An adaptive multi-agent system for dynamic detection and prediction of daily energy profiles in buildings

Efficient building energy management increasingly relies on data-driven techniques to optimize control strategies and reduce energy consumption while maintaining thermal comfort. To achieve this, building energy managers need to identify operating patterns that characterize building behavior across seasons and operation conditions, and support predictive control in non-stationary environments. This work introduces the Adaptive Multi-agent Profile Detection and Prediction (AMA-PDP) framework, an adaptive Multi-Agent System (MAS) that integrates dynamic daily profile detection with incremental day-ahead profile prediction. AMA-PDP continuously operates on streaming data and automatically identifies distinct 24-hour operational energy consumption profiles using Dynamic Time Warping (DTW) similarity analysis. The framework also adapts the clusters over time in response to changes in building operation. In addition to these dynamically detected profiles, an Incremental One-vs-Rest Neural Network ensemble predicts the next day’s cluster label using recent heating consumption data, weather variables, and temporal features. The AMA-PDP method was evaluated on both simulated and real-world data from office buildings located in different climates. The clustering process adapts to each case, with the number of clusters evolving over time to reflect new operating configurations. For the prediction task, the Incremental One-vs-Rest Neural Network outperforms a batch neural model and three incremental baseline methods (Stochastic Gradient Descent (SGD), Naive Bayes, and Hoeffding Tree), achieving higher average accuracy, lower variability across updates, and fewer cluster misclassifications. By focusing on daily energy consumption profiles rather than hourly loads, AMA-PDP is particularly well suited for scenario-based control of HVAC (Heating, Ventilation and Air-Conditioning) systems.