quantity of the target material currently available in inventory; No. The annual forecast of material demand (output by step S3); This represents the safety stock value, set at 1.2 to 1.5 times the average annual projected consumption. This indicates the maximum number of procurement cycles within the permissible shelf life of the materials; The shelf life of materials is in years. The average procurement cycle for materials, in years.

[0064] Furthermore, the early warning inventory value T1 in step S6 satisfies the following formula:

[0065]

[0066] in, This represents the average consumption of materials per unit of time. It is the standard deviation of material consumption per unit time, used to reflect demand volatility; The average delivery cycle of goods from order placement to warehousing; The service level coefficient is determined by the required service confidence level. Commonly used values ​​are k=1.28 (90% service level), k=1.65 (95% service level), and k=2.05 (98% service level). This indicates the safety warning inventory level for the corresponding current material category.

[0067] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0068] 1. In this invention, the material demand prediction method integrates ARIMA and LSTM models. First, the linear trend of historical material consumption data is modeled using the ARIMA model, and then the residuals are input into the LSTM neural network for nonlinear feature learning. This achieves multi-dimensional prediction of material demand. Compared with traditional methods that rely on a single model, it can more comprehensively capture the trends and fluctuations in material demand and adapt to changes in material consumption patterns under complex construction scenarios.

[0069] 2. In this invention, by constructing a time-series dataset driven by the Bill of Materials (BOM) and construction tasks, and combining it with procurement rules to build an annual recommended procurement model, the automatic generation and optimized configuration of annual procurement plans for multiple future construction cycles are achieved. This invention improves the systematic nature and responsiveness of procurement planning, helps reduce material shortages or excessive stockpiling, and enhances the collaborative efficiency of material support and the rationality of inventory structure during LNG vessel construction. Attached Figure Description

[0070] Figure 1 This is a schematic flowchart of the method of the present invention;

[0071] Figure 2 A comparison chart of the predicted trends from the three models;

[0072] Figure 3 The graph shows the trend of the response of the three models to abnormal fluctuations. Detailed Implementation

[0073] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading this invention, any modifications of the invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0074] Example 1:

[0075] like Figure 1 As shown, this embodiment provides a material demand forecasting method based on ARIMA and LSTM models, which is applied to material demand forecasting in the LNG ship construction process. The specific steps include:

[0076] S1: Based on the construction tasks and BOM data of LNG ship construction, collect historical procurement and consumption data of materials, construct a time-series material demand dataset, and perform data preprocessing.

[0077] Historical procurement and consumption data includes: material procurement records, construction requisition records, project construction schedules, material inventory records, BOM configuration data, construction team operation information, warehousing and circulation information, and other auxiliary attributes strongly related to material usage. Specifically, this includes:

[0078] Material procurement data includes the quantity, time, unit price, supplier information, and delivery cycle of the target material in each project over the years.

[0079] Material consumption data includes the actual quantity used daily or weekly at each construction stage, the date of use, the construction task number and construction process, the material code, the unit conversion factor, and the working group using the material.

[0080] Mapping construction tasks to BOM data by mapping task nodes in the construction schedule With the corresponding BOM item Establish a mapping relationship and calculate the standard material demand:

[0081]

[0082] In the formula: The amount of work involved in the task; This is the material allocation coefficient corresponding to the unit quantity of the project.

[0083] Historical inventory records, including material receipt time, inventory change records, inventory upper limit settings, and early warning inventory thresholds, are used to dynamically assess in-transit inventory and allocation needs.

[0084] Auxiliary feature information, including temperature, humidity, holiday arrangements, shift work rhythm, concentrated shift change nodes, and special event labels during construction, is used as the input feature dimension of LSTM.

[0085] Methods for constructing time-series material demand datasets include: after completing historical data collection, aggregating it into a continuous data sequence with a uniform time granularity (such as "week" or "ten-day period"), and then aggregating it according to the uniform time granularity to generate the original time-series data of the target materials.

