CN115759470A

CN115759470A - Flight overall process fuel consumption prediction method based on machine learning

Info

Publication number: CN115759470A
Application number: CN202211559722.1A
Authority: CN
Inventors: 胡松亭; 李福娟; 梁哲; 丁晓华; 黄蕾; 方陈浩; 蒋银; 毛金凤; 伊涵
Original assignee: China Eastern Technology Application R & D Center Co ltd
Current assignee: China Eastern Technology Application R & D Center Co ltd
Priority date: 2022-12-06
Filing date: 2022-12-06
Publication date: 2023-03-07

Abstract

The invention discloses a flight overall process fuel consumption prediction method based on machine learning, which comprises the steps of dividing a flight overall process into a plurality of flight stages according to aircraft flight data of a flight, obtaining a flight parameter sample library of each flight stage, extracting fuel consumption characteristics of each flight stage respectively by using an integrated learning algorithm, determining a fuel characteristic data set of each flight stage according to the fuel consumption characteristics of each flight stage, and carrying out deep learning by combining historical fuel consumption of each flight stage and the fuel characteristic data set of the corresponding flight stage, thereby predicting future fuel consumption of each flight stage and obtaining a flight overall process fuel consumption prediction result.

Description

Flight overall-process fuel consumption prediction method based on machine learning

Technical Field

The invention relates to the field of flight fuel consumption prediction, in particular to a flight overall-process fuel consumption prediction method based on machine learning.

Background

With the increase of environmental protection pressure and the requirement of sustainable development, the responsibility and the significance of making energy conservation and emission reduction work are obvious no matter the civil aviation stands in the aspects of cost reduction of enterprises or effective utilization of national energy.

In the work of energy conservation and emission reduction, the planning of fuel loading before the departure of a flight is a key circle. Too much fuel load can lead to too heavy fuel load, and then the problem of fuel consumption is caused; the hidden danger of civil aviation flight can be generated when the fuel oil loading capacity is insufficient. However, in the current practical operation process of civil aviation industry, how to accurately predict the oil consumption required by the flight in the whole flight process in advance is a pain point.

QAR is an abbreviation for Quick access recorder, quick access recorder. The QAR recording equipment is an important recording electronic equipment for recording the flight parameters of the airplane, can continuously record original information data of the flight for 600 hours, and can acquire hundreds of different flight parameter data or even thousands of different flight parameter data at the same time. The data recorded by the QAR cover most of the flight parameters of the airplane, such as longitude and latitude, altitude, wind speed, wind direction attack angle, oil consumption, temperature, air pressure and the like. In recent years, application research on QAR (quick access recorder) full data, which can play a role in a plurality of fields such as oil quantity, engine loss, flight safety, flight quality, and the like, is becoming a trend of civil aviation operation management, and therefore, more and more airlines are beginning to research and use QAR data to predict fuel consumption of future flight flights.

However, in the current research of predicting the fuel consumption of future flight by using QAR data in domestic and overseas aviation, there are several technical "blank spots" as follows:

1) There is a lack of quantitative screening of the fuel consumption contributors recorded for the QAR big data. The flight status parameters recorded in the QAR data are over 2000, with some parameters recorded every second. The fuel consumption prediction requires parameter aggregation and also requires parameter screening. Since using high dimensional feature data for prediction results in the risk of over-fitting, thereby reducing the generalization performance of the prediction model. At present, the research on fuel consumption influence factors of an airline company is mostly based on expert interview or manual screening, information is not extracted from QAR big data by adopting an artificial intelligence method, and the obtained fuel consumption influence factors are inaccurate.

2) And a technical method for predicting the fuel consumption in advance based on the QAR big data is lacked. The fuel consumption is predicted before the flight takes off, and guidance can be provided for the flight department to make a flight plan and determine the fuel filling amount. A reasonable fuel charge is an economic and safety tradeoff. On one hand, the excessive fuel filling amount can cause the heavy load of the airplane, so that the condition of 'fuel oil burning' is generated, and the fuel oil consumption cost is increased; on the other hand, too little fuel will cause the aircraft to fail to safely reach the designated airport and be forced to leave at another airport.

Many fuel consumption prediction studies are currently making multiple predictions of fuel flow rate. Fuel flow rate refers to the number of volumes or masses of fuel flowing through a conduit in a certain instantaneous unit of time or over a period of time. These studies often use a preamble of fuel consumption flow rate (fuel flow) during a flight to predict the subsequent fuel flow rate. Although the method has higher accuracy, the fuel consumption cannot be predicted in advance in the actual application scene, and the method has lower practical value. Meanwhile, the conventional fuel oil prediction method mostly adopts the traditional machine learning algorithm for prediction, lacks the algorithm application of integrated learning or deep learning, and has a larger promotion space in the aspects of prediction accuracy and algorithm optimization.

3) And the practical technology for predicting the whole flight process before the flight departure is lacked. Most of the current fuel prediction research focuses on the cruise phase, and less covers the phases except for cruise in the whole flight process, such as: the method comprises the following steps of taking off and sliding, climbing, descending and approaching, reaching sliding and the like, so that the practical target of predicting the fuel consumption of the whole flight process of the airplane in advance and providing reference for the advance fuel filling amount is difficult to achieve.

