Prediction of Water Level Using Machine Learning and Deep Learning Techniques

Forecasting the water levels in rivers and lakes is critical for flood warnings and water-resource management. Many soft computing techniques have been implemented for the prediction of water levels in lakes. While several deep learning models have been adopted, the comparison of the performance of these models with machine learning models is quite limited. In this study, the water level of Jezioro Kosno Lake in Poland has been predicted using 30 years of daily water level data through tree-based machine learning techniques, namely Random Forest and XGBoost, as well as deep learning techniques, namely bidirectional LSTM, convolutional 1D-BiLSTM, and recurrent neural network. The performance of the models was diagnosed using several statistical indicators such as mean square error (MSE), index of agreement (IA), root mean square error (RMSE), mean absolute percentage error (MAPE), mean absolute error (MAE), and symmetric mean absolute percentage error (SMAPE). The results suggest that XGBoost performs well in water level forecasting among the tree-based machine learning models. Among the deep learning models, the Conv1D-BiLSTM model performs unconditionally well. It is observed that XGBoost with the lowest RMSE value of 0.0066 and highest accuracy of 99.976 has outperformed all other models, and therefore, machine learning has performed better than deep learning for the current study.