Transfer Learning for Predictive Modelling of Wind Patterns in Coastal Nigeria for Off-Grid Deployment

Abstract

Reliable wind speed forecasting is essential for planning and deploying off-grid renewable energy systems in coastal regions, particularly in low-resource contexts where historical meteorological data are limited. This study presents a novel application of transfer learning to enhance predictive modeling of wind patterns in coastal Nigeria. Using a CNN-LSTM hybrid architecture, models were pretrained on extensive wind datasets from two climatologically aligned source regions—Tamil Nadu (India) and Rio Grande do Sul (Brazil)—and subsequently fine-tuned with two years of local ERA5 data from five Nigerian coastal states: Lagos, Ogun, Bayelsa, Akwa Ibom, and Cross River. Baseline models trained solely on Nigerian data exhibited limited generalization capacity, with average RMSE values exceeding 1.49 and MAPE above 19%. In contrast, the India-based transfer model achieved superior results, including an RMSE of 1.23, MAE of 0.96, and R2R^2R2 of 0.85, marking a significant improvement in predictive accuracy and convergence speed. Spatial evaluations confirmed consistent model robustness across all target zones, with the highest reliability observed in high-wind maritime regions such as Bayelsa and Akwa Ibom. Simulated turbine energy outputs based on predicted wind profiles suggested daily yields exceeding 3.2 kWh, supporting the viability of small-scale wind systems for decentralized energy applications. The study confirms that transfer learning can overcome data scarcity barriers in wind forecasting and offers a scalable framework for informed, cost-efficient deployment of wind energy technologies in coastal West Africa. This approach holds promise for advancing energy access and climate resilience across similarly under-resourced regions.