WRF Webcast. Improving the Accuracy of Short-Term Water Demand Forecasts

Similar documents
SHORT TERM LOAD FORECASTING

NSP Electric - Minnesota Annual Report Peak Demand and Annual Electric Consumption Forecast

Multivariate Regression Model Results

RD1 - Page 469 of 578

Population Research Center (PRC) Oregon Population Forecast Program

Development of Short-term Demand Forecasting Model And its Application in Analysis of Resource Adequacy. For discussion purposes only Draft

Demand Forecasting Models

Design of a Weather-Normalization Forecasting Model

STAT 212 Business Statistics II 1

NOWCASTING THE NEW TURKISH GDP

peak half-hourly New South Wales

peak half-hourly Tasmania

Introduction to Forecasting

Empirical Project, part 1, ECO 672

Proposed Changes to the PJM Load Forecast Model

Chapter 7 Forecasting Demand

FINAL REPORT EVALUATION REVIEW OF TVA'S LOAD FORECAST RISK

Page No. (and line no. if applicable):

LOADS, CUSTOMERS AND REVENUE

Presentation for the Institute of International & European Affairs

BEFORE THE PUBLIC UTILITIES COMMISSION OF THE STATE OF COLORADO * * * * *

Identifying Causal Effects in Time Series Models

The Dayton Power and Light Company Load Profiling Methodology Revised 7/1/2017

Beginning to Enjoy the Outside View A Glance at Transit Forecasting Uncertainty & Accuracy Using the Transit Forecasting Accuracy Database

WEATHER NORMALIZATION METHODS AND ISSUES. Stuart McMenamin Mark Quan David Simons

Copyright 2010 Pearson Education, Inc. Publishing as Prentice Hall.

Population and Employment Forecast

2013 WEATHER NORMALIZATION SURVEY. Industry Practices

Volume 38, Issue 2. Nowcasting the New Turkish GDP

Time Series and Forecasting

Time Series and Forecasting

Forecasting Using Time Series Models

SAN DIEGO GAS AND ELECTRIC COMPANY SOUTHERN CALIFORNIA GAS COMPANY 2013 TRIENNIAL COST ALLOCATION PROCEEDING (A ) (DATA REQUEST DRA-MPS-2)

Analysis. Components of a Time Series

NOWCASTING REPORT. Updated: August 17, 2018

Introduction to Regression Analysis. Dr. Devlina Chatterjee 11 th August, 2017

Design of a Weather- Normalization Forecasting Model

TRANSMISSION BUSINESS LOAD FORECAST AND METHODOLOGY

2013 Weather Normalization Survey. Itron, Inc El Camino Real San Diego, CA

Warwick Business School Forecasting System. Summary. Ana Galvao, Anthony Garratt and James Mitchell November, 2014

PREPARED DIRECT TESTIMONY OF GREGORY TEPLOW SOUTHERN CALIFORNIA GAS COMPANY AND SAN DIEGO GAS & ELECTRIC COMPANY

Comment on: Automated Short-Run Economic Forecast (ASEF) By Nicolas Stoffels. Bank of Canada Workshop October 25-26, 2007

Econ 300/QAC 201: Quantitative Methods in Economics/Applied Data Analysis. 17th Class 7/1/10

Seasonal Hazard Outlook

NOWCASTING REPORT. Updated: July 20, 2018

NOWCASTING REPORT. Updated: May 5, 2017

Applied Time Series Topics

Econometric Forecasting Overview

Display and analysis of weather data from NCDC using ArcGIS

Great Lakes Update. Volume 188: 2012 Annual Summary

2006 IRP Technical Workshop Load Forecasting Tuesday, January 24, :00 am 3:30 pm (Pacific) Meeting Summary

STATISTICAL LOAD MODELING

Decision 411: Class 3

FORECASTING. Methods and Applications. Third Edition. Spyros Makridakis. European Institute of Business Administration (INSEAD) Steven C Wheelwright

2006 & 2007 Pre-Hurricane Scenario Analyses

The Kentucky Mesonet: Entering a New Phase

Introduction to Econometrics

Decision 411: Class 3

Forecasting: Methods and Applications

Asitha Kodippili. Deepthika Senaratne. Department of Mathematics and Computer Science,Fayetteville State University, USA.

