Abstract Title: Artificial Intelligence-Driven Rainfall Forecasting in Flood-Prone Regions : Enhancing Prediction Accuracy Through Advanced Model Tuning Techniques

Abstract Submitted to: HYDROLOGY

Abstract Text:

Rainfall plays a decisive role, particularly in areas prone to flooding, where variations in rainfall patterns can impact water availability, transportation systems, environmental health, and short-term urban planning. The ability to accurately predict rainfall can greatly assist government bodies and private entities, enabling them to strategize and make informed short-term decisions in areas such as disaster management and early hazard warning systems, especially during the periods of flood. In this context, Artificial Intelligence (AI) is playing a significant role in enabling precise predictions of rainfall. AI is a powerful tool that can model and understand complex, non-linear relationships between inputs and outputs, often found in weather data. The relationship between meteorological factors influencing rainfall is neither linear nor simple. Artificial Neural Networks can capture these complex patterns effectively. The AI model uses historical data to learn patterns, which it then applies to predict future outcomes. The primary goal of this research is to discern the key meteorological characteristics influencing precipitation, and to forecast the subsequent day's rainfall, with a particular emphasis on areas vulnerable to flooding. The dataset employed consists of a decade's worth of daily weather readings from several locations in India that are prone to floods. This dataset encompasses a wide array of meteorological variables including minimum and maximum temperatures, rainfall, evaporation, sunshine duration, gusts of wind speed, average wind speed, humidity, atmospheric pressure, cloud coverage, and temperature readings at two specific times of the day (9 am and 3 pm). The machine learning model employed is a sequential model with four layers, incorporating dropout for regularization. The initial model, utilizing an early stopping callback, achieved an accuracy of 90.85%. To further enhance this, the study implemented a joint tuning procedure incorporating strategies such as a 'reduce learning rate on plateau' callback and a custom accuracy printing callback. This culminated in a significant boost in predictive precision, with the refined model yielding a remarkable accuracy of 94.78%.