1 © A. Kwasinski, 2015 Cyber Physical Power Systems Fall 2015 Application of a Cyber-physical Power System

1

© A. Kwasinski, 2015

Cyber Physical Power Systems

Fall 2015

Application of a Cyber-physical Power System

2


Cyber-physical power system

Introduction

• Slides contributions:• Drs. Kwasinski• Dr. Ted Song• Dr. Youngsung Kwon

3


Application

• Motivation #1: Sustainability

– Wireless networks energy consumption account in between 1% and 2% of the total annual electric energy consumption in developed countries.

– Most of the power consumption of wireless networks originate in base stations.

– Power consumption of wireless networks is increasing much faster than the total average power increase in developed countries.

– Wireless networks are powered from conventional electric power grids which, in turn, are mostly powered by non-renewable energy sources.

• Cyber-physical power system for wireless communication networks

4


• Motivation #2: Resiliency– Conventional backup power plants:

• Inefficient use of installed capital (batteries are rarely used)• Most cell sites lack a permanent genset• Battery energy storage is essential in order to reach telecom-

grade availability levels.• Power availability for air conditioners is below the minimum

required in telecom applications

– Microgrids:• Loads are primarily powered by local sources.• Renewable sources have been identified as a good solution to

power communication facilities during disasters because they do not rely on lifelines.

Application• Cyber-physical power system for wireless communication

networks

5


• Issues when integrating renewable energy in cell sites:– Power generation footprint >> load footprint

• Photovoltaic modules footprint = about 200 W/m2

• Base station footprint = a few kW/m2

– Variable output of renewable energy sources

• Solutions to these issues:– Source diversification– Use of locally stored energy (e.g. in batteries). This is

the role of energy storage in microgrids for increased use of PV systems in wireless communication networks


networks

6


• The paradox with energy storage (batteries)

• Functions of energy storage in microgrids:• Complement renewable sources (minutes to hours)• Match different dynamic characteristics of sources and loads

(seconds to minutes)• Failures ride-through/backup (seconds to hours)• Stability support (fraction of a second to seconds)• Reduced unavailability

• Issues with energy storage in microgrids• High cost• Weight and large volume issues• Demanding environmental conditions leads to lower availability

and reduced battery life (hence, air conditioning may be needed).

Microgrid Planning

7


• Concept:– Sustainable wireless areas (SWAs) are dc microgrids created by

interconnecting a few (e.g. 7) base stations.– Renewable energy sources are placed in base stations or nearby locations where there is sufficient space.– Resources (generation and energy storage) are shared among all base stations in the SWA.– traffic and electric energy management is integrated. I.e., traffic is regulated (or shaped) based on local energy resources availability and forecast.


networks

8


• Potential implementation in rural and sub-urban areas


networks

BS PV

9


• Potential implementation in urban areas

BS

BS PV


networks

10


• Goal: Use the advantages provided by integrating cybernetic and physical components in order to solve the limitations of integrating renewable energy sources into wireless communication networks without affecting other components, such as energy storage devices (e.g. batteries’ life)

• Renewable sources power output relevance:• Planning• Operation

• Time scales in renewable output power prediction• Climate forecast for long term planning• Weather forecast for short term operation.

10


networks

11



Planning

12


11 1 2

11 1

2 1

1 2

1

2 1

0

0

0

0 NN

NN N N

k p p

k p

p p

P

p p

p k

p p k

P

• Renewable energy sources• Renewable energy sources with batteries

( 1)Capacity N T

1 EA

Max SOC

Energy difference between two

states

Planning

13


Load

Wind

Solar

• Renewable energy sources with batteries

Planning

14


Solar Dist.

Solar Rnd.

Wind Dist.

Wind Rnd.

Load Dist. Load Rnd.


Planning

15


B PVP P L


Planning

16


301 301

0.3041 0.0247 0.0248

0.3041 0 0.0247

0.2890 0.0329 0

0 0.6959

0.0329 0.6959

P


Planning

17


1 3010.0042 0.0001 0.0036 0.7930

π1 π2 π3πN


Planning

18



Planning

19


• Modeling of renewable energy sources and combined energy storage based on Markov chains

PV (75%)

PV


Planning

20


Short term forecast

Hourly NOAA observation, Aug. 31st 2013Hourly NOAA 2-day ahead forecast, Aug. 31st 2013

• Solar irradiance (GHI) classification - Continuous: temperature, relative humidity, dew point - Discrete: sky coverage

