CN110288121A - Flight based on multiple time granularity attention mechanism visits rate prediction technique - Google Patents
Flight based on multiple time granularity attention mechanism visits rate prediction technique Download PDFInfo
- Publication number
- CN110288121A CN110288121A CN201910407946.2A CN201910407946A CN110288121A CN 110288121 A CN110288121 A CN 110288121A CN 201910407946 A CN201910407946 A CN 201910407946A CN 110288121 A CN110288121 A CN 110288121A
- Authority
- CN
- China
- Prior art keywords
- flight
- rate
- day
- attention mechanism
- time
- 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.)
- Pending
Links
- 230000007246 mechanism Effects 0.000 title claims abstract description 81
- 238000000034 method Methods 0.000 title claims abstract description 47
- 230000010006 flight Effects 0.000 claims abstract description 20
- 230000000306 recurrent effect Effects 0.000 claims abstract description 20
- 238000003062 neural network model Methods 0.000 claims abstract description 19
- 230000002123 temporal effect Effects 0.000 claims abstract description 12
- 238000012545 processing Methods 0.000 claims abstract description 5
- 238000012360 testing method Methods 0.000 claims description 14
- 230000006870 function Effects 0.000 claims description 12
- 238000012549 training Methods 0.000 claims description 12
- 238000002474 experimental method Methods 0.000 claims description 9
- 230000004927 fusion Effects 0.000 claims description 9
- 241000208340 Araliaceae Species 0.000 claims description 7
- 235000005035 Panax pseudoginseng ssp. pseudoginseng Nutrition 0.000 claims description 7
- 235000003140 Panax quinquefolius Nutrition 0.000 claims description 7
- 235000008434 ginseng Nutrition 0.000 claims description 7
- 238000003780 insertion Methods 0.000 claims description 7
- 230000037431 insertion Effects 0.000 claims description 7
- 230000003044 adaptive effect Effects 0.000 claims description 4
- 230000007787 long-term memory Effects 0.000 claims description 4
- 230000000630 rising effect Effects 0.000 claims description 4
- 230000008859 change Effects 0.000 claims description 3
- 239000011159 matrix material Substances 0.000 claims description 3
- 239000000284 extract Substances 0.000 claims description 2
- 125000006850 spacer group Chemical group 0.000 claims 1
- 230000000694 effects Effects 0.000 abstract description 4
- 238000010586 diagram Methods 0.000 description 7
- 230000008569 process Effects 0.000 description 5
- 238000013461 design Methods 0.000 description 4
- 238000013528 artificial neural network Methods 0.000 description 3
- NJPPVKZQTLUDBO-UHFFFAOYSA-N novaluron Chemical compound C1=C(Cl)C(OC(F)(F)C(OC(F)(F)F)F)=CC=C1NC(=O)NC(=O)C1=C(F)C=CC=C1F NJPPVKZQTLUDBO-UHFFFAOYSA-N 0.000 description 3
- 238000010276 construction Methods 0.000 description 2
- 230000008878 coupling Effects 0.000 description 2
- 238000010168 coupling process Methods 0.000 description 2
- 238000005859 coupling reaction Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 238000011156 evaluation Methods 0.000 description 2
- 230000009466 transformation Effects 0.000 description 2
- 241001269238 Data Species 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 230000000052 comparative effect Effects 0.000 description 1
- 230000003247 decreasing effect Effects 0.000 description 1
- 230000007774 longterm Effects 0.000 description 1
- 210000004218 nerve net Anatomy 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 239000000843 powder Substances 0.000 description 1
- 230000000750 progressive effect Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000006403 short-term memory Effects 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 239000011800 void material Substances 0.000 description 1
Classifications
-
- 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
- G06N3/044—Recurrent networks, e.g. Hopfield 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
- G06N3/04—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
- G06Q10/04—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
- 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
-
- G06Q50/40—
Abstract
The present invention provides a kind of flights based on multiple time granularity attention mechanism to visit rate prediction technique.This method comprises: Recognition with Recurrent Neural Network model of the building based on multiple time granularity attention mechanism, the flight history of departure times all on course line is visited into rate time series as the list entries of the encoder, encoder carries out coded treatment to list entries, decoder is decoded processing to the encoded information of encoder output, and the flight for obtaining target flight visits rate time series.Different departure time flights visit the Temporal dependency of rate in course line where the present invention captures target flight by departure time attention mechanism and other departure time flights visit the influence that rate visits rate to target flight, while the tendency and periodicity of itself visiting rate sequence using day attention mechanism capture target flight that takes off；It takes the influence of the external factor such as flight self attributes and festivals or holidays into consideration, visits this model in flight and achieve good effect in rate forecasting problem.
Description
Technical field
The present invention relates to flights to visit rate electric powder prediction, more particularly to a kind of based on multiple time granularity attention mechanism
Flight visit rate prediction technique.
Background technique
With increasing rapidly for civil aviation passenger, air passenger demand prediction increasingly by airline, nash-equilibrium quotient,
The concern of the civil aviatons such as planemaker relevant enterprise.Air passenger traffic amount requirement forecasting includes course line passenger traffic volume requirement forecasting, airport
Passenger traffic volume requirement forecasting, the prediction of airline's market share and flight passenger traffic volume requirement forecasting etc..Flight passenger traffic volume demand
Prediction is the requirement forecasting in finer grain, is the basis of airline seat optimal control and pricing strategy.
For air passenger traffic market, it is to measure an important indicator of flight passenger traffic volume demand that flight, which visits rate,.Aviation
Whether company visits rate by flight and matches between transport power and freight volume to measure, and handles a series of variations, including overflow,
Void consumption surpasses and sells；Planemaker is also required to pay close attention to visiting rate, and visiting rate is whether airline needs to increase transport power, introduces
One important indicator of aircraft.Civil aviaton market practitioner by accurate target flight visit rate, can the sensed in advance market demand,
It improves enterprise income management level, provide decision support for operation departments at different levels.
Currently, existing flight, which visits rate prediction technique, has only focused on the Variation Features that flight visits rate daily in the recent period, nothing
Other neighbouring departure times in course line where method considers the influence for other many factors that the problem is subject to, such as flight simultaneously
Flight visit rate, flight self attributes and extraneous factor etc..
Summary of the invention
The embodiment provides a kind of flights based on multiple time granularity attention mechanism to visit rate prediction technique,
To overcome problem of the prior art.
To achieve the goals above, this invention takes following technical solutions.
