HYDROLOGICAL FORECASTING TESTS ON NEURAL NETWORK MODELS CONSIDERING INPUT VARIABLE SELECTION AND DATA QUALITY

Abstract

 Accurate hydrological forecasting is essential for effective water resource management and flood prevention. This study investigates the impact of input variable selection and data quality on the performance of Artificial Neural Networks (ANNs) in predicting river discharge across three river basins: Hiyoshi Dam, Katsura, and Fukakusa. Utilizing a dataset of 50,000 training data, our findings reveal that, once we collect a relevant data pool, the choice of input variable selection methods—whether including all rainfall data, the nearest rainfall points, or variables with the highest correlation coefficients—has minimal effect on prediction accuracy when the ANN is adequately trained. However, with a limited training dataset of 5,000 data points, prediction accuracy becomes more sensitive to input selection, underscoring the importance of carefully choosing relevant variables. The study also demonstrates that the ANN model is resilient to noisy training data. Further analysis shows that the ANN can effectively disregard irrelevant data, maintaining robust performance. These insights provide valuable guidance for optimizing input variable selection and ensuring data quality in ANN-based hydrological forecasting, enhancing the reliability and effectiveness of predictive models in water resource management.