CN110782083B - Aero-engine standby demand prediction method based on deep Croston method - Google Patents

Aero-engine standby demand prediction method based on deep Croston method Download PDF

Info

Publication number
CN110782083B
CN110782083B CN201911011999.9A CN201911011999A CN110782083B CN 110782083 B CN110782083 B CN 110782083B CN 201911011999 A CN201911011999 A CN 201911011999A CN 110782083 B CN110782083 B CN 110782083B
Authority
CN
China
Prior art keywords
demand
backup
interval
sequence
prediction model
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN201911011999.9A
Other languages
Chinese (zh)
Other versions
CN110782083A (en
Inventor
林琳
刘杰
郭丰
吕彦诚
郭昊
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Harbin Institute of Technology
Original Assignee
Harbin Institute of Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Harbin Institute of Technology filed Critical Harbin Institute of Technology
Priority to CN201911011999.9A priority Critical patent/CN110782083B/en
Publication of CN110782083A publication Critical patent/CN110782083A/en
Application granted granted Critical
Publication of CN110782083B publication Critical patent/CN110782083B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
    • G06Q10/063—Operations research, analysis or management
    • G06Q10/0631—Resource planning, allocation, distributing or scheduling for enterprises or organisations
    • G06Q10/06315—Needs-based resource requirements planning or analysis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/20—Administration of product repair or maintenance

Abstract

The invention discloses an aircraft engine standby demand prediction method based on a deep Croston method, and relates to an aircraft engine standby demand prediction method. The method aims to solve the problem that the prediction accuracy of the existing method for the standby requirement of the aircraft engine is low. The process is as follows: firstly, converting an intermittent standby requirement original observation sequence into a standby requirement interval sequence and a standby requirement quantity sequence; secondly, taking the fleet state characteristic quantity as a covariate of the interval sequence of the backup transmission demand and the sequence of the backup transmission demand; thirdly, establishing a backup demand interval and backup demand prediction model; obtaining a trained backup demand interval and backup demand prediction model; respectively inputting the sample sets to be tested into the trained prediction models to obtain the prediction results of the backup demand intervals and the backup demand quantities; fourthly, converting the prediction result into an intermittent standby sending demand sequence; and fifthly, predicting the total loss of the deviation cost based on the prediction result. The invention is used in the field of aeroengines.

