Methods for constructing equipment demand forecasting model architecture and methods for equipment demand forecasting

By constructing a material demand forecasting model architecture and employing multi-source data preprocessing and multi-model fusion methods, the diverse adaptability and stability issues of material consumption forecasting in existing technologies are resolved, achieving high-precision and stable forecasting results and supporting inventory management and resource scheduling.

CN123117408APending Publication Date: 2026-11-17CHINESE PEOPLES LIBERATION ARMY AIR FORCE SERVICE ACAD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202610939752.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-06-26
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies struggle to adapt to diverse data characteristics in equipment consumption prediction, exhibiting large prediction biases, insufficient generalization ability, and poor model prediction stability. In particular, they fail to fully utilize limited data to improve model performance in scenarios with small samples and limited information.

Method used

By constructing a model architecture for predicting equipment demand, and adopting a data-driven mechanism of multi-source demand dataset preprocessing, multi-model training, testing, scoring, and fusion, a set of target prediction models is generated. By leveraging the complementary advantages of multiple models, objective quantification of model selection is achieved, avoiding the limitations of a single model and the uncertainty of manual model selection.

Benefits of technology

It significantly improves the accuracy and generalization ability of the forecasting model, provides a more reliable basis for inventory management, procurement planning and resource allocation decisions, and reduces the computational overhead and time cost of the forecasting process.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123117408A_ABST
    Figure CN123117408A_ABST
Patent Text Reader

Abstract

This application provides a method for constructing a material demand prediction model architecture and a method for predicting material demand. The method for constructing the material demand prediction model architecture includes: acquiring a multi-source demand dataset for materials and preprocessing it to generate a target multi-source demand dataset for the materials; dividing the target multi-source demand dataset based on a preset partitioning ratio; when the materials are non-occasional materials, training a preset training model set based on the training target multi-source demand dataset to generate a target training model set for the materials; testing the target training model set based on the test target multi-source demand dataset to generate a training score; obtaining a target prediction model set for the materials from the target training model set based on the training score, and fusing the target prediction model set according to the training score corresponding to each target prediction model in the target prediction model set to obtain the material demand prediction model architecture. Thus, multiple models complement each other, improving the accuracy and generalization ability of the model architecture.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of aviation service support technology, and in particular to a method for constructing an equipment demand forecasting model architecture and a method for forecasting equipment demand. Background Technology

[0002] In the field of materials support management, equipment consumption forecasting is a crucial foundation for supporting inventory optimization, procurement planning, and resource allocation. Accurate consumption forecasting can effectively reduce inventory backlog costs, improve materials support efficiency, and ensure mission execution reliability. As mission scenarios become increasingly complex and support requirements continue to rise, equipment consumption exhibits complex characteristics such as nonlinearity, multi-factor coupling, small sample size, and sparsity, placing higher demands on the accuracy and generalization ability of forecasting methods.

[0003] Currently, equipment consumption forecasting primarily employs two types of technical solutions. One type is based on statistical theory, including linear regression, autoregressive moving average models, differential autoregressive moving average models, and Markov models. These methods are simple in principle, computationally inexpensive, and highly interpretable, making them suitable for scenarios with complete data and stable demand trends. The other type is based on machine learning and deep learning, including support vector regression, random forests, extreme gradient boosting, long short-term memory networks, and gated recurrent units. These methods can fit nonlinear relationships, capture temporal dependencies, and multi-factor associations, making them suitable for medium-sized and larger sample scenarios.

[0004] However, existing technologies still have the following shortcomings: First, they struggle to adapt to the diverse data characteristics of equipment, resulting in large prediction biases and insufficient generalization ability. Second, the stability of model prediction results is poor. Third, they cannot fully utilize limited data to improve model performance in scenarios with small samples and scarce information. Summary of the Invention

[0005] This application aims to at least partially address one of the technical problems in the related art.

[0006] Therefore, the first objective of this application is to propose a method for constructing an equipment demand prediction model architecture. Through a data-driven mechanism of training, testing, scoring, optimization, and fusion, the objective quantification of model selection and the complementary advantages of multiple models are realized, effectively avoiding the limitations of a single model and the uncertainty of manual model selection, and significantly improving the accuracy and generalization ability of the prediction model.

[0007] The second objective of this application is to propose a method for forecasting equipment demand.

[0008] The third objective of this application is to propose a construction apparatus for an equipment demand forecasting model architecture.

[0009] To achieve the above objectives, the first aspect of this application proposes a method for constructing a material demand prediction model architecture, comprising: acquiring a multi-source demand dataset of materials and preprocessing the multi-source demand dataset to generate a target multi-source demand dataset of materials, wherein the multi-source demand dataset is used to represent historical data of materials generated based on actual test task record data; dividing the target multi-source demand dataset according to a preset partitioning ratio to generate a training target multi-source demand dataset and a test target multi-source demand dataset of materials; in response to the materials being non-occasional materials, training each preset training model in a preset training model set based on the training target multi-source demand dataset to generate a target training model set of materials; testing each target training model in the target training model set based on the test target multi-source demand dataset to generate a training score for each target training model; obtaining a target prediction model set of materials from the target training model set based on the training score of each target training model, and fusing the target prediction model set according to the training score corresponding to each target prediction model in the target prediction model set to obtain a demand prediction model architecture for materials.

[0010] In addition, the method for constructing the equipment demand prediction model architecture according to the above embodiments of this application may also have the following additional technical features: According to one embodiment of this application, the training target multi-source demand data includes training consumption data and training demand influencing factor data; training each preset training model in the preset training model set based on the training target multi-source demand dataset to generate a target training model set for the equipment includes: performing consumption prediction based on the training demand influencing factor data and the preset training model to generate predicted consumption data; generating consumption data deviation based on the training consumption data and the predicted consumption data; adjusting the model parameters of the preset training model based on the consumption data deviation until the consumption data deviation meets the preset training conditions; using the adjusted preset training model as the target training model; and generating the target training model set for the equipment based on the target training model.

[0011] According to one embodiment of this application, the test target multi-source demand data includes test consumption data and test demand influencing factor data; each target training model in the target training model set is tested based on the test target multi-source demand dataset to generate a training score for each target training model, including: performing consumption prediction based on the test demand influencing factor data and the target training model to generate test prediction consumption data; and generating a training score for the target training model based on the test consumption data and the test prediction consumption data.

[0012] According to one embodiment of this application, generating a training score for each target training model based on test consumption data and test prediction consumption data includes: calculating the mean absolute error, root mean square error, mean absolute percentage error, and coefficient of determination for the test consumption data and test prediction consumption data respectively; and calculating the training score for each target training model based on a preset weighting strategy, the mean absolute error, root mean square error, mean absolute percentage error, and coefficient of determination.

[0013] According to one embodiment of this application, based on the training score of each target training model, a target prediction model set for equipment is obtained from the target training model set, including: sorting the training scores of each target training model in descending order to obtain a training score ranking; taking the target training models corresponding to the top preset number of training scores in the training score ranking as target prediction models, and generating a target prediction model set based on the target prediction models.