[0086]

[0087] In the formula: Indicates the first Requisition records within the period; This represents the set of dates covered by the period; This represents the periodic consumption; this sequence constitutes the direct input for ARIMA modeling and also serves as the target variable for LSTM training samples.

[0088] Data preprocessing includes data cleaning, missing data imputation, standardization, and smoothing to reduce the impact of short-term fluctuations on modeling, particularly for sequences. The following preprocessing procedure is performed:

[0089] A1: Missing Value Imputation: Blank segments in the construction records are filled using linear interpolation or periodic mean method. The missing points are estimated as follows:

[0090] If time points are missing If two adjacent points are known, then:

[0091]

[0092] If the missing time period is long, it is supplemented using the average value of the same historical period:

[0093]

[0094] In the formula: The period length (e.g., 7 indicates a period of one week) is suitable for periodic construction tasks.

[0095] The imputation result replaces the missing positions in the original sequence and updates. ;

[0096] A2: Outlier Correction and Sequence Smoothing: Performing moving average smoothing on detected abrupt changes:

[0097]

[0098] In the formula: The width is the window width, typically 3 to 5; for points with abrupt changes, if:

[0099]

[0100] If it is identified as an outlier, it is corrected using the neighboring mean:

[0101]

[0102] Smoothed sequence More suitable for stationarity testing in ARIMA fitting;

[0103] A3: Standardization Processing: To eliminate differences in numerical scales among different material categories or different engineering stages, and to unify training scales, [the following steps are taken]. Sequence Z-score normalization:

[0104]

[0105] In the formula: This is the historical average over a period of time. Standard deviation;

[0106] Received It can be used for both ARIMA modeling and LSTM training. The standardized parameters should be recorded for future reconstruction of the predicted values.

[0107] A4: Training Sample Construction: Based on the standardized sequences processed above. Construct input-output sliding window sample pairs for the prediction model:

[0108]

[0109] In the formula: For input windows of ARIMA or LSTM models; Predict the target value for the next cycle.

[0110] This process can be further extended to include multi-dimensional feature inputs, such as adding variables like construction progress, number of work teams, climate factors, and holidays, resulting in: ;

[0111] The above steps result in two structured datasets:

[0112] Standardized stationary time series inputs are used for ARIMA model training;

[0113] : Sliding window structured sample pairs, used for LSTM residual learning.

[0114] The above process completes the derivation and preprocessing flow from "raw material procurement and consumption data" to "structured sequence input for dual-model fusion prediction". It is the key foundational link for the high-precision prediction of this invention and serves as the input basis for model training and residual learning in steps S2 and S3.

[0115] S2: Perform time series analysis on the preprocessed material demand dataset to identify whether there are periodic patterns or trends in historical procurement and consumption data. Based on the material requisition records, use the ARIMA model to model linear trends, obtain preliminary predicted values ​​and calculate prediction residuals. Select the optimal order p, d, and q of the ARIMA model through the AIC criterion and complete the model fitting.

[0116] The ARIMA model is suitable for modeling linear trends and periodic disturbances, and its mathematical form is as follows:

[0117]

[0118] In the formula: Lag operator ; These are the AR and MA partial polynomials, respectively; It is the difference order; This is the white noise term.

[0119] By minimizing the AIC / BIC information criterion, the optimal (p,d,q) combination is selected to obtain the trend forecast value: And define the prediction residual as:

[0120]

[0121] Based on several years of construction and consumption data, the periodicity and trend of historical procurement and consumption data are judged by autocorrelation plots and partial autocorrelation plots. The correlation orders p, d, and q of the ARIMA model are preliminarily determined, namely ARIMA(p,d,q); where p is the order of the autoregressive term, q is the order of the moving average term, and d is the order of the difference.

[0122] S3: Using the residual sequence calculated in step S2 as input, construct an LSTM neural network to learn the nonlinear features of the residuals, predict future residual values, make preliminary predictions of material consumption over the next few years, adjust the parameters based on the prediction results of material consumption over the next few years, use the LSTM model evaluation index to determine the optimal model, and use the optimal model to predict the material consumption of the future years.

