FFM: Convolutional neural network-based flood forecasting model

Abstract

One of the more prevalent forms of natural catastrophe, floods may wipe out whole communities and their crops, not to mention the economy and human lives. Researchers who have been trying to forecast floods for a long time face a significant obstacle when trying to do so. This article proposes a model for flood forecasting that makes use of the federated learning approach. The state-of-the-art ML method known as Federated Learning prevents data from being sent over the network for the purpose of training models, which addresses the inherent network latency trials in flood prediction and assures data availability, privacy, and security. In a federated learning setup, rather of transferring massive data sets to a central server for regional model aggregation and global data model training, onsite training for regional data models is prioritized, with an emphasis on propagation of these models throughout the network. Using data collected from 18 different customers, the suggested model can predict which stations would experience flooding and send out flood alerts to individual clients with a five-day warning. The client station anticipating the flood trains a model using a network of feed forward neural networks (FFNN). The local FFNN model's flood forecasting module uses a number of regional characteristics to make predictions about the next water level. Hydrodynamics, flow routing, snow melting, and rainfall-runoff are the four factors that were included in the data set of five distinct rivers and barrages that was compiled from 2015 to 2021. With an accuracy rate of 84%, the suggested flood-prediction model has accurately forecasted past floods that occurred in the chosen zone between 2010 and 2015.