NOWCASTING REPORT. Updated: April 15, 2016

Drew Behnke Food and Agriculture Organization of the United Nations UC Berkeley and UC Santa Barbara

Challenges to Improving the Skill of Weekly to Seasonal Climate Predictions. David DeWitt with contributions from CPC staff

Forecasting Bangladesh's Inflation through Econometric Models

Decision 411: Class 3

Forecasting. Copyright 2015 Pearson Education, Inc.

Defining Normal Weather for Energy and Peak Normalization

Empirical Approach to Modelling and Forecasting Inflation in Ghana

Forecasting: principles and practice. Rob J Hyndman 1.1 Introduction to Forecasting

UPPSALA UNIVERSITY - DEPARTMENT OF STATISTICS MIDAS. Forecasting quarterly GDP using higherfrequency

NOWCASTING REPORT. Updated: May 20, 2016

OREGON POPULATION FORECAST PROGRAM

NRC Workshop - Probabilistic Flood Hazard Assessment Jan 2013

Forecasting Regional Employment in Germany: A Review of Neural Network Approaches. Objectives:

NOWCASTING REPORT. Updated: October 21, 2016

Changes in the Spatial Distribution of Mobile Source Emissions due to the Interactions between Land-use and Regional Transportation Systems

3 Time Series Regression

VALIDATING THE RELATIONSHIP BETWEEN URBAN FORM AND TRAVEL BEHAVIOR WITH VEHICLE MILES TRAVELLED. A Thesis RAJANESH KAKUMANI

Facts and Findings. Exhibit A-1

TIME SERIES ANALYSIS AND FORECASTING USING THE STATISTICAL MODEL ARIMA

Short-Term Job Growth Impacts of Hurricane Harvey on the Gulf Coast and Texas

Forecasting Chapter 3

TOOLS AND DATA NEEDS FOR FORECASTING AND EARLY WARNING

BEFORE THE PUBLIC UTILITIES COMMISSION OF THE STATE OF COLORADO * * * *

Abram Gross Yafeng Peng Jedidiah Shirey

Enhancing Weather Information with Probability Forecasts. An Information Statement of the American Meteorological Society

NOWCASTING REPORT. Updated: September 23, 2016

The Role of Weather in Risk Management For the Market Technician s Association October 15, 2013

Oregon Population Forecast Program Rulemaking Advisory Committee (RAC) Population Research Center (PRC)

BEFORE THE PUBLIC UTILITY COMMISSION OF THE STATE OF OREGON UE 294. Load Forecast PORTLAND GENERAL ELECTRIC COMPANY. Direct Testimony and Exhibits of

An online data and consulting resource of THE UNIVERSITY OF TOLEDO THE JACK FORD URBAN AFFAIRS CENTER

Forecasting demand in the National Electricity Market. October 2017

DEPARTMENT OF QUANTITATIVE METHODS & INFORMATION SYSTEMS

Analyzing the effect of Weather on Uber Ridership

Measuring the Economic Impact of Tourism on Cities. Professor Bruce Prideaux James Cook University Australia

As included in Load Forecast Review Report (Page 1):

Ameren Missouri Peak Load Forecast Energy Forecasting Meeting, Las Vegas. April 17-18, 2013

at least 50 and preferably 100 observations should be available to build a proper model

Decision 411: Class 7

From Sales to Peak, Getting It Right Long-Term Demand Forecasting

Transcription:

No part of this presentation may be copied, reproduced, or otherwise utilized without permission. WRF Webcast Improving the Accuracy of Short-Term Water Demand Forecasts August 29, 2017

