환경 빅데이터 분석 및 서비스 개발 Ⅳ = Big Data Analysis: Application to Environmental Research and Service Ⅳ

다국어 초록 (Multilingual Abstract)

Ⅰ. Background and Aims of Research
❏ We continue to build up ‘Environmental Policy monitoring System’ dedicated to periodically identify environmental policy needs and assess timeliness and effectiveness of environmental policy as we did last year
ㅇ Environmental Policy monitoring System apply prediction accuracy and real-time data collection-analysis-diffusion capability of Machine learning to environmental policy research
ㅇ Our ‘Environmental Policy monitoring System’ consists of three components: ‘Deep Learning Based Pollution Prediction algorithm’, ‘Real Time Environmental Text Analysis algorithm’, ‘Issue Based Database’
- Deep Learning Based Pollution algorithm: Periodically update various pollution prediction
- Real Time Environmental Text Analysis algorithm: Periodically summarise environment related text data and sentiment analysis
ㆍText summary: abstract keywords and keyword network from texts produced by environmental policy provider and environmental policy consumers
ㆍSentiment analysis: Real-time collection and sentiment analysis of SNS related to all subfield of environment
- Issue Based Database: Key environmental issue network connected with data analysis for each issue updating real-time
ㅇ Policy need Identification: Detect environment policy areas and regions in need of intervention from the predictions of ‘Deep Learning Based Pollution Prediction algorithm’, the text analysis results of ‘Real Time Environmental Text Analysis algorithm’, and the data analysis results of ‘Issue Based Database’
ㅇ Timeliness assessment: check if the temporal pattern of keywords analysis result on policy provider text and the temporal pattern of keywords analysis results on policy consumer are consistent
ㅇ Effectiveness assessment: Check Pollution improvement, SNS Sentiment improvement, and Environmental Issue improvement after policy execution
❏ In 2020, we tried to improve interpretability of ‘Environmental Policy monitoring System’
ㅇ While utilizing the advantage of deep learning we found in period 1(2017~2019), we tried to reduce complexity and strengthen interpretability
ㅇ In period 1, we focused on ‘apply everything related to big data analysis to Environment policy research’ From 2020, we are going to focus on ‘Environmental Policy Research using large scale data’
ㅇ Regrading methodology, we stick to machine learning in period 1. From 2020, we are going to be more flexible and try to include traditional frequentist and Bayesian statistical methods
ㆍWe are going to use simpler models to improve our understanding on feature importance
ㆍWe are going to build models capable of longer -term prediction and models with more interpretability
❏ In 2020, we build four algorithms for ‘Environment Policy Monitoring System’ and perform two independent researches
ㅇ For ‘Environment Policy Monitoring System’, we expand the methodology and scope of previous components
- We added two fine particle estimation algorithms in `Fine particle high concentration event prediction’ and ‘PM_2.5 estimation and prediction using Graph-GRU algorithm’
- In ‘Environmental text sentiment analysis algorithm’, we expanded sentiment analysis of environment related SNS to all subfield of environment
- In ‘Climate change issue based database’, we constructed new issue based database on climate change
ㅇ On two important issues that cannot be integrated to ‘Environment Policy Monitoring System’, we did independent research
- In ‘The impact of air pollution long-term exposure to mortality of COPD patients’,we estimated the effect of long-term exposure of air pollution on the death risk of COPD patients using NHI (National Health Insurance) Data
- In ‘Solar electricity generation prediction’, we constructed an RNN based algorithm predicting solar electricity generation of F1 power plant
Ⅱ. Fine Particle High Concentration Event Prediction
❏ We built a quantile regression based prediction algorithm to predict ‘extremely bad(76+)’ event of PM_2.5 in 25 air pollution monitoring station in Seoul
ㅇ We adjusted quantile regression model to analyze data with extreme values
- We applied LASSO variable selection method to Extreme Conditional Quantile Regression Model
❏ We predicted 4 hour average(6 periods per day) PM_2.5 pollution in Seoul using air pollution data and weather data
ㅇ For independent variables, we used contemporary and 1 time earlier CO, SO₂, NO₂, O₃, PM₁₀ pollution, rainfall, temperature, humidity, wind strength, wind direction and 1 time earlier PM_2.5 pollution
ㅇ We train our model with data from Jan. 1st. 2015 to Mar. 4th. 2018 and test with data from Mar. 4th. 2018 to May. 29th. 2020
❏ We were able to achieve 89.0%~100.0% Sensitivity while limiting False Positive at 6.0%~17.1%, depending on the station.
