Development of Machine Learning Based Flood Prediction Model for Godavari River Basin

Abstract

Floods are the most prominent and lethal event of this century. Unfortunately, the lack of a robust flood prediction framework has led to the loss of human life and infrastructures. Flood prediction models are very important for assessing and managing extreme flood events. They provide robust and accurate predictions that can be utilized for policy analysis and evacuation planning. Due to the limitations of long-term observed rainfall and runoff dataset, which are the necessity of the traditional statistical and physical models, the use of data-driven models has increased. Modern data-driven models, such as those developed by machine learning, are more advantageous due to their ability to develop models with minimal inputs. They can also numerically formulate flood nonlinearity using historical data. Therefore, a support vector machine (SVM) algorithm was used to develop the flood prediction model in this study. The developed flood prediction model was then used to simulate the flood at three gauging stations namely, Purna, GR Bridge and Yelli of the Godavari river basin. The model was calibrated with four historical flood events and validated with two flood events. Historical gridded rainfall data from IMD is given as input to the model for the simulation of floods. The performance evolution of the developed SVM model was done with the NSE, Error in Peak and Error in time to Peak. In calibration, the average NSE was obtained as 0.96, 0.97 and 0.96 at the G.R. Bridge, Purna and Yelli, respectively, and in validation, it was found to be 0.93, 0.95 and 0.87 respectively. The error in time to peak was found to be zero except for one event at Yelli. The error in magnitude of peak flow at each gauging location was less than 10% in both calibration and validation, which shows that the developed SVM models were good enough to predict future flood events at respective gauging locations. The bias corrected rainfall of best performing GCM and multimodel ensemble (MME) mean of CMIP6 GCM rainfall were used to generate future flood events in the Godavari river basin. A considerable flood peak reduction around 10%–37% was observed for the future flood events under both the scenarios for the MME mean rainfall and best performing GCMs rainfall in comparison to historical floods. Future bias corrected MME mean was on low side compared to the historical rainfall and this results the reduction of future flood peaks in the study area. The HEC-RAS model was also used to prepare the flood inundation maps from the output of the SVM model. Future flood inundation maps were generated for the simulated flood events from MME mean over the period 2070–2100. The future flood inundation maps showed that the river reach from GR Bridge to Yelli is vulnerable to floods. The extent of flood maps is more for the SSP585 than SSP245 for the period of 2070–2100. Nanded town, on the bank of river Godavari is going to inundate for the above-mentioned period under high emission scenario (SSP585). The results of this study may help the local stakeholders and organizations to better understand the flood behaviour in the Godavari river basin and are useful to form mitigation strategies for the future.