Intelligent flood forecasting model using committee machine learning for early warning system

Abstract

Extreme rainfall in upstream watersheds often results in the rise of river water levels, leading to severe flood disasters in the downstream catchment. Therefore, monitoring river water level and flow are crucial for flood forecasting in early warning systems and disaster risk reduction. Although some computational models achieved good prediction accuracy in particular problems, they might not perform well in different datasets. Thus, this study proposed a novel intelligence system using an ensemble committee machine-based framework to solve the “unstable” performance of the computational model to forecast flood with individual base learners by simple averaging and weighted averaging method. In addition, the use of simple averaging in the ensemble method is compromised by the worst-performing individual models in a collective forecast. The weights of different individuals should be tuned to find the optimal weight combination. This weight tuning algorithm can be treated as an optimisation problem. Thus, the genetic algorithm (GA) and K-nearest neighbour (K-NN) optimisation method were chosen for their flexibility and performance to improve the model’s generalisability. The applied base learners using various machine learning algorithms include radial basis function neural network (RBFNN), adaptive-neuro fuzzy inference system (ANFIS), support vector machine (SVM), and long short-term memory network (LSTM). The committee machine model was employed to forecast the river water level at the downstream area in different lead times addressed for the three various datasets in different areas, including Kelantan river, Terengganu river in Malaysia, and Mekong river in Cambodia. Performance comparison of the models is evaluated and analysed using various performance metrics, including mean percentage error (MPE), root mean square error (RMSE), mean absolute error (MAE), and correlation coefficient (R). The results showed that the proposed Intelligent Committee Machine Learning (ICML) outperformed the individual base models for most performance indicators. Specifically, its MPE,RMSE, and MAE of ICML by GA produced 2% - 70% smaller than the best individual and ICML-KNN-based model in the Kelantan dataset. Likewise, R values are 0.01% - 0.24% higher than the best ANFIS model and ICML by K-NN. The proposed ICML-GA based model has improved MAEs performance in the Terengganu dataset, 0.26% - 4.5% smaller than the best individual model (LSTM). While R performance of ICML-GA model produced 0.01% - 0.06% better in all steps ahead forecasting horizons. While in the Mekong dataset, the ICML-GA model outperformed all performance indicators. Specifically, its MPEs are 2% - 11% smaller than the best ANFIS and RBF model, 2% - 7% smaller in RMSEs, and 1% - 10% smaller in MAEs than those ANFIS and RBF. In addition, R values improved 0.01% - 0.07% better than other individual models. In sum, the proposed ICML-GA model can robustly forecast river water levels to predict floods for early warning and disaster risk reduction and outperformed individual models and the ICML-KNN model for the case studies investigated in this work.