Title Page
Contents
ABSTRACT 8
Chapter 1. Introduction 10
Chapter 2. Materials and Methods 15
2.1. Site description 15
2.2. Data acquisition 16
2.3. Data pretreatment and partitioning for training and test 20
2.4. Construction of deep learning and machine learning models 21
2.4.1. 1D-CNN 21
2.4.2. ANN 23
2.4.3. Ensemble model 25
2.5. Evaluation of model performance 26
Chapter 3. Results and Discussion 29
3.1. Descriptive examination of variables related to cyanobacteria and nutrients 29
3.2. Hyperparameter configuration for three models in the overall and optimal model 42
3.3. Comparison of overall performance of the three models 46
3.4. Comparison of optimal model performance of three models 49
Chapter 4. Conclusions 55
References 57
Abstract (in Korean) 66
Table 2.1. Description of Input variables for the 9-year period (N=309) at GGW and algae alert level data information. 18
Table 3.1. Descriptive statistics for the monthly water quality, hydrological, and meteorological input variables from 2013 to 2021. 31
Table 3.2. Descriptive statistics for the yearly water quality, hydrological, and meteorological input variables from 2013 to 2021. 32
Table 3.3. Descriptive statistics for the monthly cyanobacterial and nutrient concentrations on GGW. 41
Table 3.4. Description of optimized hyperparameters obtained from 100 iterative runs for the 1D-CNN, ANN, and ensemble models. 45
Figure 2.1. Map of Gangjeong-Goryeong Weir (35° 50' 26" N and 128° 27' 39" E) with water intake stations, and water quality monitoring station. 16
Figure 2.2. Flowchart for the construction of one-dimensional convolutional neural network (1D-CNN), artificial neural network (ANN), and Ensemble models to predict... 19
Figure 2.3. The structure of 1D-CNN for prediction of algae alert levels using 1D-layer. 23
Figure 2.4. Description of a confusion matrix for the model's observed and predicted values. Figure shows recall and precision of class L-0. Recall, precision for each class... 27
Figure 3.1. (a) Temporal variations in total dissolved nitrogen, nitrate nitrogen, and ammonium nitrogen from 2013 to 2021 in water quality input variables. The red lines mark the divisions between each year. 33
Figure 3.1. (b) Temporal variations in total dissolved phosphorus, phosphate phosphorus, and conductivity from 2013 to 2021 in water quality input variables. The red lines mark the divisions between each year. 34
Figure 3.2. Temporal variations in averaged water level, averaged total discharge rate, and averaged discharge rate by the hydropower plant from 2013 to 2021 in hydrological input variables. The red lines mark the divisions between each year. 35
Figure 3.3. Temporal variations in averaged air temperature, accumulated precipitation between water quality monitoring events from 2013 to 2021 in meteorological input variables. The red lines mark the divisions between each year. 38
Figure 3.4. Confusion matrices for the average prediction of algae alert levels in overall model of 1D-CNN, ANN, and ensemble model during training and test. 46
Figure 3.5. Confusion matrices for the prediction of algae alert levels in optimal model of 1D-CNN, ANN, and RF model during training and test. 50