Enhancing weather forecasting through machine-learning models: a focus on short and medium weather patterns in Nairobi county.

Abstract

Historically, traditional data prediction techniques like Autoregressive Integrated Moving Average (ARIMA) and Numerical forecasting have experienced decreased accuracy and reliability over time, a factor attributed to the changing climatic patterns. There is thus need to develop a more robust accurate technique for weather forecasts to enhance safety and preparedness. This study employed Machine Learning models including Random Forest, Support Vector Machine and Gradient Boosting in forecasting short- and medium-term weather patterns specific to humidity, windspeed and daily temperatures targeting Nairobi County with an aim of the process effective in comparison to ARIMA. The results show that Gradient Boosting (R² = 74.30%, RMSE = 0.068) and SVM (R² = 72.72%, RMSE = 0.07) outperformed other models in predicting temperature, followed closely by Random Forest (R² = 69.43%). For humidity, SVM (R² = 68.60%) and Gradient Boosting (R² = 68.40%) also led in performance, while ARIMA performed poorly across both variables. All models struggled to predict windspeed accurately, with the best result from Gradient Boosting yielding only R² = 0.31%. These findings highlight the superior accuracy of ensemble and kernel-based machine learning algorithms for forecasting temperature and humidity in Nairobi County, while also indicating windspeed as a persistently difficult variable to model effectively. We can thus conclude that for short- and medium-term temperature and humidity, SVM (best performer) and GB (best performer) need to be utilized in forecast respectively. Nonetheless, there is need for further study to establish the best fitting ML model for forecasting wind speed.