pattern recognition technologies (prt), inc. on-line load forecasting services

Post on 09-Feb-2016

53 Views

Category:

Documents

0 Downloads

Preview:

Click to see full reader

DESCRIPTION

Pattern Recognition Technologies (PRT), Inc. On-Line Load Forecasting Services. Al Khotanzad, Ph.D., P.E. President PRT, Inc . 17950 Preston Road, Suite 916 Dallas, Texas 75252 (214) 692-5252 al@prt-inc.com www.prt-inc.com ERCOT Load Forecasting Forum January 24, 2007. - PowerPoint PPT Presentation

TRANSCRIPT

Pattern Recognition Technologies (PRT), Inc.

On-Line Load Forecasting Services

Al Khotanzad, Ph.D., P.E.President

PRT, IncPRT, Inc..17950 Preston Road, Suite 91617950 Preston Road, Suite 916

Dallas, Texas 75252Dallas, Texas 75252(214) 692-5252(214) 692-5252al@prt-inc.com

www.prt-inc.com

ERCOT Load Forecasting ForumJanuary 24, 2007

PRT, Inc. 2

Corporate Profile Founded in 1994

Products & Services Online load and price forecasting services Stand-alone load and price forecasting software Custom forecasting solutionsClients Over 90 energy firms consisting of

Utilities in North America & Overseas ISOs, Municipalities, Coops. Government Agencies Power marketing and trading organizations The Electric Power Research Institute (EPRI)

First company to develop a commercial neural network based load forecaster in the early 90’s – ANNSTLF for EPRI

PRT, Inc. 3

Load Forecasting Accurate forecast of future demand required by all entities involved

in the energy markets Electric Utilities Independent System Operators Power Marketers

Different forecast horizons Long Term: Several years out – required for planning purposes Mid Term: Several weeks to months – scheduling maintenance,

planning fuel supply, transactions Short Term: Next hour to next week – daily operation, energy

transactions, reliability studies

PRT, Inc. 4

Short Term Load Forecasting Hourly or sub-hourly forecasts starting from next hour to next seven to ten

days Forecasts used for:

Unit commitment– selection of generators in operation– start up/shut down of generation to minimize operation cost

Hydro scheduling to optimize water release from reservoirs Generator type coordination to determine the least cost operation mode

(optimum mix) Interchange scheduling and energy purchase Transmission line loading Power system security assessment

Accuracy has significant economic impact Even a 0.5% improvement in accuracy can result in thousands of dollars in

savings

PRT, Inc. 5

Factors Affecting Short Term Load Factors affecting short-term load are:

Mix of customer in the service area (residential, commercial, industrial) Weather condition (temperature, humidity, cloud cover, wind speed) Seasonal effects & recent load trends Time of day (morning, afternoon, night) Day of week (weekdays, weekends) Holidays (Christmas, New Years) Special events (popular sporting events or TV shows) Demand side management Random disturbances

Forecasts used for: Unit commitment

– selection of generators in operation– start up/shut down of generation to minimize operation cost

Hydro scheduling to optimize water release from reservoirs Generator type coordination to determine the least cost operation mode (optimum mix) Interchange scheduling and energy purchase Transmission line loading Power system security assessment

Accuracy has significant economic impact Even a 0.5% improvement in accuracy can result in thousands of dollars in savings

PRT, Inc. 6

Example of Load and Temperature

PRT, Inc. 7

Major STLF Techniques Any STLF technique attempts to model the relationship between the

load and factors that affect it – these relationships are nonlinear and complex Regression models Stochastic time series Spectral decomposition Similar-day search Intelligent system based models

Superiority of intelligent system based techniques have been demonstrated in many studies

PRT, Inc. 8

PRT’s LF Technologies Products & services are based on cutting-edge intelligent system

technologies of:

Artificial Neural Networks Fuzzy Logic Genetic Algorithms/Evolutionary Computing

PRT, Inc. 9

Artificial Neural Networks (ANNs)

Neurologically inspired systems consisting of highly interconnected elementary computational units (neurons)

Distributed processing by neurons results in intelligent outcome

ANNs learn to perform a desired task directly from examples using special training algorithms

ANNs can generalize; after training, they can produce good results for data that only broadly resembles the data they were trained on originally

ANNs are nonlinear systems, well suited for real world problems that are often nonlinear

PRT, Inc. 10

Forecasting Using ANNs A key feature of ANNs is their ability to learn a complex pattern

mapping, i.e., model the underlying relationship between a set of variables and an outcome that is a function of them –

Future Load – Function of past loads and weather, recent load trends, upcoming weather, calendar effects

