Multistep Ahead Forecasting of Electrical Conductivity in Rivers by Using a Hybrid Convolutional Neural Network-Long Short Term Memory (CNN-LSTM) Model Enhanced by Boruta-XGBoost Feature Selection Algorithm

Abstract

Electrical conductivity (EC) is a key water quality metric for predicting the salinity and mineralization. In this study, the 10-day-ahead EC of two Australian rivers, Albert River and Barratta Creek, was forecasted using a novel deep learning algorithm, i.e., the convolutional neural network combined with long short-term memory (CNN-LSTM) model. The Boruta-extreme gradient boosting (XGBoost, XGB) feature selection method was used to determine the significant inputs (time series lagged data) for the model. The performance of the proposed Boruta-XGB-CNN-LSTM model was compared with those of three machine learning approaches: multi-layer perceptron neural network (MLP), K-nearest neighbor (KNN), and XGBoost, considering different statistical metrics such as the correlation coefficient (R), root mean square error (RMSE), and mean absolute percentage error (MAPE). Ten years of data for both rivers were extracted, with data for seven (2012–2018) and three years (2019–2021) used for training and testing the models, respectively. The Boruta-XGB-CNN-LSTM algorithm outperformed the other models in forecasting the 1-day-ahead EC in both stations over the test dataset (R = 0.9429, RMSE = 45.6896, and MAPE = 5.9749 for Albert River; and R = 0.9215, RMSE = 43.8315, and MAPE = 7.6029 for Barratta Creek). In addition, the Boruta-XGB-CNN-LSTM model could effectively forecast the EC for the next 3–10 days. Nevertheless, the performance of the Boruta-XGB-CNN-LSTM model slightly deteriorated as the forecasting horizon increased from 3 to 10 days. Overall, the Boruta-XGB-CNN-LSTM model is an effective soft computing method for accurately predicting the EC fluctuation in rivers.