표제지

목차

ABSTRACT 19

제1장 서론 22

제1절 연구배경 22

1. COVID-19 대유행과 도시의 대기질 변화 22

2. 강릉시의 황사기간 전, 중, 후의 대기질 변화 24

제2절 연구방법 27

1. COVID-19 대유행과 도시의 대기질 변화 27

2. 강릉시의 황사기간 전, 중, 후의 강릉에서 대기질 변화 29

제2장 관련연구 32

제1절 국내의 연구동향 32

1. COVID-19 대유행과 도시의 대기질 변화 32

2. 강릉시의 황사기간 전, 중, 후의 대기질 변화 32

제2절 국외의 연구 동향 34

1. COVID-19 대유행과 도시의 대기질 변화 34

2. 강릉시의 황사기간 전, 중, 후의 대기질 변화 34

제3장 인공신경망모델 및 다중회귀통계모델 알고리즘 36

제1절 인공신경망모델 (Artificial neural network)의 특성 36

1. Machine learning 36

2. Deep learning 36

제2절 활성함수(Activation function)의 종류 및 특성 41

1. 선형 활성함수(Linear activation function) 41

2. 시그모이드 활성함수(Sigmoid activation function) 42

3. 쌍곡탄젠트함수(Hyperbolic tangent activation function; tanh) 44

제3절 다변량회귀모델 (Multivariate regression model) 특성 47

제4장 인공신경망 /다중회귀모델 활용 대기질 예측 실험 48

제1절 COVID-19 대유행 기간 48

1. 연구지역 48

2. 대기질 자료 및 분석 49

3. 도시별 대기질(PM10, PM2.5, NO₂)의 일별/주별 평균농도 변화[이미지참조] 49

4. 도시의 대기질 (PM10, PM2.5, NO₂) 예측식(COVID-19)-Sigmoid 활성함수[이미지참조] 54

5. ANN-Sigmoid 모델의 예측 성능 검증(COVID-19)[이미지참조] 72

6. 도시의 대기질(PM10, PM2.5, NO₂) 예측식(COVID-19)-Tanh 활성함수[이미지참조] 98

7. ANN-tanh 모델의 예측 성능 검증(COVID-19)[이미지참조] 111

8. 도시의 대기질(PM10, PM2.5, NO₂) 예측식(COVID-19)-다변량회귀모델[이미지참조] 136

9. 다변량회귀모델의 예측 성능 검증(COVID-19) 138

제2절 황사기간 151

1. 연구지역 151

2. 대기질 자료 및 분석 152

3. 황사의 이동 및 확산 지역 154

4. 매 시각별 PM, 기상, 대기오염농도 분포 156

5. 강릉의 대기질(PM10, PM2.5, NO₂) 예측식(황사)-Sigmoid 활성함수[이미지참조] 163

6. ANN-Sigmoid 모델의 검증(황사)[이미지참조] 180

7. 강릉의 대기질(PM10, PM2.5, NO₂) 예측식(황사)-tanh 활성함수[이미지참조] 207

8. ANN-tanh 모델의 예측 성능 검증(황사)[이미지참조] 220

9. 강릉의 대기질(PM10, PM2.5, NO₂) 예측식(황사)-다변량회귀모델[이미지참조] 248

10. 다변량회귀모델의 예측 성능 검증(황사) 249

제5장 결론과 제안 270

제1절 결론 270

1. COVID-19 대유행과 도시의 대기질 변화 270

2. 강릉시의 황사기간 전, 중, 후의 대기질 변화 271

제2절 제안 273

1. COVID-19 대유행과 도시의 대기질 변화 273

2. 강릉시의 황사기간 전, 중, 후의 대기질 변화 274

참고문헌 276

국문초록 288

표목차

Table 1. Maximum, minium and mean values of PM10(μg/m³), PM2.5(μg/m³) and NO₂(ppm) concentrations at Seoul, Busan and...[이미지참조] 53

