Multi-Gradient-Descent Federated Learning with Parity for Cooperative Short-Term Load Forecasting

Short-term load forecasting (STLF) is essential in power system research. In scenarios where multiple distribution system operators are involved in cooperative forecasting, it is important to uphold the fairness of multi-party cooperation and drive the parity of model gains among parties under the premise of privacy protection. In this context, this paper proposes a short-term load forecasting method namely Multi-Gradient-Descent Federated Learning with Parity Long-short Term Memory (MGD-FLP-LSTM) under a federated learning framework. First, MGD-FLP-LSTM transforms cooperative STLF into a Multi-Objective Optimization (MOO) problem with parity-driven ob-jective. The designed algorithm optimizes multiple objectives simultaneously and incorporates Karush-Kuhn-Tucker (KKT) conditions in solving the dual form of the problem, thus acceler-ating the global model convergence. Second, MGD-FLP-LSTM constitutes a parity-driven objective with cosine similarity in the MOO iteration. The objective allows all local loss functions to have similar descents on the common gradient descending direc-tion. Third, MGD-FLP-LSTM constructs a Performance Pari-ty Control (PPC) scheme that enables an active moderate trade-off between forecasting accuracy and parity. Experiments on real-measured load dataset highlight the method's enhanced ac-curacy, efficiency, and performance while maintaining consumer privacy.