[0014] According to one embodiment of this application, the target prediction model set is fused based on the training score corresponding to each target prediction model in the target prediction model set to obtain the equipment demand prediction model architecture, including: calculating the model fusion weight based on the training score corresponding to the target prediction model in the target prediction model set; and fusion of the target prediction model set based on the model fusion weight to generate the demand prediction model architecture.

[0015] According to one embodiment of this application, the method for constructing the equipment demand prediction model architecture further includes: responding to the equipment being an occasional type of equipment, iteratively verifying the preset prediction models in the preset prediction model set based on the equipment and the target multi-source demand dataset to generate the equipment model verification result, wherein the model verification result includes the verification score ranking of the preset prediction models in each round of iterative verification; based on the model verification result, obtaining the preset prediction models that meet the preset scoring conditions from the preset prediction model set, and using the preset prediction models that meet the preset scoring conditions as the target prediction models.

[0016] According to one embodiment of this application, the method for constructing the equipment demand prediction model architecture further includes: when the training target multi-source demand data in the training target multi-source demand dataset is exhausted, in response to the fact that the deviation of the consumed data does not meet the preset training conditions, adjusting the preset division ratio based on the preset adjustment strategy.

[0017] To achieve the above objectives, a second aspect of this application proposes a method for predicting equipment demand, comprising: acquiring demand influencing factor data for the current cycle of equipment, and preprocessing the demand influencing factor data to generate target demand influencing factor data; acquiring a demand prediction model architecture, wherein the demand prediction model architecture is trained based on the aforementioned method for constructing an equipment demand prediction model architecture; and performing demand prediction based on the target demand influencing factor data and the demand prediction model architecture to generate equipment consumption prediction data for the next cycle of equipment.

[0018] According to the equipment demand forecasting method of this application, by directly deploying the demand forecasting model architecture to the actual forecasting stage, the consumption forecast data and confidence interval for the next period can be quickly output after inputting the demand influencing factors data for the current period. This eliminates the need to repeatedly perform model training and validation steps, reducing computational overhead and time costs in the forecasting process. Furthermore, since the demand forecasting model architecture itself is built based on the weighted fusion of multiple models, its forecasting results combine the differentiated advantages of each component model, maintaining high accuracy while possessing stable generalization ability. The introduction of confidence intervals and risk levels allows users to quantitatively assess the reliability of the forecasting results, providing a more reliable decision-making basis for equipment inventory management, procurement planning, and resource allocation.

[0019] To achieve the above objectives, a third aspect of this application proposes a device for constructing a material demand prediction model architecture, comprising: an acquisition module for acquiring a multi-source demand dataset of materials and preprocessing the multi-source demand dataset to generate a target multi-source demand dataset of materials, wherein the multi-source demand dataset is used to represent historical data of materials generated based on actual test task record data; a partitioning module for partitioning the target multi-source demand dataset according to a preset partitioning ratio to generate a training target multi-source demand dataset and a test target multi-source demand dataset of materials; a training module for training each preset training model in a preset training model set based on the training target multi-source demand dataset, in response to the material being a non-occasional type of material, to generate a target training model set of materials; a scoring module for testing each target training model in the target training model set based on the test target multi-source demand dataset, to generate a training score for each target training model; and a generation module for obtaining a target prediction model set of materials from the target training model set based on the training score of each target training model, and fusing the target prediction model set according to the training score corresponding to each target prediction model in the target prediction model set to obtain a demand prediction model architecture of materials.

[0020] The apparatus for constructing the equipment demand prediction model architecture according to this application implements the aforementioned method for constructing the equipment demand prediction model architecture. Through a data-driven mechanism of training, testing, scoring, optimization, and fusion, objective quantification of model selection and complementary advantages of multiple models are achieved, effectively avoiding the limitations of a single model and the uncertainty of manual model selection, and significantly improving the accuracy and generalization ability of the prediction model. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating a method for constructing an equipment demand prediction model architecture according to some embodiments of this application; Figure 2 This is a flowchart illustrating a method for constructing an equipment demand prediction model architecture according to a specific embodiment of this application; Figure 3 This is a flowchart of a device demand prediction method according to some embodiments of this application; Figure 4 This is a block diagram of an apparatus for constructing an equipment demand prediction model architecture according to some embodiments of this application. Detailed Implementation

[0022] The embodiments of this application are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0023] Currently, two main technical approaches are used for predicting equipment consumption.

[0024] One category is prediction methods based on statistical theory, including linear regression, autoregressive moving average models, differential autoregressive moving average models, and Markov models. These methods have clear principles, are computationally simple, and are highly interpretable, making them suitable for scenarios with complete data and stable trends. However, they struggle to effectively capture nonlinear fluctuations and the interactive coupling relationships between multiple factors, resulting in significant prediction biases in complex scenarios.

[0025] Another category is prediction methods based on machine learning and deep learning, including support vector regression, random forests, extreme gradient boosting, long short-term memory networks, and gated recurrent units. These methods have strong nonlinear fitting and feature extraction capabilities, and can uncover deep correlations in data. However, they have high requirements for data quantity and quality, are prone to overfitting in small sample and sparse scenarios, and have high model training costs and weak interpretability.

[0026] However, existing technologies still have the following shortcomings: First, the capabilities of a single model are clearly limited. Models based on different principles have significantly different adaptability to different data features. A single model is difficult to cover the diverse consumption patterns in multiple scenarios at the same time, and its generalization ability is insufficient when migrating between different devices.

[0027] Second, the selection of models lacks a unified and objective quantitative evaluation system, and relies heavily on human experience for selection and parameter adjustment, which is highly subjective and has poor stability.

[0028] Third, the fixed and rigid division of training and test sets makes it difficult to adjust flexibly according to the actual amount of data and the model training status, and the limited data cannot be fully utilized in scenarios with small samples and scarce information.

[0029] Fourth, there is a lack of specific processing mechanisms for low-frequency sporadic equipment, and the prediction accuracy and confidence of sparse consumption sequences with a high proportion of zero values ​​are low, making it difficult to meet the actual needs of support decision-making.

[0030] The following describes in detail, with reference to the accompanying drawings, the method for constructing the equipment demand forecasting model architecture and the equipment demand forecasting method of the embodiments of this application.

[0031] This application provides a method for constructing an equipment demand forecasting model architecture. Figure 1 This is a flowchart illustrating a method for constructing an equipment demand prediction model architecture according to some embodiments of this application. (Refer to...) Figure 1 Methods for constructing equipment demand forecasting model architectures may include: S1. Obtain the multi-source demand dataset of the equipment and preprocess the multi-source demand dataset to generate the target multi-source demand dataset of the equipment. The multi-source demand dataset is used to represent the historical data of the equipment generated based on the measured task record data.

[0032] The multi-source demand dataset refers to a collection of historical data obtained from the actual test mission records of equipment. The actual test mission record data refers to the actual consumption records of equipment during aviation service support activities and the corresponding mission execution records. The consumption records include the actual quantity of equipment consumed in each historical period, and the mission execution records include the type of mission undertaken by the equipment in the corresponding period, mission intensity, number of flights, number of malfunctions, environmental conditions, and training subjects. The multi-source demand dataset can be obtained from equipment inventory management ledgers, flight training record ledgers, equipment maintenance record ledgers, and environmental monitoring records.