Forecasted &observed weather variables from National Oceanic and Atmospheric Administration (NOAA)naïve Bayes (NB) classifier

21


7AM Hour of day 20PM

Clear day

Overcastday

overlapped points

overlapped points

Step 1:Partitioning 24hr Data Set to Hourly Subsets

GHI: Global horizontal irradiance (GHI) is the total solar irradiance from the entire sky on a horizontal surface, which includes the sum of the direct- beam and diffuse and reflected solar irradiance

GH

I (W

/m2)

Short term forecast

22


• Step 3: NB Classifier

• in step 3

[𝑉 𝑇𝑒𝑚𝑝 , h𝑉 𝑅𝐻 , h𝑉 𝐷𝑃 ,h𝑉 𝑆𝐶 , h]

Input vector

Short term forecast

23


Multiple features of Bayes theorem

Estimation of ->Kernel density function (Gaussian kernel)

- observations

- bandwidth inter-quartile range + rule-of-thumb bandwidth

𝑃 (𝐶|𝑉 1 ,𝑉 2 ,𝑉 3 )=𝑃 (𝑉 1|𝐶 ) ∙𝑃 (𝑉 2|𝐶 ) ∙𝑃 (𝑉 3|𝐶 ) ∙𝑃 (𝐶 )

𝑃 (𝑉 1 ,𝑉 2 ,𝑉 3 )posterior probability

prior probability

𝐾 𝑔𝑎𝑢(𝑉 )=𝑒𝑥𝑝 (−𝑉 𝑖 ,𝑛2

2𝜎 𝑖2 )

• let

𝑅𝑖=𝑉 𝑖 ,𝑄3−𝑉 𝑖 ,𝑄 1̂𝜎 𝑖 ,𝑟𝑜𝑡=( 4𝜎 (𝑉 𝑖)5

3𝑛 )0.2

�̂� 𝑖=1.06�̂�𝑖

1.35𝑛−0.2

allows outlier data to be less sensitive

Short term forecast

24


• Final algorithm step

GH

I (W

/m2)

ES

R (

W/m

2)K

t lev

elsShort term forecast

class variable

25


8 months results

TABLE Summary of Statistical Values For Actual GHI and Two Days Ahead Prediction Errors

Month r RMBE(%)

RRMSE(%)

Aug 2013-Mar 2014

333.04 292.47 86.3 9.09 80.39 138.85 2.73 41.7

Short-term (up to 48 hours ahead) prediction evaluation

Error criteria

Short term forecast

• Method validation

26


Prediction tendency follows actual values

Short term forecast

• Method validation

27


Conf. Mean [Wh]

MBE[Wh]

RMBE (%)

C1 1071.4 -38.31 -3.58

C2 794.09 42.7 5.38

Short term forecast

• Two-day ahead battery capacity estimation

• Sizing battery capacity

•

28



Operation

29


• Concept:– Sustainable wireless areas (SWAs) are dc microgrids created by

interconnecting a few (e.g. 7) base stations.– Renewable energy sources are placed in base stations or nearby locations where there is sufficient space.– Resources (generation and energy storage) are shared among all base stations in the SWA.– traffic and electric energy management is integrated. I.e., traffic is regulated (or shaped) based on local energy resources availability and forecast.


networks

30


11 1 2

11 1

2 1

1 2

1

2 1

0

0

0

0 NN

NN N N

k p p

k p

p p

P

p p

p k

p p k

P

• Renewable energy sources• Renewable energy sources with batteries

( 1)Capacity N T

1 EA

Max SOC

Energy difference between two

states

Long Term Planning

31


• Solutions to issues with renewable energy sources:

• Combine them with local energy storage (e.g. batteries)

• Very high availability requires significant stored energy

• Diversify power sources (e.g. combine wind and PV)• Source diversification reduces energy storage capacity needs

PV (75%)

PV

• Renewable energy sources

Long Term Planning

32


• Hierarchical structure• Top Level:

• Optimizes SWA operation by coordinating all power sources, energy storage devices and loads.

• Bottom Level:• Local autonomous controller in charge of regulating local traffic and

power generation based on top-level commands.• If the top level controller fails, this controller can maintain local

operation but at a suboptimal level (i.e. like in a conventional system)

• Control architecture

Operation and Control

33


• Control architecture


34


• Load is adjusted through traffic shaping.

where T is the duration of one traffic shaping control time period, N[k] is the total number of active radio transmitters at the base station at time k (for the considered configuration with 2 transmit antennas per sector and 3 sectors per base station, N[k]=NF=6), ν[k] is the normalized traffic profile at time k with respect to the maximum traffic load, and σ[k] is the overall traffic shape factor (composed of video and data components).

