Short and long-term forecasting of wind speed in Limpopo Province using machine learning algorithm and extreme value theory

Abstract

Numerous studies have applied Extreme Value Theory (EVT) to model environmental variables like wind speed, rainfall and temperature. Recently, academic focus has shifted to machine learning algorithms for the same variables. This research study demonstrates the practical use of EVT and machine learning techniques for modelling wind speed in the Limpopo Province, with the primary goal of assessing wind power generation reliability. The data used in this research study is obtained from National Aeronautics and Space Administration (NASA), spanning the time period from 2016 to 2022. The Vanilla Long Short-Term Memory (LSTM) network exhibited remarkable accuracy, achieving 86% training and 89% testing accuracy. Additionally, Generalised Extreme Value Distribution (GEVD) for block sizes (1 to 5) revealed GEV Dm=2 as the most suitable model based on low Akaike information criteria (AIC) and Bayesian information criteria (BIC) values. The model highlighted a rare event with a 300-year return period, indicating a wind speed of 22.893 meters. This study provides valuable insights for careful power planning, economic strategy and advancement in civilisation in South Africa, with implications for future energy planning and policy decisions in the region.