[0033] Furthermore, the multi-source demand dataset is preprocessed. Preprocessing includes at least one of the following: data cleaning, missing value imputation, outlier removal, zero-value smoothing, data normalization, and time-series alignment.

[0034] As a specific embodiment of this example, missing value handling can employ a combined strategy of Lagrange interpolation, K-nearest neighbor imputation, and mean imputation. Specifically, for missing items in a multi-source demand dataset, Lagrange interpolation is preferentially used for imputation. If the interpolation conditions are not met, K-nearest neighbor imputation is switched to. If K-nearest neighbor imputation is still not applicable, mean imputation is used as a fallback strategy. Outlier detection can be based on the 3σ principle (an outlier identification criterion based on the normal distribution assumption), isolated forests, and box plots. Figure 3 This approach combines three methods to identify anomalies in the multi-source demand dataset, removing data points deemed anomalous by at least two of these methods. Zero-value smoothing targets sporadic sparse consumption sequences, marking sequences with a high proportion of zero values ​​as sparse sequences and smoothing consecutive zero-value segments to reduce their interference with model training. Data standardization can employ either Min-Max Normalization or Z-Score Standardization to uniformly scale the data across all dimensions of the multi-source demand dataset to the same numerical range, eliminating the influence of different dimensions due to varying units. Temporal alignment and feature construction align data dimensions from different collection frequencies in the multi-source demand dataset to the same time granularity, and then construct a temporal sliding window, extracting trend features, periodic features, and task-related features based on the aligned data. Through the above preprocessing, noise in the original data can be eliminated, missing information can be filled in, and the data scale can be unified, providing a target multi-source demand dataset with standardized format and reliable quality for subsequent training and testing.

[0035] Specifically, historical consumption and task data of equipment are acquired as a multi-source demand dataset, and preprocessing operations such as cleaning, missing data filling, heterogeneity removal, smoothing, alignment and feature construction are performed on the multi-source demand dataset to generate a standard target multi-source demand dataset for modeling.

[0036] S2, divide the target multi-source demand dataset according to the preset division ratio to generate the training target multi-source demand dataset and the test target multi-source demand dataset for the equipment.

[0037] The preset partitioning ratio refers to the pre-defined allocation scheme of the training target multi-source demand dataset and the test target multi-source demand dataset within the target multi-source demand dataset. This scheme is preset by technical personnel based on the historical data volume of the equipment and the actual prediction requirements, and the specific ratio value is not limited. The preset partitioning ratio is usually expressed as a percentage or ratio, such as 7:3 or 8:2.

[0038] Specifically, the complete data samples in the target multi-source demand dataset are divided into two non-overlapping subsets according to a preset division ratio, which serve as the training target multi-source demand dataset and the test target multi-source demand dataset, respectively.

[0039] The training target multi-source demand dataset refers to a set of samples used by a pre-defined training model to learn data patterns and mapping relationships. Each data sample in this set contains complete input features and the corresponding actual consumption. The test target multi-source demand dataset refers to a set of samples used to evaluate the predictive performance of the target training model after training. Each data sample in this set also contains complete input features and the corresponding actual consumption, but this set does not participate in the parameter learning process of the pre-defined training model; it only serves as a benchmark for verifying the model's generalization ability. The training target multi-source demand dataset and the test target multi-source demand dataset together constitute a complete target multi-source demand dataset, and there is no overlap between the two.

[0040] S3, in response to the fact that the equipment is a non-occasional type of equipment, trains each preset training model in the preset training model set based on the multi-source demand dataset of the training target to generate the target training model set of the equipment.

[0041] Among them, non-occasional equipment refers to equipment whose percentage of zero values ​​in historical consumption data is lower than the zero value percentage threshold. For example, when the zero value percentage threshold is set to 75%, equipment whose percentage of zero values ​​in historical consumption data is lower than 75% is judged as non-occasional equipment.

[0042] A pre-defined training model set can refer to a collection of models composed of multiple prediction models based on different principles. For example, a pre-defined training model set includes statistical models, machine learning models, and deep learning models. Statistical models include univariate regression, multivariate regression, negative binomial regression, Markov models, GM(1,1) grey models (Grey Model (1,1), GM(1,1)), AR (Autoregressive Model), MA (Moving Average Model), ARMA (Autoregressive Moving Average Model), ARIMA (Autoregressive Integrated Moving Average Model), and SARIMA (Seasonal Autoregressive Integrated Moving Average Model). Machine learning models include SVM (Support Vector Machine), BP (Back Propagation) neural networks, RF (Random Forest), XGBoost (eXtreme Gradient Boosting), and KNN (K-Nearest Neighbors). Deep learning models include the Prophet model (time series prediction model), ELM (Extreme Learning Machine), LSTM (Long Short-Term Memory Network), GRU (Gated Recurrent Unit Network), and Transformer model.

[0043] The advantage of a pre-set training model set lies in the fact that models based on different principles respond differently to data features, making it difficult for a single model to simultaneously adapt to the consumption patterns of all equipment. Statistical models excel at capturing stationary trends under small sample sizes, machine learning models are adept at fitting nonlinear interactions, and deep learning models stand out in long-term time-series dependencies and multi-factor coupling. By incorporating these three types of models into the pre-set training model set, on the one hand, the model combination that best matches the features of each piece of equipment can be automatically selected based on the actual data distribution, avoiding the subjectivity and limitations of manual model selection; on the other hand, there is a natural complementary relationship between multiple models, and their respective advantages can be integrated through weighted methods in the subsequent fusion stage, thereby improving the overall prediction accuracy, robustness, and generalization ability. In addition, the diverse model reserves also make this method more adaptable to the face of new data or equipment changes, and dynamic adjustments can be made without redesigning the model structure.

[0044] It should be noted that, in some embodiments of this application, the multi-source training target demand data includes training consumption data and training demand influencing factor data.

[0045] The training consumption data refers to the actual consumption quantity of equipment in the multi-source demand dataset of the training target during each historical period. This training consumption data serves as the target value learned by the pre-set training model during the training phase, and is used to compare it with the predicted consumption data output by the pre-set training model to calculate the deviation.

[0046] Training demand influencing factor data refers to data on various dimensions of influencing variables related to equipment consumption in the multi-source demand dataset for training objectives, including mission intensity, number of flights, number of malfunctions, environmental conditions, and training subjects. This training demand influencing factor data serves as input features to a pre-defined training model, driving the model to learn the mapping relationship between each influencing factor and equipment consumption.

[0047] In some embodiments of this application, each preset training model in the preset training model set is trained based on the training target multi-source demand dataset to generate a target training model set for the equipment. This includes: performing consumption prediction based on training demand influencing factor data and the preset training model to generate predicted consumption data; generating consumption data deviation based on training consumption data and predicted consumption data; adjusting the model parameters of the preset training model based on the consumption data deviation until the consumption data deviation meets the preset training conditions; using the adjusted preset training model as the target training model; and generating the target training model set for the equipment based on the target training model.

