HVAC optimization in residential buildings

HVAC end-users strive for comfort, but they are usually unaware of the energy consumption and GHG emission required to guarantee such comfort. End-users’ ignorance leads to energy waste, high energy costs, and unnecessary GHG emissions.

Wonkit has developed AI algorithms for HVAC optimization in residential buildings, working in partnership with Elettrone, a company specialized in system integration, smart-home device placement and installation, and home automation.

Wonkit AI algorithms support end-users in two ways:

  • reporting inefficiencies and wrong behaviours e.g. daily, weekly, monthly;
  • recommending actions that end-users can execute in real time (e.g. turn off HVAC systems when energy waste is detected, turn on HVAC systems when comfort is decreasing, program the thermostat based on desired comfort as well as renewable energy production, etc.).

AI Algorithms

(1) Inefficiency Detection.  Inefficiency Detection timely identifies suboptimal behaviours in HVAC usage, and alerts end-users in real time. The following inefficiencies are identified:

  • energy waste
  • lack of comfort
  • HVAC system fault or wrong usage (e.g. the HVAC system is on while windows are open)

(2) Low Comfort Prediction. Low Comfort Prediction turns on when comfort is below a threshold (alert threshold) and predicts whether comfort will drop below a second threshold (tolerance threshold). An alert is launched when the tolerance threshold is exceeded.

(3) Consumption Estimation. Consumption Estimation quantifies energy savings in case the HVAC system is turned on before the tolerance threshold is overcome (e.g. when Low Comfort Prediction launches an alert). The lower energy efficiency of the building the higher energy savings.

(4) Saving Estimation. Saving Estimation quantifies energy savings in case energy waste is detected and the HVAC system is turned off accordingly (e.g. when Inefficiency Detection launches an alert). The higher energy efficiency of the building the higher energy savings.

The output of the algorithms is used to both issue real-time alerts and to provide insights through periodic reports. Samples of daily, weekly and monthly reports can be viewed here.

Technology stack

The AI algorithms and supporting logic have been developed using Python with widely used open source libraries for data processing, machine learning and deep learning (NumPy, TensorFlow, …). The whole software has been deployed on Amazon Web Services and interconnected to the Elettrone stack running on Microsoft Azure. Specific solutions have been employed for both on-demand services (e.g. generation of reports) and real-time functionalities (e.g. issuing alerts to end users), as outlined in the following diagrams.

Diagram of architecture for on-demand services

Diagram of architecture for real-time services