In summary, currently, each airline company still has "pain points" such as low fuel prediction accuracy and high fuel consumption cost, and the existing technologies and studies for fuel consumption prediction are not enough to deal with these pain points, and there are several following technologies "blank points": lack of feature extraction techniques for QAR big data; the application of prediction methods based on ensemble learning, deep learning and the like is lacked; and the practical technology for predicting the whole flight process before the flight departure is lacked.

Disclosure of Invention

The following presents a simplified summary of one or more aspects in order to provide a basic understanding of such aspects. This summary is not an extensive overview of all contemplated aspects, and is intended to neither identify key or critical elements of all aspects nor delineate the scope of any or all aspects. Its sole purpose is to present some concepts of one or more aspects in a simplified form as a prelude to the more detailed description that is presented later.

The invention aims to solve the problems and provides a method for predicting fuel consumption of the whole flight process based on machine learning, which comprises the steps of dividing the whole flight process into a plurality of flight stages according to the flight data of a flight, obtaining a flight parameter sample library of each flight stage, extracting the fuel consumption characteristics of each flight stage by using an integrated learning algorithm, determining the fuel characteristic data set of each flight stage according to the fuel consumption characteristics of each flight stage, and carrying out deep learning by combining the historical fuel consumption of each flight stage and the fuel characteristic data set of the corresponding flight stage, so as to predict the future fuel consumption of each flight stage and obtain the prediction result of the fuel consumption of the whole flight process.

The technical scheme of the invention is as follows:

the invention provides a flight whole-process fuel consumption prediction method based on machine learning, which comprises the following steps:

acquiring airplane flight data of a plurality of flights, and dividing flight phases based on the airplane flight data;

summarizing the airplane flight data corresponding to each flight stage according to flights, and acquiring a flight parameter sample library of each flight stage;

extracting fuel consumption characteristics of each flight phase based on the flight parameter sample library of each flight phase, and determining a fuel characteristic data set of each flight phase;

and predicting the future fuel consumption of each flight stage based on the historical fuel consumption of each flight stage and the fuel characteristic data set of the corresponding flight stage, and obtaining the fuel consumption prediction result of the whole flight process.

According to an embodiment of the method for predicting the fuel consumption of the whole process of flight based on machine learning, after the method for predicting the fuel consumption of the whole process of flight based on machine learning obtains the flight data of the airplane, the flight data of the airplane is subjected to data cleaning, and then the flight stages are divided according to the flight data of the airplane after the data cleaning, so that the detailed data of the flight parameters of each flight stage are obtained; wherein the flight phases include a roll-out phase, a climb phase, a cruise phase, a descent phase, and a roll-in phase.

According to an embodiment of the method for predicting the fuel consumption of the whole flight based on the machine learning, after flight parameter detail data of each flight stage are obtained, parameter summarization is respectively carried out on the flight parameter detail data of different flight stages of each flight to obtain flight parameter sample data of the corresponding flight stage, and the obtained flight parameter sample data are stored in a flight parameter sample library of the corresponding flight stage, so that a flight parameter sample library of each flight stage is obtained.

According to an embodiment of the method for predicting the fuel consumption of the whole flight process based on machine learning, the flight parameter sample library comprises a plurality of flight parameters corresponding to flight stages; after the flight parameter sample library of each flight stage is obtained, a plurality of flight parameters are extracted from the flight parameter sample library of each flight stage respectively to serve as fuel consumption characteristics of the corresponding flight stage.

According to an embodiment of the flight whole-process fuel consumption prediction method based on machine learning, the flight whole-process fuel consumption prediction method based on machine learning adopts a random forest algorithm to screen flight parameters of each flight stage, so that fuel consumption characteristics of each flight stage are obtained; the method for extracting the fuel consumption characteristics by adopting the random forest algorithm comprises the following steps:

constructing a plurality of sampling sets based on a flight parameter sample library; each flight phase comprises a plurality of sampling sets, and the number of samples of the sampling sets in the same flight phase is the same;

respectively constructing a plurality of base learners based on each sampling set; the base learner is used for sequencing all flight parameters, and then weighting and summarizing all flight parameters according to the sequence of the nodes corresponding to all flight parameters to obtain the importance of all flight parameters;

and summarizing the importance of each flight parameter obtained by each base learner in the same flight stage, and extracting the fuel consumption characteristics of the corresponding flight stage according to the summarized result of the characteristics.

According to one embodiment of the method for predicting the fuel consumption of the whole flight process based on the machine learning, the random forest algorithm adopts an autonomous sampling method to respectively sample a flight parameter sample library of each flight phase to obtain a sampling set of each flight phase, and then a plurality of base learners are respectively constructed based on each sampling set; and each flight phase comprises a plurality of sampling sets, and the number of samples of each sampling set is the number of samples of the flight parameter sample library of the corresponding flight phase.

According to one embodiment of the method for predicting the fuel consumption of the whole flight process based on machine learning, the CART decision tree is adopted as a base learner in the random forest algorithm, and CART decision tree regression is utilized to train a sampling set, so that the CART decision tree corresponding to a flight stage is obtained; the CART decision tree comprises a plurality of nodes, each node corresponds to one flight parameter, and after the training of the base learner is completed, the weight of each flight parameter is determined according to the flight parameters corresponding to each node of the CART decision tree.

