Transfer learning framework for rural electrification: A GRU-based approach with custom loss for demand prediction

Electrical energy is a key driver of global socio-economic and techno-economic development. Achieving universal access while advancing climate goals requires shifting toward decentralised, renewable-powered energy systems, particularly in underserved regions. However, limited data availability, especially in unelectrified remote communities, poses a significant challenge for accurately predicting electricity demand and planning rural mini-grids. This study addresses this gap by proposing a transfer learning model based on a Gated Recurrent Unit (GRU) with a custom MSE-JSD loss function. Historical data from three communities and simulated data were used in two experimental scenarios, including pretraining-finetuning strategies across domains. To rigorously evaluate model performance, paired t-tests and two-way ANOVA were conducted. Results showed that GRU-MSE-JSD significantly outperformed the other models (p < 0.05), achieving an R² of 0.9947 on historical data and demonstrating strong generalisation to simulated data. Recurrent models with domain adaptation, particularly GRU-MSE-JSD and R-DANN, consistently outperformed traditional CNN-based architectures, which struggled to capture long-term temporal dependencies across different domains. Statistical analysis confirmed the robustness and transferability of the proposed model across different community profiles.