275 / 2024-02-28 22:02:39

Multi-point water level forecasting by a hybrid type flood forecasting model and evaluation of the impact of missing observation data

flood forecasting; multi-point water level forecasting; hybrid model; deep learning; distributed rainfall-runoff model; missing data

Since flood disasters have become more severe in recent years due to climate change, flood forecasting technology has become increasingly important. Both improved forecast accuracy and longer forecast lead times are required for the practical development of such technologies. As real-time water level forecasting methods, hybrid forecasting methods that combine the advantages of rainfall-runoff models and machine learning methods have been actively studied recently. Hybrid forecasting methods using distributed rainfall-runoff models and deep learning models are promising approaches that can be expected to provide highly accurate water level forecasts by reflecting the characteristics of rainfall distribution and accumulated observation data. However, many previous reports have verified the model performance at a single gauge site in a watershed, and there are many unknowns regarding the performance of simultaneous multi-point water level forecasting, which is expected to be used in actual systems. Additionally, actual real-time forecasting systems often face missing water level observation data, so it is important to forecast water level stably even when observation data is partially missing. This study applied a hybrid method for water level forecasting using a distributed rainfall-runoff model (RRI model) and a deep learning model (DNN) to multi-point simultaneous water level forecasting at five sites in the Oyodo River system, and the forecast accuracy was evaluated for a forecast lead time up to 24 hours ahead. In addition, to evaluate the impact of missing data on forecast accuracy, we examined 1) the case in which some water level gauging stations are entirely unavailable and 2) the case in which the data is missing at random in the time series. The results showed that in the multi-point simultaneous forecast, the effect of hybridization on forecast accuracy became larger as one progressed from upstream to downstream, and root mean squared error (RMSE) was improved by up to 10% or more compared to the simple DNN model. Furthermore, the hybrid model showed more stable prediction accuracy in both missing data cases 1) and 2), and when trained with masked data simulating missing data, the prediction accuracy degradation due to missing data was suppressed to less than 6% in RMSE. These results indicate that this hybrid model is suitable for real-time forecasting systems since it shows good forecasting accuracy even for simultaneous forecasting of multiple locations and is robust against missing data.