[0123] Residual sequence This demonstrates the nonlinear fluctuations that are difficult for ARIMA to fit. Training samples are constructed using a sliding window as the unit. The recursive modeling is performed using an LSTM network. Its core state transfer structure is as follows:

[0124]

[0125] In the formula: express Input features at all times; This indicates the hidden state at the previous moment; Indicates the current state of the memory unit; This represents the weight matrix for each gate; Indicates the bias term. This represents the Sigmoid activation function; Represents the hyperbolic tangent activation function;

[0126] The output is a nonlinear disturbance prediction value. .

[0127] The parameters are adjusted based on the projected annual material requisition volumes for the next few years, including the following steps:

[0128] B1: Divide the dataset into a training set and a test set. Train the model on the training set and predict the values ​​on the test set. Then, use the mean squared error or root mean square error to compare the actual values ​​with the predicted values. Compare the difference between the predicted values ​​and the actual values, calculate the prediction error, and evaluate the accuracy of the prediction.

[0129] B2: Based on the prediction error results, adjust the order of the autoregressive term, the order of the differencing, and the parameters of the moving average term in the ARIMA model.

[0130] The optimal model is determined using the AIC model evaluation metrics, including the following steps:

[0131] C1: Calculate the AIC value of the ARIMA model at each order;

[0132] Step C1 specifically includes:

[0133] C1-1: Preliminary Order Identification: Based on the material consumption time series data constructed in step S2 The autocorrelation function (ACF) and partial autocorrelation function (PACF) are used to plot graphs, analyze the lag characteristics of the series, and preliminarily identify the possible order of the autoregressive term. and the order of the moving average term And determine whether a differential test is needed. Order differencing is used to make the sequence stationary, forming a set of candidate model combinations:

[0134]

[0135] Supplement: Time series data of material consumption constructed based on step S2 The autocorrelation function (ACF) and partial autocorrelation function (PACF) are used to plot graphs, analyze the lag characteristics of the series, and preliminarily identify the possible order of the autoregressive term. and the order of the moving average term And determine whether a differential test is needed. The order difference is used to make the sequence stationary, forming a set of candidate model combinations. Where: This is the maximum order search boundary, usually set to 3~5.

[0136] C1-2: Fitting candidate models one by one: For each group in the model set... The ARIMA model was fitted to each combination separately to obtain the log-likelihood function value for each group. and residual sum of squares ;

[0137] C1-3: AIC Value Calculation and Sorting:

[0138] Based on the Akaike Information Criterion (AIC), the fitting effect of each model is comprehensively evaluated, and its AIC value is calculated:

[0139]

[0140] Or, the equivalent form is:

[0141]

[0142] In the formula: The length of the time series; This represents the number of free parameters in the model (including constant terms). It is the sum of squared residuals; Likelihood function value.

[0143] C2: Compare the AIC values ​​of the models at different orders, and select the order that minimizes the AIC value as the optimal order of the model.

[0144] Step C2 specifically includes:

[0145] C2-1: Optimal Model Selection: Sort all candidate models by their AIC values ​​and select the combination with the lowest AIC value as the optimal model structure.

[0146]

[0147] As an information criterion, AIC considers not only the model fitting accuracy (the smaller the residual, the better) but also the model complexity (the more parameters, the greater the penalty). Therefore, its minimum value corresponds to the optimal structure that balances accuracy and generalization ability.

[0148] C2-2: Optimal Model Fitting and Prediction

[0149] Based on the selected The structure is remodeled completely on the training set, and the prediction residuals are passed to the LSTM residual compensation module to complete the subsequent fusion prediction.

[0150] Based on traditional models, this invention proposes a hybrid prediction architecture with independent modeling and progressive fusion characteristics. The core innovations are as follows:

[0151] 1. Trend-Perturbation Decoupling Assumption:

[0152] Suppose the original demand sequence consists of a linear trend term. With nonlinear disturbance term Composed of multiple layers:

[0153]

[0154] In the formula: , This enables structural separation modeling.

