Nothing Special   »   [go: up one dir, main page]

CN107909206B - PM2.5 prediction method based on deep structure recurrent neural network - Google Patents

PM2.5 prediction method based on deep structure recurrent neural network Download PDF

Info

Publication number
CN107909206B
CN107909206B CN201711130537.XA CN201711130537A CN107909206B CN 107909206 B CN107909206 B CN 107909206B CN 201711130537 A CN201711130537 A CN 201711130537A CN 107909206 B CN107909206 B CN 107909206B
Authority
CN
China
Prior art keywords
data
prediction model
training
historical weather
weather data
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN201711130537.XA
Other languages
Chinese (zh)
Other versions
CN107909206A (en
Inventor
刘珊
杨波
郑文锋
宋利红
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
University of Electronic Science and Technology of China
Original Assignee
University of Electronic Science and Technology of China
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by University of Electronic Science and Technology of China filed Critical University of Electronic Science and Technology of China
Priority to CN201711130537.XA priority Critical patent/CN107909206B/en
Publication of CN107909206A publication Critical patent/CN107909206A/en
Application granted granted Critical
Publication of CN107909206B publication Critical patent/CN107909206B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Business, Economics & Management (AREA)
  • Economics (AREA)
  • General Physics & Mathematics (AREA)
  • Human Resources & Organizations (AREA)
  • Strategic Management (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Biomedical Technology (AREA)
  • Quality & Reliability (AREA)
  • Tourism & Hospitality (AREA)
  • Marketing (AREA)
  • General Business, Economics & Management (AREA)
  • Game Theory and Decision Science (AREA)
  • Development Economics (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Operations Research (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

本发明公开了一种基于深层结构循环神经网络的PM2.5预测方法，利用采集的大量数据，根据深度学习和循环神经网络理论构建深层结构的PM2.5的预测模型，通过数据特征的提取和训练，实现雾霾天气的预测，旨在提高雾霾预测的效率和精度，为雾霾预防和治理提出有说服力的决策依据。预测模型对于数据结构几乎没什么要求，只要数据足够大时能自学习，使得深度学习非常适合当下互联网大数据应用的需要。

Figure 201711130537

The invention discloses a PM2.5 prediction method based on a deep structure cyclic neural network. A large amount of collected data is used to construct a deep structure PM2.5 prediction model according to deep learning and cyclic neural network theory. Training to realize the prediction of haze weather, aiming to improve the efficiency and accuracy of haze prediction, and to provide a persuasive decision basis for haze prevention and governance. The prediction model has almost no requirements on the data structure. As long as the data is large enough, it can learn by itself, making deep learning very suitable for the needs of current Internet big data applications.

