Privacy-Preserving Federated Learning with Causal Multiscale Decomposition for Regional Short-Term Load Forecasting in Smart Grids

Accurate regional short-term load forecasting is essential to the reliable operation and efficient energy scheduling of smart grids. However, the data needed for collaborative model training are often distributed across regional utilities and grid operators, creating data silos that limit cross-regional learning. To address these challenges, we propose a hybrid regional load forecasting model, EMA-GRU-N-HiTS, and further extend it into a privacy-preserving federated learning framework, FL-EMA-GRU-N-HiTS. Specifically, a causal additive exponential moving average (EMA) filter bank decomposes each historical load window into low- and high-frequency component groups without using future observations. The low-frequency components are modeled by a gated recurrent unit (GRU), whereas the high-frequency components are learned by an adapted neural hierarchical interpolation for time-series forecasting (N-HiTS) network. In the federated setting, aggregation weights jointly reflect the client sample size, local training error, and update similarity. Paillier encryption enables fixed-point parameter aggregation while protecting individual uploads from an honest-but-curious cloud server under a trusted, non-colluding third party (TTP). Experiments on regional load data from the New York Independent System Operator (NYISO) show that both methods achieve the lowest test errors among the evaluated baselines across the four forecasting horizons. Under sample imbalance, contribution-aware aggregation mitigates prediction degradation caused by the evaluated label-scaling and model-update attacks. Encrypted aggregation incurs substantial computational and communication overhead, while small-scale CPU parallelism reduces the cryptographic stage time.