Modelling of uncertainties in Water Demand Forecasting Presented to Decision Tools Conference Presentation Tuesday, 2 nd November 2015 Sofitel Downtown Hotel Dubai, U.A.E Abu Dhabi Water & Electricity Company www.adwec.ae Tel: +971-2-6943333 Fax: +971-2-6425773
ADWEC is responsible for Water Demand Forecasting of Abu Dhabi Forecasting is a research of the future which involve, examining the alternative futures and identifying the most probable. It helps decision making and planning for the future in the present. ADWEC is mandated to undertake the forecast of Electricity and water demands in the Emirate of Abu Dhabi. The major objectives of ADWEC s Water Demand Forecast are; To plan the water production capacities and transmission infrastructure. To calculate the Fuel Requirement. Calculation of BST (Bulk Supply Tariff) Prior to 2006, ADWEC was using conventional forecasting techniques based on the trend analysis. The demand forecast prepared was deterministic based on a single set of assumptions. Since 2006 onwards, ADWEC adopted a probabilistic approach which will consider a number of possibilities of future demand development. ADWEC uses the @Risk software of Palisade in order to model the uncertainty parameters in the demand forecast calculations. 2
ADWEC's Water Demand Forecast is a Probabilistic Forecast. Why Probabilistic Forecast? Why @Risk? ADWEC developed the probabilistic forecasting model for dealing with the high growth and high uncertainty situation in the region. The probabilistic model enables ADWEC ; To develop a broad and comprehensive framework by incorporating all major factors that affect the demand growth. To understand and analyze the overall level of uncertainty To achieve an appropriate level of flexibility in required capacity calculation and sensitivity analysis. The probabilistic model using @Risk directly calculate the Capacity requirement as stipulated by RSB, 2% (a DSS of 1 in 50 Years). The generation capacity required to avoid the chances of demand exceeding the capacity only in 2% of the time can be directly calculated as the 98 th percentile of (demand +outage) with all uncertainties. The @Risk allows the model building in the familiar platform of Excel and at the same time it offer the facility to deal with uncertain variables. 3
ADWEC Water Demand Forecasting Data Inputs Selection of a water demand forecast methodology is driven in part by the data that can be made available. The major data inputs to the demand forecast model are ; Demographic Data (Current and Future) Current supply, Consumption rates etc.. Data on proposed and under construction developments (Mega projects data) GDP and Employment growth rate ADWEA/ Government Policies on Tariff and water conservation etc.. Part of the above data are direct inputs to the model whereas some of the data such as Government policies and Demand Management efforts etc. are indirectly built into the model. It is amply clear that the future outlook of most of these data inputs mentioned above have a lot of uncertainties. Hence the forecasting accuracy and usefulness of the forecast depends greatly on how these uncertainties are treated in the forecasting model. 4
FACTORS EFFECTING FUTURE DEMAND GROWTH Population Growth Water Tariff Consumption Rate GDP/ Industrial Growth FUTURE DEMAND GROWTH Water Policy / DSM Agricultural Policy Environment Policy/ ESTIDAMA
ADWEC Water Demand Forecasting Data Inputs as Probability Distribution Functions Traditionally the total water demand is calculated by the formula Number of users X Unit consumption rate of a user Here the users include; The population (Residential, retail, Office etc..) Housing Industries Farms etc.. To forecast the future size of the users the following data inputs are used; Census and other statistics data from the Statistics department and UPC (current and the future population, housing and other users such as farms, industry etc..) Data from the Developers, UPC, IDB etc. on new megaprojects and industries. Similarly the unit Consumption rate is also a variable depending on various factors such as: Demand side management Tariff Government policies Public Awareness etc.. ADWEC s Risk based model is trying to include these above variables (Demand users and the unit rates) as probability distribution functions. 6
Build-up of the forecasting model - A Bottom-up approach.. Model building.. Various demand parameters listed in the previous slides for the existing areas and new projects are identified and used as data inputs to calculate the demands for the individual localities. The input parameters for existing area demands include Current Population, Population Growth rates, Housing, bulk demands, Consumption rates, Transmission and distribution losses etc. Megaproject demands forecast inputs include the UPC plan data, the developers demand, current status, project phasing and market conditions etc The total Abu Dhabi Emirate Forecast is calculated from Demand Forecasts prepared separately for Abu Dhabi, WR and Al Ain down to locality levels and projects. To the Total of Abu Dhabi Emirate Demand the export commitments to Northern Emirates are added to get the Global Forecast. Simulation.. The global demand calculated by simulation for various demand combinations that may occur and the mean value of number of such possibilities is the Most Likely Demand. The Required Capacity is calculated such that it can meet the requirement up to 98th percentile of the (peak demand + outages). The flow chart of the demand forecast model is shown in the next slide where the uncertain (Risk) variables are indicated as circles. 7