Presenters Maureen Hodgins, Research Manager Water Research Foundation Thomas M. Fullerton, Jr., Ph.D. Professor & Trade in the Americas Chair University of Texas at El Paso (UTEP) David Torres, M.S. Economist El Paso Water Adam Walke, M.S. Economist UTEP Border Region Modeling Project

Webcast Agenda Introductions Maureen Hodgins 5 min Overview Tom Fullerton 5 min Key Research Findings Adam Walke 30 min El Paso Water Modeling Efforts David Torres 5 min Question & Answers 15 min

WRF Research Focus Area - Demand 2012 to 2018/19 10 projects 8 funded Forecasting

Demand Forecasting Summary

Water use estimates -Residential end uses, 4309 (2016) -Multi-family, 4554 (est 2018) -CII, 4375 (2015) & 4619 (est 2019) Forecasting methods -Customer data, 4527 (2016) -Short term, 4501, (2017) -Long term, 4667 (est 2020) Factors impacting demand -Recession, 4458 (2016) -United Kingdom - Behavioral changes, 4649 (2016) -Passive efficiency, 4495 (est 2018) Planning with uncertainty -Uncertainty & long term forecasts, 4558 (2016) Urban Landscapes -Irrigation controllers, 4227 (2016) -Urban landscape research needs (est 2017) Sizing Infrastructure -Demand patterns for sizing meters & service lines, 4689 (est 2018)

Finding the 4501 Products

4501 Products

Please Type Your Question Here! Slides and recording will be available to WRF subscribers WITHIN 24 hours after the webcast!

Agenda Overview of project 4501 Key findings: Survey of water utility planners & forecasters Short-term water demand forecasting manual Final Report: Accuracy of utility-generated forecasts Forecasting case study: El Paso Water Brief overview of add-on case studies El Paso Water modeling efforts Questions & answers

Acknowledgements Project Advisory Committee: Jack Kiefer (Hazen and Sawyer) Chris Meenan (Las Vegas Valley Water District) Paul Merchant (South West Water) Graduate Research Assistants: Juan P. Cardenas Alejandro Ceballos Omar Solis

Acknowledgements Water Research Foundation El Paso Water City of Phoenix Tampa Bay Water Three anonymous utilities Hunt Communities

Survey Targeted to utility staff involved in shortterm planning and forecasting 198 unique survey responses 42 states represented 70% of respondents use water demand forecasts - only those respondents answered the subsequent questions

Approximately how often are forecasts generated or updated? Source: Fullerton, T.F., Jr. and Walke, A.G. (2017) Improving the Accuracy of Short-Term Water Demand Forecasts. Denver, Colo.: Water Research Foundation.

If forecasts are generated for different water demand scenarios, what types of scenarios are considered? Source: Fullerton, T.F., Jr. and Walke, A.G. (2017) Improving the Accuracy of Short-Term Water Demand Forecasts. Denver, Colo.: Water Research Foundation.

What types of forecasting methodologies are used to generate your utility's forecasts? Source: Fullerton, T.F., Jr. and Walke, A.G. (2017) Improving the Accuracy of Short-Term Water Demand Forecasts. Denver, Colo.: Water Research Foundation.

Forecasting Manual Organized around:

Factors to Consider in Choosing a Forecasting Methodology Accuracy track record of the methodology Cost of implementing the methodology Data and computational requirements Organizational goals in forecasting Prediction only versus analyzing scenarios & alternative policies Forecast horizon (short- medium- or longterm forecasts)

Source: Fullerton, T.F., Jr. and Walke, A.G. (2017) Improving the Accuracy of Short-Term Water Demand Forecasts. Denver, Colo.: Water Research Foundation.

Factors to Consider in Developing a Forecast Sample size & handling of missing data Time between forecasts & vintage of data Which demand variable to forecast: Demand in each customer category? Demand in each part of the service area? Separate forecasts for the customer base and water demand per customer? Which (if any) predictor variables to use

Predictor variables used in 53 published studies for short-term water demand forecasting Source: Fullerton, T.F., Jr. and Walke, A.G. (2017) Improving the Accuracy of Short-Term Water Demand Forecasts. Denver, Colo.: Water Research Foundation.

