Flood Prediction using Hybrid ML-DL

Abstract

Flood prediction relies heavily on timely, accurate forecasting of flood risk as an important factor in the reduction of disaster and the provides benefits to the management of sustainable water resources. Existing hydrological and statistical models used to forecast the risk of flooding have shown limitations in being able to accurately assess the nonlinear and temporal relationships in the actual data derived from flood events. Recently, deep learning techniques such as RNNs, LSTMs, and GRUs have performed well on forecasting based time-series data; however, these techniques are often ineffective in the presence of noise, leading to high volatility and decreased generalizability. Therefore, this research proposes a hybrid flood prediction model that combines RForest ML for feature enhancement with DL solutions for time series forecasting. To validate the effectiveness of the developed hybrid model, a time-series dataset containing 36,500 hourly data points for total rainfall, river water level and discharge was created. First, the features were improved through ML, followed by DL sequentially predicting flood risk. Our analysis of the hybrid ML–DL solution suggests this approach is superior to single DL models on all performance indicators (MAE, MSE, RMSE, and R²) and supports our conclusion that the hybridization of RForest and recurrent DL models increases the accuracy and stability of flood forecasting.