Modern smart grids increasingly rely on accurate load forecasting to manage energy resources efficiently amid the growing integration of renewable and distributed energy sources. While federated learning offers a promising approach by enabling decentralized model training without raw data sharing, it still exposes users to privacy risks through model updates and remains vulnerable to malicious participants. This paper presents a secure and privacy-preserving federated learning framework tailored for smart grid load forecasting. The proposed system enhances privacy by employing anonymous update submissions, zero-knowledge proofs, and blind signatures, thereby unlinking user identities from their model contributions. To ensure model integrity, a blockchain-based reputation mechanism is introduced, incentivizing honest participation while penalizing malicious behavior without unfairly targeting legitimate users. The system also incorporates a distributed clustering process to address data heterogeneity, enabling the training of specialized models across user clusters. The framework is evaluated in terms of privacy, security, and performance, showing that it preserves user confidentiality, ensures model integrity, and operates efficiently on resource-constrained IoT devices.