Table 2. Definition of variables in ANN-Sigmoid and ANN-tanh models[이미지참조] 55

Table 3. Input and output variables of ANN-Sigmoid model(COVID-19)[이미지참조] 59

Table 4-1. Parameter estimate(ANN-Sigmoid, Node＝15, PM10-S), 2020[이미지참조] 63

Table 4-2. As shown in Table 4-1, except for PM2.5-S, 2020[이미지참조] 64

Table 4-3. As shown in Table 4-1, except for NO2-S, 2020[이미지참조] 65

Table 5-1. Parameter estimate(ANN-Sigmoid, Node＝15, PM10-B), Busan, 2020[이미지참조] 66

Table 5-2. As shown in Table 5-1, except for PM2.5-B, 2020[이미지참조] 67

Table 5-3. As shown in Table 5-1, except for NO2-B, 2020[이미지참조] 68

Table 6-1. Parameter estimate(ANN-Sigmoid, Node＝15, PM10-G), Gwangju, 2020[이미지참조] 69

Table 6-2. As shown in Table 6-1, except for PM2.5-G, 2020[이미지참조] 70

Table 6-3. As shown in Table 6-1, except for NO2-G, 2020[이미지참조] 71

Table 7. Classification of Pearson Correlation Coefficient r 73

Table 8. Validation of ANN-sigmoid model(RMSE and R²)[이미지참조] 74

Table 9-1. Importance of independent variables in ANN-Sigmoid model, Seoul, 2020[이미지참조] 76

Table 9-2. As shown in Table 9-1, except for Busan, 2020 77

Table 9-3. As shown in Table 9-1, except for Gwangju, 2020 78

Table 10-1. Parameter estimate(ANN-tanh, Node＝15, PM10-S), Seoul, 2020[이미지참조] 102

Table 10-2. As shown in Table 10-1, except for PM2.5-S, 2020[이미지참조] 103

Table 10-3. As shown in Table 10-1, except for NO2-S, 2020[이미지참조] 104

Table 11-1. Parameter estimate(ANN-tanh, Node＝15, PM10-B), Busan, 2020[이미지참조] 105

Table 11-2. As shown in Table 11-1, except for PM2.5-B, 2020[이미지참조] 106

Table 11-3. As shown in Table 11-1, except for NO2-B, 2020[이미지참조] 107

Table 12-1. Parameter estimate(ANN-tanh, Node＝15, PM10-G, Gwangju, 2020[이미지참조] 108

Table 12-2. As shown in Table 12-1, except for PM2.5-G, 2020[이미지참조] 109

Table 12-3. As shown in Table 12-1, except for NO2-G, 2020[이미지참조] 110

Table 13. Validation of ANN-tanh model(RMSE and R²)(COVID-19)[이미지참조] 111

Table 14-1. Importance of independent variables in ANN-tanh model, Seoul, 2020[이미지참조] 114

Table 14-2. As shown in Table 14-1, except for Busan, 2020 115

Table 14-3. As shown in Table 14-1, except for Gwangju, 2020 116

Table 15-1. Multivariate regression equation for daily mean PM10, PM2.5 and NO₂ at Seoul and Busan, 2020[이미지참조] 139

Table 15-2. As shown in Table 15-1, except for Gwangju, 2020 140

Table 16-1. Correlation coefficient matrix of all variables in a Multivariate regression model at Seoul, Busan and Gwangju, 2020 141

Table 16-2. As shown in Table 16-1, except for Busan, 2020 142

Table 16-3. As shown in Table 16-1, except for Gwangju, 2020 143

Table 17. Prediction performance of ANN-Sigmoid, ANN-tanh, and Mutivariate regression model by Pearson r on PM10, PM2.5, NO₂ at...[이미지참조] 148

Table 18. Prediction performance (Pearson r) of models for PM and gas for COVID-19, 2020 150