According to an embodiment of the method for predicting the fuel consumption of the whole flight process based on the machine learning, the CART decision tree selects the node characteristics of each node by adopting information gain, and the node characteristics corresponding to each node are determined by calculating the information gain of each flight parameter in each node.

According to an embodiment of the machine learning-based fuel consumption prediction method for the whole flight process, the CART decision tree adopts a mean square error coefficient of each node as a node characteristic selection basis, calculates the information gain of each flight parameter at each node through the following formula, then selects the flight parameter with the maximum information gain as a node characteristic of a corresponding node, establishes a child node of the corresponding node by using the node with the maximum information gain until the node characteristic corresponding to each node is determined, and completes CART decision tree training:

；

where N represents the number of samples in the sample set, N _t Indicating the number of samples of the current node,

representing the number of samples of the left branch of the current node,

representing the number of samples of the right branch of the current node.

According to an embodiment of the method for predicting the fuel consumption of the whole flight process based on machine learning, after the CART decision tree is trained, information gains of all node characteristics are normalized, so that the importance of all the node characteristics is obtained; the random forest algorithm is used for summarizing the importance of the characteristics of all nodes of the plurality of CART decision trees in the same flight stage, so that the fuel consumption characteristics of the corresponding flight stage are obtained.

According to one embodiment of the method for predicting the fuel consumption of the whole process of flight based on machine learning, the feature summarizing algorithm comprises a feature weighting algorithm, a feature ranking algorithm and a feature subset screening algorithm; the random forest algorithm adopts a feature summarization algorithm to summarize the importance of each node feature of a plurality of CART decision trees in the same flight stage, sequences the feature summarization results, and selects a plurality of node features as fuel consumption features of the corresponding flight stage according to the sequencing results.

According to an embodiment of the machine learning-based flight overall process fuel consumption prediction method, after fuel consumption characteristics of different flight stages are obtained, flight parameter sample data corresponding to the fuel consumption characteristics are extracted from a flight parameter sample library corresponding to the flight stages to serve as a fuel characteristic data set, and then future fuel consumption of each flight stage is predicted based on historical fuel consumption of each flight stage and the fuel characteristic data set corresponding to the flight stages, and a flight overall process fuel consumption prediction result is obtained.

According to an embodiment of the flight overall process fuel consumption prediction method based on machine learning, after the flight overall process fuel consumption prediction method based on machine learning obtains the fuel characteristic data sets of all flight stages, the fuel consumption prediction model of each flight stage is constructed based on the historical fuel consumption of each flight stage and the fuel characteristic data sets of the corresponding flight stages, model training is carried out on the fuel consumption prediction model of each flight stage, and therefore the fuel consumption prediction model for predicting the future fuel consumption of each flight stage is obtained.

According to an embodiment of the flight whole-process fuel consumption prediction method based on machine learning, the flight whole-process fuel consumption prediction method based on machine learning adopts an LSTM neural network to construct a fuel consumption prediction model of each flight stage, and model training is carried out on the LSTM neural network of each flight stage by using a fuel characteristic data set of the corresponding flight stage of each flight stage and historical fuel consumption of the corresponding flight stage, so that a fuel consumption prediction model based on deep learning is obtained.

According to an embodiment of the flight overall process fuel consumption prediction method based on machine learning, after model training of fuel consumption prediction models of all flight stages is completed, the fuel consumption prediction models of all flight stages are used for predicting future fuel consumption of all flight stages of a target flight, and the future fuel consumption of all flight stages is summed to obtain a flight overall process fuel consumption prediction result.

Compared with the prior art, the invention has the following beneficial effects: the method comprises the steps of dividing the whole flight process of the flight into a plurality of flight stages according to the flight data of the flight, summarizing the flight data of the flight stages to obtain a flight parameter sample library of the flight stages, and extracting the fuel consumption characteristics of the flight stages based on the flight parameter sample library of the flight stages. Compared with the existing feature extraction method, the method provided by the invention has the advantages that aiming at flight features of different flight stages, the fuel consumption features of corresponding flight stages are respectively extracted from the flight parameter sample library of each flight stage by adopting integrated learning, so that the accuracy of extracting the fuel consumption features is improved, and the accuracy of fuel consumption is further improved. In addition, the method also utilizes a deep learning algorithm to learn and train the extracted fuel consumption characteristics of each flight stage according to the historical fuel consumption of each flight stage, and captures the fuel consumption time sequence characteristics of each flight stage, so that the future fuel consumption of each flight stage is predicted, the fuel consumption prediction of the whole flight process is finally realized, the fuel consumption is reduced, and the energy conservation and emission reduction are realized.

Drawings

The above features and advantages of the present disclosure will be better understood upon reading the detailed description of embodiments of the disclosure in conjunction with the following drawings. In the drawings, components are not necessarily drawn to scale, and components having similar relative characteristics or features may have the same or similar reference numerals.

Fig. 1 is a flow chart illustrating an embodiment of the method for predicting fuel consumption of the whole process of flight based on machine learning according to the invention.

