High-Fidelity Building Energy Monitoring

Network

Computer Science DepartmentUniversity of California - Berkeley

LoCal Retreat 2009

Xiaofan Jiang and David Culler

In collaboration with Stephen Dawson-Haggerty, Prabal Dutta, Minh Van Ly, Jay Taneja

My PG&E Statement

Current level of visibility Delayed Aggregated over

time Aggregated over

space Inaccessible

Want Real-time Per-appliance

[Stern92], [Raaii83]

2

Aggregate is Not Enough

What percent is plug-load

What percent is wasted by idle PCs at night?

3

What’s the effect of server load on energy?

What’s the effect of turning off A?

What caused the spike at 7:00AM?

This would be nice…4

Architecture

ACme application Standard networking

tools Python driver + DB +

web ACme network

IPv6 wireless mesh Transparent connectivity

between nodes and applications

ACme node Plug-through Small form factor High fidelity energy

metering Control Simple API

5

ACme Node6

Two Designs7

ACme-A ACme-B

ACme-A vs ACme-B

Resistor + direct rectification + energy metering chip

Real, reactive, apparent power (power factor)

Idle power 1W Low CPU utilization

Hall-Effect + step-down transformer + software

Apparent power Idle power 0.1W Medium CPU

utilization

8

ACme-A ACme-B

A tradeoff between fidelity and efficiency

ACme Node API9

ASCII shell component running on UDP port provides direct access to individual ACme node: Adjust sampling parameter Debug network connection Over-the-air reprogramming

Separate binary UDP port for data Periodic report to ip_addr at frequency rate

Node API function Purpose

read() -> (energy, power) Read current measurements

report(ip_addr, rate) -> Null Begin sending data

switch(state) -> Null Control the SSR

ACme Network

IPv6 mesh routing Each ACme is an IP

router Header compression

using 6loWPAN/IPv6 (open implementation -blip)

Modded Meraki/OpenMesh as “edge router”

Diagnostics using ping6/tracert6

ACme send per-minute digest / no in-network aggregation

10

internetinternet

backhaul linksedge routersAcme nodes

data repository

app 1app 2

Network Performance

49 nodes in 5 floors

Single edge router

6 month to-date 802.11

interference (on channel 19)

11

ACme Application

N-tier web application ACme is just like

any data feed Python daemon

listening on UDP port and feed to MySQL database

Web application queries DB and visualize

UDP Packets

Python Daemon

MySQL DB

ApacheACme Driver

6loWPAN

12

Visualization http://acme.cs.berkeley.edu/

13

Building Energy Monitoring14

1. Understanding the load tree

2. Disaggregation Measurements Estimations

3. Re-aggregation Functional Spatial Individual

Understanding the Load Tree

15

Deployment16

Edge router obtaining IPv6 address

Ad-hoc deployment Un-planned

Online “registration” using ID and KEY Meta data collection Security

Online for 6 month and counting

10 million rows

Deployment17

Raw Data18

Additivity using Time Correlated Data

19

Multi-Resolution20

Appliance Signature21

Functional Re-aggregation22

Correlate with Meta-data23

Spatial Re-aggregation24

Individual Re-aggregation25

Improvements in Energy Usage

26

Reducing Desktop Idle Power

27

Discussion and Conclusion

Measurement fidelity vs coverage

Non-intrusive Load Monitoring (NILM)

IP node level API vs application layer gateway

Easy of deployment is key

DB design Multiple input

channel / power strip

ACme is a fine-grained AC metering network that provides real-time high-fidelity energy measurement and it’s easy to deploy

3 steps to building energy monitoring – understanding load tree; disaggregation; re-aggregation

28

Discussion Conclusion

Discussion29

LoCal web site: http://local.cs.berkeley.edu ACme web site: http://acme.cs.berkeley.edu Contact: fxjiang@cs.berkeley.edu