Physically Consistent Neural Networks for building thermal modeling: theory and analysis

Near-optimal Deep Reinforcement Learning Policies from Data for Zone Temperature Control

“…While classical NNs have been proven to be very effective function approximators in various fields, they remain agnostic to the underlying physical laws when modeling physical systems and might fail to grasp fundamental principles [14]. For example, when modeling the temperature in a zone, turning the heating on should always have the effect of increasing the temperature, as DRL agents could otherwise learn spurious behaviors.…”

“…For example, when modeling the temperature in a zone, turning the heating on should always have the effect of increasing the temperature, as DRL agents could otherwise learn spurious behaviors. To still leverage the expressiveness of NNs while retaining physical consistency with respect to some inputs, PCNNs were proposed and analyzed in [14], and we refer the reader to the original paper for the details. The main feature of PCNNs is that they are physically consistent with respect to power inputs, external temperatures, and temperatures in neighboring zones by design.…”

“…It consists of two bedrooms separated by a living room, all powered by heating/cooling panels letting hot/cold water flow through their ceiling. More than three years of data is available and was processed as explained in [14], including the normalization procedure.…”

“…Moreover, [10] trains discrete DRL agents and [13] assumes access to solar photovoltaic electricity generation and battery storage. Furthermore, in this paper, we use Physically Consistent Neural Networks (PCNNs) [14] as simulation environments instead of first principles models. PCNNs are trained from historical data, hence bypassing the complex design stage of physics-based methods, but remain physically consistent with respect to control inputs.…”

Experimental Data-Driven Model Predictive Control of a Hospital HVAC System During Regular Use

“…The experimental section of the work however only reports real experiments on a single zone over a total period of 24 hours. Feedforward neural netwokrs (NN) are another popular machine learning tool among certain researchers in the building and HVAC fields [19][20][21][22][23]. Given an adequate architecture, these models certainly have great representation capabilities and are nowadays rather straightforward to train thanks to the availability of reliable frameworks such as Pytorch [24] and TensorFlow [25].…”