Download as pdf or txt
Download as pdf or txt
You are on page 1of 21

Taking Control of the Chiller Plant

with Artificial Intelligence and


Machine Learning

28 July 2022

1
Why we need a chiller plant optimiser?

• COP26 conference in Glasgow, limiting CO2 emissions-driven global


warming is one the biggest challenges of our time
• Individual Pieces of Equipment are efficient in their own rights
• Variable Speed Drives (VSD) have been introduced to improve the
part-load efficiency of chillers, pumps, and cooling towers

For this reason you need a ‘conductor’ to carefully consider each aspect of
the plant solution and how tweaking and adjusting one part will affect the
other parts of the system.

2
Where does the chiller plant sit within the operation?

• Chiller plant is the ‘heart


beat’ of the overall building
operation
• Optimisation and Energy
Performance starts here
• Downstream systems can
operate optimally if consistent
performance is achieved from
the chiller plant

3
Typical building management system design

• Standard Building Management Systems (BMS) are generic in nature and not equipment or system specific
• BMS focus on the whole of building operation and brings all disparate building systems together
• Standard BMS are designed based on predefined rules and are not usually built to compare design
operational data vs actual onsite performance
• Chiller Plant is ‘just another’ system as part of the overall Building Management System

BMS

HVAC
Building Control

• Lighting Airside Chiller Plant


• Pumps
• Exhaust Fans • Fans
• Chillers
• Supply Fans • Valves
• Cooling Towers
• Fire Alarms • Dampers
• Pumps
• Sensors

4
Why a bespoke chiller plant system is required?

Chiller Plant has a number


of equipment and is often
spread throughout the
building
BMS Controls are
engineered for each
building design and the Focus needs to be
overall building placed past the initial
requirements installation to provide
proactive plant insights

Lack of Automation
and Control Expertise
with limited time
onsite due to
construction
timeframe pressures 5
What is iPlantManager?

iPlant Manager (iPM) is a dedicated plant room optimization solution and control software – capable of delivering efficiency
beyond the individual components
✓ Differentiates from traditional static BMS, achieve optimization in REAL TIME
✓ It utilizes on-board, real-time analytics, diagnostics and M&V system to continually readjust chiller plant for optimal
performance

Chiller Plant Chiller Plant Advanced Chiller


Performance Monitoring Analytics and FDD Plant Optimisation

Chiller Plant Chiller Plant


Chiller Plant
Measurement and Verification Continuous Commissioning and
Control and Automation
Tuning

PERFORMANCE RELIABILITY OPTIMIZATION 6


What is iPlantManager?

Hardware Dashboard Monthly Reports


• The software is embedded in

+ +
EDGE gateways hardware and
ready for on site data
integration

• Can control up to 20 chillers

Machine learning and Optimization algorithms have been implemented within iPM on an embedded computer specialised for
building automation to implement the below 2 key strategy’s:
1) Smart sequencing focuses on combining the optimal chiller configuration.
2) Smart Variable Flow optimization strategy achieves the balance between chiller input power and pump power

PERFORMANCE RELIABILITY OPTIMIZATION 7


iPM Optimisation Cycle using AI and ML

Optimisation platform
Output:
single values, or time series
Model Predictive Control (MPC)

Measurements
Onboard Local to iPM

AI-based Prediction Engine

Data preparation

Onboard Local to iPM iPM (Chiller Plant Controller) Control signals


Measurements Database

Measurements Control signals


Existing infrastructure on
site Water
Central Cooling System Buildings
Supply
Water Return

8
Machine Learning approach to the chiller plant
iPM learns the power usage of the Chillers, Pumps and Cooling Towers at
various inputs such as cooling loads, temperatures, or flows

Key target characteristics:


▪ Learn in real-time actual equipment performance from recorded site
dataset
▪ Run autonomously on site on our embedded controller
▪ Capture equipment performance degradation over time
Predicted saving graph with ML for a 1800kW Chiller

Machine Learning (ML) method:


Constrained least-square multivariate polynomial regression, with Interior-point
method to minimize the error between actual data and prediction Site A: recorded site data comparison between optimal and
standard CW Flow
▪ Embeds constraints to induce expected behaviour from domain expert
knowledge Case Load CW Chiller Pump Total
▪ Fast computations (kW) Flow Power Power Power
(L/s) (kW) (kW) (kW)
Optimal
325 52.1 30.4 5.3 35.7
(iPM)
Standard 322 80.9 29.1 17.0 46.1
Results in a saving of 10.4kW

9
Total chiller plant optimisation approach

Holistically review to reduce the power usage of the plant:


(I) (II) (III)
σ𝑛 𝑚 𝑝
Minimize 𝑃𝑜𝑤𝑒𝑟𝑝𝑙𝑎𝑛𝑡 = 𝑖=1 𝑃𝑐ℎ𝑖𝑙𝑙𝑒𝑟,𝑖 + σ𝑗=1 𝑃𝑐𝑤 𝑝𝑢𝑚𝑝,𝑗 + σk=1 𝑃𝑐𝑜𝑜𝑙𝑖𝑛𝑔𝑡𝑜𝑤𝑒𝑟,k
𝐿𝑜𝑎𝑑𝑖 , 𝐹𝑙𝑜𝑤 𝐶𝑊,𝑖 𝑇𝐿𝐶𝑇
𝑛
s. t. : ෍ 𝐿𝑜𝑎𝑑𝑖 = 𝐿𝑜𝑎𝑑𝑃𝑙𝑎𝑛𝑡,𝑇𝑎𝑟𝑔𝑒𝑡
𝑖=0
𝐿𝑜𝑎𝑑𝑀𝑖𝑛,𝑖 ≤
𝐿𝑜𝑎𝑑𝑖 ≤ 𝐿𝑜𝑎𝑑 𝑀𝑎𝑥,𝑖
𝐿𝐶𝑇𝑇𝑀𝑖𝑛 ≤ LCTT ≤ 𝐿𝐶𝑇𝑇 𝑀𝑎𝑥
𝐹𝑙𝑜𝑤𝐶𝑊,𝑀𝑖𝑛,𝑖 ≤ 𝐹𝑙𝑜𝑤𝐶𝑊,𝑖 ≤ 𝐹𝑙𝑜𝑤 𝐶𝑊,𝑀𝑎𝑥,𝑖

Decision variables: Leaving


Cooling Key constraints:
Tower
Temperature ▪ Maintain overall plant cooling load and leaving chilled water
setpoint Condenser temperature as per design
Water Flow
setpoint ▪ Maintain equipment within their safe operating zone
Cooling ▪ All relevant pieces of equipment in the system are considered
Load per
and accounts for interrelations between the sub-systems
Chiller

*Note: In this approach, the plant-level cooling production is not affected, and therefore nor the CHW pumps or AHU fan power usage 10
How is the optimisation cycle completed?

Measure the key data points and


calculate the efficiency of each
chiller and ancillary devices.
Record
Check the instantaneous
efficiency against the
Model manufacturers design data
and best case efficiency
Analyse the data and run in a modeling
diagnostics engine to
determine the cause of any
gaps between actual, design Identify
and best case chiller and plant
efficiency

Adjust the plant through

Optimise proactive automation to correct


back to optimized plant
efficiency.
11
Onboard dashboard to view AI and ML optimisation

1. Interior-point method for


Non-Linear Programming
selected as fast-converging
solver for both:

• Machine Learning

• Optimization

2. Implemented in the
commercial product with
full-suite of GUI and
flexible configuration for
various chiller plant
configurations

12
Case Study: Site Energy Savings

Sites Variables Savings (kWh) Savings (%)


Site A CW Flows 15,948 3.6
CW Flows and
Site B 7,358 7.1
Loads
Site C Loads 4,056 5.1

Daily Plant COP at Site B conventional controls period compared Instantaneous Plant COP at Site C with conventional controls period
to optimal controls period, from measured data compared to optimal controls period, from measured data

Measurement and verification (M&V) studies were carried out, following the guiding principles from the International Performance Measurement and Verification Protocol (IPMVP) to
assess the actual energy savings achieved while accounting for changes in conditions appropriately 13
How do you achieve total plant performance?

iPM is wizard based, Compatibility with any


with simplified GUI chiller brand Integration to the
iBMS Onsite

Daikin Chiller Plant


Manager Solution
Compatibility with
Highly configurable Touchscreen within iPM Controller
multiple Plantroom for
Implementation and Maintenance
communication Techs/ Engineers
network design
protocols
IP Network

BACnet/Modbus

Embedded Intelligent
Proactive service and Low Level IO
control on-board Connections to
maintenance Field Devices
without the need of
diagnostic
Cloud Connectivity
Chiller 1 Chiller n
Pump 1 Pump n
CT 1 CT n

14
Case Study Recent: Hotel
• Plant initially commissioned in 2014, running at an efficiency of ~0.9kW/RT,
• No major capital investment – iPM Controller Integrated to existing BMS Onsite
• Once iPM was installed 1 week of operation, Plant efficiency jumped ~0.63kW/RT, with improved comfort and operation

Original Solution: iPM Outcomes:


Plant Initially Commissioned 2014 Ⅱ Ⅲ
I. Plant efficiency ~0.9kW/RT
-> 2x Daikin WC (120RT and 200RT) I
iPM
-> 2x CHWP (12kW and 30kW)
II. Installation of iPM
-> 2x CWP (2 x 4kW)
III. iPM identified System Flow Ⅳ
-> 2x CT (15kW or 1307 kWr)
Issues through the inbuilt
-> BMS Schneider Electric
FDD which has now improve
comfort and operation for
the field FCUs

IV. Plant efficiency jumped to


~0.63kW/RT, with optimized
sequencing of chillers and
overnight operation for 24/7
plant operation
15
Case Study Recent: Hospital and Commercial Building

Sites Variables Duration Savings kWh Savings %


Site A LCTT 1 week 93 kWh 9.20 %
Site B LCTT, CW Flow, and Loads 6 months 174,166 kWh 18.75 %

Site B: Daily actual energy usage at Site B compared to M&V baseline


representative of previous operation and performance
Site A: Instantaneous 5-minutes Plant COP at Site A – limited optimized
period compared to optimal controls period, from measured data

16
iPlant Manager in summary

o Reduction in Electricity Bill

o Longer Life of Chiller Plant

o Less Human Resource Requirement

o Automated and Simplified Diagnostics, Analytics, Fault


Identification, Tuning, Measurement & Reporting

o Reduced Complaints from Building Users

17
THANKYOU
Questions please ?

18
(I) Learning the performance of chillers

Chiller power model equation:


𝑃𝑜𝑤𝑒𝑟𝑐ℎ𝑖𝑙𝑙𝑒𝑟 = 𝑃(𝐿𝑜𝑎𝑑, 𝑇𝐸𝐶𝑊 , 𝑇𝐿𝐶𝐻𝑊 , 𝐹𝑙𝑜𝑤𝐶𝑊 )

Table: Chiller learning results with site data


Site A Site B Site C
Chiller n° 2 1 2 3 1 2
MAE (kW) 4.1 2.6 2.8 3.9 11.0 10.6
MAE/Mean 4.6 % 2.4 % 2.4 % 3.4 % 6.5 % 6.4 %
R-squared 0.949 0.990 0.984 0.979 0.949 0.938

Site B, Chiller 1 predicted power compared to the actual measured power Site B, Chiller 1 predicted COP from chiller model for a range of conditions

19
(II) Learning the performance of condenser pumps

Condenser pump power model equation:


𝑃𝑜𝑤𝑒𝑟𝐶𝑊 𝑝𝑢𝑚𝑝 = 𝑃(𝐹𝑙𝑜𝑤 𝐶𝑊 , 𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑐ℎ𝑖𝑙𝑙𝑒𝑟𝑠, 𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑝𝑢𝑚𝑝𝑠)

Table: Condenser pump learning results with site data


Site B Site C Site D Site E
1, 2 & 3 1, 2 & 3
Pump n° 1 2 1 2
averaged averaged
MAE (kW) 0.7 0.7 0.5 0.6 0.3 0.6
MAE/Mean 4.7 % 2.7 % 2.4 % 4.8 % 3.5 % 2.7 %
R-squared 0.241 0.891 0.986 0.978 0.988 0.971

Site C, pump n°1 predicted power compared to the actual measured power

20
(III) Learning the performance of cooling towers

Evaporative Cooling Tower model equation:


𝑃𝑜𝑤𝑒𝑟𝑒𝑣𝑎𝑝𝑜𝑟𝑎𝑡𝑖𝑣𝑒−𝑐𝑜𝑜𝑙𝑒𝑟 = 𝑃(𝐻𝑒𝑎𝑡𝑟𝑒𝑗𝑒𝑐𝑡𝑒𝑑 , 𝑇𝐿𝐶𝑇 , 𝑇𝑂𝐴,𝑤𝑒𝑡𝑏𝑢𝑙𝑏 )
Non-evaporative Cooling Tower (also known as “Dry-cooler”):
𝑃𝑜𝑤𝑒𝑟𝑑𝑟𝑦−𝑐𝑜𝑜𝑙𝑒𝑟 = 𝑃(𝐻𝑒𝑎𝑡𝑟𝑒𝑗𝑒𝑐𝑡𝑒𝑑 , 𝑇𝐿𝐶𝑇 , 𝑇𝑂𝐴,𝑑𝑟𝑦𝑏𝑢𝑙𝑏 )

Table: Cooling Towers learning results with site data


Site A Site B Site C Site E Site F
Type Evap. Dry Evap. Evap. Evap.
Rated Power (kW) 30 132 12.5 380 5.6
MAE (kW) 1.1 7.9 1.0 10.9 0.2
R-squared 0.829 0.627 0.714 0.919 0.701

Site C, cooling tower predicted power compared to the actual measured power Site C, predicted power from cooling tower model for a range of conditions

21

You might also like