Issues to Consider in Evaluating a Forecast Evaluation serves to: Assess whether previous methods are working Choose models with good chances of success Good compared to what? Use benchmarks Evaluation criteria Forecast error summary statistics Tests of forecast error differentials Tests of directional forecast accuracy

Final Report Analysis of utility-generated forecasts: Tampa Bay Water City of Phoenix Three anonymous utilities Analysis of forecasts developed by the research team with utility data: El Paso Water City of Phoenix One anonymous utility

Analysis of Utility-Generated Forecasts Accuracy analyses were conducted for 5 sets of utility-generated forecasts; in one case there were not enough observations for statistical tests of forecast accuracy. Utility forecasts were compared to random walk benchmark forecasts. Forecasts were grouped in different ways (e.g. by step-length, by geography, by customer class).

Utility Forecast Accuracy Summary: Root Mean Squared Error (RMSE) # Horizon (Frequency) Methodology Utility % Better 1 2 Weeks (Daily) 2 1 Week (Weekly) 3 1 Year (Monthly) 4 2 Years (Monthly) 5 4 Years (Annual) Regression with weather and demographic explanatory variables and lagged demand Regression with weather explanatory variables and lagged demand Econometric model including price, weather, employment, unemployment rate, and lagged demand Expert judgment taking into account climatic & economic conditions End-use model based on survey data, and data on demographic trends, prices, and conservation policies 71% 100% 75% Source: Fullerton, T.F., Jr. and Walke, A.G. (2017) Improving the Accuracy of Short-Term Water Demand Forecasts. Denver, Colo.: Water Research Foundation. 0% 0%

Utility Forecast Accuracy Summary: Error Differential Regression Test # Horizon (Frequency) Methodology Utility % Significant 1 2 Weeks (Daily) 2 1 Week (Weekly) 3 1 Year (Monthly) 4 2 Years (Monthly) 5 4 Years (Annual) Regression with weather and demographic explanatory variables and lagged demand Regression with weather explanatory variables and lagged demand Econometric model including price, weather, employment, unemployment rate, and lagged demand Expert judgment taking into account climatic & economic conditions End-use model based on survey data, and data on demographic trends, prices, and conservation policies 29% 83% 50% Source: Fullerton, T.F., Jr. and Walke, A.G. (2017) Improving the Accuracy of Short-Term Water Demand Forecasts. Denver, Colo.: Water Research Foundation. 0% NA

Utility Forecast Accuracy Summary: Chi-Square Test of Independence # Horizon (Frequency) Methodology Utility % Significant 1 2 Weeks (Daily) 2 1 Week (Weekly) 3 1 Year (Monthly) 4 2 Years (Monthly) 5 4 Years (Annual) Regression with weather and demographic explanatory variables and lagged demand Regression with weather explanatory variables and lagged demand Econometric model including price, weather, employment, unemployment rate, and lagged demand Expert judgment taking into account climatic & economic conditions End-use model based on survey data, and data on demographic trends, prices, and conservation policies 57% 78% 75% Source: Fullerton, T.F., Jr. and Walke, A.G. (2017) Improving the Accuracy of Short-Term Water Demand Forecasts. Denver, Colo.: Water Research Foundation. 0% NA

Characteristics of Successful Utility-Generated Forecasts Combine statistical methods like regression with expert judgment Use time series data on key predictor variables like weather, prices, and economic indicators In one case, forecasts are generated without forecasting predictor variables All successful models harness the predictive power of lagged demand

El Paso Case Study El Paso Water serves El Paso, Texas Located in the Chihuahuan Desert Historically faced water supply constraints Adopted a comprehensive water conservation strategy in 1991

Total Water Consumption in El Paso: 1959-2015 Source: Fullerton, T.F., Jr. and Walke, A.G. (2017) Improving the Accuracy of Short-Term Water Demand Forecasts. Denver, Colo.: Water Research Foundation.

