Optimized Deep Learning Models for Monthly weather Forecast in Kogi, Nigeria

Abstract

Accurate weather forecasting is essential for agriculture, disaster preparedness, and economic planning in Nigeria, yet existing approaches such as Numerical Weather Prediction (NWP) face challenges of computational intensity and limited accuracy for localized monthly predictions. This study develops optimized deep learning models for monthly weather forecasting using Nigerian meteorological data from 2014 to 2023, with a focus on the Kogi region. Three deep learning architectures: Artificial Neural Networks (ANN), Recurrent Neural Networks (RNN), and Long Short-Term Memory (LSTM), were implemented and optimized using Bayesian hyperparameter tuning. To further enhance predictive performance, a Boosting Ensemble approach integrating the three models was proposed. Model evaluation employed Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) as benchmarks. Results showed that while RNN outperformed ANN and LSTM individually, the Boosting Ensemble achieved the best accuracy, with the lowest RMSE (52.351) and MAE (29.475), consistently capturing both stable and transitional weather patterns. The findings demonstrate that ensemble deep learning methods significantly improve monthly weather forecasting accuracy compared to standalone models. This study contributes a scalable, data-driven framework tailored to Nigeria’s climatic conditions, offering practical value for farmers, policymakers, and disaster management agencies, while also providing a foundation for future research incorporating additional climatic variables and advanced attention-based models