The study focuses on two rivers in Australia, the Albert River and Barratta Creek, and analyzes their electrical conductivity (EC) data collected from specific stations. The study uses the Boruta-XGBoost feature selection technique and develops a hybrid expert system combining Boruta-XGBoost with the CNN-LSTM model for forecasting EC values in the rivers. The Boruta-XGBoost technique selects important predictors based on Z-scores, and the CNN-LSTM model utilizes convolutional and long short-term memory (LSTM) layers for prediction.
The study also compares the performance of four machine learning models (Boruta-XGB-MLP, Boruta-XGB-XGBoost, Boruta-XGB-KNN, and Boruta-XGB-CNN-LSTM) in forecasting daily EC values in different time horizons for the two rivers. The models are optimized using grid search, and the data is pre-processed through normalization. Statistical metrics such as RMSE, correlation coefficient, mean absolute percentage error, and Nash–Sutcliffe model efficiency coefficient are used to evaluate the model performance.
The research methodology involves splitting the dataset into training and testing sets using a holdout strategy and implementing cross-validation techniques. The study provides detailed insights into the structure and functioning of the Boruta-XGBoost and CNN-LSTM models, explaining the processes involved in feature selection and prediction. The results of the study, including the optimal lagged-time components for input to the models and the statistical metrics for model evaluation, are presented and discussed.
Overall, the study aims to develop an effective forecasting model for EC values in rivers using machine learning techniques and advanced deep learning architectures. The research contributes to the field of water resource management and environmental monitoring by providing a novel approach to predicting water quality parameters in rivers.
Source link
Source link: https://www.nature.com/articles/s41598-024-65837-0