Methodology I: Linear Transfer Function ARIMA In a Linear Transfer Function (LTF) approach, demand is first modeled as a function of explanatory variable lags. Any remaining unexplained systematic variation in demand is then modeled using autoregressive (AR) and moving average (MA) parameters.

Methodology II: Vector Autoregression In the most basic version of the Vector Autoregression (VAR) approach, each variable is modeled as a function of lags of itself and lags of all of the other variables. Instead of choosing specific lags of each variable to include in the model, the analyst only has to select one lag order for all of the variables.

Methodology III: Random Walk A simple random walk (RW) forecast posits no change in demand (forecast equals actual demand one period earlier): F t = A t-1 A random walk with drift (RWD) simply adds the average annual change in demand to the random walk: F t = A t-1 + d In the case of highly seasonal variables measured at a monthly frequency, a random walk can be defined as: F t = A t-12

Data (January 1994 December 2013) From El Paso Water: Total Demand = Per-Customer Demand Customer Base Real Average Price = Total Revenues/(Total Demand CPI) From the Bureau of Labor Statistics: Nonfarm Employment in El Paso County From the National Oceanic and Atmospheric Administration: Days per Month with Temperatures above 90 F Total Monthly Precipitation in Inches The customer base is first-differenced; the other variables are first- and twelfth-differenced.

Estimation Results LTF and VAR methods can provide insights into the relationships between variables: Slope coefficients indicate how demand changes when explanatory variables change. In the case of the VAR model, there is a very large number of coefficients, so an alternative means of deciphering relationships is desirable. An impulse response function is a convenient way of depicting the reactions of endogenous variables to shocks in other endogenous variables.

LTF Estimation Results Per-Customer Usage Customers Constant 0.002 345.198** Average Price t-3-2.899** Days over 90 F t 0.089** Days over 90 F t-1 0.131** Rainfall t-1-0.396** Nonfarm Employment t 0.121** Nonfarm Employment t-32 42.440** AR t-12-0.295** 0.369** AR t-18-0.175** MA t-1-0.706** -0.299** MA t-3 0.157* MA t-12-0.265** R-Squared 0.702 0.237 F-Statistic 59.808** 10.602** * Probability value <.05; ** Probability value <.01 Source: Fullerton, T.F., Jr. and Walke, A.G. (2017) Improving the Accuracy of Short-Term Water Demand Forecasts. Denver, Colo.: Water Research Foundation.

Elasticity Estimates The estimated elasticity of demand with respect to price is -0.32. This suggests that a 10% price increase would result in a 3.2% decline in demand.

Change in Deseasonalized Per Capita Demand (Thousands of Gallons per Month) Per-Customer Demand: VAR Impulse Response Function 1.5 1.0 0.5 0.0-0.5-1.0 1 2 3 4 5 6 7 8 9 10 The figure shows the response of demand to a one standard deviation shock in the residual of price Source: Fullerton, T.F., Jr. and Walke, A.G. (2017) Improving the Accuracy of Short-Term Water Demand Forecasts. Denver, Colo.: Water Research Foundation.

Forecast Evaluation Criteria Criteria based on the size of the forecast errors: Root Mean Squared Error Forecast error differential regression test H 0 : Both sets of forecasts are equally accurate Criteria for evaluating directional accuracy: Chi-square test H 0 : Forecasted and actual events are independent (i.e. forecasts don t provide useful information for predicting the direction of change)

Most Accurate Forecast for El Paso based on Root Mean Squared Error, 2011-2013 Step-length Per-customer demand Customer base 1 month LTF VAR 2 months LTF LTF 3 months LTF LTF 4 months LTF RWD 5 months LTF RWD 6 months LTF RWD 7 months LTF RWD 8 months LTF RWD 9 months LTF VAR 10 months LTF VAR 11 months LTF VAR 12 months LTF VAR Source: Fullerton, T.F., Jr. and Walke, A.G. (2017) Improving the Accuracy of Short-Term Water Demand Forecasts. Denver, Colo.: Water Research Foundation.