FIG. 2 is a flow chart illustrating one embodiment of the present invention for extracting fuel consumption characteristics using a random forest algorithm.

FIG. 3 is a block diagram illustrating one embodiment of an LSTM neural network of the present invention.

FIG. 4 is a block diagram illustrating one embodiment of a three-layer stacked LSTM neural network of the present invention.

Detailed Description

The invention is described in detail below with reference to the figures and specific embodiments. It is noted that the aspects described below in connection with the figures and the specific embodiments are only illustrative and should not be construed as imposing any limitation on the scope of the present invention.

Fig. 1 is a structural diagram showing an embodiment of the flight overall process fuel consumption prediction method based on machine learning according to the present invention, and please refer to fig. 1, which is a detailed description of each step of the flight overall process fuel consumption prediction method based on machine learning.

Step S1: acquiring the airplane flight data of a plurality of flights, and dividing flight phases based on the airplane flight data.

In the embodiment, flight QAR data of a period of time is extracted according to the fuel consumption prediction requirement to serve as airplane flight data, and the extracted QAR data is subjected to data cleaning, so that high-quality QAR data is obtained. Because the flight forms of the airplanes are different in different flight stages, the influence factors and fluctuation characteristics of fuel consumption are also different, the whole flight process needs to be divided, and then the QAR of each flight stage is used for extracting the influence characteristics of the fuel consumption of the corresponding flight stage, so that the accuracy of extracting the influence characteristics of the fuel consumption of each flight stage is improved, and the accuracy of predicting the fuel consumption is improved.

In one embodiment, the flight overall process is divided into five stages of sliding out, climbing, cruising, descending and sliding in according to QAR data, and flight parameter detail data of each flight stage is obtained. The sliding-out stage refers to the processes of sliding and accelerating on the ground, the climbing stage refers to the process of converting kinetic energy into potential energy, the cruising stage refers to the process of stabilizing flight, the descending stage refers to the process of converting potential energy into kinetic energy, and the sliding-in stage refers to the processes of decelerating and sliding on the ground.

Specifically, in this embodiment, according to the flight phase parameter (flight phase) in the QAR data, the whole flight process of the flight is divided: (1) a stage in which the flight stage is "slide out" (flight phase = "taxi out") is classified as a slide out stage; (2) the flight phase from the end of the slide-out phase to the first cruise (flight phase = cruise) is classified as a climb phase; (3) dividing a flight phase from the end of the climbing phase to the appearance of the first time of 'descent' (flight phase = 'depth') into a cruising phase; (4) dividing the stage from the cruise stage to the stage before the cruise stage falls to the ground into a descending stage; (5) the stage from landing to engine shutdown is divided into a slide-in stage. Therefore, the whole flight stage of the flight is divided into five flight stages, the five flight stages form a complete flight process, and the QAR data is divided according to the five flight stages, so that the flight parameter detail data of each flight stage is determined.

Step S2: and summarizing the airplane flight data corresponding to each flight stage according to flights, and acquiring a flight parameter sample library of each flight stage.

In this embodiment, after the flight parameter detail data of each flight phase is obtained in step S1, the flight parameter detail data of each flight phase is subjected to parameter summarization according to flights, so as to obtain a flight parameter sample library of each flight phase.

Specifically, in this embodiment, the flight parameter detail data of each flight phase is QAR data, and the time granularity of the QAR data is a parameter record for each flight and each second. In order to facilitate subsequent machine learning training of fuel consumption prediction, the QAR data of each flight stage is summarized into a flight parameter sample data, and then the flight parameter sample data is stored into a flight parameter sample library of the corresponding flight stage, so that the flight parameter sample library of each flight stage is finally obtained. The parameter summarizing mode comprises three modes of minimum value, maximum value taking and average value taking. Since the integral of the engine fuel flow rate is equal to the engine oil consumption and does not belong to a factor affecting the engine oil consumption, it is necessary to delete parameters such as the engine fuel flow rate directly related to the predicted target value, i.e., the flight fuel consumption, when summarizing the parameters for each flight phase.

And step S3: and extracting the fuel consumption characteristics of each flight phase based on the flight parameter sample library of each flight phase, and determining the fuel characteristic data set of each flight phase.

In this embodiment, after the flight parameter sample library of each flight phase is obtained in step S2, in order to be different from the conventional feature extraction methods such as literature review and expert interview, the fuel consumption features of each flight phase are extracted by using a data-driven method.

Specifically, in this embodiment, the flight parameter sample library of each flight phase includes flight parameters of a plurality of corresponding phases, and in order to obtain the fuel consumption characteristics highly related to the fuel consumption of the flight phase, the flight parameters are screened by using a random forest algorithm, so as to extract the fuel consumption characteristics of the corresponding flight phase. Fig. 2 is a flowchart illustrating an embodiment of extracting fuel consumption characteristics by using a random forest algorithm according to the present invention, and please refer to fig. 2, which is a detailed description of each step of extracting fuel consumption characteristics by using a random forest algorithm.

Step C1: constructing a plurality of sampling sets based on a flight parameter sample library; wherein each flight phase comprises a plurality of sampling sets, and the sampling sets of the same flight phase have the same number of samples.

