CN106485353A - Air pollutant concentration forecasting procedure and system - Google Patents
Air pollutant concentration forecasting procedure and system Download PDFInfo
- Publication number
- CN106485353A CN106485353A CN201610875403.XA CN201610875403A CN106485353A CN 106485353 A CN106485353 A CN 106485353A CN 201610875403 A CN201610875403 A CN 201610875403A CN 106485353 A CN106485353 A CN 106485353A
- Authority
- CN
- China
- Prior art keywords
- model
- undetermined
- data
- forecast
- layer
- 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.)
- Granted
Links
- 239000000809 air pollutant Substances 0.000 title claims abstract description 87
- 231100001243 air pollutant Toxicity 0.000 title claims abstract description 80
- 238000000034 method Methods 0.000 title claims abstract description 78
- 238000012549 training Methods 0.000 claims abstract description 70
- 238000012544 monitoring process Methods 0.000 claims abstract description 28
- 238000013480 data collection Methods 0.000 claims abstract description 16
- 239000000284 extract Substances 0.000 claims abstract description 9
- 239000011159 matrix material Substances 0.000 claims description 29
- RAHZWNYVWXNFOC-UHFFFAOYSA-N Sulphur dioxide Chemical compound O=S=O RAHZWNYVWXNFOC-UHFFFAOYSA-N 0.000 claims description 8
- 239000000203 mixture Substances 0.000 claims description 8
- 230000008569 process Effects 0.000 claims description 8
- 238000012545 processing Methods 0.000 claims description 6
- UGFAIRIUMAVXCW-UHFFFAOYSA-N Carbon monoxide Chemical compound [O+]#[C-] UGFAIRIUMAVXCW-UHFFFAOYSA-N 0.000 claims description 5
- 230000004913 activation Effects 0.000 claims description 5
- 229910002091 carbon monoxide Inorganic materials 0.000 claims description 5
- CBENFWSGALASAD-UHFFFAOYSA-N Ozone Chemical compound [O-][O+]=O CBENFWSGALASAD-UHFFFAOYSA-N 0.000 claims description 4
- 239000002245 particle Substances 0.000 claims description 4
- 230000009467 reduction Effects 0.000 claims description 4
- MGWGWNFMUOTEHG-UHFFFAOYSA-N 4-(3,5-dimethylphenyl)-1,3-thiazol-2-amine Chemical compound CC1=CC(C)=CC(C=2N=C(N)SC=2)=C1 MGWGWNFMUOTEHG-UHFFFAOYSA-N 0.000 claims description 3
- JCXJVPUVTGWSNB-UHFFFAOYSA-N nitrogen dioxide Inorganic materials O=[N]=O JCXJVPUVTGWSNB-UHFFFAOYSA-N 0.000 claims description 3
- 230000001360 synchronised effect Effects 0.000 abstract description 5
- 230000002123 temporal effect Effects 0.000 abstract description 5
- 101001095088 Homo sapiens Melanoma antigen preferentially expressed in tumors Proteins 0.000 description 5
- 102100037020 Melanoma antigen preferentially expressed in tumors Human genes 0.000 description 5
- 238000003915 air pollution Methods 0.000 description 5
- 238000007477 logistic regression Methods 0.000 description 5
- 241001123248 Arma Species 0.000 description 4
- IJGRMHOSHXDMSA-UHFFFAOYSA-N Atomic nitrogen Chemical compound N#N IJGRMHOSHXDMSA-UHFFFAOYSA-N 0.000 description 4
- 238000013136 deep learning model Methods 0.000 description 4
- 239000010419 fine particle Substances 0.000 description 4
- 238000005457 optimization Methods 0.000 description 4
- 230000008859 change Effects 0.000 description 3
- 239000003344 environmental pollutant Substances 0.000 description 3
- 230000006870 function Effects 0.000 description 3
- 229910052757 nitrogen Inorganic materials 0.000 description 3
- 231100000719 pollutant Toxicity 0.000 description 3
- 238000012360 testing method Methods 0.000 description 3
- GWEVSGVZZGPLCZ-UHFFFAOYSA-N Titan oxide Chemical compound O=[Ti]=O GWEVSGVZZGPLCZ-UHFFFAOYSA-N 0.000 description 2
- 238000013528 artificial neural network Methods 0.000 description 2
- 230000015572 biosynthetic process Effects 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 230000007613 environmental effect Effects 0.000 description 2
- 238000012706 support-vector machine Methods 0.000 description 2
- 238000003786 synthesis reaction Methods 0.000 description 2
- 241000790917 Dioxys <bee> Species 0.000 description 1
- NINIDFKCEFEMDL-UHFFFAOYSA-N Sulfur Chemical compound [S] NINIDFKCEFEMDL-UHFFFAOYSA-N 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 230000009514 concussion Effects 0.000 description 1
- 238000011109 contamination Methods 0.000 description 1
- 230000007812 deficiency Effects 0.000 description 1
- 238000009792 diffusion process Methods 0.000 description 1
- 239000006185 dispersion Substances 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 230000002452 interceptive effect Effects 0.000 description 1
- 238000012417 linear regression Methods 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 238000003062 neural network model Methods 0.000 description 1
- 230000007935 neutral effect Effects 0.000 description 1
- 230000005180 public health Effects 0.000 description 1
- 230000001373 regressive effect Effects 0.000 description 1
- 238000005728 strengthening Methods 0.000 description 1
- 239000011593 sulfur Substances 0.000 description 1
- 229910052717 sulfur Inorganic materials 0.000 description 1
- 239000004408 titanium dioxide Substances 0.000 description 1
- 238000012546 transfer Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- 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
- G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
- G06Q50/10—Services
- G06Q50/26—Government or public services
Landscapes
- Business, Economics & Management (AREA)
- Engineering & Computer Science (AREA)
- Strategic Management (AREA)
- Economics (AREA)
- Tourism & Hospitality (AREA)
- Human Resources & Organizations (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Marketing (AREA)
- Development Economics (AREA)
- General Physics & Mathematics (AREA)
- General Business, Economics & Management (AREA)
- Quality & Reliability (AREA)
- Game Theory and Decision Science (AREA)
- Entrepreneurship & Innovation (AREA)
- Operations Research (AREA)
- Educational Administration (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Primary Health Care (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a kind of air pollutant concentration forecasting procedure, including：Determine multiple models undetermined, the node number arranging the output layer of each model undetermined is m；Training dataset and checking data set is extracted from described history concentration data；Each model undetermined is trained until model undetermined is restrained；The minimum corresponding model undetermined of synthetic error is defined as forecasting model；Described forecast data collection is inputted to described forecasting model, using the output result of described forecasting model as the forecast result forecasting that time delay is r.This air pollutant concentration forecasting procedure carries out the forecast of air pollutant concentration based on STDL model, can extract implicit temporal correlation in air pollutant concentration data, and the synchronous forecast data obtaining many monitoring stations, and forecast precision is high；The precision in the forecast of extreme air pollutant levels can also be improved.
Description
Technical field
The present invention relates to field of environment engineering technology, more particularly, to a kind of air pollutant concentration forecasting procedure and system.