Are the LTF forecasts significantly better than the VAR forecasts? Step-length Per-customer demand Customer base 1 month no no 2 months no no 3 months yes no 4 months yes no 5 months yes no 6 months yes no 7 months yes no 8 months yes no 9 months yes no 10 months yes no 11 months yes no 12 months yes no Source: Fullerton, T.F., Jr. and Walke, A.G. (2017) Improving the Accuracy of Short-Term Water Demand Forecasts. Denver, Colo.: Water Research Foundation.

Are the LTF forecasts significantly better than the Random Walk with Drift forecasts? Step-length Per-customer demand Customer base 1 month yes no 2 months yes no 3 months yes no 4 months yes no 5 months yes no 6 months yes no 7 months yes no 8 months yes no 9 months yes no 10 months yes yes 11 months yes yes 12 months yes yes Source: Fullerton, T.F., Jr. and Walke, A.G. (2017) Improving the Accuracy of Short-Term Water Demand Forecasts. Denver, Colo.: Water Research Foundation.

Do the forecasts provide useful information on directional changes in per-customer demand? Step-length LTF VAR 1 month yes yes 2 months yes yes 3 months yes yes 4 months yes no 5 months yes yes 6 months yes yes 7 months yes yes 8 months yes yes 9 months yes yes 10 months yes no 11 months yes no 12 months yes no Source: Fullerton, T.F., Jr. and Walke, A.G. (2017) Improving the Accuracy of Short-Term Water Demand Forecasts. Denver, Colo.: Water Research Foundation.

Combining Forecasts If separate forecasts each contribute complementary information that is useful for prediction, it may make sense to combine the forecasts as a strategy for improving accuracy. Forecasts can be combined by assigning a separate weight to each one based on it s relative accuracy. Demand = b 0 + b 1 LTF + b 2 VAR + b 3 RWD + e

Per-Customer Demand: Combined Results 2008-2013 Variable Coefficient t-statistic Probability Constant 1.5737 3.0507 0.0033 LTF 0.5893 6.4515 0.0000 VAR 0.2957 3.0559 0.0033 RWD 0.0259 0.3905 0.6974 AR t-1 0.3035 2.4850 0.0155 MA t-3 0.4215 3.6689 0.0005 R-squared 0.9732 Durbin-Watson 2.0305 F-statistic 472.8418 Probability (F-stat) 0.0000

Customer Base: Combined Results 2008-2013 Variable Coefficient t-statistic Probability Constant 1,446.399 1.4732 0.1454 LTF 0.0749 0.5023 0.6171 VAR 0.0360 0.2256 0.8222 RWD 0.8817 20.2273 0.0000 MA t-5-0.3338-2.6728 0.0094 R-squared 0.9985 Durbin-Watson 2.0701 F-statistic 11,119.20 Probability (F-stat) 0.0000

Conclusions of El Paso Case Study The LTF approach offers an improvement over the alternatives considered in forecasting percustomer water demand for El Paso. Random walks are competitive in forecasting the number of customers. Experimentation with alternative forecasting methods and comparisons of forecast accuracy can help inform decisions about what is the best forecasting approach to use in a given context.

Results of Add-On Case Study I Are there benefits to separately modeling and forecasting each sub-component of total water demand? Data from the City of Phoenix were analyzed. Disaggregation by customer category is the best alternative for a majority of the steplengths considered, but it is not significantly better than directly modeling aggregate demand for this sample.

Results of Add-On Case Study II Can proxies serve as predictor variables in place of price data when the latter are unavailable? Data from an anonymous utility were analyzed to address this question. During a period when water rates were changing significantly, proxy variables were poor substitutes for actual price data in forecasting water demand.

Overarching Conclusion Forecast evaluation is not the final word in the forecasting process but a tool for continual reassessment and improvement of prediction strategies.

Questions?

Thank You Comments or questions, please contact: mhodgins@waterrf.org tomf@utep.edu agwalke@utep.edu david.torres@epwu.org For more information visit: www.waterrf.org