Train with historical data (examples of the underlying relationship)

A properly trained ANN can predict the outcome of the modeled process based on the available observations

ANN based predictors employed in a wide variety of forecasting applications such as prediction of: electric load, weather, gas consumption, stock market, economic trends time series data, future sales, traffic patterns and grade point average of students

PRT, Inc. 11

Unique Aspects of PRT ANN Forecasters

Architecture of ANN specifically designed for electric load forecasting

Optimal set of inputs selected for load forecasting application

No need for frequent re-training

Quick response to deviations between forecast and actual load

Special algorithms for unusual days, e.g., weekday holidays

PRT, Inc. 12

Fuzzy Logic Fuzzy logic (FL) is a means to transform subjective/expert

knowledge about a process expressed in the form of linguistic rules into computer algorithms.

FL employs fuzzy sets, fuzzy membership functions and fuzzy if-then rules to model the uncertainty in nature, and express the knowledge

A fuzzy set is a set without a crisp, clearly defined boundary, and can contain fuzzy variables with a partial degree of membership

PRT, Inc. 13

Fuzzy Rule & Membership Function An example of a typical fuzzy IF-Then rule :

IF next-day temperature is hot, and today’s temperature is hot, THEN next-day load is high

Subjective interpretation of “hot temperature” or “high load”

• Characterized by fuzzy membership function – an example shown

PRT, Inc. 14

Fuzzy Logic Based Load Forecaster Develop applicable fuzzy membership functions

Extract relevant IF-THEN rules from historical data – There could be hundreds of such rules

During the forecasting phase several of the rules become activated along with some of the fuzzy membership function

Fuzzy inference engine converts all this information into a final crisp forecast

PRT, Inc. 15

Genetic Algorithms (GAs) Genetic Algorithms (GAs) are optimization algorithms that are

based on the concept of natural evolution

GAs can find the optimal solution quickly and efficiently, especially when there is little information about the solution available.

GAs emulate natural evolution, and make use of four operators, including reproduction, crossover, mutation, and survival of the fittest to produce and keep the optimal solutions.

PRT, Inc. 16

GA Based Forecaster Create M sets of forecasts (in random) for a given set of actual

historical data Sort based on accuracy Retain the top K most accurate sets (stronger solutions) and discard

the rest (weaker solutions) – survival of the fittest Use the retained K sets as parents to create a second generation of

M solutions through mutation & crossover – repeat the process After several generation, the top K solutions converge toward a

single solution – Strongest solution This is the optimal solution used as the final forecasting model

PRT, Inc. 17

PRT’s e-ISOForecast Price & Load Forecasting Service

e-ISOForecast is an on-line real-time price & load forecasting service that has been set up for all wholesale power markets/ISOs in North America ERCOT, PJM, NY-ISO, ISO-NE, MISO, CA-ISO, ONTARIO

IESO, ALBERTA AESO Hourly forecasts for current day and six days beyond Hourly load forecasts for one year out using various simulated

weather scenarios Forecasts are posted on www.onlineforecast.com Subscribers use a Web browser to access and download the

forecasts – available 24/7 Forecasts are updated every hour or faster based on the most

recent price/load/weather data that become available Weather forecasts are used in the models - updated several times

per day

PRT, Inc. 18

e-ISOForecast Price & Load Forecasts ERCOT

System-Wide & Congestion Zone Load Forecasts Zonal Market Clearing Price Forecasts

PJM System-Wide, Regional and Zonal Load Forecasts Real-Time & Day-Ahead LMP Price Forecasts

ISO New England (ISO-NE) System-Wide & Zonal Load Forecasts Zonal Real-Time & Day-Ahead LMP Price Forecasts

New York ISO (NYISO) System-Wide & Zonal Load Forecasts Zonal Real-Time & Day-Ahead LMP Price Forecasts

PRT, Inc. 19

e-ISOForecast Price & Load Forecasts Midwest ISO (MISO)

System-Wide Load Forecast Real-Time & Day-Ahead LMP Price Forecasts for Five Hubs and

Various CPNs

California ISO System-Wide Load Forecast Zonal Supplemental Real-Time Price Forecast s

ONTARIO EISO System-Wide Load Forecast System-Wide Price Forecast

ALBERTA AESO System-Wide Load Forecast System-Wide Price Forecast

PRT, Inc. 20

Forecasting Engines

Multiple models based on different technologies run in parallel generating independent forecasts

A top layer of intelligence decides to: Select one of the forecasts as the final forecast Combine multiple forecasts (“Combination of Experts”) into a

final forecast

Accuracy is improved over use of a single modeling technique

PRT, Inc. 21

Weather Data PRT has affiliations with two major weather service providers, WSI and