[0048] Specifically, consumption prediction refers to using training demand influencing factor data as input to a pre-set training model. The pre-set training model calculates the input data based on its internal parameters and outputs the corresponding prediction results. Consumption prediction is performed based on the training demand influencing factor data and the pre-set training model to generate predicted consumption data. That is, for each pre-set training model in the set of pre-set training models, the training demand influencing factor data for that equipment is input into the pre-set training model, and the pre-set training model calculates and outputs predicted consumption data corresponding to that training demand influencing factor data. Predicted consumption data can refer to the predicted consumption value of the equipment within the corresponding period output by the pre-set training model. This predicted consumption data has the same data format and units as the training consumption data to facilitate subsequent deviation calculations.

[0049] Furthermore, a consumption data bias is generated based on the training consumption data and the predicted consumption data. Consumption data bias refers to a measure of the difference between the training consumption data and the predicted consumption data. It is a quantitative indicator used to measure the degree to which the predicted consumption data output by the pre-trained model deviates from the actual consumption data. The calculation methods for consumption data bias include calculating at least one of the following: mean squared error, mean absolute error, or mean absolute percentage error between the training consumption data and the predicted consumption data.

[0050] Furthermore, the model parameters of the preset training model are adjusted based on the consumed data deviation until the consumed data deviation meets the preset training conditions. The preset training conditions refer to pre-set termination criteria used to determine whether the preset training model has completed training. Preset training conditions include a consumed data deviation less than or equal to a preset deviation threshold, or the number of iterations reaching a preset maximum number of iterations. These conditions are pre-set by technical personnel based on the prediction accuracy requirements of the application scenario, and the specific values ​​are not limited. For example, the preset training conditions can be set to a mean absolute percentage error not exceeding 10% and a coefficient of determination not less than 0.8.

[0051] When the consumption data deviation does not meet the preset training conditions, the consumption data deviation is fed back to the preset training model via backpropagation. The weight parameters of each node in the preset training model are adjusted layer by layer along the direction of decreasing consumption data deviation, so that the deviation between the predicted consumption data output by the corrected preset training model and the training consumption data is reduced in the next round of training. The above steps of consumption prediction, deviation calculation, and parameter adjustment are repeated until the consumption data deviation meets the preset training conditions. Training then stops, and the adjusted preset training model is used as the target training model. All target training models that have been trained are aggregated to generate the target training model set for the equipment. Each target training model in the target training model set corresponds to a preset training model in the preset training model set, and each target training model has met the performance requirements specified by the preset training conditions.

[0052] It is understood that, in some embodiments of this application, the training method of the equipment demand prediction model further includes: when the training target multi-source demand data in the training target multi-source demand dataset is exhausted, in response to the fact that the deviation of the consumed data does not meet the preset training conditions, adjusting the preset division ratio based on the preset adjustment strategy.

[0053] The phrase "training target multi-source demand data exhausted" refers to the situation where all data samples in the training target multi-source demand dataset have been used to adjust the parameters of the preset training model during the training process. If, after all the training target multi-source demand data has been used in training, the data consumption deviation of the preset training model still does not meet the performance requirements specified by the preset training conditions, it indicates that the sample size provided by the training target multi-source demand dataset under the current preset partitioning ratio is insufficient for the preset training model to learn a sufficient mapping relationship.

[0054] A preset adjustment strategy refers to an adjustment scheme used to increase the preset partition ratio to expand the sample size of the multi-source dataset required for training. Preset adjustment strategies include increasing the proportion of the training set in the preset partition ratio by a fixed step size, and decreasing the proportion of the test set. The fixed step size is preset by technical personnel based on the total amount of data, and the specific value is not limited. For example, when the preset adjustment strategy sets the fixed step size to 5%, the preset partition ratio is adjusted from 7:3 to 75:25, then to 8:2, and so on.

[0055] Based on the preset adjustment strategy, the preset division ratio is adjusted. For example, the proportion of the training target multi-source demand dataset in the target multi-source demand dataset is gradually increased by a fixed step size, and the proportion of the test target multi-source demand dataset is reduced accordingly. The preset training model is then retrained using the expanded training target multi-source demand dataset until the data deviation is consumed and the preset training conditions are met.

[0056] S4, based on the multi-source demand dataset of the test target, test each target training model in the target training model set to generate a training score for each target training model.

[0057] To provide a detailed description of this embodiment, in some embodiments of this application, the multi-source requirement data for the test target includes test consumption data and test requirement influencing factor data.

[0058] Among them, the test consumption data can refer to the actual consumption quantity of equipment in the multi-source demand data of the test target in each historical period, which is used as a benchmark answer to evaluate the prediction accuracy of the target training model.

[0059] Data on factors influencing test requirements can refer to data on various dimensions of variables related to equipment consumption in the multi-source requirement dataset of the test target, including mission intensity, number of flights, number of failures, environmental conditions, and training subjects. These data are used as input features for training the target model to drive the model to output corresponding prediction results.

[0060] In some embodiments of this application, each target training model in the target training model set is tested based on the multi-source demand dataset of the test target to generate a training score for each target training model. This includes: performing consumption prediction based on test demand influencing factor data and the target training model to generate test prediction consumption data; and generating a training score for the target training model based on the test consumption data and the test prediction consumption data.

[0061] Specifically, the predicted consumption data for testing refers to the predicted consumption value of equipment within a corresponding period, output by the target training model based on the data of factors influencing testing demand. The predicted consumption data has the same dimensions as the actual consumption data for testing and is used for point-by-point comparison to calculate the deviation.

[0062] To provide a detailed explanation of the training score calculation process, in some embodiments of this application, a training score for each target training model is generated based on test consumption data and test prediction consumption data. This includes: calculating the mean absolute error, root mean square error, mean absolute percentage error, and coefficient of determination for the test consumption data and test prediction consumption data, respectively; and calculating the training score for each target training model based on a preset weighting strategy, the mean absolute error, root mean square error, mean absolute percentage error, and coefficient of determination.

[0063] The mean absolute error (MAE) refers to the average absolute deviation of the test prediction data relative to the actual test data. A smaller MAE indicates a lower average deviation between the predicted values ​​and the true values ​​of the target training model, and thus higher prediction accuracy.

[0064] The root mean square error (RMSE) is the square root of the average of the squared deviations of the test prediction data relative to the test data. A larger RMSE indicates that the target training model produces more significant biases on individual samples; a smaller RMSE indicates that the target training model's prediction results are more stable and the probability of large errors is lower.

[0065] Mean absolute percentage error (MASE) refers to the average percentage of the absolute deviation of the predicted data from the actual data consumed in the test, relative to the actual data consumed in the test. A smaller MSE indicates a smaller relative deviation between the predicted values ​​and the true values ​​of the target training model, and thus higher prediction accuracy.

[0066] The coefficient of determination (COD) refers to the degree to which the test prediction data explains the variance of the test data. The COD ranges from negative infinity to 1. A COD closer to 1 indicates a stronger explanatory power and higher goodness of fit of the target training model on the test data; a COD closer to 0 or a negative value indicates that the prediction performance of the target training model is closer to or worse than directly using the mean as the prediction value.

[0067] Furthermore, based on the preset weighting strategy, mean absolute error, root mean square error, mean absolute percentage error, and coefficient of determination, the training score of each target training model is calculated.