Table 19. Maximum, minimum and mean concentrations(μg/m³) of PM10, PM2.5 before, during and after the Yellow Dust event at...[이미지참조] 156

Table 20. Definition of variables in ANN-Sigmoid and ANN-tanh models[이미지참조] 164

Table 21. Input and output variables in ANN-sigmoid model(Dust)[이미지참조] 167

Table 22-1. Parameter estimate(ANN-sigmoid, Node＝15, PM10-G) before the Yellow Dust event at Gangneung, 2015[이미지참조] 171

Table 22-2. As shown in Table 22-1, except for PM2.5-G, 2015[이미지참조] 172

Table 22-3. As shown in Table 22-1, except for NO2-G, 2015[이미지참조] 173

Table 23-1. Parameter estimate(ANN-sigmoid, Node＝15, PM10-G) during the Yellow Dust event at Gangneung, 2015[이미지참조] 174

Table 23-2. As shown in Table 23-1, except for PM2.5-G, 2015[이미지참조] 175

Table 23-3. As shown in Table 23-1, except for NO2-G, 2015[이미지참조] 176

Table 24-1. Parameter estimate(ANN-sigmoid, Node＝15, PM10-G) after the Yellow Dust event at Gangneung, 2015[이미지참조] 177

Table 24-2. As shown in Table 24-1, except for PM2.5-G, 2015[이미지참조] 178

Table 24-3. As shown in Table 24-1, except for NO2-G, 2015[이미지참조] 179

Table 25-1. Validation of ANN-Sigmoid model(RMSE and R²)[이미지참조] 182

Table 25-2. Determination coefficient(R²) and Pearson correlation coefficient r in ANN-Sigmoid model at Gangneung, 2015[이미지참조] 183

Table 26-1. Importance of independent variables in ANN-sigmoid model at Gangneung, before the Yellow Dust event, 2015[이미지참조] 184

Table 26-2. As shown in Table 26-1, except for during the Yellow Dust event, 2015 185

Table 26-3. As shown in Table 26-1, except for after the Yellow Dust event, 2015 186

Table 27-1. Parameter estimate(ANN-tanh, Node＝15, PM10-G) before the Yellow Dust event at Gangneung, 2015[이미지참조] 211

Table 27-2. As shown in Table 27-1, except for PM2.5-G, 2015[이미지참조] 212

Table 27-3. As shown in Table 27-1, except for NO2-G, 2015[이미지참조] 213

Table 28-1. Parameter estimate(ANN-tanh, Node＝15, PM10-G) during the Yellow Dust event at Gangneung, 2015[이미지참조] 214

Table 28-2. As shown in Table 28-1, except for PM2.5-G, 2015[이미지참조] 215

Table 28-3. As shown in Table 28-1, except for NO2-G, 2015[이미지참조] 216

Table 29-1. Parameter estimate(ANN-tanh, Node＝15, PM10-G) after the Yellow Dust event at Gangneung, 2015[이미지참조] 217

Table 29-2. As shown in Table 29-1, except for PM2.5-G, 2015[이미지참조] 218

Table 29-3. As shown in Table 29-1, except for NO2-G, 2015[이미지참조] 219

Table 30-1. Validation of ANN-tanh model (RMSE and R²) before, during and after the Yellow Sand event at Gangneung, 2015[이미지참조] 221

Table 30-2. Coefficient of determination(R²) and Pearson r in validation from ANN-tanh model at Gangneung, 2015[이미지참조] 222

Table 31-1. Importance of independent variables in ANN-tanh model at Gangneung, before the Yellow Dust event, 2015[이미지참조] 224

Table 31-2. As shown in Table 31-1, except for during the Yellow Dust event, 2015 225

Table 31-3. As shown in Table 31-1, except for after the Yellow Dust event, 2015 226

