GridPulse: A Federated Learning Architecture for Decentralized Real-Time Energy Load Forecasting in Microgrid Networks

Accurate and privacy-preserving energy forecasting has become increasingly critical as electric power systems transition toward decentralized, data-constrained architectures. This paper presents GridPulse, a federated learning (FL) framework designed to enable Nigeria’s Distribution Companies (DISCOs) to collaboratively train predictive models without exchanging raw customer data. Using a cleaned, leakage-controlled DISCO-level dataset and a multi-stage evaluation pipeline, we compare centralized LightGBM forecasting against a federated multilayer perceptron (MLP) model trained using FedAvg and FedAdam. Results show that the federated global model reconstructs approximately 85–90% of the predictive capacity of the centralized benchmark, with only a modest degradation of 2–3% in RMSE across all DISCOs. Localized fine-tuning provides minimal improvements, indicating that the global FL model generalizes well across heterogeneous clients. Because the real DISCO dataset lacks temporal continuity, a synthetic 60-month DISCO panel was generated to evaluate long-horizon forecasting. On this extended benchmark, federated learning again performs at parity with centralized LightGBM. Collectively, these findings demonstrate that FL introduces no performance disadvantage for any DISCO, preserves data privacy, maintains equitable participation, and is suitable for supporting forecasting, optimization, and automation tasks in decentralized microgrid networks.