CN110543931A - PM2.5 concentration value prediction method based on time correlation network - Google Patents
PM2.5 concentration value prediction method based on time correlation network Download PDFInfo
- Publication number
- CN110543931A CN110543931A CN201910680968.6A CN201910680968A CN110543931A CN 110543931 A CN110543931 A CN 110543931A CN 201910680968 A CN201910680968 A CN 201910680968A CN 110543931 A CN110543931 A CN 110543931A
- Authority
- CN
- China
- Prior art keywords
- time
- data
- concentration value
- network
- 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.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
a PM2.5 concentration value prediction method based on a time correlation network is characterized in that non-linear correlation analysis is carried out on PM2.5 concentration value historical data, historical data of relevant indexes of PM2.5 concentration values and historical data of meteorological data, month, date and time label data are subjected to non-differential quantization according to a proposed coding mode, and time change rules of the PM2.5 concentration values can be accurately described aiming at time characteristics of sequence data of periodic sampling by constructing an input layer and an output layer which are correlated with each other. The method can fully learn the time sequence characteristics of the PM2.5 concentration data, can deeply mine the time characteristics on the data acquisition period, effectively improve the prediction precision and the training speed of the current PM2.5 concentration value, widen the limitation of neural network prediction, and realize the accurate prediction of the periodically acquired long-term data.
Description
Technical Field
the invention relates to the technical field of prediction of PM2.5 concentration values of air particulate matters, in particular to a PM2.5 concentration value prediction method based on a time correlation network.
Background
PM2.5 refers to particulate matters with the diameter less than or equal to 2.5 microns in the atmosphere, is rich in a large amount of toxic and harmful substances, has long retention time in the atmosphere and long conveying distance, so that the influence on human health and atmospheric environment quality is greater, and another influence, namely dust haze weather, is caused when PM2.5 exceeds the standard. Air pollution is now the focus of attention, and among air pollution indexes, the PM2.5 concentration value is a symbolic detection index for measuring air quality. Nowadays, prediction of the concentration value of PM2.5 in a future time period according to historical data becomes a research problem with strong academic significance and application value.
in order to solve the above problem, wanmin et al, in the article "urban PM2.5 concentration spatial prediction based on BP artificial neural network", apply a BP neural network model to predict the PM2.5 concentration. Yang Qikai et al in the article PM2.5 evolution model based on genetic algorithm and BP neural network, utilize the genetic algorithm to improve the BP neural network and simulate the generation and prediction of PM2.5 concentration. In the 'PM 2.5 forecasting method based on ensemble learning', Li Xiang et al constructs a plurality of weak learning machines by selecting neural networks of different types and structures, and then synthesizes the weak learning machines into a strong learning machine by using an ensemble learning AdaBoost algorithm to complete PM2.5 forecasting work. Jojunfei et al put forward a PM2.5 prediction method based on a T-S fuzzy neural network in a paper PM2.5 prediction research based on the T-S fuzzy neural network. Su Ying et al in the patent PM2.5 concentration prediction method based on unscented Kalman neural network provides a PM2.5 concentration prediction method based on unscented Kalman neural network, aiming at the characteristic of nonlinear dynamic change of PM2.5 concentration. Houjunxi et al in the article PM2.5 concentration prediction method of threshold repeat unit, proposed a PM2.5 concentration real-time prediction method based on long-short term memory neural network. The leucin et al in the article "PM 2.5 prediction based on LSTM recurrent neural network" put forward a PM2.5 prediction model based on LSTM recurrent neural network.
Through literature research and analysis, currently proposed methods for predicting PM2.5 concentration values all use a neural network as a core architecture to perform nonlinear regression analysis on PM2.5 concentration values and other related indexes (such as AQI, PM10, NO2, CO, SO2 and O3). The neural network model comprises ANN, DNN, FNN, BPNN and the like, and a mixing method after optimization by combining genetic algorithm, random forest and other optimization algorithms. Furthermore, learners have begun to learn the time-series characteristics of the PM2.5 concentration data by using a method such as a recurrent neural network. However, according to the research and study of the literature, when the time-series characteristics of the PM2.5 concentration data are learned, the conventional PM2.5 concentration value neural network prediction method only performs training learning based on the PM2.5 concentration historical data, the historical weather data and the like only by mining based on the nonlinear relationship, and cannot perform targeted learning on the time-series characteristics of the same period data, so that the prediction accuracy is insufficient and the convergence rate is too slow.
disclosure of Invention
in order to solve the problems that the time sequence characteristics of the PM2.5 concentration value data cannot be subjected to targeted learning when the time sequence characteristics of the PM2.5 concentration value data are learned by the conventional PM2.5 concentration value prediction mode, so that the prediction precision is insufficient, the convergence speed is too slow and the like, the method introduces month, date and time label data in addition to the PM2.5 concentration value historical data, the PM2.5 concentration value related index historical data and the weather historical data, performs undifferentiated quantization according to the proposed coding mode, and then performs learning training by using a time association network, so that the time change rule of the PM2.5 concentration value can be accurately described.
The technical scheme adopted by the invention for solving the technical problems is as follows:
A PM2.5 concentration value prediction method based on a time-of-day correlation network comprises the following steps:
step 1, collecting original data, wherein the original data comprises PM2.5 concentration value historical data, PM2.5 concentration value index historical data, meteorological historical data, month data, date data and time data;
step 2, carrying out undifferentiated quantization processing on the month, date and time data by adopting an encoding mode, wherein the process is as follows:
step 2.1, for 12 months, 12 encodings are used to represent 12 feature values: month 1 is (0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1), month 2 is (0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,1, 0), month 3 is (0, 0, 0, 0, 0, 0, 0, 0,1, 0, 0), and so on;
step 2.2, as shown in step 2.1, 31 codes and 24 codes are respectively adopted for 31 dates and 24 moments to obtain respective characteristic values;
Step 3, training by adopting a time association network, wherein the process is as follows:
step 3.1, creating a three-layer neural network comprising an input layer, a hidden layer and an output layer, and setting the number of nodes of the hidden layer and the output layer, wherein the number of the nodes of the hidden layer adopts an empirical formula to give an estimated value, and the empirical formula is as follows:
In the above formula, a and b are the numbers of neurons of the input layer and the output layer, respectively, and c is a constant between [0 and 10 ];
step 3.2, respectively setting the training functions, the connection functions and the output functions of the hidden layer, the connection layer and the output layer:
h＝σ(z)＝σ(Ux+Wh+b) (2)
o＝Vh+c (3)
wherein t is 0,1,2, …, n +1 represents time, and x (t) represents input of training sample at time t; h (t) represents the hidden state of the model at the time t, and h (t) is jointly determined by x (t) and h (t-1); o (t) represents the output of the model at time t, and o (t) is determined only by the current hidden state h (t) of the model; y (t) represents the true output of the training sample sequence at time t; the three matrixes of U, W and V are linear relation parameters of the model;
step 3.3, setting the minimum value of the expected error, the maximum iteration times and the learning rate of the network;
step 3.4, training the recurrent neural network, and calculating a loss function:
and 3.5, adjusting the weight of each layer of the recurrent neural network according to the error, wherein the gradient of V and c is calculated as follows:
In the backward propagation, the gradient loss at a certain sequence position n is determined by the gradient loss corresponding to the output of the current position and the gradient loss at a sequence index position n +1, the gradient loss at a certain sequence position n for W needs to be calculated in a step-by-step manner by backward propagation, and the gradient of the hidden state at the sequence index position n +1 is defined as:
gradient calculation expression of W, U, b:
c and b are offset values;
step 3.6, judging whether the circulating network is converged, when the error is smaller than the minimum value of the expected error, the algorithm is converged, and when the maximum iteration times is reached, the algorithm is ended, and the moment associated network training is finished;
and 3.7, inputting the acquired data and the time data into the trained time correlation network, and outputting a final predicted value of the PM2.5 concentration value.
further, the PM2.5 concentration value indexes include AQI, PM10, NO2, CO, SO2, and O3 concentrations, and the time data includes month, date, and time.
in the present invention, in the step 2, the temporal information is subjected to the undifferentiated quantization processing by the coding method.
in the step 3, a time association network is constructed through association processing of the input dimension and the output dimension, and time sequence training and prediction are performed on data.
The technical conception of the invention is as follows: besides carrying out nonlinear correlation analysis on historical data of PM2.5 concentration values, historical data (AQI, PM10, NO2, CO, SO2 and O3) of indexes related to PM2.5 concentration values and historical data of weather data (air temperature, relative humidity, air pressure, air speed, precipitation and the like), date and time label data are subjected to undifferentiated quantization according to the proposed coding mode, and an input layer and an output layer which are related in time sequence are constructed, SO that the PM2.5 concentration value prediction method based on the time-related network, which aims at the time characteristic of sequence data sampled periodically and can accurately describe the time change rule of the PM2.5 concentration values, is provided.
The invention has the following beneficial effects: according to the technical scheme, the time sequence characteristics of the PM2.5 concentration data can be fully learned, the time characteristics on the data acquisition period can be deeply mined, the prediction precision and the training speed of the current PM2.5 concentration value are effectively improved, the limitation of neural network prediction is widened, and the long-term data accurate prediction of periodic acquisition can be realized.
drawings
fig. 1 is a schematic diagram of a method for predicting a PM2.5 concentration value based on a time-of-day correlation network.
Fig. 2 is a network training flow diagram of a time of day associative network.
Detailed Description
the invention is further described below with reference to the accompanying drawings.
referring to fig. 1 to 2, a method for predicting a PM2.5 concentration value based on a time-of-day correlation network includes the following steps:
step 1, collecting original data, wherein the original data comprises PM2.5 concentration value historical data, PM2.5 concentration value index historical data, meteorological historical data, month data, date data and time data;
step 2, carrying out undifferentiated quantization processing on the month, date and time data by adopting an encoding mode, wherein the process is as follows:
Step 2.1, for 12 months, 12 encodings are used to represent 12 feature values: (0, 0, 0, 0, 0, 0, 0, 0, 1), 2 (0, 0, 0, 0, 0, 0, 0, 0,1, 0), 3 (0, 0, 0, 0, 0, 0, 0, 4 (0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,1, 0, 0, 0), 5 (0, 0, 0, 0, 0, 0, 0, 0,1, 0, 0, 0, 0, 0, 6 (0, 0, 0, 0, 0, 0, 0, 0,1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 7 (0, 0, 0, 0, 0, 0, 0, 0, 0, 0) month 10 ═ 0, 0,1, 0, 0, 0, 0, 0, 0, 0, 0), month 11 ═ 0,1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 12 ═ 1, 0, 0, 0, 0, 0, 0, 0, 0;
Step 2.2, as shown in step 2.1, 31 codes and 24 codes are respectively adopted for 31 dates and 24 moments to obtain respective characteristic values;
step 3, training by adopting a time association network, wherein the process is as follows:
step 3.1, creating a three-layer neural network comprising an input layer, a hidden layer and an output layer, and setting the number of nodes of the hidden layer and the output layer, wherein the number of the nodes of the hidden layer adopts an empirical formula to give an estimated value, and the empirical formula is as follows:
In the above formula, a and b are the numbers of neurons of the input layer and the output layer, respectively, and c is a constant between [0 and 10 ];
Step 3.2, respectively setting the training functions, the connection functions and the output functions of the hidden layer, the connection layer and the output layer:
h＝σ(z)＝σ(Ux+Wh+b) (2)
o＝Vh+c (3)
Wherein t is 0,1,2, …, n +1 represents time, and x (t) represents input of training sample at time t; h (t) represents the hidden state of the model at the time t, and h (t) is jointly determined by x (t) and h (t-1); o (t) represents the output of the model at time t, and o (t) is determined only by the current hidden state h (t) of the model; y (t) represents the true output of the training sample sequence at time t; the three matrixes of U, W and V are linear relation parameters of the model;
and 3.3, setting the minimum value of the expected error of the network, the maximum iteration times and the learning rate.
Step 3.4, training the recurrent neural network, and calculating a loss function:
and 3.5, adjusting the weight of each layer of the recurrent neural network according to the error, wherein the gradient of V and c is calculated as follows:
in the backward propagation, the gradient loss at a certain sequence position n is determined by the gradient loss corresponding to the output of the current position and the gradient loss at a sequence index position n +1, the gradient loss at a certain sequence position n for W needs to be calculated in a step-by-step manner by backward propagation, and the gradient of the hidden state at the sequence index position n +1 is defined as:
gradient calculation expression of W, U, b:
c and b are offset values;
step 3.6, judging whether the circulating network is converged, when the error is smaller than the minimum value of the expected error, the algorithm is converged, and when the maximum iteration times is reached, the algorithm is ended, and the moment associated network training is finished;
and 3.7, inputting the acquired data and the time data into the trained time correlation network, and outputting a final predicted value of the PM2.5 concentration value.
Further, the PM2.5 concentration value indicators include AQI, PM10, NO2, CO, SO2, and O3 concentrations. The time data includes month, date and time.
in the present invention, in the step 2, the temporal information is subjected to the undifferentiated quantization processing by the coding method.
in the step 3, a time association network is constructed through association processing of the input dimension and the output dimension, and time sequence training and prediction are performed on data.
Claims (4)
1. A PM2.5 concentration value prediction method based on a time-of-day correlation network is characterized by comprising the following steps:
Step 1, collecting original data, wherein the original data comprises PM2.5 concentration value historical data, PM2.5 concentration value index historical data, meteorological historical data, month data, date data and time data;
Step 2, carrying out undifferentiated quantization processing on the month, date and time data by adopting an encoding mode, wherein the process is as follows:
step 2.1, for 12 months, adopting 12 codes to represent 12 characteristic values;
step 2.2, as shown in step 2.1, 31 codes and 24 codes are respectively adopted for 31 dates and 24 moments to obtain respective characteristic values;
Step 3, training by adopting a time association network, wherein the process is as follows:
Step 3.1, creating a three-layer neural network comprising an input layer, a hidden layer and an output layer, and setting the number of nodes of the hidden layer and the output layer, wherein the number of the nodes of the hidden layer adopts an empirical formula to give an estimated value, and the empirical formula is as follows:
in the above formula, a and b are the numbers of neurons of the input layer and the output layer, respectively, and c is a constant between [0 and 10 ];
step 3.2, respectively setting the training functions, the connection functions and the output functions of the hidden layer, the connection layer and the output layer:
h＝σ(z)＝σ(Ux+Wh+b) (2)
o＝Vh+c (3)
wherein t is 0,1,2, …, n +1 represents time, and x (t) represents input of training sample at time t; h (t) represents the hidden state of the model at the time t, and h (t) is jointly determined by x (t) and h (t-1); o (t) represents the output of the model at time t, and o (t) is determined only by the current hidden state h (t) of the model; y (t) represents the true output of the training sample sequence at time t; the three matrixes of U, W and V are linear relation parameters of the model;
Step 3.3, setting the minimum value of the expected error, the maximum iteration times and the learning rate of the network;
Step 3.4, training the recurrent neural network, and calculating a loss function:
and 3.5, adjusting the weight of each layer of the recurrent neural network according to the error, wherein the gradient of V and c is calculated as follows:
in the backward propagation, the gradient loss at a certain sequence position n is determined by the gradient loss corresponding to the output of the current position and the gradient loss at a sequence index position n +1, the gradient loss at a certain sequence position n for W needs to be calculated in a step-by-step manner by backward propagation, and the gradient of the hidden state at the sequence index position n +1 is defined as:
gradient calculation expression of W, U, b:
c and b are offset values;
step 3.6, judging whether the circulating network is converged, when the error is smaller than the minimum value of the expected error, the algorithm is converged, and when the maximum iteration times is reached, the algorithm is ended, and the moment associated network training is finished;
And 3.7, inputting the acquired data and the time data into the trained time correlation network, and outputting a final predicted value of the PM2.5 concentration value.
2. the PM2.5 concentration value prediction method based on the time of day correlation network as claimed in claim 1, wherein said PM2.5 concentration value indicators include AQI, PM10, NO2, CO, SO2 and O3 concentrations, and said time of day data includes month, date and time of day.
3. The method according to claim 1 or 2, wherein in step 2, the time information is encoded to be quantized indifferently.
4. the method for predicting the concentration value of PM2.5 based on the time-of-day correlation network as claimed in claim 1 or 2, wherein in the step 3, the time-of-day correlation network is constructed through the correlation processing of the input dimension and the output dimension, and the data is subjected to time sequence training and prediction.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910680968.6A CN110543931A (en) | 2019-07-26 | 2019-07-26 | PM2.5 concentration value prediction method based on time correlation network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910680968.6A CN110543931A (en) | 2019-07-26 | 2019-07-26 | PM2.5 concentration value prediction method based on time correlation network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110543931A true CN110543931A (en) | 2019-12-06 |
Family
ID=68710334
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910680968.6A Pending CN110543931A (en) | 2019-07-26 | 2019-07-26 | PM2.5 concentration value prediction method based on time correlation network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110543931A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112649337A (en) * | 2020-12-21 | 2021-04-13 | 张家口市杰星电子科技有限公司 | Oil smoke online monitoring method and device |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106228030A (en) * | 2016-08-31 | 2016-12-14 | 广东旭诚科技有限公司 | A kind of fine granulation self study integrated prediction method based on wavelet transformation |
| CN107909206A (en) * | 2017-11-15 | 2018-04-13 | 电子科技大学 | A kind of PM2.5 Forecasting Methodologies based on deep structure Recognition with Recurrent Neural Network |
| CN108268935A (en) * | 2018-01-11 | 2018-07-10 | 浙江工业大学 | A kind of PM2.5 concentration values Forecasting Methodology and system based on sequential Recognition with Recurrent Neural Network |
-
2019
- 2019-07-26 CN CN201910680968.6A patent/CN110543931A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106228030A (en) * | 2016-08-31 | 2016-12-14 | 广东旭诚科技有限公司 | A kind of fine granulation self study integrated prediction method based on wavelet transformation |
| CN107909206A (en) * | 2017-11-15 | 2018-04-13 | 电子科技大学 | A kind of PM2.5 Forecasting Methodologies based on deep structure Recognition with Recurrent Neural Network |
| CN108268935A (en) * | 2018-01-11 | 2018-07-10 | 浙江工业大学 | A kind of PM2.5 concentration values Forecasting Methodology and system based on sequential Recognition with Recurrent Neural Network |
Non-Patent Citations (2)
| Title |
|---|
| JUNXIANG FAN ET.AL: "A Spatiotemporal Prediction Framework for Air Pollution Based on Deep RNN", 《ISPRS ANNALS OF THE PHOTOGRAMMETRY》 * |
| XIANG LI ET.AL: "Long short-term memory neural network for air pollutant concentration predictions: Method development and evaluation", 《ENVIRONMENTAL POLLUTION 》 * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112649337A (en) * | 2020-12-21 | 2021-04-13 | 张家口市杰星电子科技有限公司 | Oil smoke online monitoring method and device |
| CN112649337B (en) * | 2020-12-21 | 2022-06-03 | 张家口市杰星电子科技有限公司 | Oil smoke online monitoring method and device |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109492822B (en) | Air pollutant concentration time-space domain correlation prediction method | |
| CN108426812B (en) | PM2.5 concentration value prediction method based on memory neural network | |
| CN108268935B (en) | PM2.5 concentration value prediction method and system based on time sequence recurrent neural network | |
| CN108009674A (en) | Air PM2.5 concentration prediction methods based on CNN and LSTM fused neural networks | |
| CN112085163A (en) | Air quality prediction method based on attention enhancement graph convolutional neural network AGC and gated cyclic unit GRU | |
| CN110070224A (en) | A kind of Air Quality Forecast method based on multi-step recursive prediction | |
| Huang et al. | A new air quality forecasting model using data mining and artificial neural network | |
| CN113554466B (en) | Short-term electricity consumption prediction model construction method, prediction method and device | |
| CN110555551B (en) | Air quality big data management method and system for smart city | |
| Zhang et al. | Prediction of urban pm 2.5 concentration based on wavelet neural network | |
| CN116013426A (en) | Site ozone concentration prediction method with high space-time resolution | |
| CN108537336A (en) | A kind of Air Quality Forecast method based on deep neural network | |
| CN109376903B (en) | PM2.5 concentration value prediction method based on game neural network | |
| CN110543931A (en) | PM2.5 concentration value prediction method based on time correlation network | |
| CN116341763B (en) | Air quality prediction method | |
| CN112580859A (en) | Haze prediction method based on global attention mechanism | |
| Gomathi et al. | Real time air pollution prediction in urban cities using deep learning algorithms and IoT | |
| CN114648095A (en) | Air quality concentration inversion method based on deep learning | |
| Kang et al. | Research on forecasting method for effluent ammonia nitrogen concentration based on GRA-TCN | |
| CN103678953A (en) | Biological fermentation yield on-line forecasting method based on Bayes combination neural network | |
| Qu et al. | Short-term mathematical prediction model of air quality based on CEEMD-ELM-PSO | |
| CN109840613B (en) | Short-term wind speed prediction method integrating coding and decoding and linear regression | |
| Liu et al. | Polluted gas quantitative detection in multi-gas sensor based on bidirectional long-short term memory network | |
| CN109670646B (en) | PM2.5 concentration value prediction method based on mixed threshold neural network | |
| Cheng et al. | Air quality forecast and evaluation based on long short-term memory network and fuzzy algorithm |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20191206 |
|
| RJ01 | Rejection of invention patent application after publication |