A kind of visiting rate prediction technique of the flight based on multiple time granularity attention mechanism, comprising:
Construct the Recognition with Recurrent Neural Network model based on multiple time granularity attention mechanism, the Recognition with Recurrent Neural Network model packet
It includes: in conjunction with the encoder of departure time attention mechanism and in conjunction with day decoder for attention mechanism that takes off；
The flight history of departure times all on course line is visited into rate time series as the list entries of the encoder,
The encoder carries out coded treatment to the list entries, and the decoder carries out the encoded information of the encoder output
Decoding process, the flight for obtaining target flight visit rate time series.
Preferably, the encoder of the combination departure time attention mechanism includes the first layer shot and long term note of hierarchical structure
Recall network LSTM unit and second layer LSTM unit, by target flight on-course all flights history visit rate time sequence
Column are input to each time step of the first layer LSTM unit, the first layer LSTM unit output integrated according to departure time
Hidden layer state value under each departure time for considering departure time timing；The output of the first layer LSTM unit is made
For the input of departure time attention mechanism, departure time attention mechanism by reference to before the second layer LSTM unit together
The hiding layer state for flying day is adaptively extracting hiding for the first layer LSTM unit related departure time each day of taking off
Stratiform state value, hidden layer state value of the second layer LSTM unit output target flight in day of respectively taking off, comprehensive described first
The flight of different departure times visits rate in course line where layer LSTM unit and the second layer LSTM elements capture target flight
Temporal dependency and other departure time flights influence of the visiting rate to the visiting rate of target flight.
Preferably, the method includes:
Given time length of window is D, is used
Indicate the T time series that the visiting rate that the flight of all departure times on a course line is gone over D days is constituted, whereinIndicate the visiting rate time series that the flight of t-th of departure time is gone over D days,Indicate that visiting rate in the flight of d-th of the past day of taking off all departure times is constituted
Vector；
Given list entries X=(x1, x2..., xT), whereinBy xtExist as the first layer LSTM unit
The input of t-th of departure time flight, and use ht=feb(ht-1, xt) update the first layer LSTM unit and take off at t-th
The hidden layer state value at moment, wherein febThe renewal function for representing the first layer LSTM unit, by the first layer LSTM
The hidden layer state value h=(h under each departure time is calculated in unit1, h2..., hT), whereinIt is described hidden
Hiding stratiform state value h considers the Temporal dependency of the visiting rate of flight between all departure times on course line；
To the hidden layer state value h for t-th of departure time that the first layer LSTM unit obtainst, using taking off as follows
The calculating of moment attention mechanism:
Wherein,It is to measure the visiting rate time series of the flight of t-th of departure time when taking off at the d days day to target
The same day of flight visits the attention weight of rate influence degree, and the parameter for needing to learn is
It is rightIt is normalized, making the sum of all attention weights is 1, then for day d that takes off, the first layer LSTM is mono-
Member combines the output vector of departure time attention mechanism as follows:
The hidden layer state value output of day d that takes off of the second layer LSTM unit output is as follows:
qd=fea(qd-1, zd)
Wherein feaIt is the renewal function of the second layer LSTM unit.
Preferably, the method further include:
Using flight self attributes and festivals or holidays attribute as extraneous factor, the flight self attributes include target flight institute
Belong to airline, type, play landing GDP, airport of rising and falling, departure time and coach cabin class seats number, the festivals or holidays attribute and includes
Whether it is festivals or holidays, whether is whether working day, week attribute and the same day open up high-speed rail；
By in the extraneous factor each flight self attributes and each festivals or holidays attribute be converted to low-dimensional vector, will be each
A low-dimensional vector inputs different embeding layers respectively to generate corresponding insertion vector, will obtain after the processing of each insertion Vector Fusion
It arrivesD ' is to take off day in the future of decoder prediction, will be describedIt is transferred to the decoder.
Preferably, the method further include:
Correlation in all time steps of second layer LSTM unit is adaptive selected using day attention mechanism of taking off
Hidden layer state value, day of taking off to calculate the corresponding encoder of prediction output valve in the day decoder the d ' at d-th hide
The attention weight of stratiform state value, is defined as follows:
The parameter for wherein needing to learn isIt is rightIt carries out
Normalization, gainedAs attention weight, cd′Indicate the defeated of day attention mechanism that take off for the decoder future day the d '
Outgoing vector；
In a decoder, the context vector c of the weighted sum for the future day the d ' is calculatedd′It later, will be upper and lower
Literary vector cd′With obtain after the insertion Vector Fusion according to external factorAnd decoder last moment is defeated convexCombine that come more new decoder hidden layer state, formula as follows:
Wherein, fdLSTM unit renewal function used in decoder is indicated, then by context vector cd′With hidden layer
State gd′Vector splicing is carried out, the new hiding layer state finally predicted is as follows:
Wherein, matrixAnd vectorVector will be spliced
Dimension map uses linear transformation to decoder hidden layer dimensionTo generate decoder in d '
The final output of a time stepDecoder is integrated to obtain the future one of target flight in the output of all time steps
The flight of section time visits rate.
Preferably, the method further include:
The visiting rate data of flight history that all departure times on certain course line are chosen in rate data set are visited in history flight
As experimental data set, experiment sample is constructed based on the experimental data set, the experiment sample is divided into training set, verifying
Collection and test set train the Recognition with Recurrent Neural Network model based on multiple time granularity attention mechanism using the training set,
Optimal super ginseng is chosen, the experimental data set drag optimized parameter out is learnt, is obtained trained based on multiple time granularity note
The Recognition with Recurrent Neural Network model for power mechanism of anticipating, using the test set to described trained based on multiple time granularity attention machine
The Recognition with Recurrent Neural Network model of system is tested, using verifying collection to described trained based on multiple time granularity attention
The test result of the Recognition with Recurrent Neural Network model of mechanism is verified.
As can be seen from the technical scheme provided by the above-mentioned embodiment of the present invention, the MTA-RNN model of the embodiment of the present invention
Multistage attention mechanism, respectively departure time attention mechanism and day attention mechanism of taking off are constructed according to time granularity, is led to
The combination of two-stage attention mechanism is crossed, flight is obtained and visits timing dependence of the rate under different time granularity.The present invention passes through
Departure time attention mechanism capture where target flight different departure time flights in course line visit rates Temporal dependency and
Other departure time flights visit the influence that rate visits rate to target flight, while using a day attention mechanism capture target of taking off
The tendency and periodicity of itself visiting rate sequence of flight；Furthermore the model considers flight using external factor Fusion Module
The influence of the external factor such as self attributes and festivals or holidays is visited this model in flight and is achieved very in rate forecasting problem
Good effect.
The additional aspect of the present invention and advantage will be set forth in part in the description, these will become from the following description
Obviously, or practice through the invention is recognized.
Detailed description of the invention
In order to illustrate the technical solution of the embodiments of the present invention more clearly, required use in being described below to embodiment
Attached drawing be briefly described, it should be apparent that, drawings in the following description are only some embodiments of the invention, for this
For the those of ordinary skill of field, without creative efforts, it can also be obtained according to these attached drawings others
Attached drawing.
Fig. 1 is a kind of Recognition with Recurrent Neural Network model based on multiple time granularity attention mechanism provided in an embodiment of the present invention
(MTA-RNN) architecture diagram；
Fig. 2 is that the different departure time flight economy classes of in March, 2010 Beijing-Shanghai route visit rate comparison diagram；
Fig. 3 is that a kind of flight based on multiple time granularity attention provided in an embodiment of the present invention visits rate prediction technique
Process flow diagram.
Specific embodiment
Embodiments of the present invention are described below in detail, the example of the embodiment is shown in the accompanying drawings, wherein from beginning
Same or similar element or element with the same or similar functions are indicated to same or similar label eventually.Below by ginseng
The embodiment for examining attached drawing description is exemplary, and for explaining only the invention, and is not construed as limiting the claims.
Those skilled in the art of the present technique are appreciated that unless expressly stated, singular " one " used herein, " one
It is a ", " described " and "the" may also comprise plural form.It is to be further understood that being arranged used in specification of the invention
Diction " comprising " refer to that there are the feature, integer, step, operation, element and/or component, but it is not excluded that in the presence of or addition
One or more of the other feature, integer, step, operation, element, component and/or their group.It should be understood that when we claim member
Part is " connected " or when " coupled " to another element, it can be directly connected or coupled to other elements, or there may also be
Intermediary element.In addition, " connection " used herein or " coupling " may include being wirelessly connected or coupling.Wording used herein
"and/or" includes one or more associated any cells for listing item and all combinations.
Those skilled in the art of the present technique are appreciated that unless otherwise defined, all terms used herein (including technology art
Language and scientific term) there is meaning identical with the general understanding of those of ordinary skill in fields of the present invention.Should also
Understand, those terms such as defined in the general dictionary, which should be understood that, to be had and the meaning in the context of the prior art
The consistent meaning of justice, and unless defined as here, it will not be explained in an idealized or overly formal meaning.
In order to facilitate understanding of embodiments of the present invention, it is done by taking several specific embodiments as an example below in conjunction with attached drawing further
Explanation, and each embodiment does not constitute the restriction to the embodiment of the present invention.
Embodiment one
The embodiment of the present invention proposes a kind of Recognition with Recurrent Neural Network model (MTA- based on multiple time granularity attention mechanism
RNN), Fig. 1 is a kind of architecture diagram of above-mentioned MTA-RNN model provided in an embodiment of the present invention, the model construction multistage attention
Mechanism obtains flight and visits timing dependence of the rate under different time granularity, which includes: in conjunction with departure time attention
The encoder of mechanism and combine day decoder for attention mechanism that takes off.
The encoder of the combination departure time attention mechanism includes the first layer shot and long term memory network of hierarchical structure
LSTM unit and second layer LSTM unit, first layer LSTM unit are the encoder lower section LSTM unit in Fig. 1, second layer LSTM
Unit is the encoder top LSTM unit in Fig. 1.By target flight on-course all flights history visit the rate time
Sequence is input to each time step of the first layer LSTM unit according to departure time, and the first layer LSTM unit output is comprehensive
Close the hidden layer state value under each departure time for considering departure time timing；The output of the first layer LSTM unit
As the input of departure time attention mechanism, departure time attention mechanism is previous by reference to the second layer LSTM unit
Take off the hiding layer state of day, is adaptively extracting the hidden of the first layer LSTM unit related departure time each day of taking off
Hide stratiform state value, hidden layer state value of the second layer LSTM unit output target flight in day of respectively taking off, comprehensive described the
The flight of different departure times is visiting in course line where one layer of LSTM unit and the second layer LSTM elements capture target flight
Influence of the visiting rate of the Temporal dependency of rate and other departure time flights to the visiting rate of target flight.
Based on above-mentioned MTA-RNN model, a kind of flight based on multiple time granularity attention provided in an embodiment of the present invention
Visiting rate prediction technique, this method construct multistage attention mechanism and obtain visiting timing phase of the rate under different time granularity of flight
Guan Xing.In the encoder, different departure time flight visitors in course line where model uses LSTM elements capture prediction flight first
The Temporal dependency of seat rate, introduces departure time attention mechanism later, by reference to before LSTM unit above encoder together
The hiding layer state for flying day is adaptively extracting hiding for the related departure time of LSTM unit below encoder each day of taking off
Stratiform state value；In a decoder, present invention introduces take off day attention mechanism in all days of taking off to select associated encoder
The hiding layer state of top LSTM unit, and the extraneous factors such as flight self attributes and festivals or holidays are combined, finally obtain future one
The target flight economy class of section time visits rate.The treatment process of this method the following steps are included:
S1. the definition of formalization representation of the present invention and the forecasting problem solved is provided；
S1.1 formalization representation of the present invention is defined as follows:
In most cases, the departure time of flight immobilizes, and the departure time in course line and specific flight one are a pair of
It answers.Therefore, the flight quantity in course line is approximately equal with departure time quantity.Assuming that in some course line there are T departure time (or
Flight), it selects the flight of a departure time to visit rate time series as target sequence in course line, which is used
In model prediction, the flight of remaining departure time visits rate time series as feature.
Given time length of window is D, and the present invention uses
Indicate the T time series that the visiting rate that the flight of all departure times on the course line is gone over D days is constituted, whereinIndicate the visiting rate time series that the flight of t-th of departure time is gone over D days,Indicate that visiting rate in the flight of d-th of the past day of taking off all departure times is constituted
Vector；The embodiment of the present invention usesIndicate the mesh that departure time is i in D days in the past
Mark the visiting rate time series of flight.
S1.2 forecasting problem solved by the invention is defined as follows:
According to the flight history T days visiting rate observation X and flight itself of departure times all on given course line
Attribute and extraneous factor
Wherein τ is the number of days for the target flight future visiting rate for predicting that departure time is i,Indicate target flight not
Carry out the τ days extraneous factors.
It predicts that departure time is the visiting rate in the τ days flight future of target of i, is denoted as
S2. the Recognition with Recurrent Neural Network model based on multiple time granularity attention mechanism is constructed, above-mentioned circulation nerve net is utilized
Network model carries out flight and visits rate prediction.