[0155] 2. Information load sharing mechanism: ARIMA handles long-term trend backbones, while LSTM focuses on short-period perturbations, avoiding performance degradation caused by long-term dependency modeling in deep neural networks, and improving training efficiency and generalization ability.

[0156] 3. Residual progressive remodeling:

[0157] The LSTM receives the prediction residuals from ARIMA and remodels them, forming an error feedback mechanism. The final fused output is as follows:

[0158]

[0159] This structure effectively absorbs the information that is insufficient in ARIMA predictions, thus achieving prediction closure.

[0160] 4. Decoupled modular fusion structure: ARIMA and LSTM are trained in separate modules, and data interaction is achieved through the "residual channel". It supports independent replacement and parameter tuning, and has good deployment flexibility and engineering adaptability.

[0161] S4: Based on historical procurement and consumption data, analyze changes in procurement demand, determine the calculation range for annual procurement quantities, construct procurement rules, and establish an annual recommended procurement model for procurement planning.

[0162] Assuming that the quantity purchased each time is not less than M0, based on the current annual material quantity P, the future annual material consumption Pi (i = 1, 2, ..., i...n), the average procurement cycle Ta, and the shelf life T, a formula for calculating the annual procurement quantity M is established. The procurement quantity M satisfies the following formula:

[0163]

[0164] in, Minimum order quantity for a single purchase (contract or supply chain constraints). The quantity of the target material currently available in inventory; No. The annual forecast of material demand (output by step S3); This represents the safety stock value, set at 1.2 to 1.5 times the average annual projected consumption. This indicates the maximum number of procurement cycles within the permissible shelf life of the materials; The shelf life of materials is in years. The average procurement cycle for materials, in years.

[0165] S5: Based on the annual material consumption predicted in step S3, and combined with the recommended procurement model in step S4, calculate the recommended annual procurement quantity (annual procurement quantity M) for the next few years, and form an annual procurement plan table.

[0166] The model makes predictions at the monthly (or weekly / ten-day) granularity and summarizes them by the annual dimension to form the annual material consumption and annual procurement volume.

[0167] Material consumption period coverage constraints:

[0168]

[0169] Ensure the procurement of materials in the future It can be completely consumed within the year.

[0170] Order quantity constraints:

[0171]

[0172] Inventory capacity limit (system optional):

[0173]

[0174] in, The upper limit of storage capacity is a strong constraint when an intelligent warehousing platform is introduced to manage inventory space.

[0175] Application instructions for procurement logic:

[0176] This model is used to predict future demand at the start of each procurement cycle. Based on current inventory levels and supply schedules, the system automatically calculates and recommends the purchase quantity. Its objectives are: to avoid over-purchasing due to shelf-life limitations; to ensure uninterrupted supply of materials (through a safety stock mechanism); and to optimize the procurement pace without increasing excessive inventory pressure.

[0177] S6: Set an early warning inventory value and monitor the inventory level in real time.

[0178] The warning inventory value T1 satisfies the following formula:

[0179]

[0180] in, This represents the average consumption of materials per unit of time. It is the standard deviation of material consumption per unit time, used to reflect demand volatility; The average delivery cycle of goods from order placement to warehousing; The service level coefficient is determined by the required service confidence level. Commonly used values ​​are k=1.28 (90% service level), k=1.65 (95% service level), and k=2.05 (98% service level). This indicates the safety warning inventory level for the corresponding current material category.

[0181] The logic is as follows:

[0182] First item: Used to cover normal consumption, i.e., the average usage expected during the supply cycle;

[0183] Second item: This serves as a risk buffer zone to address uncertainty or sudden fluctuations in consumption and to reflect demand.

[0184] Early warning strategy triggering mechanism:

[0185] The system monitors actual inventory levels in real time. The emergency replenishment alert mechanism will be triggered when the following conditions are met:

[0186]

[0187] in, This represents the total inventory of materials monitored by the system during the current cycle. Once the inventory level falls below [a certain threshold]... The system will automatically issue replenishment suggestions and generate corresponding purchase request records to ensure a continuous supply of materials during the construction process.

