2024 Volume 12 Issue 2 Article ID: 23-16127
Flood forecasting is one of the most challenging aspects in the field of hydrology. With new developments in computational intelligence, data-driven methods are gradually becoming popular and Long Short-Term Memory (LSTM) approach has shown great potential in accurate river stage forecasting due to its ability to learn long term dependencies. This study investigated methods for further improving accuracy of flood forecasts provided by LSTM models, especially when the prediction lead time is increased. LSTM model simulations were carried out for data obtained from nine river basins in Sri Lanka and forecasts made from 1 hour to 24 hours lead times were analyzed. Different scenarios were used to train the model and the results indicated that using selected data from separate river basins has improved the forecasting accuracy, significantly for longer lead times. A better feature selection method (to be used as input data to train the model) was investigated in this study by evaluating the strength and relationship of the river water level variation among different river gauging stations. Using this feature selection method for selecting optimum water level data combined with rainfall data provided the highest accuracy in the predictions.
