Presented by: Anna Scaglione

Coauthor: Mahnoosh Alizadeh

UC Davis

WORKSHOP ON “TRANSACTIVE ENERGY” MAY 18 2011

Green Energy Management with Digital Direct Load Scheduling

Outline Background work Pricing and Direct load Management Limitations of the prior art

Proposed Architecture Arrival model for smart loads Load characterization Optimal load dispatch Quantization Communication Architecture: Uplink and Downlink

CEMS real-time DDLS

Problem statement

To meet their peak-demand, utilities invest billions of dollars in equipment that is underutilized most of the time

Green energy is volatile and mostly uncontrollable needs huge amounts of reserves

Solution: exploit demand “elasticity”Combine demand side management and demand

response with scheduling opportunistically green resources

End-use demand Management: Background

Priced Based Load Control Time Of Use predetermined but variable rates during the

dayReal-time prices: cheapest way of managing demand, requires

communication of a price signal Needs an Automated System Home Energy Management Systems

Load Control Through Curtailment Interruptible Load ProgramsDevices like ACs are signaled to turn off for a pre-determined

interval from a central control center

Drawbacks of pricing

Problems with TOU? Designed years beforehand – no actual real-time control over

demand While TOU is a major step in the right direction it is not

designed for volatile generation

Problems with RTP? Not actually real-time: prices have to be delivered to the

customer beforehand to allow some planning time

Feedback loop due to customer response rebound peaks Needs extensive knowledge about customer behavior: On flat

rates, customers are much more predictable

Drawbacks of load control Problems with Load Curtailment?Not a market based program – cannot bid into any

marketOnly good for appliances that can be

interrupted (ACs or water heaters) or in emergency situations to avoid involuntary cascaded blackouts

Possible Solution? Centralized Direct Load Scheduling

Direct Load SchedulingBenefits: Economy of Scale Actual real-time control – no feedback loops due to customer

behavior Can provide ancillary services

Central optimized scheduling can use extensive real-time market and local generation data available at hand Scalability problems..

Cellular structure Community Energy Management Systems (CEMS) in charge of subscribers drawing power from a certain substation

The general idea

Model for the control architecture

Proposed Architecture

Cellular access model: HEMS CEMS GEMS

Demand elasticity Control of the activation time of smart loads (DLS) + dynamic pricing for non-smart load

Opportunistic green energy dispatch

Introduces a cyber-physical control architecture

Modular classification of loads queues for energy

Green Energy Management System

Control Layers

Network Management Layer – End to End

Local Decision and Control Layer - Access

Digital Direct Load Scheduling - Data

Physical Load Layer

Transmission assets and control Renewable Capacity

GEMS

CEMS CEMS CEMS

HEMS HEMSHEMSHEMS HEMS HEMS

HEMSHEMSHEMS

Grid Control and Assets RES

Key idea: Unbundling the load process The load offered to the grid (complex phasor) is the sum of random

contributions from each appliance

Basic assumptions Smart loads last a finite time, with random shape g(t;Cij)=gI(t;Cij)+jgQ(t;Cij) Their arrivals are Poisson (non stationary) and without control, the random

load process would be

Basic idea: Delay and reassemble optimally the contributions to “shape it”

CapacityL(t) unscheduled

Ls(t) scheduled

Poisson arrival model for smart loads

Arrival of H different types of load Nonhomogenous Poisson Arrival Processes (statistics fully described by arrival rates λj(t))

tija :time of arrival for ith device of the jth class u(t) is the unit step

Assumptions: Each Arrival has a coefficient vector, Cij, that describes the evolution of

its demand in time, i.e. g(t;Cij) fj(C) is the PDF of Cij and is equal for every arrival in the same class

Full information about demand: the tuple (j, tija,Cij)

Required Information to predict future load: statistics λj(t) and fj(C)

Optimal SMART Load dispatch

Direct Load Scheduling: optimally map tija into a switch-on time tij

d>tija

Queue departure/Switch on process:

Questions that need to be addressed: What needs to be communicated? How? How to choose the tij

d : what are the associated costs?

IDEA: Two simplifications steps lead to a Digital Direct Load Scheduling model (part encoder, part communication protocol, part controller)

Digital Direct Load Scheduling

Model for the network control interaction

Digital Direct Load Scheduling:Step 1 - analog to digital load mapping

Goal: Find a tractable model to communicate control information and reassemble the optimal load process

Assumption for smart loads: finite duration and approximately band-limited

load pattern represented with a finite number of coefficients (Fourier coefficients, samples, etc.) quantization/digital conversion Ĉij,

Example: Electric Vehicles – Cij is a scalar representing the hours of battery charge needed, Ĉij is a finite precision representation 4 bits, 16 classes give > 10% accuracy

Digital Direct Load Scheduling:Step 2 - Discrete time model

By quantizing the Cij ‘s , we can divide the appliances into queues,

* denotes the convolution operation

Computational Complexity use a predetermined set of discrete time decision intervals:

Step 2- Discrete time model (cont.)

Then we can write the smart load as,

Or simply, after quantization

replace the previously defined quantities with their discrete counterparts

a1,j(t) d1,j(t)

aQ,j(t) dQ,j(t)

Mapping requests in queues into power

gqj(t)

gQj(t)

Smart loaddue to appliances oftype j – LS

j(t)D

D

First difference operator

Sche

dulin

g

Arrival and Departure process

aj(t)

dj(t)

HEMS to CEMS uplink:The digital load request structure

The tuple (j, tija,Cij) is mapped onto a discrete triplet

(j, pija, qij)

pija :time-stamp index with respect to a time unit ∆

qij :index of Q(Cij), which is quantized in Qj different codes

Optimized Quantization of Cij - 1

Minimize Qj while satisfying a desired maximum distortion W in load reconstruction,

E(Wj) is the largest expected distortion by the j-th type:

Optimized Quantization of Cij - 2 Under a maximum bit rate constraint, minimize the distortion,

RHEMS-MAC : required rate for the LAN

RCEMS-GEMS : required rate for the backbone

It is clear that optimum is reached when constraints are tight

Data traffic in HEMS to CEMS Uplink

Each arrival generates an independent triplet (j, pija, qij)

The network has a maximum delay D we can send pija

modulo D using side-information of receive time log2D bits

H different types j requires log2H bits

Qj queues for type j log2 Qj bits Thus,

CEMS – GEMS Uplink

The traffic is divides into queues by the CEMS coalesce the arrival times into information about the arrival vectors for different queues

The communication rate required to send this vector to the GEMS is,

Broadcasting Decisions: CEMS to HEMS Downlink

A method that preserves the anonymity of the scheduled user

H feedback vectors are broadcast

the j-th feedback consists of a Qj ×1 vector Tj(ℓ), Any appliance of type j in the corresponding classes

q = 1, . . . , Qj that arrived before time Tjq(ℓ), can

enter the system.

Determining Tjq(ℓ)

Decision model

Model for the optimization of the schedule and market

Cellular Green Energy Management: Control layers

The CEMSs need to schedule smart appliances in real time

Optimal DDLS using pricing

The GEMS interacts with the whole-sale market It needs to solve optimizations to determine: The day-ahead bid

The dynamic price to communicate to customers in order to control the non-smart load

CEMS decision variables The optimization includes: Wholesale market real-time bidding cost Inconvenience cost paid to customers for delays Carbon taxing utility (always cheaper to use DG)

The following are determined by changing input variables or from the output of the DLS optimization:

Available capacity to provide wholesale market ancillary services

Participation plan for Local market of DGs Modified solution to include distribution grid constraints

Delay Cost formulation It is a well known result in traffic flow theory that the total

delay experienced by the customers in a queue, i.e. Σi (tijd−

tija), is equal to the area of the queue polygon,

After quantizing the decision set and the Cij’s

Typical DDLS formulation

D is the entire decision space, i.e. D = {dq,j(l)|l ≥ l0, j = 1, . . . ,H} A is the space defined by the arrival process, = {aq,j(l)|l ≥ l0, j = 1, . . . ,H} EA , the expected value operator, takes the effect of future arrivals into

account Cl (.) represents the cost of deviating from the day-ahead bid and the

available DG DCI is called the delay cost increment, it penalizes the scheduler for

delaying customers

Ideal Case: Stationary arrivals Two appliance types, constant power demand Stationary arrival rate λ1 and λ2, constant departure rate μ1 and μ2

constant supply P Lq q=1,2 is the average queue length, from classic queuing theory:

Delay and power cost are decoupled

μq → ∞ will set delay costto zero but won’t affect

deviation cost Actual departure rate = λ1 and λ2

Blue line α deviation from available power cost Purple line α delay cost Red line: constant contour of available power P

Ideal case solution

Challenges in selecting the optimal decision Arrival rates and planned supply profile vary over time The minimum delay and cost are now coupled The optimum decision in this case is no longer trivial

and requires computation

Simplification Departures allowed in discrete times Finite horizon Integer program for observable future arrivals In practice it is a dynamic program Relaxations? …

Preliminary results

A few simulations….

Numerical Results – Linear Programming approximation

a single type of smart appliance

1 ≤ Job size ≤ 4 units of time

Arrival process is Poisson with constant rate λ =3

Solver: one-step look-ahead rollout algorithm on acertainty equivalent controller that uses linear programming to schedule the appliances

Look-ahead horizon = 5 units of time

For fairness, the number of scheduled appliances is equal in the two profiles and no arriving appliances is delayed beyond t = 50

%50 reduction in deviation from the generation profile

Smart load forecasting

Even if we do not control the load in a centralized fashion, observability of LS(t) will improve our short-term load forecasting ability

Modeling future demand requires statistics of arrivals and of load profiles through fj (C)

The arrival rate λj(t) is a non-stationary time-series ARIMA process with multiple seasonalities

The digital conversion of the load is a model to design an efficient encoder for load metering as well

The rate requirement is very modest…

Smart load forecasting (cont.) Short term forecast with considerable number of Smart EVs

Number of future smart loads of type j:

Njnew(t) is a stochastic term but νj(t) is deterministic due to

observability EV load Parameters:

λ(t) Doubly periodic

ARIMA (110)(110)48(110)336

EV added to load profile of substation (real data transport network normalized)

CLASSICAL

Load bundled, ARMA

SMART FORECAST

Track EV rate λ(t) ~ C(t)/T using classical ARIMA

Mt/GI/∞ model

Classical Prediction for base load

Conclusions Efficient markets sail on good opportunities We described a method to realize efficiency for the customer by

allowing their request to be scheduled so that the community realizes the greatest energy efficiency

Basic principle: Unbundle and digitize the load – do not store energy, store requests! By product – The rate requirement is not obscure! No esoteric network control result: The rate requirement stems naturally

from the accuracy of the digital load description/complexity of controller trade off

The volatility of RES resources, the interaction with the Transmission grid and the market pricing are all considered in the decision model

Where do we go from here Even if occasionally we think we can do everything we really

cannot…

We need a broad base of collaboration to make this happen

Possible Model: starting an “Open Standard” DDLS is an encoder, a protocol of communication and a

controller at once The optimization needs to incorporate models for RES,

learning models for the traffic statistics and state of the art algorithms

The end to end global management strategy by GEMS needs to be defined

Thank you!!