In this embodiment, the adopted random forest algorithm is an ensemble learning algorithm, and the algorithm generates a group of base learners through a bagging algorithm, trains each base learner, and then summarizes the training results of each base learner as the basis for feature extraction, thereby improving the stability and robustness of fuel consumption feature extraction. Because one sampling set corresponds to one base learner, for each flight phase, a plurality of sampling sets with the same number of samples need to be constructed by using the flight parameter sample library of the flight phase, and each sampling set is trained to obtain one base learner, so that each flight phase generates a plurality of base learners which are recorded as S base learners.

In addition, in this embodiment, the flight parameter sample base of each flight phase is sampled by using an autonomous sampling method, so as to obtain a sampling set of each flight phase, and then a plurality of basis learners are respectively constructed based on each sampling set. And each flight phase adopts max { 1/10 of the total parameter number } as the number of sampling sets, and the number of samples of each sampling set is the number of samples of the flight parameter sample library corresponding to the flight phase.

Specifically, in this embodiment, one flight parameter sample library includes N samples, and when the sampling autonomous sampling method is used to sample the flight parameter sample library, one sample is randomly extracted each time and placed in the sampling set, and the sample is placed back. The process is repeated for N times, so that a sampling set containing N samples can be obtained, and the samples of the sampling set meet two conditions of enough quantity and certain difference, thereby improving the accuracy of feature extraction.

And C2: respectively constructing a plurality of base learners based on each sampling set; the base learner is used for sequencing all flight parameters, and then weighting and summarizing all flight parameters according to the sequence of the nodes corresponding to all flight parameters to obtain the importance of all flight parameters.

In this embodiment, a CART decision tree is used as a base learner, and a CART decision tree regression is used to train a sampling set, so as to obtain a CART decision tree corresponding to a flight phase. The CART decision tree comprises a plurality of nodes, each node corresponds to a flight parameter, and after the training of the base learner is completed, the weight of each flight parameter is determined according to the flight parameters corresponding to each node of the CART decision tree.

The CART (classification and regression tree) classification regression tree algorithm is a decision tree algorithm that can be used for regression, and reduces the impurity degree of a data set prediction target by continuously dividing nodes. In the present embodiment, the prediction target is a fuel consumption value. The average value of the fuel consumption is used as a predicted value by each division node, and the impurity degree of each node in the example is constructed by using a mean squared error coefficient mse (mean squared error). Each node corresponds to a flight parameter and the division of the value range of the parameter. The feature selection and division points of each node are obtained by taking the minimization of the weighted average of the divided mses as a target. The CART decision tree continuously divides nodes into binary trees until the mse reduction value is lower than a set threshold value after the nodes are continuously divided. The more mse is reduced at each division, indicating the more accurate the division using that node. Therefore, the importance of each node is calculated by the reduction of the impurity level after division of each node and the weight weighting of each node.

In one embodiment, the CART decision tree performs node feature selection on each node by using information gain, and determines a node feature corresponding to each node by calculating the information gain of each flight parameter at each node.

Specifically, in this embodiment, the CART decision tree uses a mean square error coefficient mse (mean squared error) of each node as a node feature selection basis, and calculates information gain of each flight parameter at each node according to the following formula:

；

representing the number of samples of the left branch of the current node,

representing the number of samples of the right branch of the current node. mse refers to the mean square error of the node and is a quantitative measure of the impurity of the node. After the information gain of each flight parameter at the node is obtained through calculation of the formula, the flight parameter with the largest information gain is selected as the node characteristic of the corresponding node, the sub-node of the node is established by utilizing the node division rule with the largest information gain, then the selected flight parameter is deleted, the node characteristic of the sub-node is selected based on the rest flight parameters, and therefore the CART decision tree is generated.

Step C3: and summarizing the importance of each flight parameter obtained by each base learner in the same flight stage, and extracting the fuel consumption characteristics of the corresponding flight stage according to the summarized result of the characteristics.

In this embodiment, after the CART decision tree is generated through the above steps, the information gain of each node feature in the CART decision tree is normalized, so as to obtain the importance of each node feature. The random forest algorithm is used for summarizing the importance of the characteristics of all nodes of the plurality of CART decision trees in the same flight stage, so that the fuel consumption characteristics of the corresponding flight stage are obtained.

Specifically, in the present embodiment, the feature summarizing algorithm includes a feature weighting algorithm (feature selection algorithm), a feature ranking algorithm (feature ranking algorithm), and a feature subset selection algorithm (feature subset selection algorithm). The random forest algorithm collects the characteristics of the importance of each node characteristic of a plurality of CART decision trees in the same flight stage through any one characteristic collection algorithm, sorts the characteristic collection results, and selects a plurality of node characteristics as the fuel consumption characteristics of the corresponding flight stage according to the sorting results.

In one embodiment, the random forest algorithm employs a feature ranking algorithm as the feature summarization algorithm. Specifically, in this embodiment, one flight phase includes S basis learners, and for each basis learner, K node-specific learners are screenedAnd characterizing the fuel consumption characteristics corresponding to the base learner. Definition of FS _i Denotes the ith basis learner, i =1, 2. Wherein, FS _i For a vector consisting of a set of 0-1 variables, a variable value of 0 or 1 is used to indicate whether the feature is selected, where 1 indicates selected and 0 indicates unselected.

