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

Abstract

Since weather forecasts are extensively relied upon in every aspect of human life—from agriculture to business, from travel to daily commuting—weather conditions around the world change quickly and frequently. Accurate forecasts are therefore crucial. In order to anticipate the weather, a variety of techniques are used, including trend forecasting and numerical weather prediction. However, these techniques are capital intensive, time-consuming, and have low accuracy. As an enhancement over current methods, this study presented a Long Short-Term Memory (LSTM) neural network model for forecasting meteorological parameters. This study used weather data (including dew point, pressure, relative humidity, temperature, wind speed, and rainfall) gathered by the Nigerian Meteorological Agency (NiMet), Abuja, for weather stations/four cities in Nigeria: Bauchi, Minna, Calabar, and Ikeja from 1 January 2015 to 30 December 2019. On the basis of the chosen multivariate weather variables, the model's performance was validated for the daily and weekly time-steps. The results show that for Mean Square Error, the proposed model performs better for short-range forecasts (values by 20.10% to 79.90%) than for medium-range forecasts (values by 26.94% to 73.06%). (MSE). Again, due to the relative consistency in meteorological variables measured at the station, the suggested model performs best for daily forecasts in Bauchi, Calabar, and Ikeja, and poorest in Minna City. Due to the relative volatility in the meteorological variables at the station, Ikeja city had the lowest results for the weekly forecasts made using the model, whereas Bauchi city had the best results. The study discovered that the proposed model's capacity for learning is influenced by the relative stability of the weather variable spread across the period. These results can be attributed to the LSTM model's memory capacity and feedback computation loop.