Abstract:
The rapid growth of computer computing power has made it possible for researchers to perform complex
and labor-intensive tasks efficiently. Weather system is one of the most important naturally occurring
phenomenon with both positive and negative consequences for the world's population. In particular, rainfall
is one of the major phenomena of a weather system. Which is one of the fundamental things for life to exit.
However, because of heavy rain floods create several damages around the world. These happenings of
various types of accidents directly or indirectly linked to lack of uptime and up-to-date information and
early warning. Recent world statistical report indicates that natural disaster kills an average 60,000 peoples
per year globally. According to Zevin (1994) reports, 80%-90% of annual flood deaths are caused by flash
flood. Floods contribute more to these socio-economic disasters and cause prolonged loss, results in loss of
life and homelessness. In general, it harms personal properties and lives. By the year 1998 and 2017, More
than 2 billion peoples affected by floods in different parts of the world. According to a recent report in
Ethiopia, 2,285 households (13,710) peoples are affected and 853 households displaced by
floods. Furthermore, by the year 2019, In Ethiopia, 1,041 hectares of farming land are flooded and 987
animals killed by floods. There are varieties of studies done to elevate the issue, but the problem persists
until now. This indicates there is lack of dedicated studies that can precisely predict the weather conditions.
It is important to know when and where it is going to be rain ahead of time in-order-to save crops, human
and animal life, and countries economy. As a result, this study developed various regression and
classification models using state-of-the-art deep learning algorithms with improved performance to deliver
uptime and up-to-date weather information to the people to minimize such kind of losses and damages. In
this study, deep learning models have developed by applying Bidirectional Long Short Term Memory (Bi LSTM), Recurrent Neural Network (RNN), Gated Recurrent Unit (GRU), and Multi-Layer Perceptron
(MLP) for time series forecasting of rainfall using Arba Minch, Jinka and Mirab Abaya dataset. After that,
performance of the regression model evaluated using Mean Absolute Error (MAE), Mean Squared Error
(MSE), and Root Mean Squared Error (RMSE). Experimental result shows that GRU model outperforms