Specifically, for the ith base learner, if the jth node feature is selected (importance ranking K top), then

If not, then,

after training S basis learners, S vectors FS can be obtained _i Then averaging each node feature to obtain the probability of each node feature being selected as

Finally, FS is obtained through calculation _j Sorting from big to small, and selecting the first K FS _j The corresponding node characteristic serves as the fuel consumption characteristic for the flight phase.

And step S4: and predicting the future fuel consumption of each flight stage based on the historical fuel consumption of each flight stage and the fuel characteristic data set of the corresponding flight stage, and obtaining the fuel consumption prediction result of the whole flight process.

In this embodiment, after the fuel consumption characteristics of each flight phase are extracted and obtained in the step S3, the corresponding flight parameter sample data is extracted from the flight parameter sample library of the corresponding flight phase according to the fuel consumption characteristics as a fuel characteristic data set, and then the future fuel consumption of each flight phase is predicted based on the historical fuel consumption of each flight phase and the fuel characteristic data set of the corresponding flight phase, so as to obtain the prediction result of the fuel consumption of the whole flight process.

Specifically, in this embodiment, the historical fuel consumption of each flight stage and the fuel characteristic data set corresponding to the flight stage are used to construct a fuel consumption prediction model of each flight stage, model training is performed on the fuel consumption prediction model of each flight stage, so as to obtain the fuel consumption prediction model of each flight stage, then the future fuel consumption of each flight stage of the future flight is predicted by using the models, and the future fuel consumption of each flight stage is summed up, so as to obtain a prediction result of the fuel consumption of the whole flight process.

Further, in this embodiment, because data such as meteorological environment in aviation flight are related to time series characteristics, the fuel consumption prediction model of each flight phase is constructed by using the LSTM neural network, and the LSTM neural network of each flight phase is model-trained by using the fuel characteristic data set of the corresponding flight phase of each flight phase and the historical fuel consumption of the corresponding flight phase, so as to obtain the fuel consumption prediction model based on deep learning. In particular, the LSTM neural network is a Recurrent Neural Network (RNN). Conventional neural networks cannot utilize information of preamble states, while RNN neural networks can learn information of preamble states, but RNN cannot learn long-term dependencies in time series when the interval between the current sequence information and the current prediction is very long. The LSTM neural network is a special RNN network specially designed for solving the problems of gradient loss and gradient explosion in the long sequence training process. Having only one delivery state h compared to RNN _t The LSTM has two transmission states, one is c _t (cell state), the other is h _t (hidden state). Fig. 3 is a block diagram showing an embodiment of the LSTM neural network according to the present invention, and the present embodiment will be further described with reference to fig. 3.

As shown in fig. 3, the LSTM neural network contains an input threshold, a forgetting threshold, and an output threshold. Where the LSTM neural network processes the current input information through the input gate. When the input gate processes the current input information, the value to be updated and the new information to be added are i _t ＝σ(W _i [h _t-1 ,x _t ]+b _i ) And

wherein i _t Is an input gate, W _i And W _c Is a weight, b _i And b _c Offset terms for input gate and input node, h _t-1 Is the output at time t-1, x _t For the new variable value input at time t, tanh is the hyperbolic tangent function. The information of the comprehensive input door and the forgetting door is updated to obtain

Wherein, C _t Cell state at time t, C _t-1 The cell state at time t-1,

is the input state of the memory cell.

Further, in this embodiment, the LSTM neural network determines which history information will be discarded from the memory module through the forgetting gate, and f is adopted _t ＝σ(W _f [h _t-1 ,x _t ]+b _f ) To screen h _t-1 And x _t And (4) information. Wherein, f _t For forgetting gates, σ () is sigmoid function, W _f Is a weight, b _f Is the bias term for the forgetting gate.

Further, in this embodiment, the LSTM neural network determines which information will be output under the memory module through the output gate. When the output result of the model is processed, the layer processing value is determined to be O through the hyperbolic tangent function _t ＝σ(W _o [h _t-1 ,x _t ]+b _o )；h _t ＝O _t ·tanh(C _t ). Wherein, O _t Is an output gate, W _o Is a weight, b _o To output the offset term of the gate, h _t Is the output at time t.

In summary, in the embodiment, the LSTM neural network controls the transmission state through each gating state, and memorizes the information that needs to be memorized for a long time in the model training process, and forgets unimportant information, so that the trained LSTM neural network adapts to the characteristics of the demand prediction based on the historical data of flight demand, i.e. the time series correlation and the long sequence. In the embodiment, the LSTM neural network is used as a fuel consumption prediction model of each flight stage, and the fuel consumption of the whole flight process can be predicted in advance through the time sequence characteristics of flight fuel consumption of the LSTM neural network.

