ScaleONet: Scalable and control-oriented modeling of building cluster thermal dynamics using deep operator networks — A practical case study for a Belgian district

Delivering energy flexibility at the district scale entails coordinating control actions across many buildings to shape aggregate demand; this coordination depends on training and deploying control policies and optimization routines, which in turn require predictive models that can be queried efficiently over large building clusters. However, conventional physics-based simulators are computationally prohibitive for large-scale control training, and simple data-driven surrogates often lack the generalization needed for heterogeneous clusters. This paper introduces ScaleONet , a deep operator network framework designed for scalable, control-oriented modeling of building-cluster thermal dynamics. ScaleONet leverages the DeepONet paradigm to decouple and share learning across buildings: an LSTM-based branch network encodes outdoor climate and individual HVAC control signals, while a multilayer perceptron (MLP)-based trunk network embeds prediction timestamps, enabling fast predictions for growing clusters with negligible extra cost for each additional building or timestep. To the authors’ knowledge, this is the first operator-learning method tailored to indoor air temperature forecasting in heterogeneous building clusters. Validation on thirty Belgian buildings (GenkNet) simulated in Dymola shows that, although a non-operator-learning LSTM baseline slightly outperforms ScaleONet for single-building cases, its error grows monotonically with cluster size. In contrast, ScaleONet’s median per-building-per-day RMSE decreases from 0.59 °C at three buildings to 0.53 °C at ten and 0.47 °C at thirty, compared to 0.95 °C for the LSTM at thirty buildings — a 51% reduction in prediction error. Error analysis across envelope heat-loss coefficients ( UA building ) further reveals that while the LSTM’s RMSE increases for high- U A structures, ScaleONet maintains uniformly low error. With millisecond-scale inference (approximately 4 ms per sample for thirty buildings), ScaleONet is well suited for large-scale reinforcement learning, receding-horizon optimization, and real-time model predictive control. • First DeepONet surrogate for day-ahead thermal forecasting in diverse buildings. • Scalable branch–trunk model generalizes from 1 to 30 buildings with 50% lower RMSE. • Predictive accuracy stays robust across varied heat-loss rates and data resolutions. • Millisecond inference (4 ms/sample for 30 buildings) enables real-time control. • Interpretable error patterns reveal physically consistent feature influences.