Description

Aero-engine standby demand prediction method based on deep Croston method
Technical Field
The invention relates to a method for predicting the standby requirement of an aircraft engine.
Background
In the actual operation and maintenance of civil aviation engine fleet, when a certain engine needs to be disassembled and maintained, in order to ensure the utilization rate of the airplane, the engine which needs to be disassembled and maintained is generally required to be replaced by a standby engine. When the reserve capacity of the fleet cannot meet the reserve demand, long-term parking of the airplane may occur and the operation and maintenance plan of the fleet is affected, thereby causing economic and public praise losses of airplane operation enterprises. The imbalance of standby requirements of the civil aviation engine fleet is a common problem of most of domestic and foreign airlines, and the following three aspects of influences can be generated on the fleet due to the large change of the standby requirements: (1) because the civil aviation engine is typical high-ownership-cost equipment, the backup and distribution requirement is greatly changed, and the backup and distribution inventory cost is increased; (2) the difficulty of renting and backup is caused, and the average cost of renting and backup is increased; (3) the maintenance workload of the repair facility is increased in a centralized manner. Therefore, prediction of the reserve demand of the aircraft engine is a key problem in the management of the aircraft engine fleet. Through the research on the reserve demand of the finite-scale fleet, the original sequence of the reserve demand of the civil aviation engine fleet is a typical discontinuous demand sequence. In the discontinuous spare part demand prediction, a large amount of '0' demand in the original demand sequence impacts the conventional continuous spare part demand prediction method, so that the conventional continuous spare part demand prediction method is generally difficult to obtain higher precision in the discontinuous spare part demand prediction.
Disclosure of Invention
The invention aims to solve the problem that the prediction accuracy of the spare requirement of the aero-engine is low in the existing method, and provides a method for predicting the spare requirement of the aero-engine based on a deep Croston method.
The method for predicting the standby demand of the aircraft engine based on the deep Croston method comprises the following specific processes:
converting an intermittent standby requirement original observation sequence into a standby requirement interval sequence and a standby requirement quantity sequence based on a Croston method framework;
acquiring a fleet state characteristic quantity influencing the reserve transmission demand of the fleet, and taking the fleet state characteristic quantity as a covariate of a reserve transmission demand interval sequence and a reserve transmission demand sequence;
step three, establishing a backup demand interval prediction model based on the LSTM deep learning network based on the covariates of the backup demand interval sequence obtained in the step two;
establishing a backup demand prediction model based on the LSTM deep learning network based on the covariates of the backup demand sequence obtained in the step two;
training an LSTM deep learning network backup demand interval prediction model and an LSTM deep learning network backup demand prediction model to obtain a trained LSTM deep learning network backup demand interval prediction model and an LSTM deep learning network backup demand prediction model;
respectively inputting the sample set to be tested into the trained LSTM deep learning network backup demand interval prediction model and the LSTM deep learning network backup demand prediction model to obtain backup demand interval and backup demand prediction results;
step four, converting the backup demand interval and the backup demand prediction result obtained in the step three into an intermittent backup demand sequence;
step five, predicting total loss L of the deviation cost based on the backup demand interval and the backup demand prediction result obtained in the step threetotal。
The invention has the beneficial effects that:
the invention provides a method for predicting the standby demand of an aircraft engine based on a deep Croston method in order to improve the prediction accuracy. The intermittent demand characteristic is processed by adopting a Croston framework, a reserve demand interval prediction model and a reserve demand prediction model are established by utilizing a long-short term memory deep learning network, and a comprehensive evaluation method for intermittent demand prediction is provided by combining the proposed aircraft engine spare part demand prediction model, so that the problem of low prediction accuracy of the existing method for the aircraft engine spare demand is solved, and the prediction accuracy of the aircraft engine spare demand is improved;
taking actual aero-engine demand data of an airline company as an example, the proposed aero-engine demand prediction model is verified, a traditional Croston method is regarded as a reference comparison model, and machine learning methods such as a BP neural network, a support vector machine, a gradient lifting decision tree, an extreme gradient lifting tree and a multilayer perceptron network are adopted as comparison models in a Croston framework. The prediction result shows that the prediction model of the spare demand of the aircraft engine based on the deep Croston method has obvious advantages.
Drawings
FIG. 1 is a schematic diagram of a model for predicting the reserve demand of an aircraft fleet based on a deep Croston method, x0Is the 1 st covariate, x1Is the 2 nd covariate, xnIs the (n +1) th covariate, oy,1To a standby transmission interval y1Prediction of (a)y,2To a standby transmission interval y2Prediction of (a)y,nTo a standby transmission interval ynPrediction of (a)d,1To reserve demand d1Prediction of (a)d,2To reserve demand d2Prediction of (a)d,nTo reserve demand dnPredicting;
FIG. 2 is a schematic diagram of an LSTM deep learning network structure of the present invention, where an Input layer is an Input layer, an output layer is an output layer, and h is0Is an initial value of LSTM cell, ht-kFor the output of LSTM cell at time t, ht-k+1For the output of LSTM cell at time t, ht-2For the output of LSTM cell at time t, ht-1For the output of LSTM cell at time t, htFor the output of the LSTM cell at time t, c0LSTM cell state at initial time, ct-kLSTM cell state at time t-k, ct-k+1LSTM cell state at time t-k +1, ct-2LSTM cell state at time t-2, ct-1LSTM cell state at time t-1, ctIs the LSTM unit state at time t, h is the LSTM hidden layer vector, C is the LSTM unit state, h and C represent an LSTM unit, xt-kIs the input vector at time t-k, xt-k+1Is the input vector at time t-k +1, xt-1Is the input vector at time t-1, xtIs the input vector at the time t;
