Beruflich Dokumente
Kultur Dokumente
Demand Forecasting
Mohammad Karamouz
Professor, Amir Kabir University (Tehran Polytechnic)
Banafsheh Zahraie
Assistant Professor, Tehran University
September 6-17, 2003 Tehran, Iran
Points to be Addressed
Data and Information Needed for Long Lead Forecasting of the Regional Water Demand Municipal Water Demand Population Forecasting Methods: Short-term and Long-term Time Series Modeling: Basic Steps Agricultural Water Demands Climate Signals: Prediction of Dry and Wet Spells Environmental Water Demands
Projected trends of water use serve as measures to guide planners as they propose:
New water resources facilities Modification of existing systems New or revised operating rules Regulatory changes Revised laws Organizational changes Research Projects
Past histories of water use are valuable information in making estimates of future water use. This information indicates the principle factors in determining the future water use.
Data and Information Needed for Lead Forecasting of the Regional Water Demand Long
Population projections based on demographic studies and studies done by the agencies responsible for multiple-sector investment decisions Distribution of urban and rural population among subregions Gross national product (GNP) Expected use of brackish or ocean water
Projected outputs of agricultural, mining, electric power, and major manufacturing sectors for determining the regional distribution of activities based on the projected GNP Projected rates of per capita water use based on technological advancements and relative share of instream and offstream uses
Disaggregate Estimation of Water Use Precise estimation of municipal water use can be obtained by breaking down the total delivery of water to urban areas into a number of classes of water use and determining separate average rates of water use for each class. Water use within some homogenous sectors is less variable compared to the total water use, therefore, greater accuracy in water use estimation can be obtained.
1. Domestic Washing and cooking Toilets Bath and/or shower Laundry House cleaning Yard irrigation Swimming pool Car washing Other personal uses (hobbies, etc.)
2. Public services
Public swimming pools Governmental agencies and private firms Educational services (such as schools, universities) Firefighting Irrigation of parks, golf courses, etc. Health services (such as hospitals) Public baths, public toilets, etc. Cultural public services (such as libraries and museums) Street cleaning and sewer washing Entertainment and sport complexes (such as cinemas) Food and beverage services Accommodation services Barber shops and beauty parlors
3. Small industries (such as laundry stores) 4. Construction and public works 5. Water losses Leakage from pipes, valves, meters, etc. Evaporation in open reservoirs Overflow of reservoirs Defective elements of a water distribution network, such as cracked reservoirs, flow back through oneway valves and pumps, etc. Loss in production process (cooling, pumping, etc.) 6. Transportation Taxies, buses, and other conveyances stations Ports and airports Railways (stations and workshops)
1. Population Changes 2. Climate Variations 3. Hydraulic Characteristics of the Water Distribution Network 4. Price and Economic Incentives 5. Living Standards Demand= f (population, price, standards, past use, etc.) + g(climate variations, )
Price Elasticity: The percent of decrease in quantity demanded due to 1 percent increase in price
California 1989 Weber -0.06 / -0.23 winter US 1982 & 1967 Howe et al -0.57 / -0.86 summer US east 1982 & 1967 Howe et al -0.43 / -0.52 summer US west 1982 & 1967 Howe et al
Short-term Estimates for 1 to 10 Years Graphical Extension Method 1. Plot the population of the past census years against time 2. Sketch a curve that fits the data 3. Extend this curve into the future to obtain the projected population
18 6 18 18 18 9 19 0 19 1 19 2 19 19 4 19 5 19 19 0 60 19 8 70 19 9 20 0 0 0 0 0 30 0 0 0 0 80 70 0
18 60 18 70 18 80 18 90 19 00 19 10 19 20 19 30 19 40 19 50 19 60 19 70 19 80 19 90 20 00
dP/dt = Ka
Short-term Estimates for 1 to 10 Years Declining Growth Rate Method: This method
considers that the city has a saturation population and the rate of growth becomes less as the population approaches the saturation level.
b a
Time (years)
Population
ty Ci
ity
City
City
Years
Long-term Forecasting for 10 to 50 years or more Mathematical Logistic Curve Method: This
method is suitable for the study of large population centers such as large cities, states, or nations. On the basis of the study of growth curve, certain mathematical equations of an empirical curve conforming to this shape (S-shape) is proposed.