Table 32-1. Multivariate regression equation for hourly PM10, PM2.5 and NO₂ at Gangneung, before and during the Yellow Dust event, 2015[이미지참조] 250

Table 32-2. As shown in Table 32-1, except for after the Yellow Dust event, 2015 251

Table 33-1. Correlation coefficient matrix of variables to PM10-G(N) in a Multivariate regression model at Gangneung, before the Yellow Sand event, 2015[이미지참조] 254

Table 33-2. As shown in Table 33-1, except for PM2.5-G(N), 2015[이미지참조] 255

Table 33-3. As shown in Table 33-1, except for NO2-G(N), 2015[이미지참조] 256

Table 34-1. Correlation coefficient matrix of variables to PM10-G(N) in a Multivariate regression model at Gangneung, during the Yellow Sand event, 2015[이미지참조] 257

Table 34-2. As shown in Table 34-1, except for PM2.5-G(N), 2015[이미지참조] 258

Table 34-3. As shown in Table 34-1, except for NO2-G(N), 2015[이미지참조] 259

Table 35-1. Correlation coefficient matrix of variables to PM10-G(N) in a Multivariate regression model at Gangneung, after the Yellow Sand event, 2015[이미지참조] 260

Table 35-2. As shown in Table 35-1, except for PM2.5-G(N), 2015[이미지참조] 261

Table 35-3. As shown in Table 35-1, except for NO2-G(N), 2015[이미지참조] 262

Table 36-1. Prediction performance of ANN-sigmoid, ANN-tanh, and Mutivariate regression model by Pearson r on PM10, PM2.5, NO₂ at...[이미지참조] 267

Table 36-2. Prediction performance of models on the Yellow Sand event-Pearson r 269

그림목차

Figure 1. Machine learning algorithm 36

Figure 2. Neural structure and single layer perceptron 37

Figure 3. Multiple layer perceptron-machine and deep learning algorithm 39

Figure 4. Feedforward shallow artificial neural network(ANN) 39

Figure 5. Feedforward deep artificial neural network 40

Figure 6. Linear activation function 42

Figure 7. Sigmoid activation function with restricted inputs(yellow-gradient) 43

Figure 8. Artificial neural network: Back- and forward-propagation 43

Figure 9. Hyperbolic tangent activation function 45

Figure 10. ReLU(Rectified linear unit activation function) 46

Figure 11. LReLU(Leaky rectified linear unit activation function) 46

Figure 12. Location of Seoul(S), Busan(B), Gwangju(G) in Korea 48

Figure 13-1. Daily/weekly mean of PM10(μg/m³), PM2.5(μg/m³), NO₂(ppm x 1000) at Seoul(2019~2020)[이미지참조] 50

Figure 13-2. As shown in Fig. 13-1, except for Busan(2019~2020) 51

Figure 13-3. As shown in Fig. 13-1, except for Gwangju(2019~2020) 52

Figure 14. Feedforward artificial Neural Network-17 nodes 56

Figure 15-1. Scattered plot(ANN-Sigmoid, No＝15, 17, 19), PM10-S, Seoul, 2020[이미지참조] 80

Figure 15-2. As shown in Fig. 15-1, except for PM2.5-S, 2020[이미지참조] 81

Figure 15-3. As shown in Fig. 15-1, except for NO2-S, 2020[이미지참조] 82

Figure 16-1. Scattered plot(ANN-Sigmoid, No＝15, 17, 19), PM10-B, Busan, 2020[이미지참조] 83

Figure 16-2. As shown in Fig. 16-1, except for PM2.5-B, 2020[이미지참조] 84

Figure 16-3. As shown in Fig. 16-1, except for NO2-B, 2020)[이미지참조] 85

Figure 17-1. Scattered plot(ANN-Sigmoid, No＝15, 17, 19), PM10-G, Gwangju, 2020[이미지참조] 86