Background technology
In recent years, with the high speed development of national economy, the quickening of urbanization process and plant-scale expansion, air is dirty
Dye problem is increasingly serious, causes the extensive concern in global range.Air with fine particle PM2.5 as primary pollutant is dirty
Dye phenomenon becomes increasingly conspicuous, and constitutes larger threat to public health.Therefore, it is necessary to development prediction of air quality, preferably reflect
The variation tendency of air pollution, and provide rapid and comprehensive environmental quality information is environmental management and avoids serious air
Contamination accident provides decision support.
At present, generally to carry out air quality using numerical forecast model and statistics forecasting model pre- for prediction of air quality
Report.Numerical Prediction Method comes the discharge of simulating pollution thing, transfer, diffusion and the process dissipating using using atmospheric dynamics theory,
With the method based on model-driven, air quality is modeled and forecasts.However, due to insecure pollutant emission
Data, complicated underlying surface (earth surface of air lower floor directly contact) situation and incomplete theoretical basiss, analog result
Precision is relatively low.
Statistics forecasting procedure then forecasts air quality with data-driven version using statistical modeling means, such as polynary line
Property return (Multi-variable Linear Regression, abbreviation MLR) model and auto regressive moving average (Auto
Regression Moving Average, abbreviation ARMA) model is commonly used for prediction of air quality.However, these methods are not because
Can nonlinear model in simulated air pollutant levels and precision is relatively low, the precision in the forecast of extreme air pollutant levels
Especially low.
Content of the invention
It is an object of the invention to overcoming above-mentioned deficiency present in prior art, and provide a kind of air pollutant concentration
Forecasting procedure and system, relatively low to solve air pollutant concentration forecast precision, in the forecast of extreme air pollutant levels
The especially low problem of precision.
In a first aspect, the invention provides a kind of air pollutant concentration forecasting procedure, including：By the node of input layer
The value multiple to be selected of number, the value multiple to be selected of the number of plies of stack self-encoding encoder, every layer of node number of stack self-encoding encoder
Multiple values to be selected are combined determining multiple models undetermined, and the node number arranging the output layer of each model undetermined is m；Its
In, the node number of input layer is n times of node number m of output layer, and n is Step Parameters；Many according to Step Parameters n
Individual value to be selected and forecast time delay r specified, obtain multigroup history concentration data of the air pollutants from m monitoring station,
Training dataset and checking data set is extracted from described history concentration data；Respectively treat cover half using described training data set pair
Type is trained until model undetermined convergence, the model each undetermined corresponding Model Weight matrix that record training completes and model are inclined
Put vector；The data input being used for inputting in described checking data set is extremely trained the model each undetermined completing, calculates each undetermined
The output result of model and the described synthetic error verifying the data for checking in data set, minimum synthetic error is corresponded to
Model undetermined be defined as forecasting model, wherein, the value of the Step Parameters of described forecasting model is nr；To monitor from m
Website, Step Parameters are nrAir pollutant concentration observation data composition forecast data collection, by described forecast data collection
Input to described forecasting model, using the output result of described forecasting model as the forecast result forecasting that time delay is r.
Said method can also have the characteristics that：Described instructed using each model undetermined of described training data set pair
Practice until model undetermined convergence includes：Using described training dataset using the successively weight square to stack self-encoding encoder for the coaching method
Battle array and bias vector are trained；Using the output vector of last layer of stack self-encoding encoder as the input of output layer, and adopt
With back-propagation algorithm, the weight matrix of stack self-encoding encoder and output layer and bias vector are adjusted from back to front until
Model convergence undetermined.
Said method can also have the characteristics that：Using successively coaching method to the weight matrix of stack self-encoding encoder and partially
When putting vector and being trained, using the weight minimizing the reconstructed error with sparse restrictive condition and adjusting every layer of self-encoding encoder
Matrix W1And W2And bias vector b and c, the described reconstructed error such as following formula with sparse restrictive condition：
Wherein, yj=f (W1x+bj), x={ x(1),...,x(i),...,x(N), z(i)=g
(W2y+ci), i=1 ..., N, j=1 ..., HD, λ is the weight of regularization term, and μ is the weight of sparse item, and N is criticizing of input
Training sample number, HDIt is the node number of current layer self-encoding encoder, | | W1||2It is W1L2Norm, | | W2||2It is W2L2Model
Number, ρ is Sparse parameter, x(i)For i-th input vector of current layer self-encoding encoder, yjFor j-th element of vector after coding, z(i)For the corresponding decoded vector of i-th input vector of current layer self-encoding encoder.
Said method can also have the characteristics that：Using successively coaching method to the weight matrix of stack self-encoding encoder and partially
When putting vector and being trained, it is that every layer of self-encoding encoder arranges single iterationses and learning rate, and in the training process not
Disconnected reduction learning rate；Using back-propagation algorithm to the weight matrix of stack self-encoding encoder and output layer and bias vector from after
When being carried forward adjustment, be in course of adjustment continuous reduction learning rate.
Said method can also have the characteristics that：The output result of described each model undetermined is verified in data set with described
Synthetic error for the data of checking includes root-mean-square error, mean absolute error and average absolute percent error.
Said method can also have the characteristics that：Described acquisition is multigroup from the air pollutants of m monitoring station
History concentration data also includes：The physical meaning of the concentration data according to air pollutants, rejects history concentration using box traction substation
Exceptional value in data；And/or using linear interpolation method, the missing values in history concentration data are filled up；And/or adopt
Use minimax method for normalizing, history concentration data is normalized.
Said method can also have the characteristics that：In described stack self-encoding encoder, every layer of encoder from code device is as follows
Formula：Decoder such as following formula：
Said method can also have the characteristics that：The activation primitive of the output layer of each described model undetermined is
Said method can also have the characteristics that：Described air pollutants are fine particle, inhalable particles, titanium dioxide
Any one in sulfur, nitrogen dioxide, carbon monoxide and ozone.
The air pollutant concentration forecasting procedure that the present invention provides carries out the pre- of air pollutant concentration based on STDL model
Report, specifically, is come to multiple monitoring stations using stack own coding (Stacked Auto-Encoder, abbreviation SAE) model
Air pollutant concentration data is modeled, such that it is able to extract profound feature and data in air pollutant concentration data
In implicit temporal correlation, and the synchronous forecast data obtaining many monitoring stations, forecast precision is high；In view of STDL model is non-
Linear model, thus the precision in the forecast of extreme air pollutant levels can also be improved.