Meteorlogix

Most free internet based weather forecast services simply provide forecasts generated by NWS or other computer models

Weather service providers bring human meteorologists in the loop who scrutinize/edit computer generated forecasts

Weather forecasts updated several times throughout the day

Actual temperature updated every hour and with every update, new load forecasts are generated

PRT, Inc. 22

Access via the Web Forecasts are posted to a dedicated password protected page Can be accessed using any standard Web browser from any

computer Provides easy access for all in the company Forecasts are displayed in tabular and graphical forms Actual data of previous day and any available data of current day

are displayed Forecasts can be downloaded in EXCEL format Other statistics including actual prices of past week, similar day

comparisons and price bands are provided

PRT, Inc. 23

e-ISOForecast Main Page

PRT, Inc. 24

e-ISOForecast PJM Segment

PRT, Inc. 25

e-ISOForecast Load Forecast View

PRT, Inc. 26

e-ISOForecast LF View - Graphical

Market After-the-Fact

CurrentDay

Next Day

Day 3 Day 4 Day 5

ERCOT Load 1.51/1.84 1.81/2.69 2.99/3.48 3.50/3.87 3.99/4.59 4.54/5.13Temp 1.38/1.73 2.14/1.98 2.41/2.13 2.86/2.68 3.21/3.02

PJM East

Load 1.24/1.33 1.32/1.91 2.49/2.64 2.88/2.93 3.25/3.45 3.67/3.81Temp 1.48/2.00 2.56/2.46 2.73/2.78 3.20/3.46 3.68/4.05

ISONE Load 1.70/1.60 1.24/1.88 2.30/2.32 2.73/2.91 3.04/3.31 3.34/3.60Temp 1.49/2.00 2.47/2.57 2.76/2.93 3.27/3.62 3.78/4.32

NYISO Load 1.15/1.28 2.21/2.40 2.56/2.59 2.73/2.74 2.93/3.05 3.26/3.47Temp 2.72/2.81 3.04/2.99 3.22/3.23 3.53/3.60 3.98/4.30

MISO Load 1.21/1.39 1.21/1.83 2.29/2.54 2.79/3.05 3.20/3.53 3.54/3.87Temp 1.29/1.70 2.15/2.32 2.37/2.48 2.82/3.05 3.32/3.57

e-ISOForecast LF Performance for 2006Forecasts Recorded at 8 am CT

Load: Hourly MAPE/ Daily Peak Load MAPETemperature: Hourly MAD/Daily Peak Temp MAD

PRT, Inc. 28

e-ISOForecast Performance for Forecast of Next-Day ERCOT Total Load – 2006

i

Forecasts Recorded at 3:00 PM of Previous Day

Month Load MAPE

Temp MAD

Month Load MAPE

Temp MAD

Jan 2.10/2.02 2.13 Jul 2.51/2.77 1.71

Feb 3.01/3.18 3.14 Aug 2.70/2.81 2.06

Mar 2.52/2.85 2.42 Sep 3.86/5.03 2.35

Apr 3.17/4.11 1.78 Oct 3.53/5.06 2.36

May 3.05/3.57 1.83 Nov 2.21/2.25 2.21

Jun 2.70/2.92 1.62 Dec 2.45/3.49 2.14

Market Period By Current Day

Next Day

Day 3 Day 4 Day 5

ERCOT 1/1-12/31

PRT 1.81/2.69 2.99/3.48 3.50/3.87 3.99/4.59 4.54/5.13ISO 2.28/2.96 3.22/3.45 3.75/4.08 4.56/4.79 5.04/5.40

PJM East

1/1-12/31

PRT 1.32/1.91 2.49/2.64 2.88/2.93 3.25/3.45 3.67/3.81ISO 1.70/2.00 3.30/2.91 3.32/3.23 3.79/3.60 4.27/4.13

ISONE 1/1-12/31

PRT 1.24/1.88 2.30/2.32 2.73/2.91 3.04/3.31 3.34/3.60

ISO 3.08/1.77 3.23/2.15 - - -NYISO 12/5-

12/31PRT 0.64/1.06 1.51/1.54 1.74/1.85 1.99/2.29 2.10/2.34

ISO 2.54/2.62 2.54/2.62 2.22/2.09 2.12/1.59 2.08/1.43

Comparison of PRT and ISO LF PerformanceForecasts Recorded at 8 am CT

Hourly MAPE/ Daily Peak Load MAPE

PRT, Inc. 30

e-LoadForecast Service An online load forecast service for company-specific load data Standard Service: Hourly/sub hourly forecasts for current day and six

days beyond Extended Service: Additional Hourly/sub hourly forecasts for several

months and years out User only needs to:

Provide historical load data for initial model training Upload the most recent actual load data as it becomes available (via

FTP, e-mail, provided Excel interface) All the required actual and forecast weather data acquired by PRT from Load and weather data quality checked and validated Forecasts posted to a dedicated and secure website in tabular and

graphical forms

PRT, Inc. 31

e-LoadForecast Service, Cont’ Forecasts are updated every hour with preceding hour’s actual

observed weather Forecasts are updated any time an actual load data is uploaded by

user 24/7 access through

Via Internet at any location An Excel Interface with built-in functions enabling user to

remotely interact with the forecasting system FTP E-Mail

ERCOT uses this service for forecast of its eight weather zones

PRT, Inc. 32

Forecasting Engines

Multiple models based on different technologies run in parallel generating independent forecasts

A top layer of intelligence decides to: Select one of the forecasts as the final forecast Combine multiple forecasts (“Combination of Experts”) into a

final forecast

Accuracy is improved over use of a single modeling technique

PRT, Inc. 33

Weather Data PRT has affiliations with two major weather service providers, WSI and

Meteorlogix

Most free internet based weather forecast services simply provide forecasts generated by NWS or other computer models

Weather service providers bring human meteorologists in the loop who scrutinize/edit computer generated forecasts

Weather forecasts updated several times throughout the day

Actual temperature updated every hour and with every update, new load forecasts are generated

PRT, Inc. 34

Other Features The provided Excel Interface allows user to:

Modify forecasted temperatures and generate corresponding load forecasts – “What-If” scenarios

Modify predicted morning and/or afternoon peak loads. Forecasts for other hours are reshaped accordingly

View load and temperature of three most similar days (temperature wise) in the history

PRT, Inc. 35

Access via the Web - View & Download Forecasts are posted to a dedicated password protected page Can be accessed using any standard Web browser from any

computer Provides easy access for all in the company Forecasts are displayed in tabular and graphical forms Actual data of previous day and any available data of current day

are displayed Forecasts can be downloaded in EXCEL format

PRT, Inc. 36

Access via Excel Interface – View, Download & Interact

An Excel interface with easy-to-use built-in features Download and view most current load and temperature

forecasts in tabular and graphical forms Modify forecasted temperatures and generate corresponding

load forecasts Modify predicted peak loads and reshape load forecasts

accordingly Download and view three most similar days Upload actual load updates to PRT’s servers

PRT, Inc. 37

Profile Based Forecasting Retailers operating in deregulated markets work with individual accounts that may

not be metered hourly (e.g., residential load) Energy transactions and settlements are done based on hourly demand Hourly load is simulated using pre-specified standard load profiles for client type To forecast their retail load, load profile for each account must be scaled

appropriately to account for pattern of usage by that account Profile based module of e-LoadForecast – User provides:

List of current accounts in the portfolio along with their corresponding profile type The historical usage data for each account

Backcasted profiles for corresponding profile types are used to develop a profile multiplier (scale factor) for each account using historical meter reads.

Forecasted standard profiles are multiplied by the scale factor to get the final hourly forecast

PRT, Inc. 38

Mid-Term/Long-Term Module Optional service includes mid-term/long-term hourly load forecast Forecast horizon can be extended to five years out ANN technology is used – Models are different from those used for short-term forecasting Impact of load growth is considered Weather forecast is needed for the forecast horizon

Simulated using historical weather data Three scenarios of “Normal”, “Hot”, and “Cold” available for each month in forecast

horizon Additional scenarios for generating “High Load” and “Low Load” cases Two statistical methods available for simulation of scenarios from historical weather

data Tools are provided for easy manipulation of simulated weather – user can build heat

waves/cold fronts

PRT, Inc. 39

Quality Control Extensive quality control system in place

Actual load and temperature data continually quality checked

Detected anomalies such as spikes and gaps corrected

Every day accuracy of load and temperature forecasts for various forecasts horizons are computed and reviewed by our experienced staff

Corrective action taken if degradation in quality detected Analysis of the cause Calibrate forecasting models Use of different kind of forecasting engines

PRT, Inc. 40

Forecasting Service Advantages Uses state-of-the-art load forecasting models More accurate forecasts than in-house systems More economical than maintaining an in-house system Frees up valuable manpower & resources No data hassles, IT overhead, software maintenance & upgrade Performance continuously monitored by specialists with extensive

experience and background in forecasting Models are continually calibrated and upgraded Convenient access to forecasts for all who need it in the

organization Unlimited use by all in the organization

top related