Project ELVIRA
Natural Gas Consumption Forecasting
Contents:
1 Introduction
There is no need to emphasize strongly the economical aspect of gas consumption forecasting in the current conditions of gas market deregulation. Even a small improvement of a forecasting error can bring significant economic savings.

One of the ways how to improve the forecasting quality is the use of computer systems both for automatic time series forecasting and also for decision support systems with an interactive feedback connection that can help experts (dispatchers, traders) in their decision, planning and control processes.
2 Gas Consumption Forecasting
As an example of a decision support system for gas consumption forecasting we introduce the forecasting system ELVIRA. Prediction mathematical models were developed in the Institute of Computer Science and were implemented in several utility companies across the Czech Republic and in the Slovak Republic.

ELVIRA is a complex modular system which solves several problems from the point of view of the load distributing utility. It includes three prediction modules for different time resolutions hours, days and months. The forecasting horizon varies form one-hour ahead up to one-year ahead.

Fig. 1: The ELVIRA modelling system flow chart

Fig. 2: The forecasting system ELVIRA
2.1 Forecasting System Components
2.1.1 Short-Term Load Forecasting Module
This module forecasts a daily value or a daily diagram (1 value per day or 24 values per day or 48 half-hour values per day) from one to seven days ahead with respect to:

  • real values of consumption, pressures, temperatures etc. gathered on from gas utility information systems;
  • expected temperatures.
The module includes an intelligent data management module for automatic detection of missing values or outliers.
2.1.2 Meteorological Module
The module performs load forecast corrections (daily diagrams, daily maximal values, week maximal values) with respect to the different weather forecast scenarios including the expected maximal and minimal temperatures, wind speeds, cloud cover variables, etc.
2.1.3 Middle Term Load Forecasting Module
The module forecasts the gas consumption with a forecasting horizon from one week up to one year (daily averages, daily maximal values, week maximal values, week averages, month maximal values, month averages) using the decomposition methods and the expected long-term meteorological situations.
2.2 Forecasting Methods
The forecasting system consists of several different models based on dozen of forecasting methods. Different methods are suitable for different forecasting horizons or different forecasting resolutions (hours, days, months). Moreover, several different forecasting models are deployed in forecast of one value. The best forecast is then automatically selected by forecasting system based on internal heuristics.

Examples of forecasting methods used are:
  • Box-Jenkins models,
  • Case based reasoning,
  • Rule based systems,
  • Artificial neural networks,
  • Decomposition time series.
The ELVIRA system automatically loads necessary data from the customer information system and automatically selects the best forecasting method.
2.3 Forecasting errors
For the risk management process it is necessary to characterize the uncertainty included in the forecasting models. The mean average absolute percentage error (MAPE) of the ELVIRA system is about 1%-3% (the concrete value depends on the region size, evaluation period, etc.). The typical duration of the residual process (time series of the prediction errors) is shown in Fig. 2. The empirical probability distribution of the residuals is shown in Fig. 3.

Fig. 3: The prediction errors (residual process)

Fig. 4: Empirical probability distribution of residuals
References
[1] Goldberg, D.E.: Genetic Algorithms in Search, Optimization and Machine Learning, Addison-Wesley, NY, 1989.
[2] Payne F. William: User´s Guide to Natural Gas Purchasing and Risk Management, The Fairmont Press, Prentice Hall, ISBN: 0-13-017931-0, 2000.
[3] Vose D.: Risk Analysis, John Wiley, ISBN 0-471-99765-X, New York, 2001.
[4] Pelikan E., Petrak L.: Electric Load Forecasting with Theory of Extreme Values, Neural Network World Journal, Vol. 9, Number 4, 1999, pp. 297-304.
[5] Pelikan E., Simunek M., Eben K., Jurus P.: Short-Term Forecasting of Natural Gas Consumption, a poster presented at the 6th World Symposium on Information and Communication Technology in the Gas Industry, Prague, Czech Republic, April 23-25, 2002.
[6] Pelikan E., Simunek M., Brabec T.: Load Forecasting Using The system Elvira, in Proceedings of the 3rd SIMONE workshop (CD) , Cesky Krumlov, May 11-14, 2004.