In addition, in the embodiment, because the characteristics of each route are different, the LSTM neural network model is respectively constructed for each route for training. From the historical steps, a list of key contributors to fuel consumption can be derived. In addition, the historical fuel consumption is also a factor that affects the future fuel consumption. Specifically, in the present embodiment, the fuel consumption characteristics and the fuel consumption values of each flight phase are extracted in the history step S3, and the history data is arranged in a time series data format as an input for predicting the fuel consumption value of the next phase. Multivariate LSTM can take into account time series variations of external features such as weather and the like, as well as the impact on future fuel consumption in the model. Meanwhile, the example normalizes all input parameters to ensure the training stability of the LSTM model.

Further, in this embodiment, in order to increase the accuracy of the prediction model, a stacked LSTM neural network is used to predict the fuel consumption of each flight phase. Wherein the stacked LSTM neural network comprises a multi-layer network, the output of the L-th layer

The method is used for inputting in the next round, and by increasing the depth of the network, the training efficiency is improved, and further higher accuracy is obtained. In one embodiment, a three-layer stacked LSTM neural network as shown in fig. 4 is used to construct a deep learning network, thereby increasing the training speed of the fuel consumption prediction model for each flight phase.

The previous description of the disclosure is provided to enable any person skilled in the art to make or use the disclosure. Various modifications to the disclosure will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other variations without departing from the spirit or scope of the disclosure. Thus, the disclosure is not intended to be limited to the examples and designs described herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Those of skill would further appreciate that the various illustrative logical blocks, modules, circuits, and algorithm steps described in connection with the embodiments disclosed herein may be implemented as electronic hardware, computer software, or combinations of both. To clearly illustrate this interchangeability of hardware and software, various illustrative components, blocks, modules, circuits, and steps have been described above generally in terms of their functionality. Whether such functionality is implemented as hardware or software depends upon the particular application and design constraints imposed on the overall system. Skilled artisans may implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present invention.

The various illustrative logical blocks, modules, and circuits described in connection with the embodiments disclosed herein may be implemented or performed with a general purpose processor, a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field Programmable Gate Array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination thereof designed to perform the functions described herein. A general-purpose processor may be a microprocessor, but in the alternative, the processor may be any conventional processor, controller, microcontroller, or state machine. A processor may also be implemented as a combination of computing devices, e.g., a combination of a DSP and a microprocessor, a plurality of microprocessors, one or more microprocessors in conjunction with a DSP core, or any other such configuration.

The steps of a method or algorithm described in connection with the embodiments disclosed herein may be embodied directly in hardware, in a software module executed by a processor, or in a combination of the two. A software module may reside in RAM memory, flash memory, ROM memory, EPROM memory, EEPROM memory, registers, hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art. An exemplary storage medium is coupled to the processor such the processor can read information from, and write information to, the storage medium. In the alternative, the storage medium may be integral to the processor. The processor and the storage medium may reside in an ASIC. The ASIC may reside in a user terminal. In the alternative, the processor and the storage medium may reside as discrete components in a user terminal.

In one or more exemplary embodiments, the functions described may be implemented in hardware, software, firmware, or any combination thereof. If implemented in software as a computer program product, the functions may be stored on or transmitted over as one or more instructions or code on a computer-readable medium. Computer-readable media includes both computer storage media and communication media including any medium that facilitates transfer of a computer program from one place to another. A storage media may be any available media that can be accessed by a computer. By way of example, and not limitation, such computer-readable media can comprise RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage or other magnetic storage devices, or any other medium that can be used to carry or store desired program code in the form of instructions or data structures and that can be accessed by a computer. Any connection is properly termed a computer-readable medium. For example, if the software is transmitted from a web site, server, or other remote source using a coaxial cable, fiber optic cable, twisted pair, digital Subscriber Line (DSL), or wireless technologies such as infrared, radio, and microwave, then the coaxial cable, fiber optic cable, twisted pair, DSL, or wireless technologies such as infrared, radio, and microwave are included in the definition of medium. Disk (disk) and disc (disc), as used herein, includes Compact Disc (CD), laser disc, optical disc, digital Versatile Disc (DVD), floppy disk and blu-ray disc where disks (disks) usually reproduce data magnetically, while discs (discs) reproduce data optically with lasers. Combinations of the above should also be included within the scope of computer-readable media.

Claims

1. A flight whole-process fuel consumption prediction method based on machine learning is characterized by comprising the following steps:

2. The machine learning-based flight overall process fuel consumption prediction method according to claim 1, characterized in that after the machine learning-based flight overall process fuel consumption prediction method acquires aircraft flight data, the aircraft flight data is subjected to data cleaning, and then flight stages are divided according to the aircraft flight data after the data cleaning, so that flight parameter detail data of each flight stage is obtained; wherein the flight phases include a roll-out phase, a climb phase, a cruise phase, a descent phase, and a roll-in phase.

3. The machine learning-based flight overall process fuel consumption prediction method according to claim 2, wherein after obtaining the flight parameter detail data of each flight phase, the machine learning-based flight overall process fuel consumption prediction method performs parameter summarization on the flight parameter detail data of different flight phases of each flight to obtain flight parameter sample data of the corresponding flight phase, and stores the obtained flight parameter sample data into the flight parameter sample library of the corresponding flight phase, thereby obtaining the flight parameter sample library of each flight phase.