EXISTING AREAS Census Data- Population, Housing, Farms etc 5-Year Statements by DISCOS and Transco Govt Policies- Eg DSM etc Rate of Consumption Trans. & Dist. losses ADWEC Existing Areas Demand Forecast Calculation Population/ Housing / bulk demands etc. Verification of Demand Forecast Reduction factor for 8 DSM Actual Supply Data ADWEC Existing Area Demand Forecast
Developers Submissions (Letters, Reports) MEGA PROJECTS UPC 2030 Plan Demand Data from Distribution Companies Lag Factor Phasing Factor Megaprojects Demand Data processing / Calculations Occupancy Factor Trans. & Dist. losses Verification of Demand Forecast Site Visits Actual Supply Data ADWEC Megaproject Demand Forecast
Selecting a Distribution Function. One of the major task involved in the modelling is to select appropriate distribution functions for the uncertainty parameters. The creation of a distribution function for the per-capita domestic consumption rate is discussed as an example. Water consumption pattern varies from person to person. ADWEC used the actual field data to formulate the Probability Distribution curve for the rate of consumption (ROC). The consumption rate data from a study involving 150 villas in gated communities is used for this purpose. The data from the Distribution companies are not available on a daily basis and the exact details on the number of occupants etc. are not available with them. Hence the data from the study where the daily consumption and the number of users etc. are available, is used to fit the distribution. 10
Distribution Curve Fitted from Consumption Rate Data Four different Distributions are Fitted with the data as shown in the above figure. It is found that based on AIV criteria Log Logistic distribution is having the highest ranking. Hence this distribution may be selected for domestic consumption rate. This distribution function was written to the cell and used in the forecasting calculation. 11
Comparison of Distribution Curve Consumption Rates It is interesting to see the characteristics of the distribution curves of different consumption rate parameters. The domestic consumption (inside the house) shows a distribution closer to Normal Distribution with a slight positive skew. However the landscape demand is highly skewed. The range is wide with no demand (zero) to exceptionally high demand in a few cases. This indicate wide variation in Irrigation consumption among different users. The total demand ie. The combination of inside domestic demand and outside landscaping demand is also shown. It has positive skew in between the two cases. 12
MEGA PROJECT RISK INPUT FUNCTIONS (Example) Triangular Distribution for phasing factor Triangular Distribution for Occupancy factor MEGA PROJECT RISK INPUT FUNCTIONS Phasing Occupancy rate Lag (Delay) Triangular Distribution for Lag (project delay) 13
Other uncertainty parameters.. Other parameters such as Demand Patterns, Losses, Bulk demand growth rate etc. are also modelled using various techniques. The Monthly Demand Pattern is represented by distribution functions fitted from the historical supply data. The transmission and distribution losses are modelled as normal distribution with the mean as targeted values and appropriate SD. By using the inputs as distribution functions, the out put (Global demand) is calculated with a range of values from the minimum to maximum possible. The band around the most likely value indicate the range of possible outputs. With this curve we can state the global demand with any required confidence percentage. 14
Sensitivity Analysis Charts 15
CONCLUSIONS It is a common practice to use arbitrary distribution functions in the model building. Selection of the appropriate distribution curve to represent the actual characteristics of the risk variable is always a challenge for the model building. Effective use of historical data and field study data can solve this problem to a great extent. Making use of actual data to construct the distribution functions for @Risk model helps to increase the reliability of the model. Use of historical data for the modelling of Monthly Demand Pattern (Variations in the demand on Monthly basis) helps in incorporating the actual historical trends in the model. Similarly the distribution function fitted from the consumption rate study helps to represent the full range of values of consumption rate. ADWEC is continuously working on various methods to make the water demand forecast model closer to the actual (real life) scenarios. 16
Thank You 17 17