Population
Time (years)
Long-term Forecasting for 10 to 50 years or more Ratio and Correlation Method: This
method is suitable for an area, which is a part of a region, state, nation, or larger area. It is assumed that the growth of the smaller area has some relation to the growth of the larger area.
Long-term Forecasting for 10 to 50 years or more Component Method: In this method, population
change is disaggregated to the changes due to: 1. Birth 2. Death B = K1 P0 t
D = K2 P0 t
3. Migration(M) Pt = P0 + B D M
A time series is a sequence of values arranged in their order of occurrence in time. A process is a mathematical description of the behavior of a phenomenon in one or more dimensions in space and/or time. Because all hydrologic phenomena change in space or time, they are hydrologic processes. If a process contains a random component, it is a stochastic process , which is a family of random variables, defined on a probability space. Stochastic processes are subdivided into stationary and non-stationary.
A stochastic process is stationary if the expected values of statistical descriptors do not change over time. If a time series is stationary, the series should be divided into a number of no-overlapping sub-series and the expected values of statistical descriptors of each series should be the same for each of the subseries.
Trend is a unidirectional gradual change (increasing or decreasing) in the average value of the variable. Changes in nature caused by human activities are the main reason for the over-several-years trends. Trends are usually smooth, and we should be able to represent it by a continuous and differentiable function of time. Trend is usually considered to be deterministic and it can be modeled by linear or polynomial functions: Linear function Polynomial function Power functions
7000 6000 5000 4000 3000 2000 1000 0 1 20 39 58 77 96 115 134 153 172 191 210 229
Jump is a sudden positive or negative change in the observed values. Human activities and natural disruptions are the main reasons for jumps in time series.
9000 8000 7000 6000 5000 4000 3000 2000 1000 0 1 20 39 58 77 96 115 134 153 172 191 210 229
Periodicity represents cyclic variations in a time series. These variations are repetitive over fixed intervals of time.
Randomness represents variations due to the uncertain nature of the stochastic process. The random component of time series can be classified as autoregressive or purely random. 1.2
1 0.8 0.6 0.4 0.2 0 1 51 101 151 201 251
TREND REMOVAL
TREND REMOVAL
2000 115 134 153 172 191 210 229 1500 1000 500 0 -500 -1000 -1500 -2000 1 20 39 58 77 96 115 134 153 172 191 210 229
APPLYING TRANSFORMATIONS
In most of the parameter estimation methods, it is assumed that the time series probability distribution is normal, but in many cases the time series do not follow normal distribution, are asymmetrically distributed, or are bounded by zero. BOX-COX Transformation:
I I
(1 , 2 ) t ( 2 ) t
( I t + 2 ) 1.0 = 1 1 1 m
= m ln( I t + 2 )
BOX-COX TRANSFORMATION
BOX-COX TRANSFORMATION
Forecasted Series
The next step is to decide upon the use of a univariate or multivariate model, or a combination, with desegregation models. This decision can be made based on the characteristics of the water resources system and existing information. information
Z t = j =1 j Z t j + t
p
Autoregressive Moving Average (ARMA) models Time series could be better forecasted by adding a moving average component to the AR models. The combination of an autoregressive model of order p and a moving average model of order q makes an ARMA(p,q) model, which is formulated as follows:
Z t = 1 Z t 1 + ... + p Z t p + t 1 t 1 ... q t q
(B )(1 B )d z t = (B ) t
Where: (B ) = 1 1 B 2 B 2 .... p B p
(B ) = 1 1 B 2 B 2 .... q B q
Formulation of multiplicative ARIMA model is as follows;
B (B ) 1 B
w
( ) (
w D
(1 B )d z t = (B w ) (B ) t
Regression based Models Regression based models has been widely used for demand forecasting: 1-Linear Regression 2- Multiple Regression
The method of least squares finds that particular line where the aggregate deviation of data points above or below it is minimized. Rather than measuring it's separation in terms of the physical distance, the procedure is instead upon vertical deviations, which are then squared. This not only eliminates in measuring perpendicular line segments, but it provides summary statistics having desirable properties
Y ( X i ) = a + bX i
100(1 )%
B = b t
SYX 1 2 X ( X ) 2 n
Inferences regarding the intercept aren't as common as for the slope. The following expression shows the confidence interval estimate of the true intercept:
A = a + t SYX
2
1 X2 + n ( X X )2
Linear multiple regression analysis extends simple linear regression by considering two or more independent variables.