[0188] Example 2:

[0189] To verify the effectiveness and impact of the present invention, the following comparative experiments and analyses were conducted in this embodiment:

[0190] Figure 2 A comparison chart of the predicted trends of the three models is presented. A typical high-frequency consumable commodity is selected, and the predicted curves of each model are compared with the actual demand over a 12-month test period. It can be seen that the ARIMA model can follow the overall upward or downward trend of demand, but it shows significant lag and bias at inflection points, especially failing to accurately predict peaks and troughs where trend changes occur. The LSTM model shows higher sensitivity to changes in demand and can capture the upward trend in time at peaks, but its grasp of the overall level is slightly biased. The ARIMA-LSTM fusion model's predicted curves are in good agreement with the actual observation sequence in multiple stages, especially showing a significant fitting effect in the high-frequency fluctuation range, achieving better predictive results.

[0191] Figure 3To illustrate the response trends of the three models to abnormal fluctuations, an abnormal peak point was artificially inserted into the test data to simulate a sudden surge in demand. The changing trends of each model's output were then observed. Due to its high nonlinearity, the LSTM model reacts quickly to sudden demand peaks, with its prediction curve rising sharply near the anomaly point, almost perfectly mirroring the actual peak. However, after the peak, the LSTM model exhibits some overshoot and delayed response, showing insufficient smoothness in adapting to the new equilibrium. The ARIMA model is relatively sluggish, reacting to sudden peaks with lag and insufficient amplitude, but after the peak, its prediction gradually returns to its original trend, recovering to a relatively stable state. In contrast, while the fusion model also shows some response lag near the anomaly point, the amplitude is between that of ARIMA and LSTM, more closely resembling actual demand changes. Furthermore, after the peak, the fusion model eliminates overshoot faster than LSTM, smoothly returning to normal levels. Overall, the fusion model achieves a better balance between accuracy and stability in responding to abnormal fluctuations, demonstrating stronger robustness. This demonstrates that the fusion model inherits LSTM's ability to sensitively capture drastic changes while suppressing excessive fluctuations in predictions through the ARIMA component, thus maintaining the reliability of predictions even under sudden events.

[0192] In summary, the method of the present invention has the following advantages compared with existing methods:

[0193] 1. Significantly improves prediction accuracy. By combining the linear trend capture strength of the ARIMA model with the nonlinear pattern learning strength of the LSTM network, the method of this invention can predict changes in material demand more accurately than a single model. Especially when material demand experiences inflection points or drastic fluctuations, the fusion model tracks actual changes better, significantly reducing prediction errors. Figure 2 The experimental results shown demonstrate that the method of this invention significantly reduces both the mean absolute error and the mean squared error compared to traditional ARIMA or single neural network models.

[0194] 2. Multi-factor Coupling Analysis. The method of this invention obtains demand data through construction schedule-BOM mapping and can integrate external influencing factors such as construction environment and shift schedules as feature inputs, improving the adaptability of the prediction model to real-world scenarios from multiple dimensions. This allows the prediction results to not only consider the inherent patterns of the time series but also respond to changes in construction schedule and environment, achieving coordinated linkage between demand and construction plans.

[0195] 3. Improved system stability. The ARIMA-LSTM fusion model improves prediction stability and robustness while maintaining accuracy. For example... Figure 2The experimental results show that the ARIMA part provides robust basic predictions, while the LSTM part corrects complex fluctuations. The two complement each other, reducing the model's sensitivity to noise and anomalies and avoiding the inaccuracies that may occur with a single model.

[0196] 4. Practical Engineering Value. This invention's method can be embedded in a materials management information system to achieve automated demand forecasting, reducing the impact of human factors. Accurate materials demand forecasting helps in formulating reasonable procurement and distribution plans, avoiding excessive stockpiling of materials that tie up capital or shortages that cause work delays, thus improving supply chain efficiency. For materials management in large-scale projects such as LNG ships, this method can ensure the continuity and economy of construction, demonstrating promising application prospects.

