HVAC Diagnosis and Predictive Analysis

HVAC.

Heating, ventilation and air conditioning (HVAC) systems account for up to 50% of energy costs of a typical building in the United States. These complex systems require expert knowledge to maintain, and optimizing their performance can depend on factors that are unrelated to the systems themselves. 

Overview

Heating, ventilation and air conditioning (HVAC) systems account for up to 50% of energy costs of a typical building in the United States. These complex systems require expert knowledge to maintain, and optimizing their performance can depend on factors that are unrelated to the systems themselves. Factors may include the environment, the systems use profile, the cost of energy and interactions with other systems. Intelligent scheduling has been used to significantly reduce energy costs; systems running at peak efficiency can result in further savings and extend their life.

Unfortunately, it is difficult to optimize traditional hand-developed, rule-based approaches or personalize them for a given environment because of these non-linear complexities and variations. As an alternative, systems that gather data either periodically or continuously are candidates for more recent advances in statistical machine learning known as deep learning. Given data in the form of measurements over time from a variety of sensors, along with examples of what effect various changes have on the system, deep learning has the potential to help develop a personalized strategy for operating and maintaining these systems.

Deep learning techniques have been developed for a wide variety of classification tasks, but more recently have been used to approach complex optimization tasks.

The project is addressing the following tasks:

  • Carrying out a study of HVAC system data captured by the current and planned sensors to understand their predictive capabilities
  • Surveying the literature for current state-of-the-art approaches to HVAC diagnosis and optimization in the context of existing products
  • Creating a simulation tool based on the current rule-based system that can be used to generate additional data points and recommendation scenarios
  • Creating a data set for the training and evaluation of deep neural network (DNN) models
  • Developing a baseline deep learning framework in support of diagnosis and predictive analysis