Figure 17-2. As shown in Fig. 17-1, except for PM2.5-G, 2020[이미지참조] 87

Figure 17-3. As shown in Fig. 17-1, except for NO2-G, 2020[이미지참조] 88

Figure 18-1. Comparison of daily/monthly observed and predicted PM10-S(ANN-Sigmoid), Seoul, 2020[이미지참조] 89

Figure 18-2. As shown in Fig. 18-1, except for PM2.5-S, 2020[이미지참조] 90

Figure 18-3. As shown in Fig. 18-1, except for NO2-S, 2020[이미지참조] 91

Figure 19-1. Comparison of daily/monthly observed and predicted PM10-B(ANN-Sigmoid), Busan, 2020[이미지참조] 92

Figure 19-2. As shown in Fig. 19-1, except for PM2.5-B, 2020[이미지참조] 93

Figure 19-3. As showm in Fig. 19-1, except for NO2-B, 2020[이미지참조] 94

Figure 20-1. Comparison of daily/monthly observed and predicted PM10-G(ANN-Sigmoid), Gwangju, 2020[이미지참조] 95

Figure 20-2. As shown in Fig. 20-1, except for PM2.5-G, 2020[이미지참조] 96

Figure 20-3. As shown in Fig. 20-1, except for NO2-G, 2020[이미지참조] 97

Figure 21-1. Scattered plot(ANN-tanh, No＝15, 17, 19), PM10-S, Seoul, 2020[이미지참조] 118

Figure 21-2. As shown in Fig. 21-1, except for PM2.5-S, 2020[이미지참조] 119

Figure 21-3. As shown in Fig. 21-1, except for NO2-S, 2020[이미지참조] 120

Figure 22-1. Scattered plot(ANN-tanh, No＝15, 17, 19), PM10-B, Busan, 2020[이미지참조] 121

Figure 22-2. As shown in Fig. 22-1, except for PM2.5-B, 2020[이미지참조] 122

Figure 22-3. As shown in Fig. 22-1, except for NO2-B, 2020[이미지참조] 123

Figure 23-1. Scattered plot(ANN-tanh, No＝15, 17, 19), PM10-G, Gwangju, 2020[이미지참조] 124

Figure 23-2. As shown in Fig. 23-1, except for PM2.5-G, 2020[이미지참조] 125

Figure 23-3. As shown in Fig. 23-1, except for NO2-G, 2020[이미지참조] 126

Figure 24-1. Comparison of daily/monthly observed and predicted PM10-S(ANN-tanh), Seoul, 2020[이미지참조] 127

Figure 24-2. As shown in Fig, 24-1, except for PM2.5-S, 2020[이미지참조] 128

Figure 24-3. As shown in Fig, 24-1, except for NO2-S, 2020[이미지참조] 129

Figure 25-1. Comparison of daily/monthly observed and predicted PM10-B(ANN-tanh), Busan, 2020[이미지참조] 130

Figure 25-2. As shown in Fig. 25-1, except for PM2.5-B, 2020[이미지참조] 131

Figure 25-3. As shown in Fig. 25-1, except for NO2-B, 2020[이미지참조] 132

Figure 26-1. Comparison of daily/monthly observed and predicted PM10-G(ANN-tanh), Gwangju, 2020[이미지참조] 133

Figure 26-2. As shown in Fig. 26-1, except for PM2.5-G, 2020[이미지참조] 134

Figure 26-3. As shown in Fig. 26-1, except for NO2-G, 2020[이미지참조] 135

Figure 27-1. Comparison of daily/monthly observed and predicted PM10-S, PM2.5-S and NO2-S by Multivariate regression, Seoul, 2020[이미지참조] 145

Figure 27-2. As shown in Fig. 27-1, except for Busan, 2020 146

Figure 27-3. As shown in Fig. 27-1, except for Gwangju, 2020 147

Figure 28. Northeastern Asia-Gobi Desert, Beijing(B, China) and Gangneung(G, Korea) 151

