Smart building vendors love big numbers. Marketing decks routinely promise 30 to 55 percent energy savings from advanced HVAC controls. The peer-reviewed field evidence tells a more modest, and more useful, story: about 13 percent average cost saving in commercial buildings. That gap is the subject of this article, the first in a three-part series on energy optimization algorithms and demand response.
Two ways to run a smarter building
Model predictive control (MPC) works from a model of the building. It combines that model with forecasts of weather, occupancy and electricity prices, then optimizes setpoints over a rolling future horizon, re-solving every few minutes while respecting comfort limits. Like a chess player, it plans ahead and revises as conditions change.
Reinforcement learning (RL) takes the opposite route. Instead of a model, an RL agent learns a control policy by trial and error, maximizing a reward for energy and cost savings minus a penalty for discomfort. It needs no explicit model, which makes it attractive where models are poor or loads drift over time.
Both beat conventional rule-based control in most published studies. The question is by how much, and how reliably.
What the field evidence actually shows
The most comprehensive reality check is a 2025 review by Khabbazi and colleagues in Applied Energy, examining 104 peer-reviewed field demonstrations of MPC and reinforcement learning for heating, ventilation and air conditioning (HVAC) in residential and commercial buildings. Their finding is sobering: 71 percent of those demonstrations used experiment protocols that may produce unreliable savings estimates, short tests, cherry-picked periods, or weak baselines. Among the remaining 29 percent judged reliable, the duration-weighted average cost saving was 13 percent in commercial buildings (16 percent in residential buildings).
That 13 percent is the honest number to plan around: real money on a large commercial building's energy bill, with comfort maintained, not sacrificed.
Individual trials can show much larger numbers. In a real five-month testbed, Wang and Dong (Applied Energy, 2024) compared three advanced controllers head to head: a hybrid data-driven predictive controller saved more than 50 percent of HVAC energy versus the baseline, reinforcement learning saved 48 percent, and a data-driven predictive controller saved 30.6 percent. Impressive, but one testbed is not a portfolio average.
The longest-running demonstration comes from Lawrence Berkeley National Laboratory (LBNL). From 2020 to 2025, an MPC system ran the HVAC of two floors (65,000 square feet) of Building 59, a large office building in Berkeley, California. It achieved 45 percent energy savings in efficiency mode and an estimated 61 percent annual cost reduction under experimental dynamic prices, far more volatile than any real tariff. The result shows what MPC can do when prices reward flexibility; it is not a typical outcome under today's tariffs.
Why simulations overpromise
Simulation studies routinely report 30 to 55 percent savings because they assume perfect weather forecasts, ideal equipment, and a weak baseline controller. Real buildings have faulty sensors, drifting equipment and unpredictable occupants, and their baselines are often already decent. Field trials also run for months, across seasons, while simulations cherry-pick favourable weeks.
Which should you choose?
On the balance of field evidence, MPC leads today: its model-based constraints make its behaviour predictable, auditable and safe. Reinforcement learning leads on adaptability: where no good model exists, or where loads change over time, an agent that keeps learning has the edge. The most promising results, like the hybrid controller in the Wang and Dong testbed, combine the two.
The practical takeaway is simple. Judge an algorithm, or a vendor, on field evidence, not on simulated percentages. Ask for the test duration, the baseline it was compared against, and the tariff the savings assumed. A vendor quoting 50 percent should be able to show you a field trial.
Platforms such as Tenergen OS take this evidence-first approach: they coordinate HVAC, storage and EV charging through the building's existing equipment, using open protocols such as BACnet and OpenADR, so promised savings can actually be measured against the meter.
In part two, we rank the demand response techniques themselves, from battery dispatch and thermal storage to automated demand response, and look at which ones deliver the largest verified peak cuts.