Y = a + b1 X 1 + b2 X 2
Agricultural water demands change from year to year and month to month. The parameters affecting agricultural water demands can be summarized as follows: Crop mix Irrigated and dry-land farming Period and sequence of cropping Physical characteristics of the water transfer and irrigation systems Market prices Climate variation Policies related to pricing, importing, and exporting of agricultural prod-ucts.
( ) Agricultural Water Demands (MCM/Year) 1000 1050 1100 1150 1200 850 900 950
13 49 13 51 13 53 13 55 13 57 13 59 13 61 13 63 13 65 13 67 13 69 13 71 13 73
3-year moving average of the entire region and fitted polynomials to dry and wet spells
400 rainfall (mm) 300 250 200 150 100 50 0 3 8 13 18 23 28 33 38 y = -1.7712x 3 + 30.219x 2 - 167.85x + 493.9 R2 = 0.9929 y = 0.4417x 3 - 41.429x 2 + 1276x - 12693 R2 = 0.9798 350 y = 0.3584x 3 - 15.639x 2 + 216.7x - 730.72 R2 = 0.8292 y = 5.266x 3 - 354.89x 2 + 7938.5x - 58723 R2 = 0.94
time (year)
Final Stage
El-Nino refers to appearance, around Christmas, of a warm ocean current off the South American coast, adjacent to Ecuador and extending into Peruvian waters.
over 20-30 percent of the lands are affected by ENSO (El-Nino Southern Oscillation) events Effect of ENSO on the average global precipitation is estimated as 15-25 percent. ENSO events have effects with delay on summer precipitation in East Asia. El-Nino Events usually are followed by: Summer precipitation less than normal in India and North Australia Winter precipitation higher than normal in southeast Asia
that sea surface water temperature has been significantly higher than normal level in the period of El-Nino events. In El-Nino years, highest sea surface temperatures have been occurred in Persian Golf compare to other water bodies around the world
Season 2
Dec.-Feb.
Redefinition of Seasons
0 0 1 2 3 4 5 6 7 8 9
Model Verification
500
Second Season
Dec.-Feb.
Results
Modified forecasts considering ENSO signals for the third season are as good as the official forecast for Salt River. In 42 percent of the analyzed years modified forecasts have been better than official forecasts. Significant improvement of the statistical forecasts for the third season (30 percent) was achieved by the proposed method when considering ENSO climate signals. Only in 7 percent of the seasons the algorithm was not able to improve or sustain the statistical streamflow forecast.
An environmental water demand (EWD) is the water regime required to sustain the ecological values of an aquatic ecosystem at a low level of risk. If this water requirement is adopted, then it is likely that a water body will: Be healthy. Look after the needs of animals and plants Maintain its biodiversity.
uses, river and estuarine health on current water priorities. management planning indicators, and water
2. Consult basin stakeholders and relevant scientific important values for each sub-basin experts to determine of ecosystem, recreation, aesthetics, in the categories physical landscape and consumptive/nonconsumptive use values 3. Assess environmental water requirements using the most appropriate scientific methodology on a catchment basis.
Deterministic Components:
1.
Population Changes
1. 2.
2.
3.
In order to incorporate the deterministic parameters in the long-term demand forecasting usually different scenarios based on the optimistic and pessimistic probable conditions are considered. These scenarios are defined based on the probable range of: 1. Migration 2. Birth Rate 3. Death Rate 4. Economic Policies and Incentives 5. Other Factors
Stochastic Components:
1.
Climate Variations
1. 2.
Univariate AR, ARMA, and ARIMA Models Multivariate AR, ARMA, and ARIMA Models