• Traffic shaping goal: to smoothly reduce service in a controlled way, so users notice little impact (reduced video quality and increased data delay) and within an acceptable level.

Traffic Shaping

35


• Availability can be improved without additional energy storage by modifying the transition probabilities.

• Transition probabilities can be modified by controlling the load (e.g. managing traffic) based on batteries state of charge or based on the present or expected future condition of the local power generators (including PV arrays).

• High-Level Controller


36


• Control architecture:• Central controller:

• Optimization• Local controller at each base station:

• Droop autonomous controller:• Operates the base station if central controller or communication link fails• Address constant-power loads stability issues

• Electrical architecture:• Includes power electronic circuits in distribution nodes

integrated APDNs

• Low-Level Controller


37


Central (high-level) controller

Load dependent on

traffic and antennae

power output

Communication link (necessary

anyway for base stations

communication functions)

Droop law(part of the local, lower

level controller)


38


LTE BS Power ConsumptionPeak power consumption of a LTE macro BS

• About 1350 W for 10 MHz bandwidth, 3 sectors, 2 antennas per sector with

feeder cable losses

RRU layout (no cable losses)

• About 800 W, thanks to a reduction of the power consumed by PA and

cooling system

Relation between traffic load and BS power

39


LTE BS Power Consumption

Traffic shaping approach

𝑩𝑺𝒕=𝑵𝑻𝑹𝑿 ∙(𝟗𝟕 ∙𝑻𝑷𝒕 ∙𝝈𝒕+𝟔𝟓)

Normalized average traffic profile in Europe

• 6 (# of antennas per sector, 2, times the # of sectors per BS, 3)

• hourly traffic profile• constant power term from radio resource overhead, cooling, processing, baseband interface, etc• 97 = dynamic power term according to the intensity of BS’ traffic load

The considered range of traffic shaping factor is derived considering mobile user’s QoE

affects mobile user’s Quality of Experience (QoE)

Other variables

traffic shaping factor

0.31

Power consumption

Renewable energy utilization

Mobile users notice little impact (reduced video quality and increased data delay) within an acceptable level

40


𝑩𝑺𝒕+𝟏=𝑵𝑻𝑹𝑿 ∙ (𝟗𝟕 ∙𝑻𝑷 𝒕+𝟏 ∙𝝈𝒕+𝟏+𝟔𝟓)

𝑵𝒆𝒕𝒕+𝟏=𝑺𝒐𝒍𝒂𝒓 𝒕+𝟏−𝑩𝑺𝒕+𝟏

depends on predicted renewable energy and battery bank SOC

Possible scenarios for choosing control factor

Traffic shaping algorithms

• Variables


41


Traffic shaping algorithms

• Design of ∆S:

Consider

where Snc[m] is the energy stored at the battery bank at time n if there is no traffic shaping and Slim corresponds to a critically low energy stored in the batteries/

At each decision time n the high level controller considers a target change in the energy stored in the batteries, S, given by

which can be determined by

is chosen based on the projected surplus or deficit of PV-supplied energy with no traffic shaping over a chosen time length MT, where MT is the time horizon for the calculation of energy surplus or deficit


42


Traffic shaping algorithms• Algorithm:

which leads to the logistic function:

The value Δ𝑆(𝜎ሾ𝑛ሿ) needs to account for the following observations: (a) Δ𝑆ሺ𝜎ሾ𝑛ሿሻshould take larger positive values when the relative battery SoC, 𝑆ሾ𝑛ሿ/𝐵, nears 𝑆𝑙𝑖𝑚 and/or 𝑆𝐴 is

close to 𝑆𝐴− (b)Δ𝑆ሺ𝜎ሾ𝑛ሿሻ can be set to a small value when the SoC is much larger than 0 and/or 𝑆𝐴 is close to 𝑆𝐴+and there is

less urgency to recharge the batteries.

where 𝑆𝐶 is a threshold for which traffic shaping is activated if 𝑆ሾ𝑛− 1ሿ/𝐵≤ 𝑆𝐶, 𝐾 is a constant chosen so that Δ𝑆 transitions between Δ𝑆𝑀𝐴𝑋 and Δ𝑆𝑚𝑖𝑛 when 0 ≤ 𝑆ሾ𝑛− 1ሿ/𝐵≤ 𝑆𝐶, 𝛿𝑡 is a constant between -1 and 1 that displaces

the logistic function horizontally (larger, positive, values for 𝛿𝑡 convey a more aggressive traffic shaping strategy) and 𝜓ሺ𝑆𝐴ሻ is a factor that incorporates the parameter 𝑆𝐴 and is defined as


43


Independent control of cell site operation by traffic shaping algorithms

• MDT in days: multiplying the proportion of failure time by 243 days (8 months)

SOC cumulative distribution functions (CDFs) for local operation

• Max traffic shaping results in maximum reduction in video quality and increase in data delayno traffic shaping andtraffic shaping

𝑺𝒍𝒊𝒎 𝑺𝒍𝒊𝒎


44


• More battery capacity Less battery capacity

and more PV panels and fewer PV panels

Video Quality

Data Delay

Battery State of Charge

Benefits of integrated traffic shaping


45


• With more batteries (10hrs backup) and PV panels (20x240W):– Probability of fully discharging the battery decreases from 0.11 with no

traffic shaping to 0.03 with maximum traffic shaping but video is reduced to just acceptable levels of 32.5 dB and maximum data delay is observed all the time.

– With integrated power/traffic management (e.g. ), probability of fully discharging the battery is 0.04 but video is degraded to 32.5 dB only 5% of the time and maximum data delay is observed only 14% of the time.

• With fewer batteries (8hrs) and PV panels (12x240W)– Probability of fully discharging the batteries is reduced from 0.35 with no

traffic shaping to 0.10 with maximum traffic shaping (and maximum service degradation)

– With integrated power/traffic management effects on quality of service are reduced.



46




47


𝑵𝒆𝒕𝒕=𝑺𝒐𝒍𝒂𝒓 𝒕−𝑩𝑺𝒕

𝒗 𝒕+𝑴=𝒗𝒕 𝑷′

𝒗 𝒕= [0 ,⋯ ,𝟏 ]1× 2 8

• Example

Real-time SOC estimation

Up to M hours in the future using the Naïve Bayes prediction model

M is determined by the battery bank capacity so that MT equals the time

that the BS can operate only from the battery bank

Markov-chain based battery bank charge and discharge process diagram

Expected SOC at time t+M

Real-time SOC estimation

Block diagram of M hours ahead

SOC estimation


48


Traffic shaping control

Real-time iterative process using Markov chain model

One day results for traffic shaping in Sep. 4th 2013: (a) battery bank SOC, (b) traffic shaping factor and (c) actual & predicted

renewable energy for C1

real-time M hours ahead SOC estimation


49


Integrated traffic-power control – power sharing

Multiple cell sites in a SWA

1

2

3

4

5

67

Power sharing algorithms


50


SWA resiliency

Battery bank lifetime model

𝐶𝑇𝐹=𝛼1 ∙𝑒𝑎2 ∙𝑥+𝛼3 ∙𝑒

𝑎4 ∙ 𝑥

𝑇 𝑏𝑎𝑡𝑡=∑𝑖=1

𝑛

(𝐷𝑂𝐷𝑖 ∙𝐶𝑇𝐹 𝑖 ∙𝐶 )

𝑛

𝐿𝑖𝑓𝑒𝑏𝑎𝑛𝑘=𝑻 𝒃𝒂𝒕𝒕 ∙𝑁𝑇 𝑏𝑎𝑛𝑘

Expected number of cycles to failure (CTF) & throughput vs DOD for a deep-cycle battery US 250E XC2

(C is a battery capacity: 1.35 kWh = 6V )

• From manufacturer’s data

(=3.388, =-0.223, =4415, and =-0.025)


Can be computed from a metric analogous to availability

51


Independent control of cell site operation by traffic shaping algorithms

No traffic shaping vs traffic shaping

C1 C2

83% 94% 65% 79%• The proportion of time available using renewable energy for 8 months

• Lifetime in year 0.93 1.02 339372 days

1.11 1.15405 420 days


52


SOC CDFs for global operation Local operation vs global operation

30% 0.85

Effect of power sharing algorithms

• Energy transfer from C1 to C2 by global operation:

𝑺𝒍𝒊𝒎


53


Local operation vs global operation

MDT is effectively improved more than twice (50 days to 21 days)

MDT of C2 is improved in the order of no traffic shaping, local traffic shaping, and global power sharing which indicate 86, 50, and 21 days

Dependent control of cell site by power sharing algorithms

Further improvement when compared to no traffic shaping (86 days)


Documents

1 © A. Kwasinski, 2015 Cyber Physical Power Systems Fall 2015 Application of a Cyber-physical Power System