Renewable energy forecasting—particularly for distributed solar power—is vital for ensuring grid reliability and efficient resource management. However, traditional centralized forecasting approaches often come with drawbacks like high communication overhead, privacy issues, and limited scalability. Federated Learning (FL), consists of a decentralized machine learning framework that allows edge devices to collaboratively train models without disclosing raw data. Yet, FL struggles when devices have limited or sparse data, leading to cold-start problems and weaker model performance. To mitigate these challenges, this work introduces a method that incorporates GAN-based data augmentation to enrich local datasets with synthetic time-series samples during the initialization phase of the federated training process.