[0068] The preset weighting strategy refers to the pre-defined weight allocation scheme for each error index in the training score calculation, including the weight values ​​corresponding to the mean absolute error, root mean square error, mean absolute percentage error and coefficient of determination. These weight values ​​are preset by technical personnel based on the application scenario's tolerance for prediction deviation and the importance attached to goodness of fit. The sum of each weight value is 1, and the specific weight values ​​are not limited.

[0069] As a specific embodiment of this example, the weight allocation in the preset weighted strategy can adopt the following scheme. The weight of the mean absolute percentage error is set to 40%, the weight of the root mean square error is set to 30%, the weight of the mean absolute error is set to 20%, and the weight of the coefficient of determination is set to 10%, with the sum of the four weights being 100%. Among them, the mean absolute percentage error reflects the relative deviation between the predicted value and the true value. It is not affected by the unit of measurement and can intuitively reflect the prediction accuracy, so it is given the highest weight of 40%. The root mean square error reflects the stability of the prediction error. It is sensitive to large deviations and can effectively identify whether there are abnormal predictions with serious deviations in the prediction results, so it is given the second highest weight of 30%. The mean absolute error reflects the absolute level of the prediction deviation. Its calculation method is intuitive. As a basic error indicator, it is given a weight of 20%. The coefficient of determination reflects the model's ability to explain the variance of the data. It is used to measure the goodness of fit of the model. As an auxiliary reference indicator of the model's explanatory power, it is given a weight of 10%.

[0070] The above weight allocation scheme can be adaptively adjusted by technical personnel according to the importance attached to relative error, stability, absolute deviation and goodness of fit in the actual application scenario. The adjustment range of each weight value is 35% to 45% for mean absolute percentage error, 25% to 35% for root mean square error, 15% to 25% for mean absolute error and 5% to 15% for coefficient of determination. When adjusting, ensure that the sum of the four weights is 100%.

[0071] Specifically, for each target training model in the target training model set, the mean absolute error, root mean square error, mean absolute percentage error, and coefficient of determination calculated for that target training model on the multi-source demand dataset of the test target are weighted and summed according to their respective weight values ​​in a preset weighting strategy. The result of the weighted summation is used as the training score of that target training model, which is used to uniformly measure the comprehensive performance of the target training model across different error dimensions. Each target training model corresponds to a training score, and the training scores of all target training models together constitute a training score set, providing a unified quantitative basis for subsequent model selection.

[0072] S5: Based on the training score of each target training model, obtain the target prediction model set of equipment from the target training model set, and fuse the target prediction model set according to the training score of each target prediction model in the target prediction model set to obtain the demand prediction model architecture of equipment.

[0073] In some embodiments of this application, obtaining a target prediction model set for equipment from the target training model set based on the training score of each target training model includes: sorting the training scores of each target training model in descending order to obtain a training score ranking; taking the target training models corresponding to the top preset number of training scores in the training score ranking as target prediction models, and generating a target prediction model set based on the target prediction models.

[0074] Specifically, all target training models in the target training model set are rearranged in descending order of their respective training scores to form a ranking of training scores for each target training model. The higher the training score, the better the overall prediction performance of the target training model on the multi-source demand dataset of the test target, and the higher its position in the ranking.

[0075] Furthermore, the target training model corresponding to the top preset number of training scores in the training score ranking is selected as the target prediction model.

[0076] The preset number refers to the number of target prediction models to be selected from the target training model set. This number is preset by technical personnel based on actual application needs, and the specific value is not limited. The preset number is usually a positive integer greater than 1 to balance the complementary advantages of multiple models during fusion with computational overhead. For example, the preset number can be set to 3, meaning that the top 3 training scores in the training score ranking are selected as the target prediction models.

[0077] Understandably, the predictive advantages exhibited by target training models based on different principles during testing often manifest in different aspects. By selecting multiple top-ranking target training models, the differentiated advantages of each model can be integrated in the subsequent fusion stage, resulting in a demand prediction model architecture that combines the advantages of multiple models, enhancing the stability and robustness of prediction results while maintaining high prediction accuracy.

[0078] In some embodiments of this application, the target prediction model set is fused according to the training score corresponding to each target prediction model in the target prediction model set to obtain the equipment demand prediction model architecture, including: calculating the model fusion weight based on the training score corresponding to the target prediction model in the target prediction model set; and fusion of the target prediction model set based on the model fusion weight to generate the demand prediction model architecture.

[0079] The model fusion weight refers to the proportion of each target prediction model in the fusion process, representing the influence of each target prediction model on the final fusion result. For example, the training score of each target prediction model in the target prediction model set is divided by the sum of the training scores of all target prediction models in the set; the resulting ratio is the model fusion weight of each target prediction model. Therefore, the higher the training score of a target prediction model, the greater its model fusion weight and the higher its contribution to the fusion result; conversely, the lower the training score of a target prediction model, the smaller its model fusion weight and the lower its contribution to the fusion result. The sum of the model fusion weights of all target prediction models is 1.

[0080] After each target prediction model outputs its predicted consumption for the new input test demand influencing factors, the predicted consumption of each target prediction model is multiplied by its corresponding model fusion weight, and then all products are added together. The resulting weighted sum is the final prediction output of the demand prediction model architecture.

[0081] In some embodiments of this application, the training method for the equipment demand prediction model further includes: responding to the equipment being an occasional type of equipment, iteratively verifying the preset prediction models in the preset prediction model set based on the equipment and the target multi-source demand dataset to generate model verification results for the equipment, wherein the model verification results include the verification score ranking of the preset prediction models in each round of iterative verification; based on the model verification results, obtaining the preset prediction models that meet the preset scoring conditions from the preset prediction model set, and using the preset prediction models that meet the preset scoring conditions as the target prediction models.

[0082] Among them, occasional equipment refers to equipment whose proportion of zero values ​​in historical consumption data reaches or exceeds a preset threshold. The preset threshold is set by technicians based on the distribution of historical consumption data of the equipment, and the specific value is not limited. For example, the preset threshold can be set to 75%, that is, when the proportion of zero values ​​in the historical consumption data of the equipment reaches or exceeds 75%, the equipment is identified as occasional equipment.

[0083] The preset prediction model set can refer to a pre-selected set of prediction models for occasional devices. The models in this set are selected from those that have good adaptability in scenarios with small samples and sparse data. For example, the preset prediction model set includes the KNN model, the GM(1,1) grey model, and the ELM model.

[0084] The KNN model, based on the principle of similarity matching, is suitable for finding records of occasional events similar to the current state in sparse historical data as a reference; the GM(1,1) grey model, based on cumulative generation and exponential fitting, requires only a small number of data points to build a prediction model, thus having low data requirements; the ELM model, a single-hidden-layer feedforward neural network, features fast training speed and simple parameter adjustment, and is not prone to overfitting under small sample conditions. These three types of models complement each other and can cover the prediction needs of occasional equipment under different sparsity levels.

[0085] Furthermore, based on the multi-source demand dataset of equipment and targets, the preset prediction models in the preset prediction model set are iteratively verified to generate model verification results for the equipment.

[0086] Iterative validation refers to training and testing equipment using various pre-defined prediction models under multiple sets of different values ​​of a pre-defined partition ratio. In each iteration, the validation score of each pre-defined prediction model is recorded and a ranking is generated. Specifically, in each iteration, the target multi-source demand dataset is divided into a training target multi-source demand dataset and a test target multi-source demand dataset according to the current pre-defined partition ratio. Each pre-defined prediction model is trained based on the training target multi-source demand dataset, and then each trained pre-defined prediction model is tested based on the test target multi-source demand dataset, and the validation score of each model is calculated.

[0087] The model validation results refer to the set of records showing the validation score ranking of each preset prediction model in each iteration after all iterations have been completed. The model validation results include the validation score ranking of each preset prediction model in each iteration, that is, the ranking of each model after sorting them from highest to lowest validation score in each iteration.

[0088] Furthermore, based on the model validation results, preset prediction models that meet preset scoring conditions are obtained from the preset prediction model set, and the preset prediction models that meet the preset scoring conditions are used as target prediction models.

[0089] The preset scoring criteria can refer to the model with the highest cumulative frequency of ranking first in the validation score across all iteration validation rounds as the model that meets the preset scoring criteria. Specifically, the validation score ranking of each iteration validation result is statistically analyzed, and the number of times each preset prediction model ranks first in the validation score across all iteration validation rounds is counted. The preset prediction model with the most frequent occurrences of ranking first is selected as the target prediction model.

[0090] Understandably, repeated validation with different training and test set ratios allows for a more comprehensive evaluation of the stability of each pre-defined prediction model under varying sample sizes. The model that ranks first and most frequently in multiple iterations is identified as the prediction model with the strongest adaptability to the data characteristics of occasional equipment, and is directly used as the target prediction model output. This target prediction model constitutes the demand prediction model architecture for this equipment.

[0091] In some embodiments of this application, when the multi-source demand dataset contains newly added multi-source demand data (such as three months of new actual consumption data and demand influencing factor data), the new data is preprocessed, incremental training is performed based on the updated multi-source demand dataset, the training scores of each model are recalculated and the score ranking is updated, and the target prediction model for the equipment is dynamically refreshed. Verification through incremental updates shows that the prediction performance remains stable.

[0092] As a specific embodiment of the method for constructing the equipment demand forecasting model architecture of this application, refer to Figure 2 The method for constructing the equipment demand forecasting model architecture of this application includes: S201, Obtain the multi-source demand dataset for equipment and perform preprocessing.

[0093] Sixty months of equipment consumption data and demand influencing factors were collected, including mission intensity, number of flights, number of malfunctions, environmental conditions, and training subjects. The collected multi-source demand dataset was preprocessed, including missing value imputation, outlier removal, zero-value labeling, data standardization, and time-series feature construction, to generate a target multi-source demand dataset for equipment.

[0094] S202, divide the target multi-source demand dataset according to the preset division ratio.

[0095] The target multi-source demand dataset is initially divided based on a preset division ratio. The initial preset division ratio is 7:3, that is, 70% of the data is used as the training target multi-source demand dataset and 30% of the data is used as the testing target multi-source demand dataset.

[0096] S203, train the preset training model based on the multi-source demand dataset of the training target.

[0097] Recognizing that this equipment is not incidental, each pre-set training model in the pre-set training model set is trained based on a multi-source demand dataset for the training target. The pre-set training model set contains 11 basic models, covering statistical models, machine learning models, and deep learning models. For each pre-set training model, consumption prediction is performed based on data influencing training demand factors, generating predicted consumption data. A consumption data deviation is generated based on the training consumption data and the predicted consumption data, and the model parameters are adjusted based on this deviation.

[0098] S204, determine whether the target training model meets the preset training conditions: if yes, execute S205, otherwise execute S208.

[0099] After each round of training, it is determined whether the data consumption deviation of the current target training model meets the preset training conditions. The preset training conditions are that the mean absolute percentage error does not exceed 10% and the coefficient of determination is not less than 0.8. If the data consumption deviation meets the preset training conditions, it indicates that the performance of the current target training model has reached the preset requirements, and the process proceeds to S205 to test the trained target training model. If the data consumption deviation does not meet the preset training conditions, it indicates that the performance of the current target training model has not yet reached the preset requirements, and the process proceeds to S208 to adjust the preset partition ratio and retrain.

[0100] S205, Test the target-trained model based on the multi-source demand dataset of the test target.

[0101] Based on the multi-source requirement dataset for the test targets, each target training model in the target training model set was tested. The mean absolute error, root mean square error, mean absolute percentage error, and coefficient of determination were calculated for each target training model. A training score was calculated for each target training model according to a preset weighting strategy: mean absolute percentage error 40%, root mean square error 30%, mean absolute error 20%, and coefficient of determination 10%. All target training models were then sorted in descending order of their training scores.

[0102] S206, Determine the target prediction model set for equipment based on training scores.

[0103] The top three training models with the highest training scores in the training score ranking are used as target prediction models to generate a target prediction model set. The ranking results are: XGBoost ranks first, Random Forest ranks second, and Backpropagation Neural Network ranks third. Therefore, XGBoost, Random Forest, and Backpropagation Neural Network are selected as target prediction models.

[0104] S207, Construct a demand prediction model architecture based on the target prediction model set and training scores.

[0105] The model fusion weights are calculated based on the training scores of each target prediction model in the target prediction model set. These fusion weights are then used to fuse the target prediction models, generating a demand prediction model architecture for equipment. After calculating the fusion weights for each target prediction model based on its training score, the models are weighted and fused to output the final predicted value and confidence interval.

[0106] S208, adjust the preset division ratio according to the preset adjustment strategy.

[0107] The default adjustment strategy is to gradually increase the proportion of the training set in 5% increments. That is, the default split ratio is adjusted from 7:3 to 75:25, then to 8:2, and so on. Return to S202 to redistribute the target multi-source requirement dataset with the adjusted default split ratio, and continue to perform subsequent training and testing steps until the model performance meets the target or all samples are used up.

[0108] For the case of adding 3 months of data, in response to the existence of new multi-source demand data in the multi-source demand dataset, the new data is preprocessed, incremental training is performed based on the updated multi-source demand dataset, the training scores of each model are recalculated and the score ranking is updated, the equipment demand prediction model architecture is dynamically refreshed, and the prediction performance remains stable.

[0109] The method for constructing the equipment demand forecasting model architecture in this application has the following technical advantages compared to existing technologies: First, by constructing a pre-set training model set covering statistical models, machine learning models, and deep learning models, it can adapt to the diverse data characteristics of different devices, eliminating the need to design and verify model structures separately for each device, thus significantly improving the efficiency and universality of model selection.

[0110] Second, by introducing a training scoring mechanism to uniformly and quantitatively evaluate the target training model set, multi-dimensional error indicators are integrated into a single training score, eliminating the subjectivity and uncertainty of manual model selection and realizing the objectivity and repeatability of model selection.

[0111] Third, by selecting multiple target training models with high training scores and weighting and fusing them according to their respective scores, the final demand prediction model architecture combines the differentiated advantages of multiple models, effectively overcoming the technical defect of unstable performance of a single model under different data features, and improving the overall prediction accuracy and robustness.

[0112] Fourth, by using a dynamic adjustment mechanism with a preset partitioning ratio, the proportion of the training set is gradually increased when the model performance does not meet the standard, so that the limited historical data can be fully utilized, effectively improving the problems of insufficient model training and inadequate generalization ability in scenarios with small samples and poor information.

[0113] Fifth, by setting up a separate set of preset prediction models and an iterative verification mechanism for occasional equipment, and selecting the optimal model through a competitive selection method, the problem of prediction bias expansion caused by poor model interference in sparse data scenarios by conventional weighted fusion methods is avoided, which significantly improves the prediction accuracy and reliability of occasional sparse consumable equipment.

[0114] Figure 3 This is a flowchart of a material demand forecasting method according to some embodiments of this application. This application provides a material demand forecasting method, referring to... Figure 3 ,include: S301: Obtain the current cycle demand influencing factor data for the equipment, and preprocess the demand influencing factor data to generate target demand influencing factor data.

[0115] The current period can refer to the time interval for which demand forecasting is required. This time interval is preset by technicians based on the actual needs of equipment support, and the specific duration is not limited.

[0116] The demand influencing factors data include at least one of the following: mission intensity, number of flights, number of failures, environmental conditions, and training subjects. The same preprocessing operations as in the training phase are performed on the demand influencing factor data, including data cleaning, missing value imputation, outlier removal, data normalization, and time-series alignment, to generate target demand influencing factor data with a standardized format and distribution consistent with the training phase data.

[0117] S302, Obtain the demand forecasting model architecture, wherein the demand forecasting model architecture is trained based on the aforementioned method for constructing the equipment demand forecasting model architecture.

[0118] The demand forecasting model architecture contains multiple target forecasting models and their corresponding model fusion weights, and is already in a deployable and applicable state.

[0119] S303 performs demand forecasting based on target demand influencing factor data and demand forecasting model architecture to generate equipment consumption forecast data for the next cycle.

[0120] The "next period" refers to the forecast time interval immediately following the current period. Equipment consumption forecast data may include point forecast values ​​and confidence intervals. A point forecast value is a single numerical prediction result output by the demand forecasting model architecture after weighted fusion, used as a baseline estimate of equipment demand in the next period. A confidence interval is a numerical range around the point forecast value, encompassing the actual equipment consumption in the next period at a certain confidence level, characterizing the degree of uncertainty in the forecast result. As another specific embodiment, equipment consumption forecast data also includes risk levels, which can be categorized based on the deviation between the point forecast value and the current inventory level, providing a reference for inventory management and assurance decisions.

[0121] It should be noted that details not disclosed in the embodiments of the equipment demand forecasting method of this application refer to the details disclosed in the embodiments of the construction method of the equipment demand forecasting model architecture, and will not be repeated here.

[0122] The equipment demand forecasting method provided in this application directly deploys the demand forecasting model architecture to the actual forecasting process. After inputting the demand influencing factors data for the current period, it can quickly output the consumption forecast data and confidence intervals for the next period, eliminating the need to repeatedly perform model training and validation steps, thus reducing computational overhead and time costs in the forecasting process. Furthermore, since the demand forecasting model architecture itself is built based on the weighted fusion of multiple models, its forecasting results combine the differentiated advantages of each component model, maintaining high accuracy while possessing stable generalization ability. The introduction of confidence intervals and risk levels allows users to quantitatively assess the reliability of the forecasting results, providing a more reliable decision-making basis for equipment inventory management, procurement planning, and resource allocation.

[0123] Corresponding to the embodiments of the above-described method for constructing the equipment demand forecasting model architecture, this application also provides an apparatus for constructing the equipment demand forecasting model architecture, referring to... Figure 4 The device 400 includes: an acquisition module 410, a division module 420, a training module 430, a scoring module 440, and a generation module 450.

[0124] The acquisition module 410 is used to acquire the multi-source demand dataset of the equipment and preprocess it to generate the target multi-source demand dataset of the equipment. The multi-source demand dataset is used to represent the historical data of the equipment generated based on the actual test task record data. The partitioning module 420 is used to partition the target multi-source demand dataset according to a preset partitioning ratio to generate the training target multi-source demand dataset and the test target multi-source demand dataset of the equipment. The training module 430 is used to train each preset training model in the preset training model set based on the training target multi-source demand dataset in response to the equipment being non-occasional equipment, to generate the target training model set of the equipment. The scoring module 440 is used to test each target training model in the target training model set based on the test target multi-source demand dataset to generate the training score of each target training model. The generation module 450 is used to obtain the target prediction model set of the equipment from the target training model set based on the training score of each target training model, and to fuse the target prediction model set according to the training score corresponding to each target prediction model in the target prediction model set to obtain the demand prediction model architecture of the equipment.

[0125] In some embodiments of this application, the training module 430 determines that the training target multi-source demand data includes training consumption data and training demand influencing factor data. The training module 430 is used to train each preset training model in the preset training model set based on the training target multi-source demand dataset to generate a target training model set for the equipment. Specifically, it is used to: predict consumption based on the training demand influencing factor data and the preset training model to generate predicted consumption data; generate consumption data deviation based on the training consumption data and the predicted consumption data; adjust the model parameters of the preset training model based on the consumption data deviation until the consumption data deviation meets the preset training conditions, use the adjusted preset training model as the target training model, and generate the target training model set for the equipment based on the target training model.

[0126] In some embodiments of this application, the training module 430 is further configured to adjust the preset division ratio based on a preset adjustment strategy when the training target multi-source demand data in the training target multi-source demand dataset is exhausted, in response to the fact that the deviation of the consumed data does not meet the preset training conditions.

[0127] In some embodiments of this application, the scoring module 440 determines that the multi-source demand data of the test target includes test consumption data and test demand influencing factor data; the scoring module 440 is used to test each target training model in the target training model set based on the multi-source demand dataset of the test target, so as to generate a training score for each target training model, specifically used to: perform consumption prediction based on the test demand influencing factor data and the target training model to generate test prediction consumption data; and generate a training score for the target training model based on the test consumption data and the test prediction consumption data.

[0128] In some embodiments of this application, the scoring module 440 is used to generate a training score for each target training model based on the test consumption data and the test prediction consumption data. Specifically, it is used to: calculate the mean absolute error, root mean square error, mean absolute percentage error, and coefficient of determination for the test consumption data and the test prediction consumption data respectively; and calculate the training score for each target training model based on a preset weighting strategy, the mean absolute error, the root mean square error, the mean absolute percentage error, and the coefficient of determination.

[0129] In some embodiments of this application, the generation module 450 obtains the target prediction model set of the equipment from the target training model set based on the training score of each target training model. Specifically, it is used to: sort the training scores of each target training model in descending order to obtain the training score ranking; take the target training models corresponding to the first preset number of training scores in the training score ranking as target prediction models, and generate the target prediction model set based on the target prediction models.

[0130] In some embodiments of this application, the generation module 450 fuses the target prediction model set according to the training score corresponding to each target prediction model in the target prediction model set to obtain the equipment demand prediction model architecture. Specifically, it is used to: calculate the model fusion weight based on the training score corresponding to the target prediction model in the target prediction model set; and fuse the target prediction model set based on the model fusion weight to generate the demand prediction model architecture.

[0131] In some embodiments of this application, the generation module 450 is further configured to: respond to the fact that the equipment is an occasional type of equipment, perform iterative verification of the preset prediction models in the preset prediction model set based on the equipment and the target multi-source demand dataset to generate the model verification result of the equipment, wherein the model verification result includes the verification score ranking of the preset prediction models in each round of iterative verification; based on the model verification result, obtain the preset prediction models that meet the preset scoring conditions from the preset prediction model set, and use the preset prediction models that meet the preset scoring conditions as the target prediction models.

[0132] It is understood that for details not disclosed in the embodiment of the equipment demand forecasting model architecture construction apparatus of this application, please refer to the details disclosed in the embodiment of the equipment demand forecasting model architecture construction method.

[0133] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0134] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0135] In this application, unless otherwise expressly specified and limited, the terms "installation," "connection," "joining," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components, unless otherwise expressly limited. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0136] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

[0137] Any process or method described in the flowchart or otherwise herein is to be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0138] It should be noted that the logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be specifically implemented in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which programs can be printed, because programs can be obtained electronically, for example, by optically scanning the paper or other media, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0139] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0140] Those skilled in the art will understand that all or part of the steps of the methods described in the above embodiments can be implemented by a program instructing related hardware, and the program can be stored in a computer-readable storage medium. When executed, the program includes one or a combination of the steps of the method embodiments.

[0141] Furthermore, the functional units in the various embodiments of this application can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0142] The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application.

[0143] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this application can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this application can be achieved, and this is not limited herein.

[0144] The specific embodiments described above do not constitute a limitation on the scope of protection of this application. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for constructing an equipment demand forecasting model architecture, characterized in that, include: A multi-source demand dataset for the equipment is acquired and preprocessed to generate a target multi-source demand dataset for the equipment, wherein the multi-source demand dataset is used to characterize the historical data of the equipment generated based on measured task record data. The target multi-source demand dataset is divided according to a preset division ratio to generate the training target multi-source demand dataset and the test target multi-source demand dataset for the equipment. In response to the fact that the equipment is a non-occasional type of equipment, each preset training model in the preset training model set is trained based on the multi-source demand dataset of the training target to generate the target training model set of the equipment. Based on the multi-source demand dataset of the test target, each target training model in the target training model set is tested to generate a training score for each target training model. Based on the training score of each target training model, a target prediction model set for the equipment is obtained from the target training model set, and the target prediction model set is fused according to the training score corresponding to each target prediction model in the target prediction model set to obtain the demand prediction model architecture for the equipment.

2. The method according to claim 1, characterized in that, The multi-source training target demand data includes training consumption data and data on factors influencing training demand. The step of training each preset training model in the preset training model set based on the multi-source demand dataset of the training target to generate the target training model set of the equipment includes: Based on the data on factors influencing training requirements and the preset training model, consumption prediction is performed to generate predicted consumption data; Based on the training consumption data and the predicted consumption data, a consumption data deviation is generated; The model parameters of the preset training model are adjusted based on the data consumption deviation until the data consumption deviation meets the preset training conditions. The adjusted preset training model is then used as the target training model, and the target training model set of the equipment is generated based on the target training model.

3. The method according to claim 1, characterized in that, The multi-source requirement data for the test targets includes test consumption data and test requirement influencing factor data; the step of testing each target training model in the target training model set based on the multi-source requirement dataset for the test targets to generate a training score for each target training model includes: Based on the data on factors influencing the test requirements and the target training model, consumption prediction is performed to generate test prediction consumption data; The training score of the target training model is generated based on the test consumption data and the test prediction consumption data.

4. The method according to claim 3, characterized in that, The step of generating a training score for each target training model based on the test consumption data and the test prediction consumption data includes: The test consumption data and the test predicted consumption data are calculated separately to obtain the mean absolute error, root mean square error, mean absolute percentage error, and coefficient of determination of the test consumption data and the test predicted consumption data. Based on the preset weighting strategy, the mean absolute error, the root mean square error, the mean absolute percentage error, and the coefficient of determination, the training score of each target training model is calculated.

5. The method according to claim 1, characterized in that, The step of obtaining the target prediction model set for the equipment from the target training model set based on the training score of each target training model includes: The training scores of each target training model are sorted in descending order to obtain the training score ranking; The target training model corresponding to the top preset number of training scores in the training score ranking is used as the target prediction model, and the target prediction model set is generated based on the target prediction model.

6. The method according to claim 1, characterized in that, The step of fusing the target prediction model set based on the training score corresponding to each target prediction model in the target prediction model set to obtain the demand prediction model architecture for the equipment includes: The model fusion weights are calculated based on the training scores corresponding to the target prediction models in the target prediction model set. The target prediction model set is fused based on the model fusion weights to generate the demand prediction model architecture.

7. The method according to claim 1, characterized in that, Also includes: In response to the fact that the equipment is an occasional type of equipment, the preset prediction models in the preset prediction model set are iteratively verified based on the equipment and the target multi-source demand dataset to generate the model verification result of the equipment, wherein the model verification result includes the verification score ranking of the preset prediction model in each round of iterative verification. Based on the model validation results, a preset prediction model that meets the preset scoring conditions is obtained from the preset prediction model set, and the preset prediction model that meets the preset scoring conditions is used as the target prediction model.

8. The method according to claim 2, characterized in that, Also includes: When the training target multi-source demand data in the training target multi-source demand dataset is exhausted, in response to the fact that the consumed data deviation does not meet the preset training conditions, the preset division ratio is adjusted based on the preset adjustment strategy.

9. A method for forecasting equipment demand, characterized in that, include: Obtain the current cycle demand influencing factor data for the equipment, and preprocess the demand influencing factor data to generate target demand influencing factor data; Obtain a demand forecasting model architecture, wherein the demand forecasting model architecture is obtained by training based on the construction method of the equipment demand forecasting model architecture as described in any one of claims 1-8; Demand forecasting is performed based on the target demand influencing factor data and the demand forecasting model architecture to generate equipment consumption forecast data for the next cycle of the equipment.

10. A device for constructing a material demand forecasting model architecture, characterized in that, include: The acquisition module is used to acquire a multi-source demand dataset of equipment and preprocess the multi-source demand dataset to generate a target multi-source demand dataset of the equipment, wherein the multi-source demand dataset is used to characterize the historical data of the equipment generated based on the measured task record data. The partitioning module is used to partition the target multi-source demand dataset based on a preset partitioning ratio to generate the training target multi-source demand dataset and the test target multi-source demand dataset of the equipment. The training module is used to train each preset training model in the preset training model set based on the training target multi-source demand dataset in response to the fact that the equipment is a non-occasional type of equipment, so as to generate the target training model set of the equipment. The scoring module is used to test each target training model in the target training model set based on the test target multi-source demand dataset, so as to generate a training score for each target training model; The generation module is used to obtain the target prediction model set of the equipment from the target training model set based on the training score of each target training model, and to fuse the target prediction model set according to the training score corresponding to each target prediction model in the target prediction model set to obtain the demand prediction model architecture of the equipment.