Flood prediction based on weather parameters using advanced machine learning-metaheuristic approaches

Abstract

Flood prediction has become more popular worldwide because of the devastating socioeconomic effects of this hazard and the predicted rise in its frequency in the near future. In India, public health, civil engineering, and agriculture are all greatly affected by flooding. Anything can be flooded, with levels ranging from a few inches to many feet. They may appear suddenly or develop gradually. The intensity and frequency of flooding will frequently increase due to human modifications to the environment. More frequent and severe weather occurrences could lead to more violent floods. Utilizing data-driven and machine-learning models to solve flow- and flood-related problems has lately gained traction as a subject of study. ML model shows two key advantages over traditional physically-based models controlled by differential equation systems. Firstly, without requiring a complete a priori understanding of the phenomenon, data-driven models are able to generate reasonably accurate predictions. The quantity, quality, and variety of data that are accessible all affect how accurate the model is. This feature shows that we can avoid the complexity of problems faced by physical-based models caused by the growing number of important components by learning from observational data. Second, data-driven flood models replace numerical integration of differential equations, which is an iterative process, with non-iterative procedures like forward propagation of neural networks. We chose to study the floods in the Barak River basin (BRB), India, a high-elevation and quickly urbanized river basin that is prone to frequent flooding because of recent evidence of the impacts of regional climate change on the hydrological cycle. Using principal component analysis (PCA), the optimal inputs were found. Decision-makers in the hydrological field of research need accurate information regarding effective predictors. This study looks into the viability of using weather input data (rainfall, humidity, evapotranspiration, temperature) to predict monthly floods using a support vector machine customized with Manta-Ray foraging optimization (SVM-MRFO). The accuracy of SVM-MRFO was assessed by comparing it against SVM tuned by the Firefly algorithm, whale optimization algorithm, Salp swarm algorithm based on mean absolute errors (MAE), root mean square errors (RMSE), determination coefficient (R2), and Nash-Sutcliffe Efficiency (NSE). Implementing the FFA, WOA, SSA, and MRFO algorithms enhances the accuracy of the SVM. The best performance metrics, NSE of 0.9914, RMSE of 0.0182, MAE of 0.0073, and BIAS of were obtained by the SVM model constructed using the MRFO training procedure, suggesting the model's potential for use in flood forecasting. The flood models in this study are significant since they were created using a mix of different inputs and AI algorithms. In conclusion, this study demonstrated the ability of AI algorithm-based models to forecast floods and produced a number of practical methods that the flood control departments of different states, regions, and nations might employ to estimate the likelihood of floods.