Time-series analysis and Flood Prediction using a Deep Learning Approach

Deep neural networks have been used successfully to solve time series prediction problems. Given their ability to automatically understand the temporal connections found in time series, they have shown to be an effective solution. In this proposed research, a Deep Learning (DL) based flood prediction model is explored and utilized for interpretation and prediction using meteorological data to reduce computational and time complexity with high accuracy. Gated Recurrent Networks (GRU) a variant of recurrent neural network model which can effectively use past data information for prediction and is faster in terms of training speed is the deep learning architecture deployed. Correlation analysis was performed on the weather parameters and the appropriate parameters were chosen. The dataset compromises 52 years (19022 records) of weather data in which 80% is used for training 20% for testing. The predictive modeling of rainfall associated with the South-west monsoon can guide the prediction of flood occurrence. The model deployed was evaluated with the performance metrics such as RMSE, MAE against LSTM model. The deployed RNN-GRU model had relatively low RMSE and MAE values when compared with LSTM architecture with improved prediction accuracy.