<p>Buildings and heating, ventilation and air conditioning (HVAC) systems are recognized as effective flexibility resources that can interact with the power grid to reduce electrical peak loads. Model predictive control (MPC) is a powerful approach for fully unlocking the energy flexibility of buildings. However, MPC relies on online optimization for practical engineering deployment, which imposes a significant computational burden and limits its widespread adoption. To address the challenge of computational burden, this study proposes a machine learning-enhanced lightweight rule-based control strategy (ML-RBC). The main idea of ML-RBC is to use a machine learning (ML) model to automatically tune the adjustable parameters of the rule-based controller (RBC). Specifically, the ML model learns the functional relationship between external inputs and adjustable parameters from a dataset generated by batch offline closed-loop MPC simulations. The proposed method retains inherent high computational efficiency of RBC while also achieving optimal control performance. The demand response (DR) control performance of the proposed method is evaluated using a high-fidelity co-simulation platform that integrates Spawn of EnergyPlus and Modelica. Simulation experiments are performed on a multi-zone office building equipped with a variable air volume (VAV) cooling system under time-of-use electricity pricing and day-ahead DR programs. The experimental results indicate that, compared to the baseline strategy, ML-RBC and traditional MPC achieve cost savings of 21.95% and 23.07%, respectively. Importantly, ML-RBC eliminates the need for online optimization while achieving a computational cost of less than one-thousandth that of MPC, with only a slight performance loss as the trade-off. Finally, the impact of the trajectory interpolation method in ML-RBC on control performance is discussed, revealing that different interpolation methods have a minor influence on the overall performance.</p>