FIG. 3 is a schematic diagram of a discontinuous backup demand forecast offset cost loss calculation;
fig. 4 is a schematic diagram of sample acquisition for a backup demand prediction experiment.
Detailed Description
The first embodiment is as follows: the method for predicting the standby demand of the aircraft engine based on the deep Croston method comprises the following specific processes:
converting an intermittent standby requirement original observation sequence into a standby requirement interval sequence and a standby requirement quantity sequence based on a Croston method framework;
acquiring a fleet state characteristic quantity influencing the reserve transmission demand of the fleet, and taking the fleet state characteristic quantity as a covariate of a reserve transmission demand interval sequence and a reserve transmission demand sequence;
step three, establishing a backup demand interval prediction model based on the LSTM deep learning network based on the covariates of the backup demand interval sequence obtained in the step two;
establishing a backup demand prediction model based on the LSTM deep learning network based on the covariates of the backup demand sequence obtained in the step two;
training an LSTM deep learning network backup demand interval prediction model and an LSTM deep learning network backup demand prediction model to obtain a trained LSTM deep learning network backup demand interval prediction model and an LSTM deep learning network backup demand prediction model;
respectively inputting the sample set to be tested into the trained LSTM deep learning network backup demand interval prediction model and the LSTM deep learning network backup demand prediction model to obtain backup demand interval and backup demand prediction results;
step four, converting the backup demand interval and the backup demand prediction result obtained in the step three into an intermittent backup demand sequence;
step five, predicting total loss L of the deviation cost based on the backup demand interval and the backup demand prediction result obtained in the step threetotal。
The second embodiment is as follows: the difference between the first embodiment and the first embodiment is that in the first step, based on a crossbar method framework, an intermittent type primary observation sequence of the backup demand is converted into a backup demand interval sequence and a backup demand sequence; the specific process is as follows:
through the research on the reserve demand of the finite-scale fleet, the original sequence of the reserve demand of the civil aviation engine fleet contains a large amount of '0' demand, and is a typical discontinuous demand sequence. So-called intermittent demand, which Silver describes as demand infrequent, in particular with demand intervals greater than demand observation intervals. The visual characteristic of the discontinuous demand is that the original demand sequence contains a large amount of '0' demand. This is an impact on the conventional continuous spare part demand forecasting method, so that it is generally difficult to obtain higher accuracy in the discontinuous spare part demand forecasting method.
Converting an intermittent primary observation sequence of the backup requirement into a backup requirement interval sequence and a backup requirement quantity sequence by using a Croston method frame, wherein the intermittent primary observation sequence of the backup requirement is expressed as follows:
Z＝{d0,0,...,0,d1,0,...,0,di,0,...,0,dn},di∈{1,2,3,..,m} (1)
in the formula (d)iDemand amount for the ith non-0 backup demand response;
in discontinuous standby power demand prediction, a demand interval is visually defined as two adjacent non-0 demand quantities diAnd di+1Two adjacent non-0 demands diAnd di+1Between is aiThe second "0" demand response, then the demand interval between the ith non-0 "demand response and the (i +1) th non-0" demand response is expressed as:
yi+1＝ai+1+1,i∈{0,1,...,n-1} (2)
in the formula, ai+1For two adjacent non-0 demands diAnd di+1The number of "0" demand responses in between;
based on the method, the discontinuous type standby transmission demand original observation sequence of the formula (1) can be decomposed into a demand interval sequence and a demand quantity sequence, which are respectively expressed as a formula (3) and a formula (4):
Y＝(Z)＝{y1,...,yn} (3)
D＝γ(Z)＝{d0,d1,...,dn} (4)
wherein Z is an interrupted primary observation sequence for the demand of reserve hair, y1For the 1 st demand interval, ynFor the nth demand interval, d0To the initial demand, d 11 st demand, dnIs the nth demand; (x) and γ (×) represent the transfer function of the demand interval sequence and the demand quantity sequence, respectively.
Due to demand d0The previous demand interval is unknown, and d is not considered when an intermittent type reserve demand prediction model based on a deep Croston method is established0。
Other steps and parameters are the same as those in the first embodiment.
The third concrete implementation mode: the second step is to obtain a fleet state characteristic quantity influencing the fleet's reserve transmission demand, and use the fleet state characteristic quantity as a covariate of a reserve transmission demand interval sequence and a reserve transmission demand sequence; the specific process is as follows:
the demand interval prediction model and the demand quantity prediction model are main parts of the deep Croston method. Because the state of a fleet of civil aviation engines has a great influence on the reserve demand of the fleet, in a forecast model of the reserve demand of the fleet based on a deep Croston method, the characteristic quantity of the state of the fleet is used as covariates of a reserve demand interval forecast model and a reserve demand forecast model (here, characteristic quantity (input variable)) of the state of the fleet;
the fleet state characterization quantity comprises: total number of engines on the wing, total time of operation on the wing, average operating time on the wing, total number of engines on the trim, total time of operation on the trim, average operating time on the trim, etc.
Other steps and parameters are the same as those in the first or second embodiment.
The fourth concrete implementation mode: the third step is to establish a backup demand interval prediction model based on the LSTM deep learning network based on the covariates of the backup demand interval sequence obtained in the second step;
establishing a backup demand prediction model based on the LSTM deep learning network based on the covariates of the backup demand sequence obtained in the step two;
the specific process is as follows:
on the basis of a Croston method, a demand interval prediction model and a demand quantity prediction model are introduced into a reserve demand prediction model in combination with the state of a civil aircraft engine fleet;
the LSTM deep learning network backup demand interval prediction model and the LSTM deep learning network backup demand prediction model both adopt three-layer structures and respectively comprise an input layer, an LSTM layer and an output layer;
LSTM units are stacked on the LSTM layer, which extracts the time series information of the input layer from the input sequence (what happens at this point in time may have an impact on the next point in time, which the LSTM network can discover and utilize); the LSTM layer mines the depth representation of the input and then the output layer calculates the demand interval or demand amount based on the depth representation (the output layer based on the LSTM layer finally calculates the predicted result of the output layer).
The structural diagram of the adopted LSTM deep learning network is shown in FIG. 2.
Other steps and parameters are the same as those in one of the first to third embodiments.
The fifth concrete implementation mode: the difference between this embodiment and one of the first to the fourth embodiments is that the number of stacked LSTM layers is 4, the time step of each layer is set to 5, which means that 5 LSTM units are stacked in the time series direction;
the output dimensions of each layer are respectively set as: 20. 15, 10 and 5;
and each layer adopts a linear activation function, and an Adam optimization algorithm is adopted to train the network.
Other steps and parameters are the same as in one of the first to fourth embodiments.
The sixth specific implementation mode: the embodiment is different from the first to the fifth embodiment in that the LSTM deep learning network backup demand interval prediction model and the LSTM deep learning network backup demand prediction model are trained to obtain the trained LSTM deep learning network backup demand interval prediction model and the trained LSTM deep learning network backup demand prediction model;
respectively inputting the sample set to be tested into the trained LSTM deep learning network backup demand interval prediction model and the LSTM deep learning network backup demand prediction model to obtain backup demand interval and backup demand prediction results;
the specific process is as follows:
dividing a data set into a training sample set and a testing sample set;
the training sample set and the testing sample set are both sets containing a standby transmission demand interval sequence, a standby transmission demand sequence (converted into the standby transmission demand interval sequence and the standby transmission demand sequence obtained in the step one) and a fleet state characterization quantity;
inputting the training sample set into an LSTM deep learning network backup demand interval prediction model for training, and testing the trained LSTM deep learning network backup demand interval prediction model by adopting a test sample set:
if the preset accuracy is reached, obtaining a trained backup demand interval prediction model of the LSTM deep learning network;
if the preset accuracy is not reached, the training sample set is continuously input into the LSTM deep learning network backup demand interval prediction model until the preset accuracy is reached, and a trained LSTM deep learning network backup demand interval prediction model is obtained;
inputting the training sample set into an LSTM deep learning network backup demand prediction model for training, and testing the trained LSTM deep learning network backup demand prediction model by adopting a test sample set:
if the preset accuracy is reached, obtaining a trained LSTM deep learning network backup demand prediction model;
if the preset accuracy is not reached, the training sample set is continuously input into the LSTM deep learning network backup demand prediction model for training until the preset accuracy is reached, and a trained LSTM deep learning network backup demand prediction model is obtained;
and respectively inputting the sample set to be tested into the trained LSTM deep learning network backup demand interval prediction model and the LSTM deep learning network backup demand prediction model to obtain backup demand interval and backup demand prediction results.
Other steps and parameters are the same as those in one of the first to fifth embodiments.
The seventh embodiment: the difference between this embodiment and the first to sixth embodiment is that, in the fourth step, the backup demand interval and the backup demand prediction result obtained in the third step are converted into an intermittent backup demand sequence; the specific process is as follows:
and (3) reversely carrying out the prediction results of the reserve demand interval and the reserve demand according to a formula (1) (the Croston method changes the original sequence into the demand interval and the demand, and the process is reversed, so that the reserve demand sequence can be constructed according to the demand interval and the demand, which is a unique result), and obtaining the discontinuous reserve demand sequence (the prediction results may be the same as or different from the original sequence in the step one).
Other steps and parameters are the same as those in one of the first to sixth embodiments.
The specific implementation mode is eight: the difference between this embodiment and one of the first to seventh embodiments is that, in the fifth step, the total loss L of the offset cost is predicted based on the backup demand interval and the backup demand prediction result obtained in the third steptotal(ii) a The specific process is as follows:
for the prediction of the reserve demand of the discontinuous type fleet, the comprehensive evaluation of the reserve demand interval prediction result and the reserve demand prediction result has more practical value. Therefore, an intermittent reserve power demand prediction comprehensive evaluation method is provided by combining the actual operation and maintenance characteristics of the civil aviation engine fleet.
When the wing civil aviation engine is disassembled for maintenance, the disassembled and maintained engine is generally required to be replaced by a standby engine. The standby engine is from the fleet stock backup or the rental backup. The stock cost is generated by the stock reserve of the fleet, and the lease cost is generated by the lease reserve. In the comprehensive evaluation method for predicting the reserve power demand of the intermittent type fleet, the fleet is supposed to prepare the reserve power according to the result of a reserve power demand prediction model, and when the prediction result deviates from the actual reserve power demand, corresponding cost loss is generated, so that the prediction precision of the intermittent type demand prediction model is indirectly represented by the cost loss.