Second aspect, the invention provides a kind of air pollutant concentration forecast system, including：Model undetermined determines single
Unit, for will be self-editing to the value multiple to be selected of the node number of input layer, the value multiple to be selected of the number of plies of stack self-encoding encoder, stack
The value multiple to be selected of every layer of node number of code device is combined determining multiple models undetermined, arranges each model undetermined
The node number of output layer is m；Wherein, the node number of input layer is n times of node number m of output layer, and n joins for time step
Number；History concentration data processing unit, for the value multiple to be selected according to Step Parameters n with forecast time delay r specified, obtains
From multigroup history concentration data of the air pollutants of m monitoring station, from described history concentration data, extract training
Data set and checking data set；Model training unit undetermined, for being instructed using each model undetermined of described training data set pair
Practice until model undetermined convergence, record model each undetermined corresponding Model Weight matrix and the model bias vector that training completes；
Forecasting model determining unit, for respectively treating cover half by the data input being used for inputting in described checking data set to what training completed
Type, calculates the output result of each model undetermined and the described synthetic error verifying the data for checking in data set, by minimum
The corresponding model undetermined of synthetic error be defined as forecasting model, wherein, the value of the Step Parameters of described forecasting model is nr；
Air pollutant concentration forecast unit, for being n from m monitoring station, Step ParametersrAir pollutant concentration
Observation data composition forecast data collection, described forecast data collection is inputted to described forecasting model, and described forecasting model is defeated
Going out result as forecast time delay is the forecast result of r.
The air pollutant concentration forecast system that the present invention provides carries out the pre- of air pollutant concentration based on STDL model
Report, can extract implicit temporal correlation in air pollutant concentration data, and the synchronous pre- count off obtaining many monitoring stations
According to forecast precision is high；The precision in the forecast of extreme air pollutant levels can also be improved.
Brief description
The accompanying drawing constituting the part of the present invention is used for providing a further understanding of the present invention, the schematic reality of the present invention
Apply example and its illustrate, for explaining the present invention, not constituting inappropriate limitation of the present invention.In the accompanying drawings：
Fig. 1 is a kind of flow chart of air pollutant concentration forecasting procedure；
Fig. 2 is a kind of composition figure of air pollutant concentration forecast system.
Specific embodiment
Purpose, technical scheme and advantage for making the embodiment of the present invention are clearer, below in conjunction with the embodiment of the present invention
In accompanying drawing, the technical scheme in the embodiment of the present invention is clearly and completely described it is clear that described embodiment is
The a part of embodiment of the present invention, rather than whole embodiments.Based on the embodiment in the present invention, those of ordinary skill in the art
The every other embodiment being obtained under the premise of not making creative work, broadly falls into the scope of protection of the invention.Need
Illustrate, in the case of not conflicting, the embodiment in the application and the feature in embodiment can mutual combination in any.
Fig. 1 is the flow chart of air pollutant concentration forecasting procedure.As shown in figure 1, the air pollutants that the present invention provides
Concentration prediction method, including：
Step S10：By the value multiple to be selected of the node number of input layer, the number of plies of stack self-encoding encoder multiple to be selected
Value, the value multiple to be selected of every layer of node number of stack self-encoding encoder are combined determining multiple models undetermined, and setting is every
The node number of the output layer of individual model undetermined is m；Wherein, the node number of input layer is the n of node number m of output layer
Times, n is Step Parameters, m and n is positive integer；
Step S20：Value multiple to be selected according to Step Parameters n and forecast time delay r specified, obtain from m monitoring
Multigroup history concentration data of the air pollutants of website, extracts training dataset and checking from described history concentration data
Data set；
Step S30：It is trained using each model undetermined of described training data set pair, until model undetermined convergence, recording instruction
Model each undetermined corresponding Model Weight matrix and model bias vector that white silk completes；
Step S40：The data input being used for inputting in described checking data set is extremely trained the model each undetermined completing, meter
Calculate the output result of each model undetermined and the described synthetic error verifying the data for checking in data set, by minimum synthesis
The corresponding model undetermined of error is defined as forecasting model, and wherein, the value of the Step Parameters of described forecasting model is nr；
Step S50：To be n from m monitoring station, Step ParametersrAir pollutant concentration observation data set
Become forecast data collection, described forecast data collection inputted to described forecasting model, using the output result of described forecasting model as
Forecast time delay is the forecast result of r.
The present invention adopts space-time deep learning (Spatio-Temporal Deep Learning, abbreviation STDL) model
Build air pollutant concentration forecasting model, this STDL model has the input layer being sequentially connected, stack self-encoding encoder and output
Layer；And using grid search (Grid Search, abbreviation GD) method, the structural parameters of this space-time deep learning model are carried out excellent
Change.Specifically, by the value multiple to be selected of the node number of input layer, the value multiple to be selected of the number of plies of stack self-encoding encoder, stack
The value multiple to be selected of the node number of every layer of self-encoding encoder is combined, so that it is determined that going out multiple models undetermined, and then from this
Finding out the minimum model to be selected of synthetic error in multiple models undetermined is forecasting model.
It should be noted that with generally can shorten required for the heuristic search of Optimizing Search time or search at random
Suo Fangfa compares, although heuristic search and random search is in hgher efficiency, effect of optimization is poor.Trellis search method is to the greatest extent
The pipe required time is longer, but its permutation and combination is carried out to multigroup parameter to be selected method, it is similar to exhaustion, therefore can be bigger
Find optimal solution, that is, the structural parameters combination of optimum STDL model probability.
Specifically, the number m according to air pollutants monitoring station, arranges the node of the output layer of each model undetermined
Number is m, then the node number of input layer is n times of node number m of output layer, and wherein, n is Step Parameters, m and n is
Positive integer.
Further, the value multiple to be selected according to Step Parameters n and forecast time delay r specified, obtains from m monitoring
Multigroup history concentration data of the air pollutants of website.Every group of history concentration data has different Step Parameters n and phase
Same forecast time delay r.Comprise for as input in each sample from single website in each group of history concentration data
N data point and for as output have forecast time delay r 1 data point；It is derived from m in each group of history concentration data
The single sample collection of individual monitoring station has identical data points p, specially：P=m × (n+1).It should be noted that it is every
One group of history concentration data includes multiple sample sets with p data point, for follow-up by the way of batch processing to every
Individual model undetermined carries out batch processing training.
Preferably, the physical meaning of air pollutant concentration data is air pollutants hour average concentration, goes through for each group
Comprise in each sample from single website in history concentration data for as input n data point in, consecutive number
Time interval between strong point is 1 hour；For the nearest data point in sequential of that data point as output
Time interval is r hour.
Preferably, the physical meaning of air pollutant concentration data is air pollutants hour average concentration, goes through for each group
Comprise in each sample from single website in history concentration data for as input n data point in, consecutive number
Time interval between strong point is r hour；For the nearest data point in sequential of that data point as output
Time interval is r hour.
Further, described history concentration data is divided into training dataset and checking data set.Wherein, training data
Collection and checking data set all include multiple sample sets with p data point.
Determine corresponding to the plurality of model undetermined training dataset and checking data set after, using described training
Data set is trained to each model undetermined until model undetermined is restrained, and records the corresponding mould of model each undetermined that training completes
Type weight matrix and model bias vector.
After the Model Weight matrix determining the plurality of model undetermined and model bias vector, by described checking data
The data input for input is concentrated extremely to train the model each undetermined completing, the output result calculating each model undetermined is tested with described
The synthetic error of the data for verifying in card data set, the minimum corresponding model undetermined of synthetic error is defined as forecasting mould
Type.
Specifically, the output result of described each model undetermined and the described synthesis verifying the data for checking in data set
Error can include root-mean-square error (Root Mean Square Error, abbreviation RMSE), mean absolute error (Mean
Absolute Error, abbreviation MAE) and average absolute percent error (Mean Absolute Percentage Error, referred to as
MAPE).Wherein, RMSE, MAE and MAPE are respectively provided with usual implication in the art, no longer list its point here in detail
Not corresponding computing formula.