Figure 29-1. COMS satellite image with dust cloud at 15 LST, March 20, 2015, Gangneung before the Yellow Dust event 155

Figure 29-2. As shown in Fig. 29-1, except for 21 LST, March 21, 2015, Gangneung during the Yellow Dust event 155

Figure 30. Hourly distribution of PM10, PM2.5 and PM₁ concentrations(μg/m³) at Gangneung from March 18 to March 27, 2015[이미지참조] 157

Figure 31. Hourly distribution of PM10 and PM2.5 concentrations(μg/m³) at Beijing, China from March 16 to March 25, 2015[이미지참조] 157

Figure 32. Hourly distributions of air temperature(℃)(a), wind speed(m/s)(b), and relative humidity(%)(c) at Gangneung, 2015 158

Figure 33. Hourly distributions of SO₂(x 1,000), CO(x 100), O₃(x 1,000), and NO₂(x 1,000; NOx ＝ NO₂ + O₃) concentrations(ppm) at...[이미지참조] 159

Figure 34. Hourly distributions of SO₂(x 1,000), CO(x 100), O₃(x 1,000), and NO₂(x 1,000) concentrations(ppm) at Beijing, China, 2015 160

Figure 35. Feedforward artificial neural network-Node＝15 166

Figure 36-1. Scattered plot of PM10-G (ANN-sigmoid, No＝13, 15, 17) at Gangneung, before the Yellow Dust event, 2015[이미지참조] 188

Figure 36-2. As shown in Fig. 36-1, except for PM2.5-G, 2015[이미지참조] 189

Figure 36-3. As shown in Fig. 36-1, except for NO2-G, 2015[이미지참조] 190

Figure 37-1. Scattered plot of PM10-G (ANN-sigmoid, No＝13, 15, 17) at Gangneung, during the Yellow Dust event, 2015[이미지참조] 191

Figure 37-2. As shown in Fig. 37-1, except for PM2.5-G, 2015[이미지참조] 192

Figure 37-3. As shown in Fig. 27-1, except for NO₂, 2015 193

Figure 38-1. Scattered plot of PM10-G (ANN-sigmoid, No＝13, 15, 17) at Gangneung, after the Yellow Dust event, 2015[이미지참조] 194

Figure 38-2. As shown in Fig. 38-1, except for PM2.5-G, 2015[이미지참조] 195

Figure 38-3. As shown in Fig. 38-1, except for NO2-G, 2015[이미지참조] 196

Figure 39-1. Comparison of hourly observed and predicted PM10-G (ANN-Sigmoid), Gangneung, before the Yellow Dust event, 2015[이미지참조] 198

Figure 39-2. As shown in Fig. 39-1, except for PM2.5-G, 2015[이미지참조] 199

Figure 39-3. As shown in Fig. 39-1, except for NO2-G, 2015[이미지참조] 200

Figure 40-1. Comparison of hourly observed and predicted PM10-G (ANN-Sigmoid), Gangneung, during the Yellow Dust event, 2015[이미지참조] 201

Figure 40-2. As shown in Fig. 40-1, except for PM2.5-G, 2015[이미지참조] 202

Figure 40-3. As shown in Fig. 40-1, except for NO2-G, 2015[이미지참조] 203

Figure 41-1. Comparison of hourly observed and predicted PM10-G (ANN-Sigmoid), Gangneung, after the Yellow Dust event, 2015[이미지참조] 204

Figure 41-2. As shown in Fig. 41-1, except for PM2.5-G, 2015[이미지참조] 205

Figure 41-3. As shown in Fig. 41-1, except for NO2-G, 2015[이미지참조] 206

Figure 42-1. Scattered plot of PM10-G (ANN-tanh, No＝13, 15, 17), Gangneung, before the Yellow Dust event, 2015[이미지참조] 229

