Abstract ID: 3.11533 | Accepted as Poster | Talk/Oral | TBA | TBA

Heavy rainfall events that trigger floods with economic and social consequences in Ecuador can be caused mainly by seasonal events such as El Nino, sub-seasonal events such as the Madden Julian Oscillation, or mesoscale convective systems that develop in a few hours. Given the significant influence of extreme rainfall, it is essential to have a system for forecasting extreme rain to help prepare more effectively to face this threat and mitigate its impact. Although the application of neural networks in weather forecasting has been going on for almost two decades, in recent years, the application and development of algorithms have advanced at a dizzying pace. However, in Ecuador, there are few studies on applying neural networks to weather forecasting and much less on forecasting extreme rainfall events, which are events with social and economic consequences of significant impact. In addition, the complexity of predicting extreme events, given the chaotic behavior of the atmosphere, is still well known. Although global and mesoscale meteorological models are currently used for weather forecasting, it is still complex to forecast the amount of precipitation with such precision and in such advance that they can help prevent the effects they may cause. This study uses a neural network technique to analyze a twenty-year time series of meteorological data from surface weather stations in the country’s coastal, inter-Andean, and Amazonian regions with different climate patterns. Firstly, an exploratory data analysis is performed to observe the relationship between the variables: rainfall, average, minimum, and maximum temperature, average relative humidity, solar radiation, wind speed, and cellophane. The study then proceeds to implement a neural network for extreme events. The neural network is optimized by adjusting the number of neurons in the hidden layer to achieve the lowest mean square error (MSE) and the highest coefficient of determination (R2). The results demonstrate that Extreme Learning Machine’s single-layer feed-forward network model effectively forecasts extreme rainfall. The implemented neural network model performs best in the inter-Andean region, with the lowest mean square errors. The Amazon region follows this, and lastly, the coastal area of Ecuador.