Abstract:
Weather forecasting is the field of making predictions of the future state of the atmosphere of a certain
location by analyzing initial values of relevant atmospheric characteristics which are obtained by
meteorological observations. Since weather prediction has substantial effect in economic sectors such
as agriculture, health, aviation, hydro power generation and even in daily lives of people, issuing
accurate weather forecasts is a major responsibility of meteorological authorities across the world.
Even though forecasting weather in mid-latitudes is uncomplicated and reliable, weather prediction in
a tropical country like Sri Lanka is notoriously difficult as sudden changes of convective tropical
weather phenomena are quite difficult to be predicted by prevailing Numerical Weather Prediction
(NWP) methods. Therefore, the current research aims to present machine learning based weather
prediction models for Sri Lanka for making short term forecasts for the most significant weather
attributes such as temperature and precipitation. This paper discusses on implementing two
multivariate Long Short-Term Memory Network models (LSTM) to make predictions on temperature
and precipitation separately for a selected weather station in Sri Lanka and review the applicability of
machine learning to solve highly nonlinear and complex weather problems. The prediction
performances of the implemented LSTM models are evaluated using standard evaluation techniques
such as Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE). The results show that two
LSTM models have made predictions with least RMSE and MAE values, evidencing the successful
applicability of machine learning for solving complex and nonlinear patterns of past observational
weather data and making accurate weather forecasts.
