Thermal comfort-constrained HVAC flexibility prediction in institutional buildings using hybrid physics and LSTM modelling

• Hybrid physics–ML framework enables scalable HVAC flexibility forecasting. • Thermal comfort constraints are integrated into demand-side flexibility modelling. • Institutional buildings can reliably provide short-term demand response capacity. • Demonstrates efficient integration of legacy buildings into flexibility markets. The commercial building sector presents a significant opportunity for demand-side flexibility, particularly in HVAC systems, which account for a substantial portion of peak electricity consumption. However, in Australia, the integration of flexibility strategies remains limited, partly due to data constraints in legacy buildings. This study presents a hybrid modelling framework that integrates a calibrated physics-based energy simulation with a long short-term memory (LSTM) neural network to forecast 24-hour thermally constrained HVAC power flexibility in an institutional building using aggregate energy meter data. A five-storey university building in Melbourne served as the case study. The simulation model, validated against measured data with CV(RMSE) = 10.90 % and NMBE = –6.94 %, was employed to generate flexibility profiles under multiple setpoint adjustment scenarios. Results show that ±1 °C deadband adjustments can provide substantial daily flexibility without compromising comfort, whereas larger set-point changes fail to maintain thermal comfort. However, a ±2 °C adjustment over only a two-hour weekday event reliably meets the 5 kWh grid flexibility target year-round. The optimal LSTM configuration: 256 units, 0.1 dropout, 0.0005 learning rate, 40 epochs, and a 4-day input window achieved a validation RMSE of 0.0981, and could reliably predict the flexible power availability, while the total HVAC consumption and diurnal temporal cycles were identified as the most influential predictors of the available flexible power. The proposed approach demonstrates a scalable, data-efficient pathway for integrating institutional buildings into flexibility markets, offering tangible load reduction potential while maintaining thermal comfort. These findings provide actionable insights for building operators and support the transition towards grid-interactive, decarbonised building operations.