Personalized Short-Term Load Forecasting for Building Clusters Based on Federated Learning and Transformer Model

Buildings are the main power consumers on the load side of power systems, and accurately forecasting building loads is crucial for proactive decision-making and maintaining stable operation of power systems. In the context of building cluster load forecasting, data is distributed across multiple building clients, making centralized model training a risk to data privacy. Additionally, building load data often exhibits randomness, volatility, and nonlinearity. To address these challenges, this paper proposes a method that combines federated learning and the Transformer model. The aim is to establish a personalized load forecasting model for building clusters that preserves data privacy while capturing the complex temporal dependencies and nonlinear features of building energy data, thereby enhancing the accuracy of energy consumption forecasting for building clusters. First, a federated learning framework is utilized to aggregate local model parameters from each building, resulting in a global forecasting model. Throughout the process, the Transformer model is employed to extract features from building load data. To further enhance the model's personalization, local data from each building is used to fine-tune specific Transformer layers, thereby reducing computational costs and improving both the model's generalization and predictive accuracy. Experimental results show that this approach effectively balances building load data privacy with forecasting model accuracy.