Developing New Data-Driven Models for CHCP Boilers, Chillers
From eSociety, September 2019
Science and Technology for the Built Environment’s 2018 best paper award winner introduced a method to develop new data-driven models for central heating and cooling plant (CHCP) boilers and chillers and demonstrated the method on four boilers and five chillers of a CHCP in Canada.
Burak Gunay, Ph.D., Weiming Shen and Chunsheng Yang won the 2018 Science and Technology for the Built Environment Best Paper Award for “Blackbox Modeling of Central Heating and Cooling Plant Equipment Performance” at the 2019 ASHRAE Annual Conference in Kansas City. Gunay is an assistant professor at Carleton University in Ottawa, Ontario, Canada; Shen is a principal research officer at the National Research Council Canada in Ottawa; and Yang is a senior research officer at the National Research Council Canada.
The article presents an analysis conducted on sensor data gathered from the distribution control system of a CHCP plant in Ottawa, Canada. The performance of this plant’s four boilers and five chillers varied substantially in time under steady-state conditions, so the researchers developed data-driven models to explain the variability from the archived sensor data, according to the paper. By using a forward stepwise regression and a repeated random sub-sampling cross-validation approach, two-layer feed-forward artificial neural network models with 7 to 15 hidden-nodes were selected for each boiler and chiller.
Gunay further explores the paper and its significance.
1. What is the significance of this research?
Heating and cooling in many building clusters are served by central heating and cooling plants (CHCPs) whereby only a few chillers and boilers provide chilled and hot water or steam to many buildings. While failures in these systems can impact the comfort, health, and productivity of many occupants, improving the maintenance routines and the control sequences of these pieces of equipment have the potential to substantially improve the energy performance.
Data-driven models of plant-level equipment developed with the sensor data available in a distribution control system enable diagnostics, performance monitoring and benchmarking and predictive controls. In this paper, we introduced a method to develop new data-driven models for CHCP boilers and chillers and demonstrated the method on four boilers and five chillers of a CHCP in Ottawa, Canada.
2. Why is it important to explore this topic now?
Improvements to the operation of CHCPs represent an untapped opportunity to reduce energy use and GHG emissions as well as to improve its reliability. Building on this research, an equipment sequencing scheme was derived for this facility to determine the optimal combination of boilers and chillers in order to meet forecasted heating and cooling loads. While implementation of this equipment sequencing scheme at this plant is still underway, we anticipate achieving a 5% to 10% reduction in natural gas and electricity use.
Again, building on this research, we analyzed common operating conditions triggering frequent boiler and chiller tripping events. We plan to work with the controls service provider of this facility to modify the controls programs to reduce the frequency of these events.
3. Were there any surprises or unforeseen challenges for you when preparing this research?
During this research, we frequently interacted with the operations staff of the case study CHCP. This helped us better understand their day-to-day workflow and challenges. To our surprise, the operators of the case study facility (a facility that manages major energy resources of 22.9 MW of heating and 17.1 MW of cooling) were expected to undertake critical operational decisions with only limited analytical guidance. Our interactions with the operators highlighted the need for improved fault detection and diagnostics capabilities and predictive control algorithms for the integration of individual pieces of equipment and coordination of the supply and demand between the buildings and the plant.
A challenge that we observed was access to quality data from equipment in use—either because sensors were not present or the data from existing sensors were interrupted.
4. What lessons, facts and/or guidance can an engineer working in the field take away from this research?
Our research demonstrated that two-layer feed-forward artificial neural network models with seven to 15 hidden-nodes are capable of explaining up to 95% of the variability in a boiler’s efficiency and a chiller’s coefficient of performance. Top three data types for the boiler model were the flue gas O2, pressure, and part-load ratio; while they were the return water flow rate, part-load ratio, and outdoor temperature for the chiller model. It is important to trend these data types continuously and develop data-driven models based on them for diagnostic, control, and monitoring and benchmarking purposes.
5. How can this research further the industry’s knowledge on this topic?
The inverse blackbox modeling method presented in this research demonstrated how common data types in distributed control systems can be leveraged to improve the operation of CHCPs. Based on this research, we later developed a near-optimal equipment sequencing scheme and prognostics-oriented models to explain underlying reasons behind frequent equipment tripping events. Similarly, the research can be adapted by those that oversee the controls and maintenance of CHCPs and improved by industry researchers by using larger datasets from other facilities.