It should be noted that above-mentioned synthetic error can be RMSE and MAE having dimension weighted mean it is also possible to
It is the value of nondimensional MAPE.If necessary to preferably go out the little forecasting model of absolute error, then need to select the error of dimension
Index；If necessary to preferably go out the little forecasting model of relative error, then need to select nondimensional index.Do not rush in physical meaning
In the case of prominent, three can mutual combination in any.
After determining forecasting model, will be n from m monitoring station, Step ParametersrAir pollutant concentration
Observation data composition forecast data collection, described forecast data collection is inputted to described forecasting model, by described forecasting model
Output result is the forecast result of r as forecast time delay, i.e. synchronization obtains the air pollution that m monitoring station forecasts that time delay is r
Thing concentration data.
The air pollutant concentration forecasting procedure that the present invention provides carries out the pre- of air pollutant concentration based on STDL model
Report, specifically, is come to multiple monitoring stations using stack own coding (Stacked Auto-Encoder, abbreviation SAE) model
Air pollutant concentration data is modeled, such that it is able to extract profound feature and data in air pollutant concentration data
In implicit temporal correlation, and the synchronous forecast data obtaining many monitoring stations, forecast precision is high；In view of STDL model is non-
Linear model, thus the precision in the forecast of extreme air pollutant levels can also be improved.
Specifically, described it is trained until model undetermined convergence bag using each model undetermined of described training data set pair
Include：The weight matrix of stack self-encoding encoder and bias vector are instructed using successively coaching method using described training dataset
Practice；Using the output vector of last layer of stack self-encoding encoder as the input vector of output layer, and adopt back-propagation algorithm
Weight matrix to stack self-encoding encoder and output layer and bias vector are adjusted from back to front up to model convergence undetermined.
In view of traditional back-propagation algorithm is easily trapped into local extremum when training space-time deep learning model, adopt
The successively coaching method (Greedy Layer-wise Training) that Hinton proposes to the weight matrix of stack self-encoding encoder and
Bias vector carries out pre-training, by successively training from bottom to top, can be restrained and global optimum stack own coding
Device.
Further, using convergence stack self-encoding encoder last layer output vector as output layer input, and
Using back-propagation algorithm, the weight matrix of stack self-encoding encoder and output layer and bias vector are adjusted directly from back to front
To model convergence undetermined.
In each model training undetermined, first pre-training is carried out to stack self-encoding encoder from bottom to top using successively coaching method
Until convergence, in conjunction with the stack self-encoding encoder of convergence, depth is learnt to entire depth from back to front by back-propagation algorithm
Network is finely adjusted until model undetermined is restrained.The Training strategy training precision that this pre-training combines with fine setting is high, training
Speed is fast.
Specifically, when the weight matrix of stack self-encoding encoder and bias vector being trained using successively coaching method, adopt
Adjust the weight matrix W of every layer of self-encoding encoder with minimizing the reconstructed error with sparse restrictive condition1And W2And bias vector
B and c, the described reconstructed error such as following formula with sparse restrictive condition：
Wherein, θ refers to all unknown parameters comprising in formula on the right of above-mentioned equation, i.e. (W1, W2, b, c), namely J (θ)
=J (W1,W2,b,c)；Common minimum mean-square error for x and z；(W1||2+||W2||2) it is regularization term,
||W1||2It is W1L2Norm, | | W2||2It is W2L2Norm, λ is the weight of regularization term；It is sparse item, μ
It is the weight of sparse item, j=1 ..., HD, HDIt is the node number of current layer self-encoding encoder；It is to sdpecific dispersion
(Kullback Leibler Divergence), for strengthening the sparse restriction of cataloged procedure, it is defined as follows formula：Wherein, ρ is Sparse parameter, generally 0 or close to 0；It is self-editing
The node j of code device corresponds to the average activation of input vector x, and it is defined as follows formula：Wherein, yjFor compiling
The value of j-th node, y after j-th element of vector, namely coding after codej=f (W1x+bj)；X is current layer self-encoding encoder
Input vector, x={ x(1),...,x(i),...x(N), x(i)∈Rd, x(i)I-th input vector for current layer self-encoding encoder；
In addition, z(i)For the corresponding decoded vector of i-th input vector of current layer self-encoding encoder, z(i)=g (W2y+ci)；I=1 ..., N,
N is batch training sample number of input；Y is the coding vector of current layer self-encoding encoder.
The air pollutant concentration forecasting procedure that the present invention provides to be improved using the sparse self-encoding encoder that Ranzato proposes
Conventional self-encoding encoder, to adjust the weight square of every layer of self-encoding encoder using the reconstructed error that minimum has sparse restrictive condition
Battle array W1And W2And bias vector b and c, to avoid " simple copy data " or " maximization interactive information ", substantially extract
Representational feature in data.
In view of the learning rate mistake when deep neural network models, in training process (including pre-training and trim process)
Conference leads to reconstructed error concussion not restrain, and learning rate is too little to be that network convergence is excessively slow, and weights are difficult to tend towards stability, and adopt
When the weight matrix of stack self-encoding encoder and bias vector being trained with successively coaching method, constantly reduce in training process and learn
Practise speed；Using back-propagation algorithm, the weight matrix of stack self-encoding encoder and output layer and bias vector are carried out from back to front
During adjustment, be in course of adjustment continuous reduction learning rate.
During pre-training, every layer of self study encoder can arrange single iterationses and learning rate, when weight
When structure error meets the condition of convergence, you can stop the pre-training of this layer, thus improving training speed.
When obtaining from multigroup history concentration data of the air pollutants of m monitoring station, need regarding the quality of data,
Optionally adopt following preprocess method：
The physical meaning of the concentration data according to air pollutants, rejects the exception in history concentration data using box traction substation
Value；Using linear interpolation method, the missing values in history concentration data are filled up；Using minimax method for normalizing, right
History concentration data is normalized.
Normalized can accelerate model training speed.Specifically, minimax method for normalizing passes through to travel through and owns
Monitoring station air pollutant concentration data, searches maximum a therein and minima b, and using following formula to original air matter
Amount X is normalized：
Preferably, the present invention adopt space-time deep learning model in, in described stack self-encoding encoder every layer from code device
Encoder such as following formula：Decoder such as following formula：The output layer of each described model undetermined
Activation primitive is
Specifically, the output layer of each described model undetermined is logistic regression layer, and its activation primitive is logistic regression function.
Logistic regression function is to realize simple and can preferably the nonlinear characteristic in data be modeled.In addition, described treat cover half
The output layer of type can also adopt SVR regression model, to realize more preferable Nonlinear Modeling, but realizes complicated, the calculating of consumption
Resource is many.
Alternatively, the present invention provide air pollutant concentration forecasting procedure in, described air pollutants be fine particle,
Any one in inhalable particles, sulfur dioxide, nitrogen dioxide, carbon monoxide and ozone.That is, the air that the present invention provides is dirty
Dye thing concentration prediction method can be respectively to fine particle (PM2.5), inhalable particles (PM10), sulfur dioxide (SO2), dioxy
Change nitrogen (NO2), carbon monoxide (CO), ozone (O3) etc. the concentration of 6 kinds of pollutant forecast.
The present invention is entered to the weight matrix of stack self-encoding encoder and bias vector using the successively coaching method that Hinton proposes
Row pre-training, after pre-training terminates, finely tunes the parameter of whole network from back to front using back-propagation algorithm.Model is instructed
The idiographic flow practiced is as follows：
A () sets the hidden layer number of plies and every node layer number, pre-training iterationses, pre-training learning rate, fine setting iteration
Number of times, fine setting learning rate, batch processing size.
(b) network pre-training：Initialization regularization term and the weight of sparse item, the weight square of each layer of random initializtion network
Battle array and bias vector；Train ground floor using training set；Train follow-up training hidden layer using the mode of successively pre-training.?
During pre-training, every layer can arrange single iterationses and learning rate, when reconstructed error meets the condition of convergence, you can
Stop the pre-training of this layer.
C () network is finely tuned：The output input as logistic regression layer of stack self-encoding encoder last hidden layer is used；
The weight matrix of random initializtion logistic regression layer and bias vector；Using back-propagation algorithm from back to front finely tuning whole net
The parameter of network.In network trim process, when model prediction error convergence, you can stop iteration.
In addition, using optimize after forecast model carry out air pollution concentration prediction after, can by predict the outcome with after
The air pollution concentration data of continuous actual measurement is compared, and if precision of prediction can not meet pre-provisioning request, can change
Each structural parameters in STDL model, according to the method for aforementioned optimization forecast model, in addition find a preferably prediction mould
Type.
Fig. 2 is the composition figure of air pollutant concentration forecast system, and air pollutant concentration forecast system is air pollution
The corresponding virtual bench of thing concentration prediction method.As shown in Fig. 2 the air pollutant concentration forecast system that the present invention provides, bag
Include：Model determining unit 100 undetermined, for by the number of plies of the value multiple to be selected of the node number of input layer, stack self-encoding encoder
Value multiple to be selected, the value multiple to be selected of every layer of node number of stack self-encoding encoder is combined determining and multiple treats cover half
Type, the node number arranging the output layer of each model undetermined is m；Wherein, the node number of input layer is the node of output layer
N times of number m, n is Step Parameters；History concentration data processing unit 200, for treating according to the multiple of Step Parameters n
Choosing value and forecast time delay r specified, obtain multigroup history concentration data of the air pollutants from m monitoring station, from institute
State and in history concentration data, extract training dataset and checking data set；Model training unit 300 undetermined, for using described
The each model undetermined of training data set pair is trained until model undetermined convergence, records and trains the model each undetermined completing corresponding
Model Weight matrix and model bias vector；Forecasting model determining unit 400, for being used for inputting in described checking data set
Data input extremely train the model each undetermined completing, calculate in the output result of each model undetermined and described checking data set and use
In the synthetic error of the data of checking, the minimum corresponding model undetermined of synthetic error is defined as forecasting model, wherein, described
The value of the Step Parameters of forecasting model is nr；Air pollutant concentration forecast unit 500, for will from m monitoring station,
Step Parameters are nrAir pollutant concentration observation data composition forecast data collection, by described forecast data collection input to
Described forecasting model, using the output result of described forecasting model as the forecast result forecasting that time delay is r.
The air pollutant concentration forecast system that the present invention provides carries out the pre- of air pollutant concentration based on STDL model
Report, can extract implicit temporal correlation in air pollutant concentration data, and the synchronous pre- count off obtaining many monitoring stations
According to forecast precision is high；The precision in the forecast of extreme air pollutant levels can also be improved.
Embodiment：
Using 12, Beijing urban district air quality monitoring station on May 28,1 day to 2016 January in 2014 by hour
PM2.5 mean concentration numbers, carry out PM2.5 hour average concentration using the air pollutant concentration forecasting procedure that the present invention provides
Data prediction.That is, forecast time delay r is one hour；In each history concentration samples, between the time between the data of adjacent sequential
Every being one hour.
After pretreatment, historical data is concentrated and is comprised 20196 PM2.5 hour average concentration records.By random
Select 60% data as training set, as checking collection, remaining 20% data is as test set for 20% data.
During model training undetermined and Model Parameter Optimization undetermined, the set to be selected of each parameter is as shown in table 1.
Rational method is treated in table 1 prediction of air quality
Through grid search optimization, Step Parameters are 8 and the number of plies of stack self-encoding encoder is 3 and every layer of own coding utensil
When having 300 nodes, the value of forecasting of PM2.5 hour average concentration is best.For a certain website, its PM2.5 hourly average is dense
Relatively, specific forecast precision index is respectively for degree predicted value and actual value：MAE=8.44 μ g/m3, RMSE=14 μ g/
m3, MAPE=18.6%.
Further, the STDL model air pollutant concentration forecasting procedure that the present invention provides being used is artificial with space-time
Neutral net (Spatio-Temporal Artificial Neural Network, abbreviation STANN) model, support vector machine
Model (Support Vector Machine, abbreviation SVR) and arma modeling carry out contrast test respectively.These models use phase
Same training set and test set, but mode input is slightly different.STANN model uses identical data, can be to multiple stations
The air quality of point synchronizes forecast.But STANN model does not use successively pre-training method, but use general neural network
Training method, namely back-propagation algorithm.For SVR model and arma modeling, they are time series models, so that
Individually modeled respectively for each website and forecast.
The forecast precision of above-mentioned 4 kinds of methods is as shown in table 2, from Table 2, it can be seen that the air pollutants that the present invention provides
The forecast precision of the space-time deep learning model that concentration prediction method uses is highest, and MAPE improves more than 5.12%.
The forecast precision of STANN model is higher than the forecast precision of other two kinds of time series models (arma modeling and SVR model), explanation
Consider when air quality data is modeled that spatial coherence is highly effective.
Table 2 prediction of air quality model accuracy contrasts
Descriptions above can combine individually or in every way enforcement, and these variant all exist
Within protection scope of the present invention.
One of ordinary skill in the art will appreciate that all or part of step in said method can be instructed by program
Related hardware completes, and described program can be stored in computer-readable recording medium, such as read only memory, disk or CD
Deng.Alternatively, all or part of step of above-described embodiment can also be realized using one or more integrated circuits, accordingly
Ground, each module/unit in above-described embodiment can be to be realized in the form of hardware, it would however also be possible to employ the shape of software function module
Formula is realized.The present invention is not restricted to the combination of the hardware and software of any particular form.
It should be noted that herein, term " inclusion ", "comprising" or its any other variant are intended to non-row
The comprising of his property, so that including a series of article of key elements or equipment not only includes those key elements, but also include not having
There are other key elements being expressly recited, or also include for this article or the intrinsic key element of equipment.There is no more limits
In the case of system, the key element that limited by sentence " include ... " is it is not excluded that in the article including described key element or equipment
Also there is other identical element.
Above example only in order to technical scheme to be described and unrestricted, reference only to preferred embodiment to this
Bright it has been described in detail.It will be understood by those within the art that, technical scheme can be modified
Or equivalent, without deviating from the spirit and scope of technical solution of the present invention, all should cover the claim model in the present invention
In the middle of enclosing.
Claims (10)
1. a kind of air pollutant concentration forecasting procedure is it is characterised in that include：
By the value multiple to be selected of the node number of input layer, the value multiple to be selected of the number of plies of stack self-encoding encoder, stack own coding
The value multiple to be selected of the node number of every layer of device is combined determining multiple models undetermined, arranges the defeated of each model undetermined
The node number going out layer is m；Wherein, the node number of input layer is n times of node number m of output layer, and n joins for time step
Number；
Value multiple to be selected according to Step Parameters n and forecast time delay r specified, the air obtaining from m monitoring station is dirty
Multigroup history concentration data of dye thing, extracts training dataset and checking data set from described history concentration data；
It is trained using each model undetermined of described training data set pair until model undetermined convergence, what record training completed respectively treats
Cover half type corresponding Model Weight matrix and model bias vector；
The data input being used for inputting in described checking data set is extremely trained the model each undetermined completing, calculates each model undetermined
Output result and described checking data set in for checking data synthetic error, treat corresponding for minimum synthetic error
Cover half type is defined as forecasting model, and wherein, the value of the Step Parameters of described forecasting model is nr；
To be n from m monitoring station, Step ParametersrAir pollutant concentration observation data composition forecast data collection,
Described forecast data collection is inputted to described forecasting model, the output result of described forecasting model is r's as forecast time delay
Forecast result.
2. air pollutant concentration forecasting procedure as claimed in claim 1 it is characterised in that described using described training data
Each model undetermined of set pair is trained until model undetermined convergence includes：
The weight matrix of stack self-encoding encoder and bias vector are instructed using successively coaching method using described training dataset
Practice；
Using the output vector of last layer of stack self-encoding encoder as output layer input, and using back-propagation algorithm to stack
The weight matrix of formula self-encoding encoder and output layer and bias vector are adjusted from back to front up to model convergence undetermined.
3. air pollutant concentration forecasting procedure as claimed in claim 2 it is characterised in that
When the weight matrix of stack self-encoding encoder and bias vector being trained using successively coaching method, had using minimum
The reconstructed error of sparse restrictive condition is adjusting the weight matrix W of every layer of self-encoding encoder1And W2And bias vector b and c, described tool
There is the reconstructed error such as following formula of sparse restrictive condition：
Wherein, yj=f (W1x+bj), x={ x(1),...,x(i),...,x(N), z(i)=g
(W2y+ci), i=1 ..., N, j=1 ..., HD, λ is the weight of regularization term, and μ is the weight of sparse item, and N is criticizing of input
Training sample number, HDIt is the node number of current layer self-encoding encoder, | | W1||2It is W1L2Norm, | | W2||2It is W2L2Model
Number, ρ is Sparse parameter, x(i)For i-th input vector of current layer self-encoding encoder, yjFor j-th element of vector after coding, z(i)For the corresponding decoded vector of i-th input vector of current layer self-encoding encoder.
4. air pollutant concentration forecasting procedure as claimed in claim 2 is it is characterised in that using successively coaching method to stack
When the weight matrix of self-encoding encoder and bias vector are trained, it is that every layer of self-encoding encoder arranges single iterationses and study
Speed, and constantly reduce learning rate in the training process；
Using back-propagation algorithm, the weight matrix of stack self-encoding encoder and output layer and bias vector are adjusted from back to front
When whole, be in course of adjustment continuous reduction learning rate.
5. air pollutant concentration forecasting procedure as claimed in claim 1 is it is characterised in that the output of described each model undetermined
Result includes root-mean-square error, mean absolute error with the described synthetic error verifying the data for checking in data set and puts down
All absolute percent error.
6. air pollutant concentration forecasting procedure as claimed in claim 1 is it is characterised in that described acquisition is derived from m monitoring
Multigroup history concentration data of the air pollutants of website also includes：
The physical meaning of the concentration data according to air pollutants, rejects the exceptional value in history concentration data using box traction substation；
And/or
Using linear interpolation method, the missing values in history concentration data are filled up；And/or
Using minimax method for normalizing, history concentration data is normalized.
7. air pollutant concentration forecasting procedure as claimed in claim 1 is it is characterised in that every in described stack self-encoding encoder
Layer is from the encoder such as following formula of code device：Decoder such as following formula：
8. air pollutant concentration forecasting procedure as claimed in claim 1 it is characterised in that each described model undetermined defeated
The activation primitive going out layer is
9. air pollutant concentration forecasting procedure as claimed in claim 1 is it is characterised in that described air pollutants are thin
Any one in grain thing, inhalable particles, sulfur dioxide, nitrogen dioxide, carbon monoxide and ozone.
10. a kind of air pollutant concentration forecast system is it is characterised in that include：
Model determining unit undetermined, for by the value multiple to be selected of the node number of input layer, the number of plies of stack self-encoding encoder
Multiple values to be selected, the value multiple to be selected of every layer of node number of stack self-encoding encoder are combined determining multiple treats cover half
Type, the node number arranging the output layer of each model undetermined is m；Wherein, the node number of input layer is the node of output layer
N times of number m, n is Step Parameters；
History concentration data processing unit, for the value multiple to be selected according to Step Parameters n with forecast time delay r specified, obtains
Fetch multigroup history concentration data of the air pollutants from m monitoring station, from described history concentration data, extract instruction
Practice data set and checking data set；
Model training unit undetermined, for being trained using each model undetermined of described training data set pair until model undetermined is received
Hold back, model each undetermined corresponding Model Weight matrix and model bias vector that record training completes；
Forecasting model determining unit, for respectively treating the data input being used for inputting in described checking data set to what training completed
Cover half type, calculates the output result of each model undetermined and the described synthetic error verifying the data for checking in data set, will
The minimum corresponding model undetermined of synthetic error is defined as forecasting model, wherein, the value of the Step Parameters of described forecasting model
For nr；
Air pollutant concentration forecast unit, for being n from m monitoring station, Step ParametersrAir pollutants dense
The observation data composition forecast data collection of degree, described forecast data collection is inputted to described forecasting model, by described forecasting model
Output result as forecast time delay be r forecast result.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610875403.XA CN106485353B (en) | 2016-09-30 | 2016-09-30 | Air pollutant concentration forecasting procedure and system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610875403.XA CN106485353B (en) | 2016-09-30 | 2016-09-30 | Air pollutant concentration forecasting procedure and system |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN106485353A true CN106485353A (en) | 2017-03-08 |
| CN106485353B CN106485353B (en) | 2019-11-29 |
Family
ID=58268474
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201610875403.XA Active CN106485353B (en) | 2016-09-30 | 2016-09-30 | Air pollutant concentration forecasting procedure and system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106485353B (en) |
Cited By (26)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106997550A (en) * | 2017-03-17 | 2017-08-01 | 常州大学 | A kind of method of the ad click rate prediction based on stack self-encoding encoder |
| CN107229916A (en) * | 2017-05-27 | 2017-10-03 | 南京航空航天大学 | A kind of airport noise Monitoring Data restorative procedure based on depth noise reduction own coding |
| CN107679743A (en) * | 2017-09-29 | 2018-02-09 | 河北百斛环保科技有限公司 | A kind of emergent schedule management method of tracing to the source of environment-protecting intelligent law enforcement based on big data |
| CN108061360A (en) * | 2017-12-06 | 2018-05-22 | 成都猴子软件有限公司 | Be conducive to family's control method with fresh air |
| CN108627326A (en) * | 2018-05-07 | 2018-10-09 | 东南大学 | A kind of elevator brake method of evaluating performance based on Bagging-RNN models |
| CN108921440A (en) * | 2018-07-11 | 2018-11-30 | 平安科技（深圳）有限公司 | Pollutant method for monitoring abnormality, system, computer equipment and storage medium |
| CN109034252A (en) * | 2018-08-01 | 2018-12-18 | 中国科学院大气物理研究所 | The automatic identification method of air quality website monitoring data exception |
| CN109146161A (en) * | 2018-08-07 | 2019-01-04 | 河海大学 | Merge PM2.5 concentration prediction method of the stack from coding and support vector regression |
| CN109521725A (en) * | 2017-09-20 | 2019-03-26 | 西门子公司 | The method, apparatus and equipment and machine readable media of detection abnormal data |
| CN109615124A (en) * | 2018-11-29 | 2019-04-12 | 中国铁路总公司 | A kind of SCADA main website load forecasting method based on deep learning |
| CN109657842A (en) * | 2018-11-27 | 2019-04-19 | 平安科技（深圳）有限公司 | The prediction technique and device of air pollutant concentration, electronic equipment |
| CN109685249A (en) * | 2018-11-24 | 2019-04-26 | 上海师范大学 | Air PM2.5 concentration prediction method based on AutoEncoder and BiLSTM fused neural network |
| CN110135057A (en) * | 2019-05-14 | 2019-08-16 | 北京工业大学 | Solid waste burning process dioxin concentration flexible measurement method based on multilayer feature selection |
| CN110209992A (en) * | 2019-05-24 | 2019-09-06 | 西北工业大学 | A kind of perceived position selection method based on space and cross-domain correlation |
| CN110261547A (en) * | 2019-07-04 | 2019-09-20 | 北京思路创新科技有限公司 | A kind of Urban Air Pollution Methods and equipment |
| CN110428053A (en) * | 2019-08-08 | 2019-11-08 | 东北大学 | A kind of steam produces the dynamic prediction method of consumption |
| CN110782093A (en) * | 2019-10-26 | 2020-02-11 | 福州大学 | PM fusing SSAE deep feature learning and LSTM network 2.5Hourly concentration prediction method and system |
| WO2020044127A1 (en) * | 2018-08-25 | 2020-03-05 | 山东诺方电子科技有限公司 | Atmospheric pollution forecasting method |
| CN110927338A (en) * | 2019-12-04 | 2020-03-27 | 华北电力科学研究院有限责任公司 | Gas concentration data complementing method and device |
| CN110939959A (en) * | 2018-09-25 | 2020-03-31 | 珠海格力电器股份有限公司 | Control method and device of range hood |
| CN111832814A (en) * | 2020-07-01 | 2020-10-27 | 北京工商大学 | Air pollutant concentration prediction method based on graph attention machine mechanism |
| CN112561191A (en) * | 2020-12-22 | 2021-03-26 | 北京百度网讯科技有限公司 | Prediction model training method, prediction method, device, apparatus, program, and medium |
| CN113888381A (en) * | 2021-09-24 | 2022-01-04 | 中科三清科技有限公司 | Pollutant concentration forecasting method and device |
| CN114282721A (en) * | 2021-12-22 | 2022-04-05 | 中科三清科技有限公司 | Pollutant forecast model training method and device, electronic equipment and storage medium |
| US11714939B2 (en) | 2018-01-09 | 2023-08-01 | Landmark Graphics Corporation | Determining gas leak flow rate in a wellbore environment |
| WO2023173987A1 (en) * | 2022-03-16 | 2023-09-21 | International Business Machines Corporation | Prediction and operational efficiency for system-wide optimization of an industrial processing system |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20110257837A1 (en) * | 2006-09-12 | 2011-10-20 | Spx Corporation | Method and Apparatus for Reading and Erasing Diagnostic Trouble Codes from a Vehicle |
| CN104166859A (en) * | 2014-08-13 | 2014-11-26 | 西安电子科技大学 | Polarization SAR image classification based on SSAE and FSALS-SVM |
| CN105825231A (en) * | 2016-03-15 | 2016-08-03 | 浙江工业大学 | Classification method for spectral features of space debris based on artificial intelligence |
-
2016
- 2016-09-30 CN CN201610875403.XA patent/CN106485353B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20110257837A1 (en) * | 2006-09-12 | 2011-10-20 | Spx Corporation | Method and Apparatus for Reading and Erasing Diagnostic Trouble Codes from a Vehicle |
| CN104166859A (en) * | 2014-08-13 | 2014-11-26 | 西安电子科技大学 | Polarization SAR image classification based on SSAE and FSALS-SVM |
| CN105825231A (en) * | 2016-03-15 | 2016-08-03 | 浙江工业大学 | Classification method for spectral features of space debris based on artificial intelligence |
Non-Patent Citations (2)
| Title |
|---|
| 尹文君等: "基于深度学习的大数据空气污染预报", 《中国环境管理》 * |
| 王知音等: "基于栈式自编码的水体提取方法", 《计算机应用》 * |
Cited By (37)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106997550A (en) * | 2017-03-17 | 2017-08-01 | 常州大学 | A kind of method of the ad click rate prediction based on stack self-encoding encoder |
| CN107229916A (en) * | 2017-05-27 | 2017-10-03 | 南京航空航天大学 | A kind of airport noise Monitoring Data restorative procedure based on depth noise reduction own coding |
| US11379745B2 (en) | 2017-09-20 | 2022-07-05 | Siemens Aktiengesellschaft | Method, apparatus, device and machine-readable medium for detecting abnormal data using an autoencoder |
| CN109521725A (en) * | 2017-09-20 | 2019-03-26 | 西门子公司 | The method, apparatus and equipment and machine readable media of detection abnormal data |
| CN107679743A (en) * | 2017-09-29 | 2018-02-09 | 河北百斛环保科技有限公司 | A kind of emergent schedule management method of tracing to the source of environment-protecting intelligent law enforcement based on big data |
| CN107679743B (en) * | 2017-09-29 | 2021-06-22 | 河北百斛环保科技有限公司 | Environment-friendly intelligent law enforcement emergency traceability scheduling management method based on big data |
| CN108061360A (en) * | 2017-12-06 | 2018-05-22 | 成都猴子软件有限公司 | Be conducive to family's control method with fresh air |
| US11714939B2 (en) | 2018-01-09 | 2023-08-01 | Landmark Graphics Corporation | Determining gas leak flow rate in a wellbore environment |
| CN108627326A (en) * | 2018-05-07 | 2018-10-09 | 东南大学 | A kind of elevator brake method of evaluating performance based on Bagging-RNN models |
| CN108921440A (en) * | 2018-07-11 | 2018-11-30 | 平安科技（深圳）有限公司 | Pollutant method for monitoring abnormality, system, computer equipment and storage medium |
| CN109034252A (en) * | 2018-08-01 | 2018-12-18 | 中国科学院大气物理研究所 | The automatic identification method of air quality website monitoring data exception |
| CN109034252B (en) * | 2018-08-01 | 2020-10-30 | 中国科学院大气物理研究所 | Automatic identification method for monitoring data abnormity of air quality station |
| CN109146161A (en) * | 2018-08-07 | 2019-01-04 | 河海大学 | Merge PM2.5 concentration prediction method of the stack from coding and support vector regression |
| GB2591886A (en) * | 2018-08-25 | 2021-08-11 | Nova Fitness Co Ltd | Atmospheric pollution forecasting method |
| WO2020044127A1 (en) * | 2018-08-25 | 2020-03-05 | 山东诺方电子科技有限公司 | Atmospheric pollution forecasting method |
| CN110939959A (en) * | 2018-09-25 | 2020-03-31 | 珠海格力电器股份有限公司 | Control method and device of range hood |
| CN109685249A (en) * | 2018-11-24 | 2019-04-26 | 上海师范大学 | Air PM2.5 concentration prediction method based on AutoEncoder and BiLSTM fused neural network |
| CN109657842A (en) * | 2018-11-27 | 2019-04-19 | 平安科技（深圳）有限公司 | The prediction technique and device of air pollutant concentration, electronic equipment |
| CN109615124B (en) * | 2018-11-29 | 2022-12-30 | 中国铁路总公司 | SCADA master station load prediction method based on deep learning |
| CN109615124A (en) * | 2018-11-29 | 2019-04-12 | 中国铁路总公司 | A kind of SCADA main website load forecasting method based on deep learning |
| CN110135057A (en) * | 2019-05-14 | 2019-08-16 | 北京工业大学 | Solid waste burning process dioxin concentration flexible measurement method based on multilayer feature selection |
| CN110209992A (en) * | 2019-05-24 | 2019-09-06 | 西北工业大学 | A kind of perceived position selection method based on space and cross-domain correlation |
| CN110261547B (en) * | 2019-07-04 | 2020-12-15 | 北京思路创新科技有限公司 | Air quality forecasting method and equipment |
| CN110261547A (en) * | 2019-07-04 | 2019-09-20 | 北京思路创新科技有限公司 | A kind of Urban Air Pollution Methods and equipment |
| CN110428053A (en) * | 2019-08-08 | 2019-11-08 | 东北大学 | A kind of steam produces the dynamic prediction method of consumption |
| CN110782093A (en) * | 2019-10-26 | 2020-02-11 | 福州大学 | PM fusing SSAE deep feature learning and LSTM network 2.5Hourly concentration prediction method and system |
| CN110782093B (en) * | 2019-10-26 | 2022-05-27 | 福州大学 | PM fusing SSAE deep feature learning and LSTM2.5Hourly concentration prediction method and system |
| CN110927338A (en) * | 2019-12-04 | 2020-03-27 | 华北电力科学研究院有限责任公司 | Gas concentration data complementing method and device |
| CN110927338B (en) * | 2019-12-04 | 2023-02-07 | 华北电力科学研究院有限责任公司 | Gas concentration data complementing method and device |
| CN111832814B (en) * | 2020-07-01 | 2023-06-23 | 北京工商大学 | Air pollutant concentration prediction method based on graph attention mechanism |
| CN111832814A (en) * | 2020-07-01 | 2020-10-27 | 北京工商大学 | Air pollutant concentration prediction method based on graph attention machine mechanism |
| CN112561191A (en) * | 2020-12-22 | 2021-03-26 | 北京百度网讯科技有限公司 | Prediction model training method, prediction method, device, apparatus, program, and medium |
| CN112561191B (en) * | 2020-12-22 | 2024-02-27 | 北京百度网讯科技有限公司 | Prediction model training method, prediction device, prediction apparatus, prediction program, and program |
| CN113888381A (en) * | 2021-09-24 | 2022-01-04 | 中科三清科技有限公司 | Pollutant concentration forecasting method and device |
| CN113888381B (en) * | 2021-09-24 | 2023-08-29 | 中科三清科技有限公司 | Pollutant Concentration Forecasting Method and Device |
| CN114282721A (en) * | 2021-12-22 | 2022-04-05 | 中科三清科技有限公司 | Pollutant forecast model training method and device, electronic equipment and storage medium |
| WO2023173987A1 (en) * | 2022-03-16 | 2023-09-21 | International Business Machines Corporation | Prediction and operational efficiency for system-wide optimization of an industrial processing system |
Also Published As
| Publication number | Publication date |
|---|---|
| CN106485353B (en) | 2019-11-29 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106485353A (en) | Air pollutant concentration forecasting procedure and system | |
| Guan et al. | An artificial-neural-network-based, constrained CA model for simulating urban growth | |
| CN106781489B (en) | A kind of road network trend prediction method based on recurrent neural network | |
| CN106971240A (en) | The short-term load forecasting method that a kind of variables choice is returned with Gaussian process | |
| CN103761138B (en) | Parameter correction method for traffic simulation software | |
| Flamm et al. | Alternative model formulations for a stochastic simulation of landscape change | |
| CN107730039A (en) | The method and system of distribution network load prediction | |
| CN105931116A (en) | Automated credit scoring system and method based on depth learning mechanism | |
| CN110826244B (en) | Conjugated gradient cellular automaton method for simulating influence of rail transit on urban growth | |
| CN110909924A (en) | City expansion multi-scenario simulation cellular automata method based on cross entropy optimizer | |
| CN110046764A (en) | The method and device of passenger flow forecast amount | |
| CN108090788A (en) | Ad conversion rates predictor method based on temporal information integrated model | |
| CN110363230A (en) | Stacking integrated sewage handling failure diagnostic method based on weighting base classifier | |
| CN110059867A (en) | A kind of wind speed forecasting method of SWLSTM combination GPR | |
| CN105930456A (en) | Method for simulating potential distribution of endangered pinaceae plants | |
| CN104361255B (en) | It is a kind of to improve cellular automata urban sprawl analogy method | |
| Sharma | Review of most used urban growth models | |
| Sun | Real estate evaluation model based on genetic algorithm optimized neural network | |
| CN108537581A (en) | Based on the GMDH energy consumption Time Series Forecasting Methods selectively combined and device | |
| CN102663562B (en) | Method for grading land resource evaluation factors based on clonal selection algorithm | |
| Duca et al. | Joint modelling wind speed and power via Bayesian Dynamical models | |
| CN106844626A (en) | Using microblogging keyword and the method and system of positional information simulated air quality | |
| Sangita et al. | An improved k-means clustering approach for teaching evaluation | |
| Meng et al. | Application of K-means algorithm based on ant clustering algorithm in macroscopic planning of highway transportation hub | |
| Srisuksomwong et al. | The Chlorophyll-a Modelling over the Andaman Sea using Bi-Directional LSTM Neural Network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |