One of the most frequent natural catastrophes that causes significant harm to property, crops, the economy, and human lives is flooding. For scholars who have been trying to forecast floods for a long time, flood prediction presents a significant difficulty. This article proposes a flood forecasting model that makes use of the federated learning approach. By preventing data from being exchanged over the network for model training, Federated Learning, the most sophisticated machine learning (ML) approach, addresses network latency issues that arise in flood prediction while guaranteeing data privacy, availability, and security. Instead of transferring large data sets to a central server for local model aggregation and global data model training at the central server, the Federated Learning approach encourages onsite training of local data models and concentrates on transmitting these local models across the network. The suggested model in this article combines locally trained models of eighteen customers, determines which station is likely to experience flooding, and provides a flood alarm with a five-day lead time for a particular client. At the client station where the flood is anticipated, a local feed forward neural network (FFNN) model is developed. The local FFNN model's flood forecasting module uses a number of regional factors as input to estimate the anticipated water level. From 2015 to 2021, a dataset of five distinct rivers and barrages was gathered, taking into account four factors: hydrodynamics, flow routing, rainfall-runoff, and snow melting. The suggested flood forecasting model has an 84% accuracy rate in predicting past floods that occurred in the chosen area between 2010 and 2015.