ㅇ Except for Gandonggu(88.9%), our algorithm achieve sensitivity higher than 90%
❏ The sensitivity of our model was higher by more than 11.3%p, compared to the sensitivity of models based on random forest, supporting vector machine and GRU
ㅇ Sensitivity of Ganseogu: our algorithm 92.3% ＞ GRU 81.0% ＞ SVM 73.3% ＞ RandomForest 65%
❏ The probability of ‘extremely bad’ event of PM_2.5 pollution increases when contemporary CO, O₃, PM₁₀ pollution, wind direction and 1 time ahead PM_2.5 pollution.
- We applied LASSO variable selection method to Extreme Conditional Quantile Regression Model
❏ We predicted 4 hour average(6 periods per day) PM_2.5 pollution in Seoul using air pollution data and weather data
ㅇ For independent variables, we used contemporary and 1 time earlier CO, SO₂, NO₂, O₃, PM₁₀ pollution, rainfall, temperature, humidity, wind strength, wind direction and 1 time earlier PM_2.5 pollution
ㅇ We train our model with data from Jan. 1st. 2015 to Mar. 4th. 2018 and test with data from Mar. 4th. 2018 to May. 29th. 2020
❏ We were able to achieve 89.0%~100.0% Sensitivity while limiting False Positive at 6.0%~17.1%, depending on the station.
ㅇ Except for Gandonggu(88.9%), our algorithm achieve sensitivity higher than 90%
❏ The sensitivity of our model was higher by more than 11.3%p, compared to the sensitivity of models based on random forest, supporting vector machine and GRU
ㅇ Sensitivity of Ganseogu: our algorithm 92.3% ＞ GRU 81.0% ＞ SVM 73.3% ＞ RandomForest 65%
❏ The probability of ‘extremely bad’ event of PM_2.5 pollution increases when contemporary CO, O₃, PM₁₀ pollution, wind direction and 1 time ahead PM_2.5 pollution.
❏ The probability of ‘extremely bad’ event of PM_2.5 pollution decreases when contemporary and 1 time ahead rainfall, windspeed increases
Ⅲ. PM_2.5 Estimation and Prediction Using Graph-GRU Algorithm
❏ We built Graph-GRU albright utilizing weather and geography information to prediction PM_2.5 pollution
ㅇ We construct a 3 dimension dataset consists of temporal and spatial data
- Our dependent variable is 3 hour average PM2.5 concentration of 0.125° × 0.125° Grid (12.5km x 12.5km) containing air-pollution monitoring stations in South Korea
- Our independent variables are weather data, air pollution data and height data
ㆍWe trained our model with data from Jan. 1st. 2015 to Dec. 31th. 2015, validated with data from Jan. 1st. 2016 to Dec. 21th. 2016, and tested with data from Jan. 1st. 2017 to Dec. 21th. 2017,
ㅇ We used Graph-GRU algorithm
- For node attributes, we used weather data, spacial data, and air pollution data
- To measure edge attributes, we constructed a function evaluating impact of air pollutant emission between air pollution monitoring stations
ㆍThis function evaluates impact of air pollutant emission between air pollution monitoring stations using distance, wind speed, and wind direction
- For adjacency Matrix, we used threshold of distance of 300km or height of 1,200m
❏ RMSE of our model in 3 hour~72 hour prediction was 4.05μg/m³ ~ 11.49μg/m³
ㅇ Temporal information reduced RMSE by 0.12μg/m³, and spacial information reduced RMSE by 0.16μg/m³
Ⅳ. Environmental Text Sentiment Analysis Algorithm
❏ ‘Environmental text sentiment analysis algorithm’ periodically performs collection-analysis-result distribution on SNS text regarding environment
ㅇ We expanded climate change sentiment analysis algorithm we build in 2018~2019 to all subfield of environments, and trained with new 140 thousands SNS text training data labeled by semi-supervised learning to improve accuracy
ㅇ We constructed web based user interface to visualize sentiment analysis results over time
❏ To improve accuracy, we increase the size of our training data from 50 thousand to 180 thousand using semi-supervised learning
ㅇ We collected 6.5 million SNS text, and applied 2018~2019 version climate change sentiment analysis algorithm. We collected 140 thousands cases with high positive/negative sentiment score, which we added to training data. We kept 10 thousand cases for testing
ㅇ With this new training data, the accuracy of sentiment analysis improved by 1%p: From 78.7% to 79.7%
❏ Sensitivity of our newly trained model was 66~92%, and Recall of our newly trained model was 73%~81%, according to subfield
ㅇ ‘Waste’ field had the lowest Sensitivity and Recall. This field needs supervised learning approach to improve overall accuracy
❏ We built web based user interface to visualize sentiment analysis results with user option of keyword search and period choice
ㅇ Our user interface also abstract keyword network of SNS of negative sentiment, which should give insights on the cause of negative sentiment
Ⅴ. Climate Change Issue Based Database
❏ We built a datamap on Climate change consists of hierarchically organized climate change issue network and data analysis linked to each issue in the network. This datamap is capable of real-time data analysis update
ㅇ We build issue collection module to extract issues from text data and data analysis module to link data analysis with extracted issues
❏ Issue collection module execute ‘Climate change text collection → Topic Extraction → Issue Identification → Issue Network Organization’ process
ㅇ Text Collection: Reports from government sponsored research institutes/ Formal speeches from higher-ranking official/Press Release from government/Assembly meeting transcripts/Materials from Climate Change From in Assembly/DBpia academic paper abstracts/NAVER paper articles of 12 major papers from 2012 to 2019
ㅇ Topic Extraction: Apply Correlated Topic Model to extract 10 topics and correlation between topics
ㅇ Issue Identification: Derive issues from key sentences extracted from TextRank algorithm
ㅇ Issue Network Organization: Assign each issue to topics and organize issues according to the relationship between topics
- We deduced relationship between topics combining three sources of informations - (1) Correlation coefficient from Correlated Topic Model (2) Similarity of time series frequency pattern from Dynamic Time Warping (3) Specialist Survey
- We re-categorized 10 CTM topics into 5 Categories: Climate Issue cooperation/Climate Change adaptation/Greenhouse Gas Reduction /Energy and Environment/Urban Environment and Citizen
❏ Data analysis module attach data source and data analysis result to each issue in Climate Change Issue Network
- We linked each data analysis to data source so that we can update data analysis in real-time
Ⅵ. The Impact of Air Pollution Long-Term Exposure on the Mortality of COPD Patients
❏ We estimate the impact of 1-year and 5-year air pollution exposure on the mortality of COPD patients
ㅇ We analyzed medical data of COPD patients older than 40 diagnosed from 2009 to 2018
ㅇ We used kriging to convert air pollution monitoring station data to small local district (Up. Myun.Dong) data and applied Cox Proportional hazard model to small local district data ❏ We combined NHI(National Health Insurance) individual patient data and air-pollution monitoring station data
ㅇ For air pollution exposure variable, we used 1-year and 5 year average of small local district PM₁₀, O₃, NO₂ pollution
- For PM₁₀ and NO₂, We used daily average. For O₃, we used average of maximum 8 hours for each day (We converted the unit of O₃ and NO₂ from ppb to ㎍/m3)
ㅇ From NHI individual medical data, we obtained gender, age, income percentile, CCI, COPD exacerbation, smoking status
ㅇ For dependent variable, we used dummy variable assigning 1 for death and 0 for survival
❏ We found that COPD patients exposed higher O₃ 1-year or 5-year had higher risk of death
ㅇ Hazard ratio of 1 year O₃ exposure was estimated as 1.003. Hazard ratio of 5 year O₃ exposure was estimated as 1.004
Ⅶ. Solar Electricity Generation Prediction
❏ We developed an LSTM algorithm predicting electricity generation of Yung -am F1 solar power plant
ㅇ We predicted hourly electricity generation and 12 hour average of electricity generation. The electricity generation was non-stationary time-series
❏ For independent variables, we used electricity generation and weather data with 1 lag
ㅇ For weather data, we used hourly temperature, rainfall, humidity, solar insolation, Total Cloud amount from Mokpo weather monitoring station
ㅇ We used data from Jan. 01.2017 to Jun. 30. 2018 for training, and data from Jun. 30. 2018 to Jun. 30. 2019 for testing
❏ We built and RNN based LSTM algorithm and compared RMSE with ARIMA model and 3-lag moving average
❏ The RMSE of our model was 36.9% of standard deviation in 1 hour prediction and 51.1% in 12 hours average prediction
ㅇ The RMSE to standard deviation of our model was 71% of the RMSE to standard deviation ratio of 3-lag moving average, and 45% of RMSE to standard deviation of ARIMA model
Ⅷ. Conclusion and Suggestions
❏ Summarizing, we improved ‘Environmental Policy Monitoring System’ and added some new results
ㅇ We supplemented and improved components of ‘Environmental Policy Monitoring System’
- ‘Deep Learning Based Pollution Prediction algorithm’: We improved interpretability and extended prediction lag
ㆍ ‘Fine particle high concentration event prediction’: We built a Quantile regression model which can produce coefficient estimates for independent variables and is capable of prediction as accurate as machine learning algorithm
ㆍ ‘PM_2.5 estimation and prediction using Graph-GRU algorithm’: We achieved 7.64g/m³ 12 hour prediction RMSE, which is equivalent to 1 hour prediction RMSE of our CNN based algorithm in 2019
- ‘Real Time Environmental Text Analysis algorithm ’: We expanded realtime sensitivity analysis and keyword network abstraction of negative sentiment for all environmental policy subfield
- ‘Issue Based Database’: We improve policy monitoring scope from fine particle issue (2019) to climate change (2020)
ㅇ We developed three new algorithms and one new issue based database
- Fine particle high concentration event prediction quantile regression model, PM_2.5 pollution prediction Graph-GRU model, RNN based Solar electricity generation prediction algorithm/ climate change datamap
ㅇ We expanded scope of environmental text analysis: Real-time environmental text analysis web interface
❏ For policy application, we strengthen environmental policy monitoring capability, quantify the health risk of air pollution, and provided items for the infrastructure of renewable energy
ㅇ Environmental Policy monitoring: We improved policy need identification and information generation for precautionary policy intervention
- Now our ‘Real Time Environmental Text Analysis algorithm’ is capable of identifying subfield of environment regarding which general public has negative sentiment.
- Now our ‘Climate Change datamap’ is capable of real-time assessment of climate change issues
- Now our ‘Deep Learning Based Pollution Estimation algorithm’ is capable of extending prediction lag of PM_2.5 and providing basic causality analysis for high concentration event of PM_2.5
ㆍGraph-GRU extended prediction lag. We can use this time for preventive policy intervention
ㆍQuantile Regression model can be used policy evaluation tool by extending control variables and policy related variables
ㅇ We provided quantified risk of air pollution on COPD patients, which can be used to quantify benefits of air-pollution reduction policy
ㅇ We provided solar electricity generation prediction algorithm, which can be used for renewable energy smart-grid infrastructure