4. The machine learning-based flight overall-process fuel consumption prediction method of claim 3, wherein the flight parameter sample library comprises flight parameters of a plurality of corresponding flight phases; after the flight parameter sample library of each flight stage is obtained, a plurality of flight parameters are extracted from the flight parameter sample library of each flight stage respectively to serve as fuel consumption characteristics of the corresponding flight stage.

5. The flight whole-process fuel consumption prediction method based on machine learning as claimed in claim 4, characterized in that the flight whole-process fuel consumption prediction method based on machine learning adopts a random forest algorithm to screen flight parameters of each flight stage, so as to obtain fuel consumption characteristics of each flight stage; the method for extracting the fuel consumption characteristics by adopting the random forest algorithm comprises the following steps:

6. A flight whole-process fuel consumption prediction method based on machine learning as claimed in claim 5, characterized in that the random forest algorithm adopts an autonomous sampling method to respectively sample the flight parameter sample library of each flight phase to obtain a sampling set of each flight phase, and then a plurality of base learners are respectively constructed based on each sampling set; and each flight phase comprises a plurality of sampling sets, and the number of samples of each sampling set is the number of samples of the flight parameter sample library corresponding to the flight phase.

7. A machine learning-based flight whole-process fuel consumption prediction method according to claim 5, characterized in that the random forest algorithm adopts a CART decision tree as a base learner, and a CART decision tree regression is used for training a sampling set, so that a CART decision tree corresponding to a flight stage is obtained; the CART decision tree comprises a plurality of nodes, each node corresponds to one flight parameter, and after the training of the base learner is completed, the weight of each flight parameter is determined according to the flight parameters corresponding to each node of the CART decision tree.

8. The machine learning-based prediction method of fuel consumption for the whole flight process based on flight according to claim 7, wherein the CART decision tree selects node characteristics of each node by using information gain, and determines the node characteristics corresponding to each node by calculating the information gain of each flight parameter at each node.

9. The machine learning-based flight whole-process fuel consumption prediction method according to claim 1, characterized in that the CART decision tree adopts the mean square error coefficient of each node as the node feature selection basis, calculates the information gain of each flight parameter at each node by the following formula, then selects the flight parameter with the maximum information gain as the node feature of the corresponding node, and establishes the child node of the corresponding node by using the node with the maximum information gain until the node feature corresponding to each node is determined, and completes CART decision tree training:

;

representing the number of samples of the left branch of the current node,

representing the number of samples of the right branch of the current node.

10. The machine learning-based flight overall-process fuel consumption prediction method according to claim 9, characterized in that after the training of the CART decision tree is completed, the information gain of each node feature is normalized, so as to obtain the importance of each node feature; the random forest algorithm is used for summarizing the importance of the characteristics of all nodes of the plurality of CART decision trees in the same flight stage, so that the fuel consumption characteristics of the corresponding flight stage are obtained.

11. The machine learning-based flight overall-process fuel consumption prediction method according to claim 10, wherein the feature summarizing algorithm comprises a feature weighting algorithm, a feature ranking algorithm and a feature subset screening algorithm; the random forest algorithm adopts a feature summarization algorithm to summarize the importance of each node feature of a plurality of CART decision trees in the same flight stage, sequences the feature summarization results, and selects a plurality of node features as fuel consumption features of the corresponding flight stage according to the sequencing results.

12. The machine learning-based flight overall process fuel consumption prediction method according to claim 5, wherein after the machine learning-based flight overall process fuel consumption prediction method obtains fuel consumption characteristics of different flight phases, corresponding flight parameter sample data is extracted from a flight parameter sample library of the corresponding flight phase according to the fuel consumption characteristics to serve as a fuel characteristic data set, and then future fuel consumption of each flight phase is predicted based on the historical fuel consumption of each flight phase and the fuel characteristic data set of the corresponding flight phase, so that a flight overall process fuel consumption prediction result is obtained.

13. The machine learning-based flight overall process fuel consumption prediction method according to claim 12, wherein after the machine learning-based flight overall process fuel consumption prediction method obtains the fuel characteristic data sets of the flight stages, a fuel consumption prediction model of each flight stage is constructed based on the historical fuel consumption of each flight stage and the fuel characteristic data sets of the corresponding flight stages, and model training is performed on the fuel consumption prediction model of each flight stage, so that a fuel consumption prediction model for predicting the future fuel consumption of each flight stage is obtained.

14. The machine-learning-based flight whole-process fuel consumption prediction method according to claim 13, wherein the machine-learning-based flight whole-process fuel consumption prediction method adopts an LSTM neural network to construct a fuel consumption prediction model of each flight stage, and model training is performed on the LSTM neural network of each flight stage by using a fuel characteristic data set of the corresponding flight stage of each flight stage and the historical fuel consumption of the corresponding flight stage, so that a deep-learning-based fuel consumption prediction model is obtained.

15. The machine learning-based flight overall process fuel consumption prediction method according to claim 1, wherein after model training of the fuel consumption prediction models of the flight stages is completed, the fuel consumption prediction models of the flight stages are used for predicting future fuel consumption of the flight stages of the target flight, and the future fuel consumption of the flight stages is summed to obtain a flight overall process fuel consumption prediction result.