Taking the ith backup demand forecast as an example, if the backup demand interval forecast value yi,predGreater than the actual value y of the interval of the reserve transmission demandi,realAccording to the model prediction result, prepare the backup to have (y)i,pred-yi,real) No reserve hair is available per unit time. At this time, if the fleet needs to temporarily rent the backup, assuming that a temporary rental cost loss occurs, the forecast deviation of the ith backup demand is represented by the rental cost loss. Similarly, ifBackup demand interval prediction value yi,predLess than actual value y of interval of backup transmission demandi,realAccording to the model prediction result, prepare the backup to have (y)i,real-yi,pred) A reserve inventory cost loss is generated per unit time, and the forecast deviation of the ith reserve demand response is characterized by the inventory cost loss. Total loss L of prediction deviation cost in comprehensive evaluation method for prediction precision of intermittent backup power generation demandtotalCan be calculated from equation (5):
Figure BDA0002244473710000071
in the formula, LiPredicting the deviation cost loss of the sample i to be tested; m is the total number of the test samples;
each sample to be tested is a set containing a demand interval, a demand quantity and a fleet state characterization quantity;
a schematic diagram of the calculation of the prediction deviation cost loss in the proposed method for comprehensively evaluating the prediction accuracy of the intermittent backup power demand is shown in fig. 3.
In summary, compared with the traditional prediction error characterization method, the provided evaluation method can utilize the demand interval prediction result and the demand quantity prediction result at the same time to predict the deviation cost loss by using one index to replace the demand interval prediction error characterization quantity and the demand quantity prediction error characterization quantity, and the provided evaluation method is more in line with the engineering practice.
Other steps and parameters are the same as those in one of the first to seventh embodiments.
The specific implementation method nine: the present embodiment is different from the first to seventh embodiments in that the predicted deviation cost loss L of the sample i to be testediCalculated from the following formula:
Figure BDA0002244473710000081
in the formula, yi,pred、yi,realRespectively a predicted value and an actual value of the backup demand interval; di,pred、di,realAre respectively preparedGenerating a predicted value and an actual value of the demand quantity;
lrent、lownoff-cost loss rates for lease intervals and inventory intervals (set, empirical values); l'rent、l′ownAre rental bias cost loss rate and inventory bias cost loss rate (set, empirical value).
Other steps and parameters are the same as those in one to eight of the embodiments.
The following examples were used to demonstrate the beneficial effects of the present invention:
the first embodiment is as follows:
and (3) a stand-by demand prediction comparison experiment of the fleet:
prediction experiment of backup requirement of deep Croston method
In order to verify a reserve demand prediction model of a fleet based on a deep Croston method, reserve demand data and fleet state data of a certain civil aviation engine fleet from 2007 to 2016 are collected and obtained. The ADI of the original sequence of the sample airplane fleet reserve requirement is calculated by taking the calendar day as an observation unit and is 11.7, and the sample airplane fleet reserve requirement is a typical intermittent requirement according to the intermittent spare part requirement judgment standard.
Firstly, verifying a model for predicting the reserve demand of the airplane fleet based on a deep Croston method by using the reserve demand data of the sample airplane fleet for 10 years. And decomposing the original sequence of the reserve demand of the sample fleet into a reserve demand interval sequence and a demand sequence based on a Croston method framework.
The standby demand interval sequence and the demand sequence of the sample fleet are arranged by 213 samples according to the standby demand response time sequence. In the backup demand interval and demand quantity prediction, the state variables of the fleet are used as covariates of 2 prediction models. In the prediction experiment of the reserve demand of the airplane fleet, the considered state variables of the airplane fleet comprise: total number of engines on wing, total time of operation on wing, average operating time on wing, total number of engines on repair, total time of operation on repair, average operating time on repair, total number of available spares, total time of operation on available spares, average operating time on available spares.
Establishing LSTM deep learning network structures of all prediction models, and determining specific node structures of LSTM stacking layers as shown in Table 1 through comparison experiments. The time step length of the LSTM stacking layer of the deep learning network is set to be 5, 5 LSTM units are stacked in the time sequence direction, the LSTM stacking layer is formed by stacking 4 layers of structures, and the output dimensionality of each layer is respectively set as: 20. 15, 10 and 5. And each layer adopts a linear activation function, and an Adam optimization algorithm is adopted to train the network.
TABLE 1 deep Croston method LSTM Stack layer Structure information
Figure BDA0002244473710000091
In order to fully utilize the reserve demand sample of the sample fleet, the reserve demand prediction experiment of the fleet adopts a sequential testing method, and the method is specifically described as follows:
when m (m > n) is usedtimestamp) When the prediction experiment is carried out on each sample of the backup demand, the first (m-1) samples are adopted to train a prediction model, and the mth sample is used as a test sample. Wherein n istimestampRepresenting the number of sequential stacks of LSTM cells. Each sample contains a demand interval, a demand and a fleet state variable set. In a deep Croston method backup demand prediction experiment, a demand interval prediction model and a demand prediction model are respectively trained by utilizing a training sample set, and the accuracy of the demand interval prediction model and the accuracy of the demand prediction model are respectively verified by utilizing a test sample set. The sample data of the standby requirement is arranged according to the time sequence of the requirement, and the test samples are obtained sequentially according to the method, so the method is called as a sequential test method. The schematic diagram of the sequential testing method to obtain the test sample and the corresponding training sample set is shown in fig. 4.
