This project is to develop deep learning technology to enhance the building energy efficiency. The new technology can decide the optimal control policy based on the operational data on the integrated building system including smart-grid, air conditioning and mechanical ventilation (ACMV), solar, lighting and occupancy. It has the benefits of energy efficiency optimization, adaptation to equipment and operation conditions and robustness against environmental uncertainty, compared with the current state-of-the-art of model-based control, which highly depends on detailed domain knowledge and many restrictive assumptions. The final target is to achieve 20% energy saving and higher comfort level. The technology can be promoted island wide through a parallel AI cloud with great significance on energy sustainability and service quality.
Deep Learning Based Integrated Energy Efficiency Optimization for Smart Building
1. Deep Learning Based Integrated Energy Efficiency
Optimization for Smart Building
1
2018 Dec
2. MOTIVATIONS
2. Limited energy efficiency of current building systems; lack of new
breakthroughs
1. Energy consumption reduction for buildings with satisfied comfort level
- Paris Agreement, Singapore promised to reduce emissions by 36%
from 2005 to 2030
- The total annual electricity consumption of commercial buildings in
2016 is 6,371GWh
3. Poor demand response. General iterative methods are unable to
satisfy the dynamic changing demand and environment.
2
3. MOTIVATIONS
4. Gaps: Electricity buyer, cooling generation, cooling transmission and
demand response always work separately in distributed sub-systems
5. Practice. The building energy efficiency products available in the
market do not fit to the practical situation with customized
requirements
6. Call for solutions: Smart buildings need integrated solutions for
maximum energy efficiency by dynamic cooling generation with better
demand strategies.
3
4. MOTIVATIONS
7. AI offers potential solutions
• Many recent AI success stories; But they are in other domains
• Substantial R & D is needed for AI innovations for its application
building systems with higher efficiency and service quality but less
manpower.
AlphaGo 2016StarCraft 2017
Robot arm 2016 Autonomous driving 20164
5. OBJECTIVES
Objectives
The proposed integrated AI solution can achieve
• energy saving
• cost saving
• Quality service with comfort
It consists of system integration and system optimization
5
6. OBJECTIVES
1. System integration
Consider all the sub-systems
• cooling,
• lighting,
• electrical devices,
• renewable energy & storage system.
Key factors such as
• efficiency model, system character
• ambient environment, weather,
• occupation rate/behavior,
• system status;
6
7. OBJECTIVES
2. System optimization
We develop reinforcement learning(RL) algorithm with site deployment.
• High efficiency optimization
• Objective function
• Data pre-processing
• Training
• Feedback control
• Generate optimal models and predictions of the changes of the
internal and external environments such as weather and occupancy,
to further improve transient behaviours of disturbance rejection.
• To handle more complex nonlinear and uncertain dynamics of
realistic systems with high performance and robustness. It is readily
applicable to different buildings with its universal model structure
and training algorithms.
7
8. OBJECTIVES
3. Economical and social impacts
Innovative and scalable energy autopilot solution enables a spin-off through
commercialization in global market. Obviously, the energy optimization
solution is applicable to UOB building and other local building with large
benefit.
A cloud subscription service open for all the building owners.
10,000 large buildings in Singapore. 10 % of equipped with centralized
air-cooling systems. Then more tenants can benefit from power
consumption saving and higher comfort level. Therefore, the productivity
can also be improved.
8
9. APPROACH
action
At
observation
• state Si
• reward Rt
agent environment
Actor-Critic
Policy gradient Value based
Q function (action-value) 𝑄 𝜇 𝑠𝑡, 𝑎 𝑡 = 𝐸 𝑟 𝑠𝑡, 𝑎 𝑡 + 𝛾𝑄 𝜇 𝑠𝑡+1, 𝜇 𝑠𝑡+1
To maximum objective: 𝐽 𝛽 𝜇 = 𝑆
𝜌 𝛽 𝑠 𝑄 𝜇(𝑠, 𝜇(𝑠))𝑑𝑠 𝜌 𝛽distribution, μ strategy
Convolutional neural network to simulate μ and Q function, deep learning method for
training.
Variant algorithms to improve efficiency and resource utilization: deep deterministic
policy gradient(DDPG), asynchronous advantage actor-critic(A3C), evolution strategy
Challenge: 1) fast real-time response 2) security measure 3) high dimensional state
control
Advantage: 1) Dynamic model 2) avoid inaccuracy from high order numerical solution
Balance explore and exploit
9
10. APPROACH
• Deep reinforcement learning autopilot smart building for
optimization control: the network will find optimal control policy in
different status.
State estimation
- states of the control model
- occupancy
- CO2 concentration
- temp/heat for individual room
Deep learning (AI)
optimal control system
Greyg-box controller
model
Object
Constraints
Control
Building
chiller FCU AHU
etc.
Monitoring
Endogenous
Forecasting
disturbance
Exdogenous
Forecasting
disturbance
- cooling constraint
Weather & forecast
- ambient temperature
- solar radiation
- wind speed
- relative humility
- min $ /energy
- comfort level
• Network & algorithm design
• Data pre-process & feature extraction
• Pruning optimization & deployment
• Off/on policy training & simulation
environment
10
11. APPROACH
• Dynamic energy modelling with future demand response: Integrated
solution will be provided based on multiple dynamic energy models
with forecasted demands response.
Deep learning control
optimization system
Niagara Siemens Desigo cc Kura IoT
Chiller plant
BACnet
OPC
Modbus
HVAC CCTV lighting
Power
management
(AMR)
Microgrid
Work space
management
Differential
pressure
transducer
electromagnetic
flowmeter
Frequency
inverter
temperature
and
humidity
sensor
water
temperature
sensor
switch valve
liquid level
switch Occupancy
Smart
meter
Camera
plug load
Schneider
Microgrid (smartgrid)
Chiller plant
Microgrid (smartgrid)
HVAC
Occupancy & enviroment
AMR
( Automatic meter reading )
Weather
HVAC
Deep learning control
optimization system
Chiller plant
Physical system architecture
• Multi sensing system
• Kura IoT platform
Subsystems architecture
11
12. APPROACH
• AI cloud for multiple building optimization by parallel inference
threads. 1) cut down maintenance cost, 2) more data more
intelligent
GPU 0
building 1
BMS decode inference
Display/
Storage
building 2
BMS decode inference
building 3
BMS decode inference
Runtime environment
Nvidia GPU platform
Linux, cuda driver
Parallel computing
acceleration
BMS stream
decode
Scalable RL inference architecture
Optimization
control
Other
applications
Data collection
BACnet
Pre-process
APP
Infenrence
Feedback control
Display&storage
Trained weights
TensorRT
Network description
Multi building subscription of AI cloud
Inference instance of one building
AI cloud architecture
12
13. APPROACH
• Real time implementation with visualization and storage
• Field testing at building
• Centralized chiller cooling system, space
over 14,000sf
• Statistical analysis; verification of energy
saving target of 20%.
• Comparing test in different time scale and
scenarios.
• Economic comparison test on cost saving by
more reasonable load and renewable
energy dispatch with electricity price
variant in day and night time.
13