Article

Why an AI heating experiment needed to model showers

For homeowners, a Madrid trial shows how showers challenged AI heating—and why comfortable rooms are not proof of lower energy bills.

Editorial illustration for Why an AI heating experiment needed to model showers: a geometric block represents a model release. Not documentary evidence.

Showers with the bathroom door open exposed a gap in an AI heating experiment: the virtual home did not reproduce moisture spreading between rooms. In a preprint submitted on 17 November 2026, researchers describe revising that model and report generally adequate temperature control in one occupied Madrid house. For homeowners considering intelligent heating, this is evidence of a limited working deployment—not of lower bills.

The work comes from Antonio Manjavacas and colleagues affiliated with the University of Granada and Georg Fischer Building Flow Solutions. Their real-home evaluation ran from 24 March to 5 April 2026, so the new development is the paper, not an experiment conducted this week. It documents both a working controller and the engineering needed to accommodate the people living with it.

Why the virtual home needed showers

The researchers built a digital twin: a computer model of the home, including its construction, heating system and occupancy schedules. They used reinforcement learning, in which a controller practises decisions and receives feedback on the results. Here, the training objective balanced indoor comfort against energy consumption. The learning took place in simulation before deployment in the house.

But a floor plan and an occupancy schedule do not capture everything people do. The authors report that moisture from showers, spreading through open bathroom doors, produced indoor humidity patterns absent from the initial model. Their analysis also identified indoor humidity as an influential input to the learned controller. A mundane bathroom routine was changing information the system used to make decisions.

The researchers added sudden moisture releases and the spread of humidity between neighbouring spaces to the model. They report that this narrowed the mismatch with the house. The AI did not diagnose or repair the omission itself, and the paper does not isolate a numerical comfort improvement attributable to that correction. This was about representing shower moisture, not controlling the shower.

The same deployment discussion records other adjustments: two measurements available in simulation were unavailable from the installed sensors, requiring retraining; local humidity and sunlight measurements also differed from the weather data used for training. The authors revised the simulated conditions. Their statement that deployment required no additional online fine-tuning therefore does not mean the system worked unchanged out of the box. Revising a model offline is different from having the controller learn while running the heating.

The practical lesson from these findings: a useful virtual home needs to represent how people inhabit it, as well as its walls and equipment. That is a different task from generating convincing views of a simulated scene—a distinction explored in our guide to choosing tools for virtual robot trials. Here, the missing detail was an everyday household activity.

What the controller managed in Madrid

The house used an air-to-water heat pump feeding water through underfloor heating circuits. The controller selected the water supply temperature and the positions of the circuit valves. According to the field results, it generally maintained the occupants’ chosen temperature ranges, with brief excursions in the bathroom-dressing area less than 1°C above the upper limit.

Water supply temperatures clustered around 34°C, and control decisions came every 30 minutes. Those are characteristics of this installation, not recommended settings for every home. Most importantly, the trial covered mild late-winter and early-spring conditions, not the coldest months. It does not establish equivalent performance across homes or through a severe winter. The authors identify seasonal testing as unfinished work.

The paper also lacks a comparable field baseline demonstrating energy savings. A comfortable home and a cheaper-to-heat home are separate results. In the authors’ simulated tests with degraded building conditions, preserving comfort actually required more energy to compensate for greater heat loss. That is a useful reminder of why stable temperatures alone cannot prove lower bills.

Why a shut valve did not stop rooms warming

Some rooms continued warming even when their heating valves were closed. The researchers chiefly attribute this to sunlight, particularly in the exposed living area during the Madrid spring. That is their explanation of the observations, not a separately controlled test isolating the solar contribution.

Turning off heat input is not the same as cooling a room. A controller can close a valve; it cannot use that valve to switch off sunshine. Nor does a warm floor instantly lose its stored heat. This makes an above-target temperature more complicated to interpret than a simple verdict that the AI chose the wrong command.

People also remain part of the control system. The authors describe successive manual thermostat adjustments while occupants waited for the slow floor heating to respond. For a household product, the implication is to design around that delay and ordinary human intervention, rather than assume residents will behave like fixed entries in a simulation.

There was a more basic obstacle, too: delivering the command. The cloud-hosted controller used smart relays, and the paper reports unreliable Wi-Fi communication associated with a recessed metal control enclosure. A disconnected relay could postpone a valve update until the next decision. Choosing an action and physically carrying it out need separate checks—the distinction behind ToyBlocks’ tests of robot planning and movement.

What homeowners and installers should ask next

For a proposed heating pilot, this study suggests three separate questions. Does it maintain acceptable temperatures through normal routines and seasonal changes? Do the valves actually carry out the commands? And does it use less energy than ordinary control under comparable conditions? Success on one question should not be used as the answer to another.

A meaningful comparison would need to account for weather, occupancy and heat already stored in the floor. The authors explain why rapidly alternating controllers can distort a comparison: the next controller inherits thermal conditions created by the previous one. Their planned longer-term comparisons remain future work, not savings already demonstrated.

The public Sinergym framework connects learning agents to building simulations; downloading it is not the same as installing this household controller. The cited sources do not establish a consumer purchase route, installed price or general equipment compatibility for the Madrid system.

The shower detail is what makes this experiment useful beyond the lab. A home is not merely a building with a target temperature. It is a place where doors stay open, people adjust thermostats and sunlight arrives on its own schedule. A convincing heating pilot has to work with that household—and measure any promised savings separately.