MULTIDIMENSIONAL TIME SERIES WEATHER PREDICTION USING LONG SHORT TERM MEMORY NEURAL NETWORK

Abstract

Weather conditions around the world change rapidly and continuously. Correct forecasts are essential in the daily living of people due to over-dependency on weather forecasts heavily; from agriculture to industry, from traveling to daily commuting. A number of approaches such as ensemble weather prediction systems are highly time-ineffective such as Numerical Weather Prediction, and Trend Forecasting. This study proposed Long Short-Term Memory (LSTM) neural network model for forecasting weather parameters as improvement over existing approaches. This study used weather variables (such as dew point, pressure, relative humidity, temperature, wind speed and rainfall) collected from first of January, 2015 to thirtieth of December, 2019 for four cities of Nigeria, including: Bauchi, Minna, Calabar and Ikeja. The performance of the model was validated for the daily and weekly time-steps on the basis of the selected multivariate weather variables. The outcomes reveal that the proposed model performed best for short-range forecasts (values by 20.10% to 79.90%) than medium-range forecasts (values by 26.94% to 73.06%) for Mean Square Error (MSE). Again, the proposed model performed best for Bauchi, Calabar and Ikeja city, and worst for Minna City for daily forecasts because of the relative stability in weather variables measured of the former. In the case for weekly forecasts performed with the model in which Ikeja city had the worst outcomes, while Bauchi city had the best outcomes due to the relative instability in the weather variables of the former. The study found that relative stability in the weather variable spread across the period influences on the learning capability of the proposed model. These outcomes can be attributed to memory capacity and feedback loop of computation of Recurrent Neural Network (RNN-LSTM) model.