Figure 201711130537

Description

PM2.5 prediction method based on deep structure recurrent neural network
Technical Field
The invention belongs to the technical field of environmental engineering detection, and particularly relates to a PM2.5 prediction method based on a deep structure cyclic neural network.
Background
The air quality is always a great problem related to the future fate of human beings, and along with the social progress and the rapid increase of the automobile holding amount, the content of particles which can absorb people in the air is greatly increased, and the problem of environmental pollution is getting more and more serious. Along with the continuous deterioration of air quality, haze weather phenomenon is more and more, and harm is bigger and bigger. Haze is a disaster weather phenomenon. The inhalable particles PM2.5 are the main cause of haze weather, have serious influence on air quality, and importantly have great threat to human health.
The air quality prediction research has many ideas and methods, and among many methods, the realization of quantitative research and effective prediction on environmental quality, especially haze, based on the ideas of system engineering and effectively combined with new theories and new methods is a main development trend.
Due to the influence of a large number of uncertainty and complexity factors such as climate, temperature, human activities and the like, time sequences of various weather data have characteristics of high nonlinearity, uncertainty and the like, and conventional analysis and prediction methods are difficult to master change rules and change characteristics.
The shallow neural network has obvious effect on solving simple or more limited problems, but has limited realization capability for some more complex problems related to natural signals in real life due to limited modeling and representation capability.
The deep neural network has a plurality of hidden layers, has more structural advantages than the traditional neural network, and has strong feature abstraction capability. The deep neural network adopts a brand-new coding mode, an algorithm and programming are not required to be directly designed for solving the problems, only the programming is required in the training process, the correct method for solving the problems can be learned by the network in the retraining process, and under the condition that the data volume is ensured, the special effect can be obtained by the simple algorithm and the complex data.
Meanwhile, due to the huge improvement of the chip processing performance, the data for training is explosively increased, and machine learning and signal and information processing research are greatly developed recently, so that the deep learning method can effectively utilize complex nonlinear functions and nonlinear compound functions to learn distributed and layered feature representations and can fully and effectively utilize labeled and non-labeled data.
A Recurrent Neural Network (RNN) is a class of deep networks that can be used for unsupervised (and supervised) learning, even to the extent of the length of the input sequence, in unsupervised learning mode, RNNs are used to predict future data sequences from previous data samples, and class information is not used in the learning process, so RNNs are well suited for sequence data modeling.
Disclosure of Invention
The invention aims to overcome the defects of the prior art and provides a PM2.5 prediction method based on a deep-structure recurrent neural network, wherein a PM2.5 prediction model is built by combining the basic theory, the network structure and the flow principle of the recurrent neural network, so that the PM2.5 prediction is realized.
In order to achieve the above object, the present invention provides a PM2.5 prediction method based on a deep structure recurrent neural network, comprising the steps of:
(1) obtaining historical weather data, including hourly temperature, illumination, wind speed, rainfall, SO2, O3, NO, PM10, PM2.5 data indexes, wherein, the temperature unit: DEG C, light unit: lm/square meter, wind speed unit: m/s, rainfall units: mm, SO2, O3, NO, PM10 and PM2.5 are concentration data;
(2) data preprocessing
(2.1) complementing missing historical weather data
And (3) complementing the missing historical weather data by using an averaging method:
Figure BDA0001469418970000021
wherein, XtIndicating missing historical weather data at the current time, Xt-1Indicating missing historical weather data at the previous time, Xt+1Representing missing historical weather data at a previous and subsequent time;
(2.2) normalizing all historical weather data
Normalizing the historical weather data to be between-1 and 1 according to the following formula;
Figure BDA0001469418970000022
wherein X' represents the historical weather data after normalization, X represents the historical weather data before normalization,
Figure BDA0001469418970000023
means for representing historical weather data, XmaxRepresenting a maximum value, X, of historical weather dataminRepresenting historical weather data minimum;
(3) dividing the preprocessed historical weather data into training data and testing data according to a proportion;
(4) PM2.5 prediction model for constructing deep structure based on deep learning theory and recurrent neural network
(4.1) constructing a deep circulation neural network prediction model: the model depth is greater than N layers, the input is training data, and the output is a predicted value of PM2.5 concentration;
(4.2) setting the dimension of an input layer as Kx (H-1), setting the dimension of an output layer as 1 xT, and adopting Tanh functions as activation functions of the input layer and the hidden layer, the hidden layer and the hidden layer, and the hidden layer and the output layer;
wherein K represents the depth of the recurrent neural network expanded according to the time sequence, namely K time frames, and each time frame inputs a group of historical weather data; h represents the number of data indexes, T represents the number of data output by a prediction model of the recurrent neural network, and represents that K pieces of historical data are used for predicting the PM2.5 concentration at T moments in the future, namely the weather data at the first K moments are input, and the PM2.5 concentration data at the later T moments are predicted;
(4.3) selecting a loss function for use in the PM2.5 prediction model
The mean square error is used as a loss function in the PM2.5 prediction model:
Figure BDA0001469418970000031
where q represents the number of iterations, t is the output vector dimension, yi,jActual value, y, representing training datai,j' represents a predicted value of training data;
(4.4) updating parameters in the PM2.5 prediction model by adopting a small batch stochastic gradient descent algorithm
(4.4.1) initializing parameter θ0；
(4.4.2) dividing the training data into a group according to each m training data of the time sequence, calculating the gradient value of each training data in the first group of training data by using a small batch stochastic gradient descent algorithm, and performing weighted average summation on the gradient values to obtain the descent gradient of the group of training data
Figure BDA0001469418970000032
i represents the ith set of training data,
Figure BDA0001469418970000033
represents input and output data corresponding to the τ -th training data in the ith group;
(4.4.3) updating parameters in the PM2.5 prediction model by the descending gradient of the training data, wherein the parameter updating formula is as follows:
Figure BDA0001469418970000034
wherein, thetai-1Represents the target parameter theta after the training of the last group of data is finishediRepresenting target parameters after the training data of the group is finished, and representing the learning rate by eta;
(4.4.4), after the target parameters of the training data are updated, returning to the step (4.4.2) to train and update the next training data group until the error value is lower than the set expected error value or the training of the last training data group is finished, and then updating and storing the final parameters to obtain a PM2.5 prediction model after the training is finished;
(5) judging whether the PM2.5 prediction model reaches the training stop condition or not
Inputting K groups of data into a trained PM2.5 prediction model according to a time sequence, outputting T predicted values, judging the error between each predicted value and the true value, if the error is within an allowable range, considering that the prediction model completes training, and if not, returning to the step (4) to retrain until a stop condition is reached;
(6) PM2.5 prediction by using PM2.5 prediction model
And inputting the current K groups of weather data into a PM2.5 prediction model, and outputting T PM2.5 prediction values.
The invention aims to realize the following steps:
according to the PM2.5 prediction method based on the deep structure recurrent neural network, the collected mass data are utilized, the prediction model of the PM2.5 of the deep structure is constructed according to deep learning and recurrent neural network theories, and the prediction of haze weather is realized through extraction and training of data characteristics, so that the efficiency and the precision of haze prediction are improved, and a convincing decision basis is provided for haze prevention and treatment. The prediction model has little requirement on a data structure, and can be learned by self as long as the data is large enough, so that the deep learning is very suitable for the requirements of current internet big data application.
Meanwhile, the PM2.5 prediction method based on the deep structure recurrent neural network further has the following beneficial effects:
(1) the prediction model transforms original data into higher-level abstract expression through some simple nonlinear models, combines multi-layer transformation, and learns and extracts a very complex function characteristic method. The problem that the traditional prediction model of the shallow structure has limited representation capability on complex functions under the condition of limited samples and calculation units is solved.
(2) The core difference of the prediction model is that a plurality of hidden layers are provided, and the feature extraction of each layer is not designed manually but obtained by self-learning from data by using a general learning process, so that the data structure is not required, the data processing process is simplified, and the efficiency is improved.
(3) The PM2.5 concentration prediction of the region can be realized by using data of different regions, parameters of a PM2.2 prediction model are redefined according to actual conditions and requirements, and a network does not need to be reconstructed, so that the method has flexibility and portability.
Drawings
FIG. 1 is a flowchart of a method for predicting PM2.5 based on a deep-structure recurrent neural network according to the present invention;
Detailed Description
The following description of the embodiments of the present invention is provided in order to better understand the present invention for those skilled in the art with reference to the accompanying drawings. It is to be expressly noted that in the following description, a detailed description of known functions and designs will be omitted when it may obscure the subject matter of the present invention.
Examples
FIG. 1 is a flowchart of a method for predicting PM2.5 based on a deep-structure recurrent neural network according to the present invention.
In this embodiment, as shown in fig. 1, a method for predicting PM2.5 based on a deep-structure recurrent neural network according to the present invention includes the following steps:
s1, obtaining historical weather data, including hourly temperature, illumination, wind speed, rainfall, SO2, O3, NO, PM10 and PM2.5 data indexes, wherein the temperature unit: DEG C, light unit: lm/square meter, wind speed unit: m/s, rainfall units: mm, SO2, O3, NO, PM10 and PM2.5 are concentration data;
in the embodiment, historical weather data from 5 months 2014 to 5 months 2017 are applied and obtained from the China weather service bureau, and data information comprises temperature, light, wind speed, rainfall, SO2, O3, NO, PM10 and PM2.5 data indexes (wherein the temperature unit is in DEG C, the light unit is lm/square meter, the wind speed unit is m/s, the rainfall unit is mm, the SO2, the O3, the NO, the PM10 and the PM2.5 are concentration data) of each hour, 9 indexes are counted in total, 26280 multiplied by 9 data are counted in total, and the data quantity of the PM2.5 model predicted by the recurrent neural network of the deep layer structure is ensured.
S2, preprocessing data
S2.1, complementing the missing historical weather data
The collected weather data is historical data based on a time sequence, few missing data exist in the collected data, the missing data are supplemented by adopting a mean value method, and the integrity of the data is ensured.
The formula for complementing the missing historical weather data by using an averaging method is as follows:
Figure BDA0001469418970000061
wherein, XtIndicating missing historical weather data at the current time, Xt-1Indicating missing historical weather data at the previous time, Xt+1Indicating the number of missing historical days at a time before and afterAccordingly;
s2.2, normalizing all historical weather data
Normalizing the historical weather data to be between-1 and 1 according to the following formula;
Figure BDA0001469418970000062
wherein X' represents the historical weather data after normalization, X represents the historical weather data before normalization,
Figure BDA0001469418970000063
means for representing historical weather data, XmaxRepresenting a maximum value, X, of historical weather dataminRepresenting historical weather data minimum;
s3, dividing the preprocessed historical weather data into training data and testing data according to the proportion of 70% to 30%;
s4 PM2.5 prediction model for constructing deep structure based on deep learning theory and recurrent neural network
S4.1, in the embodiment, constructing a deep-layer circulation neural network prediction model: the model comprises an input layer, eight hidden layers and an input layer, wherein the model depth is 10 layers, the number of nodes of the input layer is 9, the number of nodes of the hidden layer is 50, the number of nodes of the output layer is 5, the input is training data, and the output is a predicted value of PM2.5 concentration;
s4.2, setting the dimension of an input layer as Kx 9, the dimension of an output layer as 1 xT, the dimension of the input layer and a hidden layer, the dimension of the hidden layer and the hidden layer, and the activation functions of the hidden layer and the output layer as Tanh functions;
wherein K represents the depth of the recurrent neural network expanded according to the time sequence, namely K time frames, and each time frame inputs a group of historical weather data; t represents the number of output data of a prediction model of the recurrent neural network, and represents that K pieces of historical data are used for predicting the PM2.5 concentration at the T moments in the future, namely the weather data at the K moments before input and the PM2.5 concentration data at the T moments after input are predicted;
s4.3, selecting a loss function used in the PM2.5 prediction model
In this embodiment, the subset M is selected from the training data set segment by segment according to the data in the PM2.5 prediction model in time seriesiSubset M as a small batch datasetiThe number of the middle sample points is M, the middle sample points comprise input and mark output data, and the number is marked as Mi(Xi,Yi) Input data is XiTrue tag output data is YiThe neural network training output data is represented as Yi', the correspondence relationship is expressed as follows:
Figure BDA0001469418970000071
Figure BDA0001469418970000072
i.e. representing the input of the # th one in time series
Figure BDA0001469418970000073
Matrix, namely to obtain
Figure BDA0001469418970000074
A row vector, τ ═ 1, 2.., m-k + 1.
Each subset MiThe training loss function of (a) is chosen to be the mean square error, representing the measure of error of the parameter vector of the prediction model over a given subset of training data, noted as
Figure BDA0001469418970000075
Expressed by the following formula:
Figure BDA0001469418970000076
where q represents the number of iterations, t is the output vector dimension, yi,jActual value, y, representing training datai,j' represents a predicted value of training data;
s4.4, updating parameters in the PM2.5 prediction model by adopting a small-batch stochastic gradient descent algorithm
S4.4.1, initialization parameter θ0；
S4.4.2, dividing the training data into a group according to m training data of the time sequence, calculating the gradient value of each training data in the first group of training data by using a small batch random gradient descent algorithm, and then carrying out weighted average summation on q gradient values to obtain the descending gradient of the group of training data
Figure BDA0001469418970000077
i represents the ith set of training data,
Figure BDA0001469418970000078
represents input and output data corresponding to the τ -th training data in the ith group;
s4.4.3, updating parameters in the PM2.5 prediction model by the descending gradient of the training data, wherein the parameter updating formula is as follows:
Figure BDA0001469418970000081
wherein, thetai-1Represents the target parameter theta after the training of the last group of data is finishediRepresenting target parameters after the training data of the group is finished, and representing the learning rate by eta;
in the present embodiment, in combination with the embodiment in step S3, when the PM2.5 prediction model trains parameters using a small batch stochastic gradient descent algorithm, each subset MiIn (m-K +1) iterations, each using K pieces of data, for
Figure BDA0001469418970000082
Deriving to obtain the gradient of each parameter, performing weighted average summation on m-k +1 gradients to serve as a descending gradient of one-time small-batch training, and then updating the parameters, specifically:
Figure BDA0001469418970000083
s4.4.4, when the target parameters of the training data are updated, returning to step S4.4.2 to train and update the next training data group until the gradient of the descent is lower than the set expected error value or the training of the last training data group is completed, and then updating and storing the final parameters to obtain the PM2.5 prediction model after training;
s5, judging whether the PM2.5 prediction model reaches the training stop condition or not
Inputting data of the test set into K groups of data according to a time sequence to a trained PM2.5 prediction model, outputting T predicted values, judging the error between each predicted value and a true value, if the error is within an allowable range, considering that the prediction model completes training, otherwise, returning to the step S4 to retrain until a stop condition is reached;
s6, PM2.5 prediction is carried out by utilizing PM2.5 prediction model
And inputting the current K groups of weather data into a PM2.5 prediction model, and outputting T PM2.5 prediction values.
Although illustrative embodiments of the present invention have been described above to facilitate the understanding of the present invention by those skilled in the art, it should be understood that the present invention is not limited to the scope of the embodiments, and various changes may be made apparent to those skilled in the art as long as they are within the spirit and scope of the present invention as defined and defined by the appended claims, and all matters of the invention which utilize the inventive concepts are protected.

Claims (1)

1.一种基于深层结构循环神经网络的PM2.5预测方法，其特征在于，包括以下步骤：1. a PM2.5 prediction method based on deep structure recurrent neural network, is characterized in that, comprises the following steps: (1)、获取历史天气数据，包括每小时的温度，光照，风速，降雨量，SO2，O3，NO，PM10，PM2.5数据指标，其中，温度单位：℃，光照单位：lm/㎡，风速单位：m/s，降雨量单位：mm，SO2，O3，NO，PM10，PM2.5均是浓度数据；(1) Obtain historical weather data, including hourly temperature, light, wind speed, rainfall, SO2, O3, NO, PM10, PM2.5 data indicators, among which, temperature unit: °C, light unit: lm/㎡, Wind speed unit: m/s, rainfall unit: mm, SO2, O3, NO, PM10, PM2.5 are concentration data; (2)、数据预处理(2), data preprocessing (2.1)、对缺失历史天气数据进行补全处理(2.1) Completion of missing historical weather data 利用均值法补全缺失的历史天气数据：Use the mean method to fill in missing historical weather data:
Figure FDA0003000957500000011
Figure FDA0003000957500000011
 其中，Xt表示当前时刻的缺失历史天气数据，Xt-1表示前一时刻的缺失历史天气数据，Xt+1表示前后一时刻的缺失历史天气数据；Among them, X t represents the missing historical weather data at the current moment, X t-1 represents the missing historical weather data at the previous moment, and X t+1 represents the missing historical weather data at the previous moment; (2.2)、对所有历史天气数据进行归一化处理(2.2), normalize all historical weather data 按照如下公式将所以历史天气数据归一化到-1～1之间；Normalize all historical weather data to between -1 and 1 according to the following formula;
Figure FDA0003000957500000012
Figure FDA0003000957500000012
 其中，X'表示归一化后的历史天气数据，X表示归一化前的历史天气数据，
Figure FDA0003000957500000013
表示历史天气数据均值，Xmax表示历史天气数据最大值，Xmin表示历史天气数据最小值；Among them, X' represents the normalized historical weather data, X represents the historical weather data before normalization,
Figure FDA0003000957500000013
Represents the mean value of historical weather data, X max represents the maximum value of historical weather data, and X min represents the minimum value of historical weather data; (3)、将预处理完成后的历史天气数据按照比例分为训练数据和测试数据；(3), divide the historical weather data after preprocessing into training data and test data according to the proportion; (4)、基于深度学习理论和循环神经网络构造深层结构的PM2.5预测模型(4) PM2.5 prediction model with deep structure based on deep learning theory and recurrent neural network (4.1)、构建深层循环神经网络预测模型：一层输入层，多层隐藏层，一层输入层，模型深度大于N层，输入是训练数据，输出是PM2.5浓度的预测值；(4.1), build a deep recurrent neural network prediction model: one input layer, multiple hidden layers, one input layer, the model depth is greater than N layers, the input is the training data, and the output is the predicted value of PM2.5 concentration; (4.2)、设输入层维度为K×(H-1)，输出层维度为1×T，输入层与隐藏层，隐藏层与隐藏层，隐藏层和输出层的激活函数采用Tanh函数；(4.2), set the dimension of the input layer as K×(H-1), the dimension of the output layer as 1×T, the input layer and the hidden layer, the hidden layer and the hidden layer, and the activation function of the hidden layer and the output layer using the Tanh function; 其中，K表示循环神经网络按时间序列展开的深度，即K个时间帧，每一个时间帧输入一组历史天气数据；H表示数据指标数目，T表示循环神经网络的预测模型输出数据个数，表示用K条历史数据预测未来T个时刻的PM2.5浓度，即为了输入前K个时刻的天气数据，预测出之后T个时刻的PM2.5浓度数据；Among them, K represents the depth of the cyclic neural network expansion in time series, that is, K time frames, each time frame inputs a set of historical weather data; H represents the number of data indicators, T represents the number of output data of the prediction model of the cyclic neural network, Indicates that K pieces of historical data are used to predict the PM2.5 concentration at T times in the future, that is, in order to input the weather data at the first K times, the PM2.5 concentration data at the next T times are predicted; (4.3)、选择PM2.5预测模型中使用的损失函数(4.3), select the loss function used in the PM2.5 prediction model 在PM2.5预测模型中采用均方误差作为损失函数：The mean squared error is used as the loss function in the PM2.5 prediction model:
Figure FDA0003000957500000021
Figure FDA0003000957500000021
 其中，q表示迭代次数，t为输出向量维度，yi,j表示训练数据的真实值，yi,j'表示训练数据的预测值；Among them, q represents the number of iterations, t is the dimension of the output vector, y i,j represents the actual value of the training data, and y i,j ' represents the predicted value of the training data; (4.4)、采用小批量随机梯度下降算法更新PM2.5预测模型中的参数(4.4), use the mini-batch stochastic gradient descent algorithm to update the parameters in the PM2.5 prediction model (4.4.1)、初始化参数θ0；(4.4.1), initialization parameter θ 0 ; (4.4.2)、将训练数据按照时间序列每m个训练数据分为一组，再利用小批量随机梯度下降算法计算第一组训练数据中每个训练数据的梯度值，然后梯度值进行加权平均求和,得到本组训练数据的下降梯度
Figure FDA0003000957500000022
i表示第i组训练数据，
Figure FDA0003000957500000023
表示第i组中第τ个训练数据对应的输入、输出数据；(4.4.2) Divide the training data into one group for every m training data in the time series, and then use the mini-batch stochastic gradient descent algorithm to calculate the gradient value of each training data in the first group of training data, and then weight the gradient value Average summation to get the descending gradient of this set of training data
Figure FDA0003000957500000022
i represents the i-th group of training data,
Figure FDA0003000957500000023
represents the input and output data corresponding to the τth training data in the i-th group; (4.4.3)、本组训练数据的下降梯度更新PM2.5预测模型中的参数，参数更新公式为：(4.4.3), the descending gradient of this group of training data updates the parameters in the PM2.5 prediction model, and the parameter update formula is:
Figure FDA0003000957500000024
Figure FDA0003000957500000024
 其中，θi-1表示上一组数据训练完成后的目标参数，θi表示本组训练数据完成后的目标参数，η表示学习率；Among them, θ i-1 represents the target parameter after the previous set of data training is completed, θ i represents the target parameter after the completion of this set of training data, and η represents the learning rate; (4.4.4)、当本组训练数据完成后的目标参数更新完成后，返回步骤(4.4.2)进行下一组训练数据的训练及更新，直到误差值低于设定期望误差值或者最后一组训练数据训练完成时结束，然后更新并保存最终参数，得到训练完成的PM2.5预测模型；(4.4.4), when the target parameter update after the completion of this group of training data is completed, return to step (4.4.2) to train and update the next group of training data, until the error value is lower than the set expected error value or the last A set of training data is finished when the training is completed, and then the final parameters are updated and saved, and the trained PM2.5 prediction model is obtained; (5)、判断PM2.5预测模型的是否达到训练停止条件(5), determine whether the PM2.5 prediction model reaches the training stop condition 将测试集数据按照时间序列输入一个K组数据至已经训练好的PM2.5预测模型中，输出T个预测值，再将每个预测值和真实值之间进行误差判断，如果误差在允许范围内，则认为预测模型完成训练，否则返回步骤(4)重新训练，直到达到停止条件；Input the test set data into a K group of data according to the time series into the PM2.5 prediction model that has been trained, output T predicted values, and then judge the error between each predicted value and the actual value, if the error is within the allowable range within, the prediction model is considered to have completed the training, otherwise it returns to step (4) for retraining until the stopping condition is reached; (6)、利用PM2.5预测模型进行PM2.5的预测(6) Use the PM2.5 prediction model to predict PM2.5 将当前K组天气数据输入至PM2.5预测模型，输出T个PM2.5预测值。Input the current K groups of weather data into the PM2.5 prediction model, and output T PM2.5 prediction values.
CN201711130537.XA 2017-11-15 2017-11-15 PM2.5 prediction method based on deep structure recurrent neural network Active CN107909206B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201711130537.XA CN107909206B (en) 2017-11-15 2017-11-15 PM2.5 prediction method based on deep structure recurrent neural network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201711130537.XA CN107909206B (en) 2017-11-15 2017-11-15 PM2.5 prediction method based on deep structure recurrent neural network

Publications (2)

Publication Number Publication Date
CN107909206A CN107909206A (en) 2018-04-13
CN107909206B true CN107909206B (en) 2021-06-04

Family

ID=61845600

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201711130537.XA Active CN107909206B (en) 2017-11-15 2017-11-15 PM2.5 prediction method based on deep structure recurrent neural network

Country Status (1)

Country Link
CN (1) CN107909206B (en)

Families Citing this family (15)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108564326B (en) * 2018-04-19 2021-12-21 安吉汽车物流股份有限公司 Order prediction method and device, computer readable medium and logistics system
CN108957418A (en) * 2018-05-30 2018-12-07 西安电子科技大学 A kind of radar target identification method based on Recognition with Recurrent Neural Network model
CN109242166B (en) * 2018-08-25 2021-05-04 臻儒(天津)科技有限公司 Environmental prediction prevention and control system based on multi-scale deep neural network
CN109214592B (en) * 2018-10-17 2022-03-08 北京工商大学 Multi-model-fused deep learning air quality prediction method
CN109613178A (en) * 2018-11-05 2019-04-12 广东奥博信息产业股份有限公司 A kind of method and system based on recurrent neural networks prediction air pollution
CN109447373A (en) * 2018-11-16 2019-03-08 上海海事大学 Haze method is predicted based on the LSTM neural network of python platform
CN111292523B (en) * 2018-12-06 2023-04-07 中国信息通信科技集团有限公司 Network intelligent system
CN109934130A (en) * 2019-02-28 2019-06-25 中国空间技术研究院 Real-time fault diagnosis method and system for satellite faults in orbit based on deep learning
CN110059082A (en) * 2019-04-17 2019-07-26 东南大学 A kind of weather prediction method based on 1D-CNN and Bi-LSTM
CN110334382B (en) * 2019-05-09 2020-11-17 电子科技大学 Automobile window fogging condition prediction method based on cyclic neural network
CN110543931A (en) * 2019-07-26 2019-12-06 浙江工业大学 A prediction method of PM2.5 concentration value based on time correlation network
CN110766219A (en) * 2019-10-21 2020-02-07 成都理工大学工程技术学院 Haze prediction method based on BP neural network
CN111103220B (en) * 2019-12-31 2021-04-27 西安交通大学 A method and system for predicting and regulating the concentration of air pollutants
CN112990531B (en) * 2020-12-23 2021-11-19 山西大学 Haze prediction method based on feature-enhanced ConvLSTM
CN112884243B (en) * 2021-03-16 2025-01-10 杭州职业技术学院 An air quality analysis and prediction method based on deep learning and Bayesian model

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106384166A (en) * 2016-09-12 2017-02-08 中山大学 Deep learning stock market prediction method combined with financial news
CN106599520A (en) * 2016-12-31 2017-04-26 中国科学技术大学 LSTM-RNN model-based air pollutant concentration forecast method
CN106845371A (en) * 2016-12-31 2017-06-13 中国科学技术大学 A kind of city road network automotive emission remote sensing monitoring system
CN106952161A (en) * 2017-03-31 2017-07-14 洪志令 A kind of recent forward prediction method of stock based on shot and long term memory depth learning network
CN107239859A (en) * 2017-06-05 2017-10-10 国网山东省电力公司电力科学研究院 The heating load forecasting method of Recognition with Recurrent Neural Network is remembered based on series connection shot and long term

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106384166A (en) * 2016-09-12 2017-02-08 中山大学 Deep learning stock market prediction method combined with financial news
CN106599520A (en) * 2016-12-31 2017-04-26 中国科学技术大学 LSTM-RNN model-based air pollutant concentration forecast method
CN106845371A (en) * 2016-12-31 2017-06-13 中国科学技术大学 A kind of city road network automotive emission remote sensing monitoring system
CN106952161A (en) * 2017-03-31 2017-07-14 洪志令 A kind of recent forward prediction method of stock based on shot and long term memory depth learning network
CN107239859A (en) * 2017-06-05 2017-10-10 国网山东省电力公司电力科学研究院 The heating load forecasting method of Recognition with Recurrent Neural Network is remembered based on series connection shot and long term

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
Influence of Social-economic Activities on Air Pollutants in Beijing, China;Xiaolu Li 等;《Open Geosciences》;20170831;第9卷(第1期);314-321 *
基于BP人工神经网络的雾霾天气预测研究;艾洪福 等;《计算机仿真》;20150402;第32卷(第1期);402-405 *
基于RNN的空气污染时空预报模型研究;范竣翔 等;《测绘科学》;20170731;第42卷(第7期);76-83 *

Also Published As

Publication number Publication date
CN107909206A (en) 2018-04-13

Similar Documents

Publication Publication Date Title
CN107909206B (en) PM2.5 prediction method based on deep structure recurrent neural network
CN108764540B (en) Water supply network pressure prediction method based on parallel LSTM series DNN
CN112085163A (en) An air quality prediction method based on attention-enhanced graph convolutional neural network (AGC) and gated recurrent unit (GRU)
CN108009674A (en) Air PM2.5 concentration prediction methods based on CNN and LSTM fused neural networks
CN113283588B (en) Near-shore single-point wave height forecasting method based on deep learning
CN109726802B (en) A Machine Learning Prediction Method for Environmental Wind Speed of Railways and Wind Farms
CN109583565A (en) Forecasting Flood method based on the long memory network in short-term of attention model
CN112489497B (en) Airspace operation complexity evaluation method based on deep convolutional neural network
CN112001110A (en) Structural damage identification monitoring method based on vibration signal space real-time recursive graph convolutional neural network
CN112242060A (en) Traffic flow prediction method and device, computer equipment and readable storage medium
CN111860787A (en) A short-term prediction method and device for coupled directed graph structure traffic data containing missing data
CN107045785A (en) A kind of method of the short-term traffic flow forecast based on grey ELM neutral nets
CN110289987B (en) Multi-agent system network anti-attack capability assessment method based on characterization learning
CN107274030A (en) Runoff Forecast method and system based on hydrology variable year border and monthly variation characteristic
CN109725531A (en) A Continuous Learning Method Based on Gate Control Mechanism
CN112465199A (en) Airspace situation evaluation system
CN108628164A (en) A kind of semi-supervised flexible measurement method of industrial process based on Recognition with Recurrent Neural Network model
CN115359338A (en) Sea surface temperature prediction method and system based on hybrid learning model
CN113988210A (en) Distorted data restoration method, device and storage medium for structural monitoring sensor network
CN107977748A (en) A Multivariate Distorted Time Series Forecasting Method
CN113947182A (en) Traffic flow prediction model construction method based on double-stage stack graph convolution network
CN115545334B (en) Land utilization type prediction method and device, electronic equipment and storage medium
CN107563508A (en) A kind of big data failure prediction method
CN115392387B (en) A method for forecasting output of low-voltage distributed photovoltaic power generation
CN114282614B (en) Medium-long runoff prediction method for optimizing CNN-GRU based on random forest and IFDA

Legal Events

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