Self-adaptive optimization framework for water quality prediction

Abstract

The framework predicts water-quality data from observation data associated with raw features, e.g., rainfall. In the framework, an artificial neural network (ANN) computes the predicted water-quality data from feature data associated with impact features and derived from the observation data. The impact features are learnable, and are selected from impact-feature candidates comprising directional change- (DC-based) features each indicating occurrence of DC events in a corresponding raw feature. Including the DC-based features in the candidates enhances the ANN's ability of capturing significant change patterns of water quality due to extreme/unexpected events. The ANN architecture is also configurable according to model hyperparameters, which are learnable. The impact features and model hyperparameters are learnt by differential evolution for maximizing a prediction performance achieved by the ANN, thereby enabling the ANN architecture and impact features to be automatically optimized without requiring manual adjustment by domain experts in applying the ANN to different situations.