[0197] In summary, this invention overcomes the limitations of existing material demand forecasting by innovatively integrating statistical models and deep learning models. It achieves significant improvements in forecast accuracy, stability, and adaptability to complex scenarios, demonstrating remarkable beneficial effects.

Claims

1. A method for forecasting material demand based on ARIMA and LSTM models, characterized in that, Includes the following steps: S1: Based on the construction tasks and BOM data of LNG ship construction, collect historical procurement and consumption data of materials, construct a time-series material demand dataset, and perform data preprocessing. S2: Perform time series analysis on the preprocessed material demand dataset to identify whether there are periodic patterns or trends in historical procurement and consumption data. Based on the material requisition records, use the ARIMA model to model linear trends, obtain preliminary predicted values ​​and calculate prediction residuals. Select the optimal order p, d, and q of the ARIMA model through the AIC criterion and complete the model fitting. S3: Using the residual sequence calculated in step S2 as input, construct an LSTM neural network to learn the nonlinear features of the residuals, predict future residual values, make preliminary predictions of material consumption over the next few years, adjust the parameters based on the prediction results of material consumption over the next few years, use the LSTM model evaluation index to determine the optimal model, and use the optimal model to predict the material consumption of the future years. S4: Based on historical procurement and consumption data, analyze changes in procurement demand, determine the calculation range for annual procurement quantities, construct procurement rules, and establish an annual recommended procurement model for procurement planning. S5: Based on the annual material consumption predicted in step S3, and combined with the recommended procurement model in step S4, calculate the recommended annual procurement amount for the next few years and form an annual procurement plan. S6: Set an early warning inventory value and monitor the inventory level in real time.

2. The material demand forecasting method based on ARIMA and LSTM models according to claim 1, characterized in that, The historical procurement and consumption data in step S1 includes: material procurement records, construction requisition records, project construction schedule, material inventory records, BOM configuration data, construction team operation information, warehousing and circulation information, and other auxiliary attributes strongly related to material use.

3. The material demand forecasting method based on ARIMA and LSTM models according to claim 2, characterized in that, The method for constructing a time-series material demand dataset in step S1 includes: after completing the historical data collection, aggregating it into a continuous data sequence with a uniform time granularity, and generating the original time-series data of the target material by aggregating it according to the uniform time granularity. ； In the formula: Indicates the first Requisition records within the period; This represents the set of dates covered by the period; This represents the periodic consumption; this sequence constitutes the direct input for ARIMA modeling and also serves as the target variable for LSTM training samples. Data preprocessing includes data cleaning, missing data imputation, standardization, and smoothing to reduce the impact of short-term fluctuations on modeling, particularly for sequences. The following preprocessing procedure is performed: A1: Missing Value Imputation: Blank segments in the construction records are filled using linear interpolation or periodic mean method. The missing points are estimated as follows: If time points are missing If two adjacent points are known, then: ； If the missing time period is long, it is supplemented using the average value of the same historical period: ； In the formula: This refers to the cycle length, suitable for periodic construction tasks. The imputation result replaces the missing positions in the original sequence and updates. ; A2: Outlier Correction and Sequence Smoothing: Performing moving average smoothing on detected abrupt changes: ； In the formula: Let be the window width; for points with abrupt changes, if: ； If it is identified as an outlier, it is corrected using the neighboring mean: ； Smoothed sequence More suitable for stationarity testing in ARIMA fitting; A3: Standardization Processing: To eliminate differences in numerical scales among different material categories or different engineering stages, and to unify training scales, [the following steps are taken]. Sequence Z-score normalization: ； In the formula: This is the historical average over a period of time. Standard deviation; Received It can be used for both ARIMA modeling and LSTM training. The standardized parameters should be recorded for future reconstruction of the predicted values. A4: Training Sample Construction: Based on the standardized sequences processed above. Construct input-output sliding window sample pairs for the prediction model: ； In the formula: For input windows of ARIMA or LSTM models; Predict the target value for the next cycle.