Figure 42-2. As shown in Fig. 42-1, except for PM2.5-G, 2015[이미지참조] 230

Figure 42-3. As shown in Fig. 42-1, except for NO2-G, 2015[이미지참조] 231

Figure 43-1. Scattered plot of PM10-G (ANN-tanh, No＝13, 15, 17), Gangneung, during the Yellow Dust event, 2015[이미지참조] 232

Figure 43-2. As shown in Fig. 43-1, except for PM2.5-G, 2015[이미지참조] 233

Figure 43-3. As shown in Fig. 43-1, except for NO2-G, 2015[이미지참조] 234

Figure 44-1. Scattered plot of PM10-G (ANN-tanh, No＝13, 15, 17), Gangneung, after the Yellow Dust event, 2015[이미지참조] 235

Figure 44-2. As shown in Fig. 44-1, except for PM2.5-G, 2015[이미지참조] 236

Figure 44-3. As shown in Fig. 44-1, except for NO2-G, 2015[이미지참조] 237

Figure 45-1. Comparison of hourly observed and predicted PM10-G (ANN-tanh), Gangneung, before the Yellow Dust event, 2015[이미지참조] 239

Figure 45-2. As shown in Fig. 45-1, except for PM2.5-G, 2015[이미지참조] 240

Figure 45-3. As shown in Fig. 45-1, except for NO2-G, 2015[이미지참조] 241

Figure 46-1. Comparison of hourly observed and predicted PM10-G (ANN-tanh), Gangneung, during the Yellow Dust event, 2015[이미지참조] 242

Figure 46-2. As shown in Fig. 46-1, except for PM2.5-G, 2015[이미지참조] 243

Figure 46-3. As shown in Fig. 46-1, except for NO2-G, 2015[이미지참조] 244

Figure 47-1. Comparison of hourly observed and predicted PM10-G (ANN-tanh), Gangneung, after the Yellow Dust event, 2015[이미지참조] 245

Figure 47-2. As shown in Fig. 47-1, except for PM2.5-G, 2015[이미지참조] 246

Figure 47-3. As shown in Fig. 47-1, except for NO2-G, 2015[이미지참조] 247

Figure 48-1. Comparison of hourly observed and predicted PM10-G, PM2.5-G and NO2-G by Multivariate regression, Gangneung, before...[이미지참조] 264

Figure 48-2. As shown in Fig. 48-1, except for during the Yellow Dust event, 2015 265

Figure 48-3. As shown in Fig. 48-1, except for after the Yellow Dust event, 2015 266

2020년 COVI-19 대유행기간에 대표적인 주요 도시인 서울, 부산, 광주와 2015년 3월의 황사기간에 강릉의 실시간 대기질 예측을 인공신경망모델인 Artificial neural network model-Sigmoid function(ANN-Sigmoid)과 -Hyperbolic tangent function(ANN-tanh), 다변량회귀기법(Multivariate regression model; MRM)으로 수행하였다.

2020년 COVID-19가 전국적으로 대유행하고 있을 때, 도시 자체와 도시간에 사람 및 차량의 이동이 많이 감소하여, 2019년 비해 서울(부산, 광주)에서 PM10, PM2.5와 NO₂의 평균농도가 15%(17.8%, 31.4%), 16.4%(21%, 22.9%), 15%(16.8%, 21.2%)로 전국적으로 약 20% 정도 감소하였다. 3개 도시간에 대기오염물질의 이동을 고려하고, 그 도시들에서 PM10, PM2.5, NO₂의 1일 평균농도를 예측하는 모델을 고안하기 위해, 각 도시의 PM10, PM2.5, SO₂, CO, O₃ 및 NO₂ 농도의 총 18개 자료를 ANN-Sigmoid, ANN-tanh과 다변량회귀모델의 입력자료로 사용하고, 모델들의 예측 성능을 비교하였다.

인공신경망모델의 은닉층에서 노드 수를 15, 17, 19로 구분하고, 입력층-은닉층-출력층 사이에 활성함수를 Sigmoid와 Sigmoid, tanh와 tanh로 채택하였다. 서울에서, PM10(PM2.5, NO₂) 예측값과 측정값 간의 Pearson r 상관계수는 Sigmoid-0.961(0.972, 0.957), tanh-0.976(0.976, 0.959), 다변량회귀-0.959(0.964, 0.951)이었고, ANN-tanh모델의 예측 성능이 가장 높았다. 부산에서, PM10(PM2.5, NO₂)의 Pearson r은 Sigmoid-0.964(0.964, 0.942), tanh-0.974(0.982, 0.938), 다변량회귀-0.963(0.966, 0.913)이었다. NO₂의 경우에 ANN-Sigmoid의 예측 성능이 ANN-tanh보다 약간 높았고, 그 외는 ANN-tanh의 성능이 가장 높았다. 광주에서, PM10(PM2.5, NO₂)의 Pearson r은 Sigmoid-0.974(0.974, 0.956), tanh-0.989(0.966, 0.961), 다변량회귀-0.971(0.971, 0.954)로 ANN-tanh 모델의 예측 성능이 가장 높았다. 3개 도시에서 3개 모델의 예측 성능이 탁월했으며, ANN-tanh 모델이 가장 우수하였다.

고비사막에서 발생한 황사가 북서풍에 의해 중국 북동지역의 북경시로 수송되면, 북경시의 대기오염물질과 결합된다. 결합된 대기오염물질이 강릉시와 같은 한반도로 수송되어, 도시의 대기질에 크게 영향을 준다. 강릉시에서 2015년 3월 18일부터 3월 27일까지 황사 전, 중, 후로 나누어서, 강릉시의 매 시각별 PM10, PM2.5, NO₂ 농도가 예측되었다.

강릉에서 현재 시각의 PM10. PM2.5와 NO₂의 농도를 예측하기 위해, 강릉보다 2일 전의 북경의 대기질 자료(PM10, PM2.5, SO₂, CO, O₃, NO₂)와 강릉의 1시간(또는 3시간) 전의 3개의 기상변수(기온, 풍속, 상대습도)와 대기질 자료(PM10, PM2.5, SO₂, CO, O₃, NO₂), 총 15 자료를 ANN-Sigmoid, ANN-tanh, 다변량회귀모델의 입력자료로 사용하였다. ANN-Sigmoid, ANN-tanh에서 13, 15, 17개로 구분된 은닉층 노드 수가 COVID-19 경우와 같은 방법으로 활성함수로 처리하여 모델을 운용하기위해 사용되었다.

황사 전, 중, 후에 강릉에서 매 시각별 PM10, PM2.5, NO₂ 농도를 예측하는 ANN-Sigmoid, ANN-tanh, 다변량회귀모델 같은 3종류의 모델이 개발되었으며, 모델들의 성능이 상호 비교되었다. 황사 전에, PM10, 의 예측값과 측정값 간에 피어슨 r은 Sigmoid-0.955(0.867), tanh-0.982(0.959), and multivariate regression-0.961(0.909)이었으며, ANN-tanh 모델의 예측 성능이 가장 높았다. 황사 중에는 PM10(PM2.5)의 Pearson r은 Sigmoid-0.943(0.980), tanh-0.972(0.984), 다변량회귀-0.948(0.977)이었으며, ANN-tanh 모델이 최상이었다. 황사 후에, Pearson r이 sigmoid-0.931(0.956), tanh-0.955(0.962), 다변량회귀-0.920(0.947)이었고, ANN-tanh모델이 예측 성능이 가장 높았다. 전 연구기간에 모든 모델의 예측 성능이 매우 우수하였고, ANN-tanh 모델이 가장 좋았다. 본 연구에서 개발된 모델들은 평상시와 황사기간에 도시의 대기질 예측과 활용에 크게 기여할 것이다.