A Cross-Regional Load Forecasting Method Based on a Pseudo-Distributed Federated Learning Strategy

Accurate load forecasting serves as the core foundation for grid planning and operations. Traditional load forecasting methods often rely solely on historical load data from a single region for training, making the models region-specific and leading to significant accuracy degradation when applied to other regions. This limits the generalization ability of these models to cross-regional load forecasting tasks. To address this issue, this study proposed a collaborative training strategy based on pseudo-distributed federated learning. Inspired by the pseudo-distributed concept, this strategy builds multiple sub-models by serially training load datasets from different regions on the same server. After a certain number of local epochs for each sub-model, parameter aggregation was performed. The aggregated parameters are then updated into each sub-model, and this process is repeated during each global epoch until the model converges, ultimately forming a global model capable of forecasting loads across multiple regions. Experiments demonstrated that this strategy exhibited exceptional generalization ability across various deep learning models, federated learning methods, and datasets.