213 spare requirement samples are obtained from a sample fleet in a deep Croston method spare requirement prediction experiment, and ntimestampSet to 5. According to the sequential testing method, at most 208 test samples can be obtained from the sample fleet, and 208 corresponding to each test sample is obtained at the same timeA set of training samples. A backup demand prediction experiment of the test sample is performed, and the predicted backup demand interval and backup demand are shown in table 2.
TABLE 2 example backup demand intervals and demand forecast results
Figure BDA0002244473710000092
Figure BDA0002244473710000101
In the deep Croston method backup demand prediction experiment, the prediction accuracy of a backup demand interval prediction model and a backup demand prediction model is represented by adopting the traditional prediction error comprising MAE, ARE and RMSE. Because the neural network model inevitably has certain prediction fluctuation, in a prediction experiment, all prediction samples are repeatedly modeled and predicted for 10 times by using the same method, and the average value of repeated modeling prediction for multiple times is taken as a final result.
The demand interval and the demand quantity of 208 test samples of the sample fleet are predicted by using the LSTM deep learning network, and the traditional error characterization of the prediction result is shown in Table 3.
TABLE 3 backup demand interval and demand prediction error
Figure BDA0002244473710000102
The prediction experiment also adopts the proposed discontinuous type reserve hair demand prediction and evaluation method to comprehensively evaluate the airplane fleet reserve hair demand prediction model based on the deep Croston method. In order to be compared with other experimental methods conveniently, the comprehensive evaluation result of the airplane fleet reserve demand prediction model based on the deep Croston method is analyzed in a reserve demand prediction comparison experiment.
Backup demand prediction contrast experiment
In order to compare and verify a model for predicting the reserve demand of a fleet based on a deep Croston method, the traditional Croston method is used as a reference comparison experiment model. In addition, based on a cross method framework, a BPNN method, a Support Vector Machine (SVM) method, a Gradient Boosting Decision Tree (GBDT) method, an extreme Gradient Boosting (XGBoost) method, and an LSTM multilayer sensing network having a good prediction capability in the current Machine learning method are used as comparative experiment models.
In a backup demand prediction comparison experiment, a training sample set and a test sample set which are the same as those of a deep Croston method backup demand prediction experiment are adopted. In order to improve the prediction precision of the comparison experiment model, the structural parameters of each prediction model are optimized in the backup demand prediction comparison experiment.
In each comparison experiment, a Croston method frame is utilized to decompose an original sequence of the reserve demand of the sample fleet into a reserve demand interval sequence and a reserve demand sequence, and each comparison experiment model is utilized to predict the reserve demand interval and the reserve demand respectively. The following are determined by a comparative experiment method: in a BPNN method comparison experiment, a sigmoid activation function is adopted, and the number of hidden nodes is set to be 5; a sigmoid kernel function is adopted in an SVM method contrast experiment; in a GBDT method comparison experiment, a Gini coefficient impurity is used as a criterion function of cracking quality, an optimal segmentation strategy is adopted on each node, the minimum sample number required by segmenting internal nodes is set to be 2, and the minimum sample number required by leaf nodes is set to be 1; in the XGboost method comparison experiment, a linear regression objective function is adopted, RMSE is used as an evaluation index of verification data, the maximum depth of each decision tree is set to be 5, the learning rate is set to be 0.15, and the number of estimators is set to be 30; in the comparison experiment of the multilayer perception network, corresponding to the LSTM deep learning network structure, 4 LSTM units are adopted for stacking, and the output dimensions of the 4 LSTM stacking units are respectively as follows: 20. 15, 10 and 5, again using a linear activation function and Adam optimization algorithm.
And carrying out a reserve distribution demand prediction experiment on each comparative experiment model according to the structural parameters. And characterizing the prediction accuracy of the demand interval by using traditional prediction errors MAE, ARE and RMSE. The demand interval prediction error for each comparative experimental model is shown in table 4.
TABLE 4 prediction error of alternate prediction comparison experiment for demand
Figure BDA0002244473710000111
In the backup demand prediction of the backup demand prediction comparison experiment, the prediction accuracy of each comparison experiment method is also characterized by MAE, ARE and RMSE, as shown in Table 5.
TABLE 5 backup demand prediction versus experimental prediction error
Figure BDA0002244473710000112
In the comparison experiment of the reserve demand interval prediction and the reserve demand quantity prediction, the LSTM method adopted in the deep Croston method provided by the invention obtains the best prediction precision.
In the backup demand prediction comparison experiment, an intermittent backup demand prediction evaluation method is adopted to comprehensively evaluate each comparison experiment method. Wherein the predicted bias cost loss rates are respectively set as: c. Crent＝50、cown＝20、c′rent100 and c'own150. The backup demand forecast bias cost loss can be calculated by equation (6). Based on the backup demand interval prediction result and the backup demand amount prediction result of each comparative experimental method, the backup demand prediction bias cost loss of each comparative experimental method is shown in table 6. The BPNN method, the SVM method, the GBDT method and the MLP method all belong to machine learning methods, and compared with the traditional Croston method, the machine learning method can obtain better prediction accuracy in discontinuous backup demand prediction. Among them, the deep Croston method achieves the lowest predicted offset cost loss.
TABLE 6 comparison of Experimental methods backup demand prediction bias cost loss
Figure BDA0002244473710000121
The present invention is capable of other embodiments and its several details are capable of modifications in various obvious respects, all without departing from the spirit and scope of the present invention.

Claims (8)

1. The method for predicting the standby demand of the aircraft engine based on the deep Croston method is characterized by comprising the following steps: the method comprises the following specific processes:
converting an intermittent standby requirement original observation sequence into a standby requirement interval sequence and a standby requirement quantity sequence based on a Croston method framework;
acquiring a fleet state characteristic quantity influencing the reserve transmission demand of the fleet, and taking the fleet state characteristic quantity as a covariate of a reserve transmission demand interval sequence and a reserve transmission demand sequence;
step three, establishing a backup demand interval prediction model based on the LSTM deep learning network based on the covariates of the backup demand interval sequence obtained in the step two;
establishing a backup demand prediction model based on the LSTM deep learning network based on the covariates of the backup demand sequence obtained in the step two;
training an LSTM deep learning network backup demand interval prediction model and an LSTM deep learning network backup demand prediction model to obtain a trained LSTM deep learning network backup demand interval prediction model and an LSTM deep learning network backup demand prediction model;
respectively inputting the sample set to be tested into the trained LSTM deep learning network backup demand interval prediction model and the LSTM deep learning network backup demand prediction model to obtain backup demand interval and backup demand prediction results;
step four, converting the backup demand interval and the backup demand prediction result obtained in the step three into an intermittent backup demand sequence;
step five, predicting total loss L of the deviation cost based on the backup demand interval and the backup demand prediction result obtained in the step threetotal；
Converting an intermittent standby requirement original observation sequence into a standby requirement interval sequence and a standby requirement quantity sequence based on a Croston method framework in the first step; the specific process is as follows:
converting an intermittent primary observation sequence of the backup requirement into a backup requirement interval sequence and a backup requirement quantity sequence by using a Croston method frame, wherein the intermittent primary observation sequence of the backup requirement is expressed as follows:
Z＝{d0,0,...,0,d1,0,...,0,di,0,...,0,dn},di∈{1,2,3,..,m} (1)
in the formula (d)iDemand amount for the ith non-0 backup demand response;
intuitively defining the demand interval as two adjacent non-0 demands diAnd di+1Two adjacent non-0 demands diAnd di+1Between is aiThe second "0" demand response, then the demand interval between the ith non-0 "demand response and the (i +1) th non-0" demand response is expressed as:
yi+1＝ai+1+1，i∈{0，1，...，n-1} (2)
in the formula, ai+1For two adjacent non-0 demands diAnd di+1The number of "0" demand responses in between;
decomposing the discontinuous type standby transmission demand original observation sequence of the formula (1) into a demand interval sequence and a demand quantity sequence, which are respectively expressed as a formula (3) and a formula (4):
Y＝(Z)＝{y1，...，yn} (3)
D＝γ(Z)＝{d0，d1，...，dn} (4)
wherein Z is an interrupted primary observation sequence for the demand of reserve hair, y1For the 1 st demand interval, ynFor the nth demand interval, d0To the initial demand, d11 st demand, dnIs the nth demand; (x) and γ (×) represent the transfer function of the demand interval sequence and the demand quantity sequence, respectively.
2. The method for predicting the standby demand of an aircraft engine based on the deep Croston method according to claim 1, wherein: acquiring a fleet state characteristic quantity influencing the reserve transmission demand of the fleet, and taking the fleet state characteristic quantity as a covariate of a reserve transmission demand interval sequence and a reserve transmission demand sequence; the specific process is as follows:
taking the fleet state characterization quantity as covariates of a backup demand interval prediction model and a backup demand prediction model;
the fleet state characterization quantity comprises: total number of engines on wing, total time of operation on wing, average operating time on wing, total number of engines on repair, total time of operation on repair, average operating time on repair.
3. The method for predicting the standby demand of an aircraft engine based on the deep Croston method according to claim 2, wherein: in the third step, a backup demand interval prediction model based on an LSTM deep learning network is established based on the covariates of the backup demand interval sequence obtained in the second step;
establishing a backup demand prediction model based on the LSTM deep learning network based on the covariates of the backup demand sequence obtained in the step two;
the specific process is as follows:
the LSTM deep learning network backup demand interval prediction model and the LSTM deep learning network backup demand prediction model both adopt three-layer structures and respectively comprise an input layer, an LSTM layer and an output layer.
4. The method for predicting the standby demand of an aircraft engine based on the deep Croston method according to claim 3, wherein: the stacking layer number of the LSTM layers is 4, the time step of each layer is set to be 5, and 5 LSTM units are stacked in the time sequence direction;
the output dimensions of each layer are respectively set as: 20. 15, 10 and 5;
and each layer adopts a linear activation function, and an Adam optimization algorithm is adopted to train the network.
5. The method for predicting the standby demand of an aircraft engine based on the deep Croston method according to claim 4, wherein: the LSTM deep learning network backup demand interval prediction model and the LSTM deep learning network backup demand prediction model are trained to obtain a trained LSTM deep learning network backup demand interval prediction model and a trained LSTM deep learning network backup demand prediction model;
respectively inputting the sample set to be tested into the trained LSTM deep learning network backup demand interval prediction model and the LSTM deep learning network backup demand prediction model to obtain backup demand interval and backup demand prediction results;
the specific process is as follows:
dividing a data set into a training sample set and a testing sample set;
the training sample set and the testing sample set are sets comprising a standby transmission demand interval sequence, a standby transmission demand sequence and a fleet state characterization quantity;
inputting the training sample set into an LSTM deep learning network backup demand interval prediction model for training, and testing the trained LSTM deep learning network backup demand interval prediction model by adopting a test sample set:
if the preset accuracy is reached, obtaining a trained backup demand interval prediction model of the LSTM deep learning network;
if the preset accuracy is not reached, the training sample set is continuously input into the LSTM deep learning network backup demand interval prediction model until the preset accuracy is reached, and a trained LSTM deep learning network backup demand interval prediction model is obtained;
inputting the training sample set into an LSTM deep learning network backup demand prediction model for training, and testing the trained LSTM deep learning network backup demand prediction model by adopting a test sample set:
if the preset accuracy is reached, obtaining a trained LSTM deep learning network backup demand prediction model;
if the preset accuracy is not reached, the training sample set is continuously input into the LSTM deep learning network backup demand prediction model for training until the preset accuracy is reached, and a trained LSTM deep learning network backup demand prediction model is obtained;
and respectively inputting the sample set to be tested into the trained LSTM deep learning network backup demand interval prediction model and the LSTM deep learning network backup demand prediction model to obtain backup demand interval and backup demand prediction results.
6. The method for predicting the standby demand of an aircraft engine based on the deep Croston method according to claim 5, wherein: in the fourth step, the reserve demand interval and reserve demand prediction result obtained in the third step are converted into an intermittent reserve demand sequence; the specific process is as follows:
and (4) reversely carrying out the reserve demand interval and the reserve demand prediction result according to the formula (1) to obtain an intermittent reserve demand sequence.
7. The method for predicting the standby demand of an aircraft engine based on the deep Croston method according to claim 6, wherein: predicting total offset cost loss L based on backup demand interval and backup demand prediction result obtained in the step three in the step fivetotal(ii) a The specific process is as follows:
calculated from equation (5):
Figure FDA0002621226070000041
in the formula, LiPredicting the deviation cost loss of the sample i to be tested; m is the total number of the test samples;
each sample to be tested is a set containing a demand interval, a demand quantity and a fleet state characterization quantity.
8. The method for predicting the standby demand of an aircraft engine based on the deep Croston method according to claim 7, wherein: the predicted offset cost loss L of the sample i to be testediCalculated from the following formula:
Figure FDA0002621226070000042
in the formula, yi,pred、yi,realRespectively a predicted value and an actual value of the backup demand interval; di,pred、di,realRespectively a predicted value and an actual value of the backup demand;
lrent、lowna lease interval bias cost loss rate and an inventory interval bias cost loss rate; l'rent、l′ownThe rental quantity deviation cost loss rate and the inventory quantity deviation cost loss rate.
CN201911011999.9A 2019-10-23 2019-10-23 Aero-engine standby demand prediction method based on deep Croston method Active CN110782083B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201911011999.9A CN110782083B (en) 2019-10-23 2019-10-23 Aero-engine standby demand prediction method based on deep Croston method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201911011999.9A CN110782083B (en) 2019-10-23 2019-10-23 Aero-engine standby demand prediction method based on deep Croston method

Publications (2)

Publication Number Publication Date
CN110782083A CN110782083A (en) 2020-02-11
CN110782083B true CN110782083B (en) 2020-10-27

Family

ID=69386528

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201911011999.9A Active CN110782083B (en) 2019-10-23 2019-10-23 Aero-engine standby demand prediction method based on deep Croston method

Country Status (1)

Country Link
CN (1) CN110782083B (en)

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113723978B (en) * 2020-05-26 2023-10-24 华晨宝马汽车有限公司 Bearing bush demand prediction method and system
CN112329983A (en) * 2020-09-30 2021-02-05 联想(北京)有限公司 Data processing method and device
CN113127537B (en) * 2021-04-16 2023-11-24 北京交通大学 Spare part demand prediction method integrating time sequence prediction model and machine learning model

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104732287A (en) * 2013-12-19 2015-06-24 广州市地下铁道总公司 Stock control method based on optimum replenishment period of spare part
US10311129B1 (en) * 2018-02-09 2019-06-04 Banjo, Inc. Detecting events from features derived from multiple ingested signals

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6205431B1 (en) * 1998-10-29 2001-03-20 Smart Software, Inc. System and method for forecasting intermittent demand

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104732287A (en) * 2013-12-19 2015-06-24 广州市地下铁道总公司 Stock control method based on optimum replenishment period of spare part
US10311129B1 (en) * 2018-02-09 2019-06-04 Banjo, Inc. Detecting events from features derived from multiple ingested signals

Also Published As

Publication number Publication date
CN110782083A (en) 2020-02-11

Similar Documents

Publication Publication Date Title
CN110782083B (en) Aero-engine standby demand prediction method based on deep Croston method
CN106448151B (en) Short-term traffic flow prediction method
CN112989635B (en) Integrated learning soft measurement modeling method based on self-encoder diversity generation mechanism
CN108153982B (en) Aero-engine after-repair performance prediction method based on stacked self-coding deep learning network
CN109523021A (en) A kind of dynamic network Structure Prediction Methods based on long memory network in short-term
CN113723010A (en) Bridge damage early warning method based on LSTM temperature-displacement correlation model
CN113487855B (en) Traffic flow prediction method based on EMD-GAN neural network structure
CN111178585A (en) Fault reporting amount prediction method based on multi-algorithm model fusion
CN111861013A (en) Power load prediction method and device
CN109214565A (en) A kind of subregion system loading prediction technique suitable for the scheduling of bulk power grid subregion
CN112439794A (en) Hot rolling bending force prediction method based on LSTM
Huang et al. An adversarial learning approach for machine prognostic health management
CN108170945B (en) Method for predicting after-repair exhaust temperature margin of aircraft engine
CN116663419A (en) Sensorless equipment fault prediction method based on optimized Elman neural network
CN116703644A (en) Attention-RNN-based short-term power load prediction method
CN113919162B (en) Voltage sag risk early warning method based on simulation and multi-source measured data fusion
CN111008661B (en) Croston-XGboost prediction method for reserve demand of aircraft engine
Azadeh et al. Selecting optimum maintenance activity plans by a unique simulation-multivariate approach
CN114819395A (en) Industry medium and long term load prediction method based on long and short term memory neural network and support vector regression combination model
CN113393102A (en) Distribution transformer operation state trend prediction method based on data driving
Ding et al. Prediction of laboratory equipment demand based on grey relation analysis and Elman neural network
CN113675844B (en) Multi-node power load prediction method
Lu et al. Application of GA-SVM time series prediction in tax forecasting
Zhang et al. Short-term load forecasting of power system based on attention mechanism CNN-BiLSTM
Chen et al. A Bridge Health Diagnosis Approach Base on Deep Neural Networks

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant