CN111798051A - Air quality space-time prediction method based on long-short term memory neural network - Google Patents

Abstract

The invention discloses an air quality space-time prediction method based on a long-term and short-term memory neural network. The method integrates the particulate matter concentration data, the meteorological data and the gaseous pollutant data of the experimental station and the nearest adjacent station, converts the particulate matter concentration data, the meteorological data and the gaseous pollutant data into a data format for supervision and learning, normalizes the data, and obtains the prediction sequence of the air quality concentration by using the training data of the long-term and short-term memory networks. The method comprises the following steps: s1: acquiring historical air quality data and meteorological data; s2: performing data preprocessing on the historical air quality, including abnormal value elimination, missing value interpolation processing, extraction of particle concentration data of adjacent stations and data normalization; s3: converting the data format from a sequence to a pair of input and output sequences; s4: dividing a data set into a training set and a testing set and initializing various hyper-parameters of the long-term and short-term memory network; s5: the model effect was verified by prediction on the test set. The method can improve the prediction precision of the air quality data.

Description

Air quality space-time prediction method based on long-short term memory neural network

Technical Field

The invention relates to an air quality space-time prediction method based on a long-term and short-term memory neural network, and belongs to the field of air pollution prediction.

Background

In recent years, with the rapid development of society, the pressure on the environment is increasing, and some serious air pollution problems seriously threaten the health of people. Therefore, the requirement of people can not be met by accurately monitoring the air quality, people hope to predict the air quality condition in advance and timely warn and defend, and the threat to the life is maximally relieved. However, the difficulty in predicting the air mass concentration is serious, and the air mass concentration is easily influenced by other factors, such as meteorological factors (temperature, relative humidity, wind speed, precipitation and the like), traffic pollution, industrial emission and the like; in addition, the complex and unstable interactions between the pollutants in the air also present challenges to the prediction of air quality. Furthermore, the air quality prediction is not only limited by the space of the above factors, but also the air quality at the current time is affected by the accumulation of the air quality at the past time in the time dimension. We need to capture both spatial and temporal dependencies.

In recent years, deep learning is more and more widely applied to the fields of artificial intelligence and big data, and urban air quality concentration prediction research based on deep learning is popularized in interdisciplinary research. At present, an air quality prediction model is mainly divided into a classical model of time series prediction and a joint model of space-time prediction, and the VAR can well capture time dependence but cannot process space correlation in the existing time series models such as an Auto-Regressive Integrated Moving Average model (ARIMA for short), a Seasonal ARIMA model (SARIMA for short) and a Vector Auto-Regressive model (VAR for short). Researchers have also proposed a number of joint models, such as CNN and LSTM-based urban PM proposed by Dongming et al_2.5A concentration joint prediction scheme; graph convolution neural network and long-short term memory based PM proposed by Yanlin et al_2.5A space-time prediction hybrid model; a Multi-output LSTM (Deep Multi-output LSTM, DM-LSTM for short) Deep neural network model proposed by Yanlai et al; the new space-time convolution long-short term memory neural network proposed by Congcong et al. The models capture the time-space correlation of air pollution concentration data well, however, most of the models only consider the influence of meteorological factors on the air pollution concentration, but ignore the interaction among pollutants in the air; since the assistance data of all monitored stations is added, interference by unrelated stations can therefore have a negative effect on the model accuracy.

Disclosure of Invention

In order to overcome the defects of the model, the invention provides an air quality space-time prediction method based on a long-short term memory neural network_2.5The concentration data of the 5 sites with the highest concentration sequence data correlation degree can avoid the interference of the sites with small correlation degree to the model accuracy, and the meteorological data and the SO at the same moment are fused₂、NO₂、O₃And CO concentration data is used as the input of a model, then the data format is converted into a supervised learning format, the supervised learning format is input into a long-short-term memory neural network with N hidden layers and a full connection layer for training, the space-time characteristics of the air quality concentration data are extracted, and the data D hours before t is used for [ t, t +1, …, t + N ] of each station]Time of day air pollution data (PM)_2.5、CO、NO₂、O₃、SO₂) The concentration value of (2) is predicted. The specific technical scheme is as follows:

a long-short term memory neural network-based air quality space-time prediction method comprises the following specific steps:

step 1: acquiring historical air quality data and meteorological data;

step 2: performing data preprocessing on the historical air quality, including abnormal value elimination, missing value interpolation processing, extraction of particle concentration data of adjacent stations and data normalization processing;

calculating the distance between two sites according to the Haversene formula recommended by Wikipedia, measuring the linear correlation strength between continuous variables by using a Pearson correlation coefficient, wherein the Haversene formula is as follows:

wherein haversin (·) denotes the distance between stations, R is the radius of the earth, and the average value is 6371km,

indicates the latitude between two points, and Δ λ indicates the difference between the latitudes of two points.

The pearson correlation coefficient calculation formula is as follows:

where r (-) denotes the correlation coefficient of the PM2.5 concentration sequence between sites, Cov (-) is the covariance, and σ (-) is the standard deviation.

And step 3: converting data format, dividing data set. From sequence to input and output sequence pairs; dividing a data set into a training set and a test set;

and 4, step 4: initializing various parameters of a long-short term memory network (LSTM), inputting data of a training set into the long-short term memory neural network for training until the network converges, and adjusting network hyper-parameters through a comparison test to finally obtain optimal parameters;

and 5: and testing the model effect through the prediction of the test set data to obtain the predicted value of each pollution index of the air quality.

The invention adopts Mean Absolute Error (MAE for short) as a loss function, and the calculation formula is as follows:

wherein y isⁱRepresenting the true contaminant concentration value, pⁱRepresenting a predicted contaminant concentration value.

The indexes for measuring the performance of the prediction method are three, namely average absolute Error (MAE), Root Mean Square Error (RMSE) and decision coefficient (R Squared, R for short)²) RMSE and R²The calculation formula of (a) is as follows:

wherein, yⁱRepresenting the true contaminant concentration value, pⁱIs indicative of a predicted concentration value of the contaminant,

mean values of contaminant concentrations are indicated.

The air quality space-time prediction method based on the long-term and short-term memory neural network greatly improves the prediction precision of the air quality concentration.

Drawings

FIG. 1 is a flow chart of a method of the present invention;

FIG. 2 is a model architecture diagram of the global prediction method;

FIG. 3 is a graph of the predicted result for PM 2.5;

FIG. 4 is a graph of the predicted result of NO 2;

FIG. 5 is a graph of predicted results for O3;

FIG. 6 is a graph of the predicted results of CO;

fig. 7 is a diagram of the predicted result of SO 2.

Detailed Description

The invention will be further explained with reference to the drawings.

As shown in fig. 1 and 2, the air quality space-time prediction method based on the long-short term memory neural network of the invention specifically comprises the following steps:

step 1: acquiring historical air quality data and meteorological data;

step 2: performing data preprocessing on the historical air quality, including abnormal value elimination, missing value interpolation processing, extraction of particle concentration data of adjacent stations and data normalization processing;

and step 3: converting data format, dividing data set. From sequence to input and output sequence pairs; dividing a data set into a training set and a test set;

and 4, step 4: initializing various parameters of a long-short term memory network (LSTM), inputting data of a training set into the long-short term memory neural network for training until the network converges, and adjusting network hyper-parameters through a comparison test to finally obtain optimal parameters;

and 5: and testing the model effect through the prediction of the test set data to obtain the predicted value of each pollution index of the air quality.

Further, the step 1 specifically comprises:

step 1.1: obtaining historical air quality data, including PM_2.5Concentration, SO₂Concentration, O₃Concentration, CO concentration and NO₂Concentration 5 data characteristics;

step 1.2: acquiring meteorological data comprising five data characteristics of temperature, dew point, pressure, wind direction and wind speed.

Further, the step 2 specifically includes:

step 2.1: and removing abnormal values. Directly deleting records with obvious abnormal values in the data;

step 2.2: and (5) processing missing values. Filling missing values by adopting an interpolation method;

step 2.3: and (4) extracting the particle concentration data of adjacent stations. Calculating the distance between two sites according to the Haversene formula recommended by Wikipedia and the longitude and latitude of each site, measuring the linear correlation strength between continuous variables by using a Pearson correlation coefficient, selecting data of five adjacent sites with the highest correlation, adding the data into a model, and improving the prediction accuracy by using adjacent sites.

The Haverine formula is as follows:

wherein haversin (·) denotes the distance between stations, R is the radius of the earth, and the average value is 6371km,

indicates the latitude between two points, and Δ λ indicates the difference between the latitudes of two points.

The pearson correlation coefficient calculation formula is as follows:

where r (-) denotes the correlation coefficient of the PM2.5 concentration sequence between sites, Cov (-) is the covariance, and σ (-) is the standard deviation.

The fused time series dataset is denoted R, where the record at each time in R is represented as:

r＝[pm_2.5,near_1,near_2,near_3,near_4,near_5,temp，dew,press,wnd_dir,wnd_spd,SO₂,NO₂,O₃,CO]。

wherein pm_2.5Near _1, near _2, near _3, near _4 and near _5 represent the concentrations of the current site and the adjacent five sites with the highest correlation respectively; temp, ew, press, wnd _ dir and wnd _ spd respectively represent data of temperature, dew point, pressure, wind direction and wind speed at the same time; SO (SO)₂,NO₂,O₃CO represents the respective concentration value;

step 2.4: and (6) data normalization processing. Data is scaled to [0,1 ].

Further, the step 3 specifically includes:

step 3.1: and converting the data format. The raw data sequence is converted into a supervised learning sequence format, from a time series to pairs of input and output sequences. The time series used by the model is a multivariate time series that is used for sequence prediction by specifying the length of the input and output sequences. Assigning D as an input time step and N as an output time step to the model, and predicting values at t, t +1, … and t + N moments through all data characteristics at t-D, t-D +1, … and t-1;

step 3.2: the data set is partitioned. The historical time series data set is as follows 7: the scale of 3 is divided into a training set and a test set.

Further, the step 4 specifically includes:

step 4.1: the LSTM structure comprises an input layer, a hidden layer and a full connection layer (output layer). Setting a maximum epoch, the number of hidden layers and the number of neurons in each layer, and setting the number of complete connection layers and neurons thereof;

step 4.2: initializing the cell state and the hidden layer state of the LSTM neural unit;

step 4.3: calculating the weights of an input gate, a forgetting gate and an output gate of the current neuron and the current memory candidate value;

step 4.4: calculating the hidden state and the memory state of the current neuron and transmitting the hidden state and the memory state to the next neuron;

step 4.5: adopting MAE as a loss function and Adam as an optimization algorithm, and circularly reciprocating until the model converges;

step 4.6: and comparing and adjusting the parameters, and acquiring and storing the optimal parameters. Including maximum epoch, neuron number, learning rate, mini-batch size, L2 regularization coefficients, weight vectors, and bias vectors.

Further, the step 5 specifically includes:

step 5.1: forecasting air pollution indexes on a test set by using a trained forecasting model, including PM_2.5、SO₂、O₃、CO、NO₂These 5 contamination indicators;

step 5.2: comparing all the predicted results with the true values, and using MAE, RMSE, R²As an evaluation index of the model, obtainingAnd (5) evaluating the performance of the model.

Step 5.3: the final prediction model may be used for prediction of the actual air mass concentration.

The final model results are shown in table 1, and the prediction trend graphs are shown in fig. 3-7, so that the method of the invention realizes more accurate prediction effect.

TABLE 1 prediction results table

In summary, the above description is only a preferred embodiment of the present invention, and is not intended to limit the scope of the present invention. Any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims (6)

1. The air quality space-time prediction method based on the long-short term memory neural network is characterized by comprising the following steps of:

step 1: acquiring historical air quality data and meteorological data;

step 2: performing data preprocessing on the historical air quality, including abnormal value elimination, missing value interpolation processing, extraction of particle concentration data of adjacent stations and data normalization processing;

and step 3: converting data format and dividing data set; from sequence to input and output sequence pairs; dividing a data set into a training set and a test set;

and 4, step 4: initializing various parameters of a long-short term memory network (LSTM), inputting data of a training set into the long-short term memory neural network for training until the network converges, and adjusting network hyper-parameters through a comparison test to finally obtain optimal parameters;

and 5: and testing the model effect through the prediction of the test set data to obtain the predicted value of each pollution index of the air quality.