4. The material demand forecasting method based on ARIMA and LSTM models according to claim 3, characterized in that, In step S2, based on several years of construction and consumption data, the periodicity and trend of historical procurement and consumption data are judged through autocorrelation plots and partial autocorrelation plots to initially determine the correlation order p, d, and q of the ARIMA model, i.e., ARIMA(p,d,q); where p is the order of the autoregressive term, q is the order of the moving average term, and d is the order of the difference.

5. The material demand forecasting method based on ARIMA and LSTM models according to claim 4, characterized in that, Step S3 involves adjusting the parameters based on the projected annual material requisition volume for the next few years, including the following steps: B1: Divide the dataset into a training set and a test set. Train the model on the training set and predict the values ​​on the test set. Then, use the mean squared error or root mean square error to compare the actual values ​​with the predicted values. Compare the difference between the predicted values ​​and the actual values, calculate the prediction error, and evaluate the accuracy of the prediction. B2: Based on the prediction error results, adjust the order of the autoregressive term, the order of the differencing, and the parameters of the moving average term in the ARIMA model.

6. The material demand forecasting method based on ARIMA and LSTM models according to claim 5, characterized in that, Step S3 uses the AIC model evaluation index to determine the optimal model, including the following steps: C1: Calculate the AIC value of the ARIMA model at each order; C2: Compare the AIC values ​​of the models at different orders, and select the order that minimizes the AIC value as the optimal order of the model.

7. The material demand forecasting method based on ARIMA and LSTM models according to claim 6, characterized in that, Step C1 specifically includes: C1-1: Preliminary Order Identification: Based on the material consumption time series data constructed in step S2 By plotting graphs using the autocorrelation function and partial autocorrelation function, the lag characteristics of the series are analyzed, and the possible orders of autoregressive terms are preliminarily identified. and the order of the moving average term And determine whether a differential test is needed. Order differencing is used to make the sequence stationary, forming a set of candidate model combinations: ； C1-2: Fitting candidate models one by one: For each group in the model set... The ARIMA model was fitted to each combination separately to obtain the log-likelihood function value for each group. and residual sum of squares ; C1-3: AIC Value Calculation and Sorting: Based on the Akaike Information Content Criterion, the fitting effect of each model is comprehensively evaluated, and its AIC value is calculated: ； Or, the equivalent form is: ； In the formula: The length of the time series; This represents the number of free parameters in the model. It is the sum of squared residuals; Likelihood function value.

8. The material demand forecasting method based on ARIMA and LSTM models according to claim 7, characterized in that, Step C2 specifically includes: C2-1: Optimal Model Selection: Sort all candidate models by their AIC values ​​and select the combination with the lowest AIC value as the optimal model structure. ； C2-2: Optimal Model Fitting and Prediction Based on the selected The structure is remodeled completely on the training set, and the prediction residuals are passed to the LSTM residual compensation module to complete the subsequent fusion prediction.

9. The material demand forecasting method based on ARIMA and LSTM models according to claim 8, characterized in that, In step S4, it is assumed that the quantity purchased each time is not less than M0. Based on the current annual material quantity P, the future annual material consumption Pi (i = 1, 2, ... i ... n), the average procurement cycle Ta, and the shelf life T, a formula for calculating the annual procurement quantity M is established. The procurement quantity M satisfies the following formula: ； in, This is the minimum order quantity for a single purchase. The quantity of the target material currently available in inventory; No. Annual forecast of material demand; Indicates the safety stock value; This indicates the maximum number of procurement cycles within the permissible shelf life of the materials; Shelf life of materials; Average procurement cycle of materials.

10. A material demand forecasting method based on ARIMA and LSTM models according to claim 9, characterized in that, In step S6, the early warning inventory value T1 satisfies the following formula: ； in, This represents the average consumption of materials per unit of time. It is the standard deviation of material consumption per unit time, used to reflect demand volatility; The average delivery cycle of goods from order placement to warehousing; Service level coefficient; This indicates the safety warning inventory level for the corresponding current material category.

Images