Multi-agent system of an energy building for prediction and learning activities

The rise in energy prices has intensified the need to improve the energy management in office buildings. To address this challenge, building owners often require forecasting techniques capable of supporting energy-management decisions. Forecasting can help reduce energy consumption during peak periods and support optimization strategies that improve the balance between electricity demand and supply. Several forecasting algorithms can be considered for this purpose, with different advantages and limitations, including Artificial Intelligence methods such as Artificial Neural Networks (ANN) and Long Short-Term Memory (LSTM), Machine Learning methods such as K-Nearest Neighbors (KNN), ensemble learning methods such as eXtreme Gradient Boosting (XGB) and Light Gradient-Boosting Machine (LightGBM), and time-series forecasting methods such as Prophet. This paper presents a multi-agent system designed to address forecasting activities in an office building. The proposed architecture enables multiple stakeholders and specialized agents to interact while encapsulating the procedures performed by each agent For example, the building owner can request forecasts from a forecast provider without requiring access to the internal forecasting procedures performed by the forecasting agents. Other specialized agents are responsible for preparing the dataset for forecasting, calculating forecasting errors, and selecting forecasting algorithms. In addition, a learner agent uses a multi-armed bandit approach to dynamically select the forecasting algorithm for each five-minute period. The proposed architecture therefore combines modular forecasting agents, error analysis, and learning-based algorithm selection within a distributed multi-agent system.