2. The air quality space-time prediction method based on the long-short term memory neural network as claimed in claim 1, wherein the step 1 is specifically as follows:

step 1.1: obtaining historical air quality data, including PM_2.5Concentration, SO₂Concentration, O₃Concentration, CO concentration and NO₂Concentration 5 data characteristics;

step 1.2: acquiring meteorological data comprising five data characteristics of temperature, dew point, pressure, wind direction and wind speed.

3. The air quality spatiotemporal prediction method based on the long-short term memory neural network as claimed in claim 1, wherein the step 2 is specifically as follows:

step 2.1: removing abnormal values, and directly deleting records with obvious abnormal values in the data;

step 2.2: processing missing values, and filling the missing values by adopting an interpolation method;

step 2.3: extracting particle concentration data of adjacent sites, calculating the distance between two sites according to the Haverine formula recommended by Wikipedia and the longitude and latitude of each site, measuring the linear correlation strength between continuous variables by using a Pearson correlation coefficient, selecting data of five adjacent sites with the highest correlation, adding the data into a model, and improving the prediction accuracy by using the adjacent sites; the fused time series dataset is denoted R, where the record at each time in R is represented as:

r＝[pm_2.5,near_1,near_2,near_3,near_4,near_5,temp，dew,press,wnd_dir,wnd_spd,SO₂,NO₂,O₃,CO]；

wherein pm_2.5Near _1, near _2, near _3, near _4 and near _5 represent the concentrations of the current site and the adjacent five sites with the highest correlation respectively; temp, ew, press, wnd _ dir and wnd _ spd respectively represent data of temperature, dew point, pressure, wind direction and wind speed at the same time; SO (SO)₂,NO₂,O₃CO represents the respective concentration value;

step 2.4: and (5) carrying out data normalization processing, and scaling the data to [0,1 ].

4. The air quality spatiotemporal prediction method based on the long-short term memory neural network as claimed in claim 1, wherein the step 3 is specifically as follows:

step 3.1: converting a data format, namely converting an original data sequence into a format of a supervised learning sequence, from a time sequence to an input and output sequence pair, wherein the time sequence used by the model is a multivariate time sequence, and the multivariate time sequence is used for sequence prediction by specifying the lengths of the input and output sequences; assigning D as an input time step and N as an output time step to the model, and predicting values at t, t +1, … and t + N moments through all data characteristics at t-D, t-D +1, … and t-1;

step 3.2: dividing the data set, and dividing the historical time-series data set into 7: the scale of 3 is divided into a training set and a test set.

5. The air quality spatiotemporal prediction method based on the long-short term memory neural network as claimed in claim 1, wherein the step 4 is specifically as follows:

step 4.1: constructing an LSTM structure which comprises an input layer, a hidden layer and a full connection layer (output layer); setting a maximum epoch, the number of hidden layers and the number of neurons in each layer, and setting the number of complete connection layers and neurons thereof;

step 4.2: initializing the cell state and the hidden layer state of the LSTM neural unit;

step 4.3: calculating the weights of an input gate, a forgetting gate and an output gate of the current neuron and the current memory candidate value;

step 4.4: calculating the hidden state and the memory state of the current neuron and transmitting the hidden state and the memory state to the next neuron;

step 4.5: adopting MAE as a loss function and Adam as an optimization algorithm, and circularly reciprocating until the model converges;

step 4.6: comparing and adjusting parameters, and acquiring and storing optimal parameters; including maximum epoch, neuron number, learning rate, mini-batch size, L2 regularization coefficients, weight vectors, and bias vectors.

6. The air quality spatiotemporal prediction method based on the long-short term memory neural network as claimed in claim 1, wherein the step 5 is specifically as follows:

step 5.1: forecasting air pollution indexes on a test set by using a trained forecasting model, including PM_2.5、SO₂、O₃、CO、NO₂These 5 contamination indicators;

step 5.2: comparing all the predicted results with the true values, and using MAE, RMSE, R²Obtaining a performance evaluation result of the model as an evaluation index of the model;

step 5.3: the final prediction model is used for prediction of the actual air mass concentration.

Images