Dihybrid Recurrent Neural Networks for Solar Radiation Prediction: A Comparative Study in Nigeria

Alabi, N. O. & Ojo, G.O, Volume 5 Issue 2, December 2024 Pages 62-71, Published: 2024-11-19

Abstract

This research investigates solar radiation forecasting using two dihybrid recurrent neural networks architectures (Parallel and Sequential) on climate data from six Nigerian cities: Sokoto, Maiduguri, Ilorin, Ikeja, Enugu, and Port Harcourt. The dataset includes 31 years of monthly data on rainfall, relative humidity, sunlight hours, wind speed, maximum and lowest temperatures, and evaporation Piche. Each model is built using Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) layers and optimized through hyperparameter tuning. We analyzed solar radiation patterns, identified optimal locations for solar energy projects, and optimized agricultural planning. Both models showed excellent predictive accuracy, with the Parallel model outperforming the Sequential one. The Parallel dihybrid architecture achieved a lower Mean Square Error (MSE) of 0.0001 and a higher R² of 0.9995, making it more reliable for solar radiation forecasting. These findings contribute to more sustainable energy planning and agricultural optimization in Nigeria