The main framework of MTA-RNN model proposed by the present invention follows codec models, including encoder and decoder
Two large divisions.Encoder encodes list entries, and above-mentioned list entries is the flight history of all departure times on course line
Visiting rate time series.Decoder is for predicting output sequenceMore specifically, MTA-RNN model is by two masters below
Mian part is grouped as:
1) multiple time granularity attention mechanism.It is taken off day by combining the encoder of departure time attention mechanism and combining
The decoder of attention mechanism forms.In the encoder, the present invention uses two independent LSTM (Long Short-Term
Memory, shot and long term memory network) unit, Fig. 1 Encoder beneath portions LSTM unit is for course line where capturing target flight
The flight of middle difference departure time visits the Temporal dependency of rate, introduces departure time attention mechanism, the machine on this basis
System is existed by reference to the previous hidden layer state value of Encoder upper LSTM unit and Encoder beneath portions LSTM unit
The state output value of each departure time, the flight that can adaptively capture other departure times under each day of taking off visit rate
The influence of rate is visited to target flight；In a decoder, present invention introduces day attention mechanism of taking off, and target is adaptive selected
The state output value of day rank of taking off of flight is to capture itself tendency and periodicity that target flight visits rate time series.
2) external factor merges.The module is for handling the extraneous factors such as flight self attributes and festivals or holidays to target flight
The output of the influence of visiting rate, the module is supplied to decoder as part input.
In Fig. 1, FlightTimeAttn represents departure time attention mechanism, and FlightDayAttn representative is taken off day
Attention mechanism, Concat represent articulamentum；Expression departure time is predicted value of the target flight of i in the day the d ', cd′
Indicate the context vector in the day the d ', h0、q0Respectively indicate the initial hidden layer state value of two LSTM in encoder.This hair
Bright useHidden layer state value of the LSTM t-th of departure time below presentation code device,Hidden layer state value and note of the LSTM in d-th of history day of taking off above encoder in expression figure
Recall cell-like state value；Equally,Indicate decoder in the hidden layer state value and note of prediction day d '
Recall cell-like state value.
S2.1 obtains flight using multiple time granularity attention mechanism and visits timing correlation of the rate under different time granularity
Property.
In conjunction with the encoder of departure time attention mechanism, due to all departure times on same course line, there are timing
Property, the present invention is using when difference is taken off in course line where LSTM unit (Fig. 1 Encoder beneath portions LSTM) capture target flight
The flight at quarter visits the Temporal dependency of rate.Given list entries X=(x1, x2..., xT), whereinThe present invention will
xtAs the LSTM unit t-th of departure time flight input, and use ht=feb(ht-1, xt) update the LSTM unit and exist
The hidden layer state value of t-th of departure time, wherein febRepresent lower section LSTM unit renewal function in encoder.By the LSTM
The hidden layer state value h=(h under each departure time is calculated in unit1, h2..., hT), wherein The value is examined
The Temporal dependency of the visiting rate of flight between all departure times on course line is considered.In order to adaptively capture target sequence and
Correlation between the time series that the daily visiting rate of the flight of other departure times is constituted, for LSTM unit below encoder
In the hidden layer state value h of t-th of departure time outputt, the present invention is using following departure time attention mechanism:
The parameter for wherein needing to learn isDeparture time
Attention weight is LSTM unit above hidden layer state value and encoder by LSTM unit each departure time below encoder
Historic state (such as qd-1, sd-1) codetermine.To measure when taking off at the d days day, the flight of t-th of departure time its
Visiting rate time series visited the attention weight of rate influence degree to the target flight same day.The present invention uses softmax function
It is rightIt is normalized, making the sum of all attention weights is 1.Departure time attention mechanism is feedforward neural network, can be with
It is trained jointly with other components of RNN.After attention weight calculation, then for day d that takes off, departure time attention mechanism
Output vector it is as follows:
Then hidden layer state value output of the LSTM unit in day d that takes off is as follows above encoder:
qd=fea(qd-1, zd)
Wherein feaIt is LSTM unit renewal function above encoder.The departure time attention machine proposed through the invention
System, the encoder top LSTM unit property of can choose focus on LSTM unit hiding in certain departure times below encoder
Stratiform state value, rather than coequally handle the hidden layer state value of all departure times.
In conjunction with day decoder of attention mechanism is taken off in order to predict departure time be i τ days flight future of target visitor
Seat rateThe present invention decodes coded input information using another Recognition with Recurrent Neural Network based on LSTM.It is solved due to compiling
The performance of code device structure can be gradually decreased with the increase of encoder length, therefore the present invention is in a decoder using the day note that takes off
The related hidden layer state value above encoder in all time steps of LSTM unit is adaptive selected in meaning power mechanism, i.e., catches automatically
Obtain itself tendency and periodicity that target flight visits rate time series.Specifically, in order to calculate the pre- of the day decoder the d '
Survey the corresponding encoder of output valve take off at d-th day hidden layer state value attention weight, the present invention is defined as follows:
The parameter for wherein needing to learn isUsing
Softmax function pairIt is normalized, gainedAs attention weight.cd′It indicates for the decoder future day the d ',
It takes off a day output vector for attention mechanism.
In a decoder, the context vector c of the weighted sum for the future day the d ' is calculatedd′Later, of the invention
Then by the output of itself and external factor Fusion ModuleAnd the output of decoder last momentIt combines to update
Decoder hides layer state, and formula is as follows:
Wherein, fdIndicate LSTM unit renewal function used in decoder.Then by context vector cd′With hidden layer
State gd' vector splicing is carried out, this will become the new hiding layer state finally predicted, as follows:
Wherein, matrixAnd vectorVector will be spliced
Dimension map is to decoder hidden layer dimension.The last present invention uses linear transformation (i.e.) Lai Shengcheng
Final output.For decoder a time step of d ' final output, by decoder all time steps output carry out
The flight of comprehensive following a period of time for obtaining target flight visits rate.
The S2.2 present invention designs external factor Fusion Module and considers the extraneous factors such as flight self attributes and festivals or holidays
The interior forecasting accuracy to improve the visiting rate of flight.
The visiting rate of flight is influenced by extraneous factors such as flight self attributes and festivals or holidays and Liang et al. is in research work
The influence that external factor applies space-time is paid close attention in work.It is inspired by this, the present invention devises a simple and effective module and comes
Handle these factors.
When flight self attributes include the affiliated airline of target flight, type, play landing GDP, airport of rising and falling, take off
It carves, coach cabin class seats number；In extraneous factor, due to the finiteness of data acquisition, whether the present invention access time feature (is saved
Holiday, whether working day, week attribute) and the same day whether open up high-speed rail feature.As shown in Figure 1, first the present invention by flight from
Body attribute and extraneous factor are input to external factor Fusion Module simultaneously.Since most of these factors are all Category Attributes, no
Neural network can be directly inputted, each Category Attributes are converted to low-dimensional vector by the present invention, they are inputted respectively different embedding
Enter layer to generate corresponding insertion vector.Vector finally will be embedded in and remaining feature is connected to the output of the module, be denoted asD ' is to take off day in the future of decoder prediction.
Embodiment two
Fig. 2 is that the different departure time flight economy classes of in March, 2010 Beijing-Shanghai route visit rate comparison diagram, and horizontal axis is
Flight takeoff day, the longitudinal axis are that flight economy class visits rate, and the economy class that different curves represent the flight of different departure times is visiting
Rate change curve.As can be seen from Figure 2 flight is visited rate and is had a characteristic that
1. departure time correlation.A certain course line is given, the visiting rate of different departure time flights in a certain day of taking off
Between influence each other, and the visiting rate of departure time more similar flight is closer.As shown in Figure 2, departure time is more similar
The variation tendency that the daily economy class of 8:30AM flight and 9:00AM flight visits the visiting rate curve of rate and flight entirety more connects
Closely；
2. a day correlation of taking off.Tendency and periodicity is presented in the visiting rate data of flight itself on time dimension, by
Fig. 2 is it is found that the economy class of different departure time flights visits the periodicity that rate all has all ranks on the whole.
The visiting rate data of flight history that all departure times on certain course line are chosen in rate data set are visited in history flight
Training set, verifying collection, test set are divided based on this construction experiment sample as experimental data set；To the data that are related to of experiment into
Line number Data preprocess；Training MTA-RNN model, chooses optimal super ginseng, learns the data set drag optimized parameter, trained
The good Recognition with Recurrent Neural Network model based on multiple time granularity attention mechanism, using the test set to the trained base
It is tested in the Recognition with Recurrent Neural Network model of multiple time granularity attention mechanism, is collected using the verifying to described trained
The test result of Recognition with Recurrent Neural Network model based on multiple time granularity attention mechanism is verified.Then, it will be verified
MTA-RNN model visit rate prediction for flight future on the course line.
Implementation of the invention fully considers that flight visits the departure time of rate and a day correlation of taking off, as shown in figure 3, design
Multiple time granularity attention mechanism capture flight visits timing dependence of the rate under different time granularity out.Pass through outside simultaneously
Factor Fusion Module fully considers the influence of the extraneous factors such as flight self attributes and festivals or holidays, being capable of MTA- to design
RNN model realization more accurately predicts unknown ticket price in one week future.
The embodiment of offer is provided:
By taking the true history flight that middle Air China letter provides visits rate data set as an example, wherein choosing Beijing-Shanghai route
In in day of taking off in two years up to flight datas more than 584 days (730*89%) as experimental data set, if setting time window
Length D=28, prediction number of days τ=7, then statistical information is as shown in table 1.
1 data set statistical information of table
Fig. 3 is that a kind of flight based on multiple time granularity attention provided in an embodiment of the present invention visits rate prediction technique
Process flow diagram, including following processing step:
Step S1, experiment sample is constructed based on experimental data set, divides training set, verifying collection and test set；
Data set is divided into the training, verifying and test set of non-overlap, division proportion 8 by the present invention sequentially in time
: 1: 1, i.e., it is concentrated in experimental data and chooses the warp that 27 flight demand forecasting periods are 2010/01/29-2011/08/09 (totally 557 days)
The visiting rate data in cabin help as training set, prediction period is used as the data of 2011/08/10-2011/10/18 (70 days) and tests totally
Card collection, remainder data is test set.It is assumed that h represents first day needed in target τ days future of flight for predicting, then prediction period
It is [h-D, h-1] that for [h, h+ τ -1], on the course line, the history of all departure time flights, which visits rate data period, according to more than
Time segments division constructs individual data sample, constructs new prediction period according to sliding once a day, foundation aforesaid operations, each
Flight can construct 696 samples, then total number of samples is 27 × 696=18792.
Step S2, data prediction is carried out to the data that experiment is related to；
Visiting rate value range due to flight is [0,1], so being normalized without visiting rate value to flight；For
Outer input data, the present invention by the way of one-hot coding (0ne-Hot) respectively to week date attribute, whether work
Day, whether festivals or holidays, airline's type, airport of rising and falling, place course line open up high-speed rail these discrete features is encoded,
Returned for playing a landing GDP, departure time (extract hour and minute) and coach cabin class seats number these continuous features using Min-Max
One change method is normalized between [0,1], and method for normalizing formula is as follows:
X '=(x-minx)/(maxx-minx)
Wherein x indicates former data, and the expression new between [0,1] of x ' expression data, minx, maxx respectively indicate former number
Minimum value and maximum value.
Step S3, training MTA-RNN model chooses optimal super ginseng, learns the data set drag optimized parameter；
Model training stage, crowd size (Batch Size) are set as 256, and learning rate 0.0001, exercise wheel number is 500,
Using early stopping strategy prevents over-fitting.In MTA-RNN model proposed by the present invention, there are 4 super ginsengs for needing to adjust, be respectively
Time window length D, hidden layer dimension n, m of two LSTM units, decoder hidden layer dimension p in encoder.Enable D ∈ 7,
14,28,42,56 }, grid search is carried out on D select optimal value；For simplicity, the LSTM in codec is mono-
Identical hidden layer dimension is used in member, carries out grid search n=m=p ∈ { 32,64,128,256 } in the parameter；In addition,
The present invention improves model performance as the unit of codec using LSTMs (number of plies is denoted as f) is stacked.
Step S4, the MTA-RNN model trained in step S3 is subjected to model prediction Performance Evaluation on test set, with
Other pedestal methods compare；
The present invention is referred to using mean absolute error (MAE) and root-mean-square error (RMSE) as the evaluation of forecast result of model
Mark.Model proposed by the present invention is compared with other pedestal method prediction results, and the pedestal method chosen in this experiment includes
Autoregressive moving average summation model (ARIMA), shot and long term memory models (LSTM), codec models (Seq2seq), future
It is as shown in table 2 that 7 days flights visit the comparison of rate macro-forecast performance.
The following 7 days macro-forecast Comparative results of 2 distinct methods of table
In conclusion the MTA-RNN model of the embodiment of the present invention constructs multistage attention mechanism according to time granularity, respectively
It is visiting to be obtained by the combination of two-stage attention mechanism for flight for departure time attention mechanism and day attention mechanism of taking off
Timing dependence of the rate under different time granularity.Boat where the present invention captures target flight by departure time attention mechanism
Different departure time flights visit the Temporal dependency of rate in line and other departure time flights are visited rate and visited to target flight
The influence of rate, while using the tendency of itself visiting rate sequence of day attention mechanism capture target flight that takes off and period
Property；Furthermore the model considers the influence of the external factor such as flight self attributes and festivals or holidays using external factor Fusion Module,
It visits this model in flight and achieves good effect in rate forecasting problem.
The flight that the embodiment of the present invention proposes is visited rate prediction technique and is compared in real data set with other methods,
Good prediction effect is achieved, flight can be completed well and visit rate prediction task, and then enable civil aviaton market practitioner
Enough sensed in advance market demands improve enterprise income management level, provide decision support for operation departments at different levels.Meanwhile this hair
Multiple time granularity attention Mechanism Design in bright, to have the flight demand of same characteristics, flight inquiring Liang Deng civil aviaton field
Other forecasting problems provide new thinking.
Those of ordinary skill in the art will appreciate that: attached drawing is the schematic diagram of one embodiment, module in attached drawing or
Process is not necessarily implemented necessary to the present invention.
As seen through the above description of the embodiments, those skilled in the art can be understood that the present invention can
It realizes by means of software and necessary general hardware platform.Based on this understanding, technical solution of the present invention essence
On in other words the part that contributes to existing technology can be embodied in the form of software products, the computer software product
It can store in storage medium, such as ROM/RAM, magnetic disk, CD, including some instructions are used so that a computer equipment
(can be personal computer, server or the network equipment etc.) executes the certain of each embodiment or embodiment of the invention
Method described in part.
All the embodiments in this specification are described in a progressive manner, same and similar portion between each embodiment
Dividing may refer to each other, and the highlights of each of the examples are differences from other embodiments.Especially for device or
For system embodiment, since it is substantially similar to the method embodiment, so describing fairly simple, related place is referring to method
The part of embodiment illustrates.Apparatus and system embodiment described above is only schematical, wherein the conduct
The unit of separate part description may or may not be physically separated, component shown as a unit can be or
Person may not be physical unit, it can and it is in one place, or may be distributed over multiple network units.It can root
According to actual need that some or all of the modules therein is selected to achieve the purpose of the solution of this embodiment.Ordinary skill
Personnel can understand and implement without creative efforts.
The foregoing is only a preferred embodiment of the present invention, but scope of protection of the present invention is not limited thereto,
In the technical scope disclosed by the present invention, any changes or substitutions that can be easily thought of by anyone skilled in the art,
It should be covered by the protection scope of the present invention.Therefore, protection scope of the present invention should be with scope of protection of the claims
Subject to.
Claims (6)
1. a kind of flight based on multiple time granularity attention mechanism visits rate prediction technique characterized by comprising
The Recognition with Recurrent Neural Network model based on multiple time granularity attention mechanism is constructed, the Recognition with Recurrent Neural Network model includes:
In conjunction with the encoder of departure time attention mechanism and in conjunction with day decoder for attention mechanism that takes off；
It is described using the visiting rate time series of the flight history of departure times all on course line as the list entries of the encoder
Encoder carries out coded treatment to the list entries, and the decoder is decoded the encoded information of the encoder output
Processing, the flight for obtaining target flight visit rate time series.
2. the method according to claim 1, wherein the encoder packet of the combination departure time attention mechanism
The first layer shot and long term memory network LSTM unit and second layer LSTM unit for including hierarchical structure, on-course by target flight institute
The history of all flights visits each time step that rate time series is input to the first layer LSTM unit according to departure time,
Hidden layer state value under each departure time that the first layer LSTM unit output integrated considers departure time timing；
Input of the output of the first layer LSTM unit as departure time attention mechanism, departure time attention mechanism pass through ginseng
The hiding layer state for examining the second layer LSTM unit previous day of taking off adaptively extracts the first layer in each day of taking off
The hidden layer state value of LSTM unit related departure time, the second layer LSTM unit output target flight is in day of respectively taking off
Hidden layer state value, where the comprehensive first layer LSTM unit and the second layer LSTM elements capture target flight in course line
The flights of different departure times visits the Temporal dependency of rate and the visiting rate of other departure time flights to the visitor of target flight
The influence of seat rate.
3. according to the method described in claim 2, it is characterized in that, the method includes:
Given time length of window is D, is usedIt indicates
The T time series that the visiting rate that the flight of all departure times is gone over D days on one course line is constituted, whereinIndicate the visiting rate time series that the flight of t-th of departure time is gone over D days,Indicate that visiting rate in the flight of d-th of the past day of taking off all departure times is constituted
Vector；
Given list entries X=(x1, x2..., xT), whereinBy xtAs the first layer LSTM unit in t
The input of a departure time flight, and use ht=feb(ht-1, xt) the first layer LSTM unit is updated when taking off for t-th
The hidden layer state value at quarter, wherein febThe renewal function for representing the first layer LSTM unit, it is mono- by the first layer LSTM
The hidden layer state value h=(h under each departure time is calculated in member1, h2..., hT), whereinIt is described to hide
Stratiform state value h considers the Temporal dependency of the visiting rate of flight between all departure times on course line；
To the hidden layer state value h for t-th of departure time that the first layer LSTM unit obtainst, infused using following departure time
The calculating for power mechanism of anticipating:
Wherein,It is to measure the visiting rate time series of the flight of t-th of departure time when taking off at the d days day to target flight
The same day visit rate influence degree attention weight, the parameter for needing to learn is Ve,
It is rightIt is normalized, making the sum of all attention weights is 1, then for day d that takes off, the first layer LSTM unit knot
The output vector for closing departure time attention mechanism is as follows:
The hidden layer state value output of day d that takes off of the second layer LSTM unit output is as follows:
qd=fea(qd-1, zd)
Wherein feaIt is the renewal function of the second layer LSTM unit.
4. according to the method described in claim 3, it is characterized in that, the method further include:
Using flight self attributes and festivals or holidays attribute as extraneous factor, the flight self attributes include boat belonging to target flight
Empty company, type, landing GDP, airport of rising and falling, departure time and coach cabin class seats number, the festivals or holidays attribute include whether
For festivals or holidays, whether it is whether working day, week attribute and the same day open up high-speed rail；
By in the extraneous factor each flight self attributes and each festivals or holidays attribute be converted to low-dimensional vector, will be each low
Dimensional vector inputs different embeding layers respectively to generate corresponding insertion vector, will obtain after the processing of each insertion Vector FusionD ' is to take off day in the future of decoder prediction, will be describedIt is transferred to the decoder.
5. according to the method described in claim 4, it is characterized in that, the method further include:
The correlation in all time steps of second layer LSTM unit is adaptive selected using day attention mechanism of taking off to hide
Stratiform state value, the hiding stratiform for day of taking off to calculate the corresponding encoder of prediction output valve in the day decoder the d ' at d-th
The attention weight of state value, is defined as follows:
The parameter for wherein needing to learn is Vl,It is rightCarry out normalizing
Change, gainedAs attention weight, cd′Indicate for the decoder future day the d ' take off day output of attention mechanism to
Amount；
In a decoder, the context vector c of the weighted sum for the future day the d ' is calculatedd′Later, by context to
Measure cd′With obtain after the insertion Vector Fusion according to external factorAnd the output of decoder last momentPhase
In conjunction with coming, more new decoder hidden layer state, formula are as follows:
Wherein, fdLSTM unit renewal function used in decoder is indicated, then by context vector cd′With hiding layer state
gd′Vector splicing is carried out, the new hiding layer state finally predicted is as follows:
Wherein, matrixAnd vectorVector will be splicedDimension
It is mapped to decoder hidden layer dimension, uses linear transformationTo generate decoder when d ' is a
The final output of spacer stepBy decoder when the output of all time steps is integrated to obtain one section of future of target flight
Between flight visit rate.
6. method according to any one of claims 1 to 5, which is characterized in that the method further include:
The visiting rate data conduct of flight history that all departure times on certain course line are chosen in rate data set is visited in history flight
Experimental data set, based on the experimental data set construct experiment sample, by the experiment sample be divided into training set, verifying collection and
Test set is chosen using the training set training Recognition with Recurrent Neural Network model based on multiple time granularity attention mechanism
Optimal super ginseng learns the experimental data set drag optimized parameter out, obtains trained based on multiple time granularity attention
The Recognition with Recurrent Neural Network model of mechanism, using the test set to described trained based on multiple time granularity attention mechanism
Recognition with Recurrent Neural Network model is tested, using verifying collection to described trained based on multiple time granularity attention mechanism
The test result of Recognition with Recurrent Neural Network model verified.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910407946.2A CN110288121A (en) | 2019-05-16 | 2019-05-16 | Flight based on multiple time granularity attention mechanism visits rate prediction technique |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910407946.2A CN110288121A (en) | 2019-05-16 | 2019-05-16 | Flight based on multiple time granularity attention mechanism visits rate prediction technique |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110288121A true CN110288121A (en) | 2019-09-27 |
Family
ID=68002164
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910407946.2A Pending CN110288121A (en) | 2019-05-16 | 2019-05-16 | Flight based on multiple time granularity attention mechanism visits rate prediction technique |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110288121A (en) |
Cited By (18)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110751395A (en) * | 2019-10-21 | 2020-02-04 | 中国民航信息网络股份有限公司 | Passenger travel state determination method, device and server |
| CN110751576A (en) * | 2019-10-21 | 2020-02-04 | 中国民航信息网络股份有限公司 | Passenger travel determining method, device and server |
| CN110956342A (en) * | 2020-01-02 | 2020-04-03 | 中国民航大学 | CliqueNet flight delay prediction method based on attention mechanism |
| CN111091407A (en) * | 2019-10-28 | 2020-05-01 | 海南太美航空股份有限公司 | Airline passenger seat rate prediction method and system |
| CN111222555A (en) * | 2019-12-31 | 2020-06-02 | 南京天技通信技术实业有限公司 | Oil-gas-containing property detection method based on multi-granularity time sequence structure representation |
| CN111325380A (en) * | 2019-12-29 | 2020-06-23 | 航天信息股份有限公司 | Method and system for determining flight passenger seat rate based on multi-granularity time attention mechanism |
| CN111340543A (en) * | 2020-02-25 | 2020-06-26 | 上海大学 | Fine-grained taxi taking demand prediction method based on attention mechanism |
| CN111340536A (en) * | 2020-02-20 | 2020-06-26 | 上海携程商务有限公司 | Model training method, passenger seat rate progress prediction method, system, device and medium |
| CN111582919A (en) * | 2020-04-24 | 2020-08-25 | 海南太美航空股份有限公司 | Flight passenger seat rate prediction method and system |
| CN111652437A (en) * | 2020-06-04 | 2020-09-11 | 上海眼控科技股份有限公司 | Runway visual range predicted value obtaining method, computer equipment and storage medium |
| CN111739301A (en) * | 2020-08-07 | 2020-10-02 | 中航信移动科技有限公司 | Method and device for predicting trip data in short period |
| CN111915059A (en) * | 2020-06-29 | 2020-11-10 | 西安理工大学 | Method for predicting occupancy of Seq2Seq berth based on attention mechanism |
| CN112308618A (en) * | 2020-11-02 | 2021-02-02 | 沈阳民航东北凯亚有限公司 | Data processing method and device, electronic equipment and storage medium |
| CN112329993A (en) * | 2020-10-22 | 2021-02-05 | 海南太美航空股份有限公司 | Method and system for predicting number of passengers on flight and electronic equipment |
| CN112801455A (en) * | 2020-12-31 | 2021-05-14 | 携程旅游网络技术（上海）有限公司 | Flight density adjusting method, flight density adjusting system, electronic device and storage medium |
| CN112949952A (en) * | 2021-05-10 | 2021-06-11 | 北京人人云图信息技术有限公司 | Flight passenger seat rate prediction method and system based on graph neural network |
| CN113033861A (en) * | 2019-12-25 | 2021-06-25 | 广东奥博信息产业股份有限公司 | Water quality prediction method and system based on time series model |
| CN113360848A (en) * | 2021-06-04 | 2021-09-07 | 北京工商大学 | Time sequence data prediction method and device |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2007310634A (en) * | 2006-05-18 | 2007-11-29 | Japan Aerospace Exploration Agency | Online information providing system for seat selection |
| CN101464966A (en) * | 2008-12-31 | 2009-06-24 | 中山大学 | Method for planning airplane flight based on inheritance algorithm |
-
2019
- 2019-05-16 CN CN201910407946.2A patent/CN110288121A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2007310634A (en) * | 2006-05-18 | 2007-11-29 | Japan Aerospace Exploration Agency | Online information providing system for seat selection |
| CN101464966A (en) * | 2008-12-31 | 2009-06-24 | 中山大学 | Method for planning airplane flight based on inheritance algorithm |
Non-Patent Citations (1)
| Title |
|---|
| 邓玉婧: "基于多粒度时间注意力RNN的航班客座率预测", 《计算机工程》 * |
Cited By (27)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110751395B (en) * | 2019-10-21 | 2023-07-04 | 中国民航信息网络股份有限公司 | Passenger journey state determining method, device and server |
| CN110751576A (en) * | 2019-10-21 | 2020-02-04 | 中国民航信息网络股份有限公司 | Passenger travel determining method, device and server |
| CN110751576B (en) * | 2019-10-21 | 2023-07-25 | 中国民航信息网络股份有限公司 | Passenger journey determining method, device and server |
| CN110751395A (en) * | 2019-10-21 | 2020-02-04 | 中国民航信息网络股份有限公司 | Passenger travel state determination method, device and server |
| CN111091407A (en) * | 2019-10-28 | 2020-05-01 | 海南太美航空股份有限公司 | Airline passenger seat rate prediction method and system |
| CN111091407B (en) * | 2019-10-28 | 2023-06-02 | 海南太美航空股份有限公司 | Prediction method and system for passenger rate of airline |
| CN113033861A (en) * | 2019-12-25 | 2021-06-25 | 广东奥博信息产业股份有限公司 | Water quality prediction method and system based on time series model |
| CN111325380A (en) * | 2019-12-29 | 2020-06-23 | 航天信息股份有限公司 | Method and system for determining flight passenger seat rate based on multi-granularity time attention mechanism |
| CN111222555A (en) * | 2019-12-31 | 2020-06-02 | 南京天技通信技术实业有限公司 | Oil-gas-containing property detection method based on multi-granularity time sequence structure representation |
| CN111222555B (en) * | 2019-12-31 | 2023-06-20 | 南京天技通信技术实业有限公司 | Oil-gas-containing property detection method based on multi-granularity time sequence structural representation |
| CN110956342A (en) * | 2020-01-02 | 2020-04-03 | 中国民航大学 | CliqueNet flight delay prediction method based on attention mechanism |
| CN111340536A (en) * | 2020-02-20 | 2020-06-26 | 上海携程商务有限公司 | Model training method, passenger seat rate progress prediction method, system, device and medium |
| CN111340543B (en) * | 2020-02-25 | 2024-03-22 | 上海大学 | Fine granularity taxi taking demand prediction method based on attention mechanism |
| CN111340543A (en) * | 2020-02-25 | 2020-06-26 | 上海大学 | Fine-grained taxi taking demand prediction method based on attention mechanism |
| CN111582919A (en) * | 2020-04-24 | 2020-08-25 | 海南太美航空股份有限公司 | Flight passenger seat rate prediction method and system |
| CN111582919B (en) * | 2020-04-24 | 2023-09-19 | 海南太美航空股份有限公司 | Flight passenger rate prediction method and system |
| CN111652437A (en) * | 2020-06-04 | 2020-09-11 | 上海眼控科技股份有限公司 | Runway visual range predicted value obtaining method, computer equipment and storage medium |
| CN111915059A (en) * | 2020-06-29 | 2020-11-10 | 西安理工大学 | Method for predicting occupancy of Seq2Seq berth based on attention mechanism |
| CN111915059B (en) * | 2020-06-29 | 2023-06-30 | 西安理工大学 | Attention mechanism-based Seq2Seq berth occupancy prediction method |
| CN111739301A (en) * | 2020-08-07 | 2020-10-02 | 中航信移动科技有限公司 | Method and device for predicting trip data in short period |
| CN112329993B (en) * | 2020-10-22 | 2024-03-19 | 海南太美航空股份有限公司 | Method, system and electronic device for predicting number of flight passengers |
| CN112329993A (en) * | 2020-10-22 | 2021-02-05 | 海南太美航空股份有限公司 | Method and system for predicting number of passengers on flight and electronic equipment |
| CN112308618A (en) * | 2020-11-02 | 2021-02-02 | 沈阳民航东北凯亚有限公司 | Data processing method and device, electronic equipment and storage medium |
| CN112801455A (en) * | 2020-12-31 | 2021-05-14 | 携程旅游网络技术（上海）有限公司 | Flight density adjusting method, flight density adjusting system, electronic device and storage medium |
| CN112949952B (en) * | 2021-05-10 | 2021-08-17 | 北京人人云图信息技术有限公司 | Flight passenger seat rate prediction method and system based on graph neural network |
| CN112949952A (en) * | 2021-05-10 | 2021-06-11 | 北京人人云图信息技术有限公司 | Flight passenger seat rate prediction method and system based on graph neural network |
| CN113360848A (en) * | 2021-06-04 | 2021-09-07 | 北京工商大学 | Time sequence data prediction method and device |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110288121A (en) | Flight based on multiple time granularity attention mechanism visits rate prediction technique | |
| Wu et al. | Forecasting tourist daily arrivals with a hybrid Sarima–Lstm approach | |
| Yu et al. | Prediction of highway tunnel pavement performance based on digital twin and multiple time series stacking | |
| CN106952181A (en) | Electric Load Prediction System based on long Memory Neural Networks in short-term | |
| CN109902880A (en) | A kind of city stream of people's prediction technique generating confrontation network based on Seq2Seq | |
| CN106971240A (en) | The short-term load forecasting method that a kind of variables choice is returned with Gaussian process | |
| CN107103394A (en) | A kind of real-time passenger flow forecasting of track traffic based on neutral net | |
| CN111091196B (en) | Passenger flow data determination method and device, computer equipment and storage medium | |
| CN110334843A (en) | A kind of time-varying attention improves be hospitalized medial demand prediction technique and the device of Bi-LSTM | |
| CN104966105A (en) | Robust machine error retrieving method and system | |
| CN107067076A (en) | A kind of passenger flow forecasting based on time lag NARX neutral nets | |
| CN109858671A (en) | Predict the method and system of the visiting rate of target pattern | |
| CN111325380A (en) | Method and system for determining flight passenger seat rate based on multi-granularity time attention mechanism | |
| CN110535149A (en) | A kind of area Gong Biantai electric load three-phase imbalance prediction technique | |
| CN111899059A (en) | Navigation driver revenue management dynamic pricing method based on block chain | |
| CN112948412A (en) | Flight inventory updating method, system, electronic equipment and storage medium | |
| Dadashpour Moghaddam et al. | A GIS-based assessment of urban tourism potential with a branding approach utilizing hybrid modeling | |
| Wu et al. | Interpretable tourism demand forecasting with temporal fusion transformers amid COVID-19 | |
| Zhou | Sports achievement prediction and influencing factors analysis combined with deep learning model | |
| Abdelghany et al. | An integrated reinforced learning and network competition analysis for calibrating airline itinerary choice models with constrained demand | |
| Xinliang et al. | University restaurant sales forecast based on bp neural network–in shanghai jiao tong university case | |
| Alkhatib et al. | Airline capacity strategies: worldwide analysis, taxonomies and regional comparison | |
| Zhang et al. | Short-term water demand forecasting: A case study | |
| AYAYDIN et al. | Deep learning based forecasting of delay on flights | |
| HUANG | A STUDY ON ARTIFICIAL INTELLIGENCE FORECASTING OF RESORT DEMAND. |
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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190927 |
|
| RJ01 | Rejection of invention patent application after publication |