Title Page
Contents
Abstract 12
Chapter 1. Introduction 17
1.1. Motivation 17
1.2. Structure of Thesis 19
Chapter 2. Theoretical Background 20
2.1. Tele-Monitoring System (TMS) 20
2.1.1. Sewage treatment facility wastewater effluent limitation 22
2.1.2. Sewage treatment process 24
2.2. Machine Learning 27
2.2.1. RNN (Recurrent Neural Networks) 28
2.2.2. LSTM (Long short term memory) 33
2.2.3. GRU (Gated recurrent unit) 35
Chapter 3. Research Method 37
3.1. Definition of Parameters 37
3.2. Experiment Process 38
3.3. Data Preprocessing 40
3.4. Experimental Environment 41
3.5. Model Optimization 42
3.5.1. Loss function 42
3.5.2. Mini-batch training 42
3.5.3. Overfitting and gradient vanishing problem 43
Chapter 4. Forecasting Model Experiment and Evaluation 46
4.1. Forecasting Model Experiment 1 (Influent-Effluent Prediction) 46
4.1.1. Data acquisition and descriptive statistics 46
4.1.2. Model learning and test 49
4.1.3. LSTM and GRU comparison of influent-effluent BOD and COD measured values and predicted values 52
4.1.4. Optimum hyper parameter selection 56
4.1.5. Influent-effluent hyper-parameter analysis 57
4.1.6. Forecasting model evaluation 65
4.1.7. Statistical analysis 67
4.2. Forecasting Model Experiment 2 (Effluent-Future Effluent Prediction) 69
4.2.1. Data acquisition and descriptive statistics 69
4.2.2. Model learning and test 71
4.2.3. LSTM and GRU comparison of effluent-future effluent COD measured values and predicted values 72
4.2.4. Optimum hyper-parameter selection 75
4.2.5. Effluent-future effluent hyper-parameter analysis 75
4.2.6. Forecasting model evaluation 80
4.2.7. Effluent COD forecasting model evaluation 80
4.2.8. Statistical analysis 81
Chapter 5. Results 83
5.1. Influent-Effluent BOD-COD Predictive Simulation Result 83
5.1.1. LSTM and GRU comparison of influent-effluent BOD measured values and predicted values 83
5.1.2. LSTM and GRU comparison of influent-effluent COD measured values and predicted values 84
5.2. Effluent-future Effluent COD Predictive Simulation Result 85
5.2.1. LSTM and GRU comparison of effluent-future effluent COD measured values and predicted values 85
Chapter 6. Conclusion 86
References 89
Appendices 93
Appendix 1. Key Implementation Codes 93
Appendix 2. [제목없음] 97
Appendix 3. Learning rates and statistics according to hyper-parameter 100
국문요약 106
Table 1. Wastewater effluent limitation by water quality item 23
Table 2. Experimental environment 41
Table 3. Influent-effluent parameters 48
Table 4. Influent-effluent BOD and COD optimized hyper parameters 56
Table 5. Evaluation result of influent-effluent BOD, RMSE, and MAPE 66
Table 6. Evaluation result of influent-effluent COD, RMSE, and MAPE 66
Table 7. Paired comparison results of influent-effluent BOD, 68
Table 8. Paired comparison results of influent-effluent COD 68
Table 9. Parameter statistics for water quality prediction of effluent-future effluent 71
Table 10. Optimized hyper-parameters of prediction for effluent-future effluent COD 75
Table 11. RMSE and MAPE evaluation results of effluent-future effluent COD prediction 81
Table 12. Paired comparison results of effluent-future effluent BOD 82
Figure 1. Photograph of water quality TMS in B area 21
Figure 2. Overview of sewage treatment facility 24
Figure 3. Sewage flowing into sewage treatment facility 24
Figure 4. Water treatment process of the sewage treatment facility 26
Figure 5. Time-series structure and iterative input and output layers of RNN 29
Figure 6. Flowchart of weight calculation of each layer obtained during the BPTT process 31
Figure 7. Weight of RNN 33
Figure 8. LSTM structure and internal storage and status output details t 34
Figure 9. Types of LSTM 35
Figure 10. GRU structure, internal storage, and status output details 36
Figure 11. Experiment process of the forecasting model 38
Figure 12. (a) Standard network of connecting layers (b) Network with dropout applied 44
Figure 13. Skip connection diagram 45
Figure 14. Method of avoiding local minima 50
Figure 15. LSTM and GRU comparison of influent-effluent BOD measured values and predicted values (a) actual measured... 53
Figure 16. LSTM and GRU comparison of influent-effluent COD measured values and predicted values (a) actual measured... 55
Figure 17. RMSE results according to batch size (box plot), (Above: LSTM, Below: GRU) 57
Figure 18. RMSE results according to epoch (box plot), (Above: LSTM, Below: GRU) 58
Figure 19. RMSE results according to learning rate (box plot), (Above: LSTM, Below: GRU) 59
Figure 20. RMSE results according to neuron (box plot), (Above: LSTM, Below: GRU) 60
Figure 21. RMSE results obtained according to batch size (box plot), (Above: LSTM, Below: GRU) 61
Figure 22. RMSE results obtained according to epoch (box plot), (Above: LSTM, Below: GRU) 62
Figure 23. RMSE results obtained according to learning rate (box plot), (Above: LSTM, Below: GRU) 63
Figure 24. RMSE results obtained according to neuron (box plot), (Above: LSTM, Below: GRU) 64
Figure 25. LSTM and GRU comparisons of effluent-future effluent COD measured values and predicted values (a) actual... 74
Figure 26. RMSE results according to batch size (box plot), (Above: LSTM, Below: GRU) 76
Figure 27. RMSE results obtained according to epoch (box plot), (Above: LSTM, Below: GRU) 77
Figure 28. RMSE results obtained according to learning rate (box plot), (Above: LSTM, Below: GRU) 78
Figure 29. RMSE results obtained according to neuron (box plot), (Above: LSTM, Below: GRU) 79
Figure 30. LSTM and GRU overlap graphs for influent-effluent BOD actual measured values and predicted values 83
Figure 31. LSTM and GRU overlap graphs for influent-effluent COD actual measured values and predicted values 84
Figure 32. LSTM and GRU overlap graphs for effluent-future effluent COD actual measured values and predicted values 85