Pipeline networks constitute the major bulk carriers for crude oil, natural gas, water and petroleum products in Western Canada. Each day millions of barrels ofcrude oil billions of cubic feet of natural gas as well as millions of gallons of water are transferred through pipeline networks from the sources to the users. Optimum operations scheduling of these pipeline. Systems can he used to increase the performance and reduce the energy consumption of pumping stations. A key factor in making such systems truly successful is the accuracy a/their demand prediction which is normally implemented with time series analysis. However, for many areas especially in Saskatchewan, weather, special events and other related parameters have major effects on the demand patterns which cannot simply be modeled with the time series techniques.

This paper demonstrates how artificial neural networks (ANN) improve the demand prediction of pipeline networks. The water demand patterns were modeled by historical data and related variables using a single continuous perceptron (SCP) and multilayer perceptron (MLP). The implementation was based on real-world data from the City of Regina's pipeline networks. Weights of the SCP model were interpreted to determine relationships between demand patterns and related variables. Finally, comparisons between the models and performance improving techniques were discussed.


The accuracy of demand prediction is an important parameter for the operation planning of pipeline network systems for crude oil natural gas and water transportation. For example the prediction of demand patterns can be used to find optimum pumping schedules. In the implementation of the intelligent system on monitoring and control of the water distribution system at the City of Regina, the knowledge of future demand has great effects on the operation performance of pipeline networks.

Most of the existing demand predictors are based on time series analysis, especially the ARIMA (Autoregressive Integrated Moving Average) model. For example. Quevedo et al1 used Box-Jenkins methodology to predict the water demand of Barcelona city's distribution network. However, it was found that the Box-]enkins's model is very sensitive to noise and is not appropriate for small data set.2

In addition there arc a few other approaches such as were examined (The intervention components Ie were omitted in Equations 5.2 and 5.3 for clarity):

Equation (5.1) (Available in full paper)

Equation (5.2) (Available in full paper)

Equation (5.3 (Available in full paper)

Equation (5.4) (Available in full paper)

Equation (5.5) (Available in full paper)

Then all the components were combined to predict the water demand as shown below:

Equation (5.6) (Available in full paper)

For Equation 5.6, the prediction error was reduced to 3.5% of the mean demand value. However, the errors from the other pipeline networks were lying between 4.16% to 6.74%. A Disadvantage of this method is that model testing and modification is time consuming. By contrast, the ANN approach docs not involve human intervention transfer function analysis and model selection because it can automatically capture the model from appropriate training data: autonomous updating is also possible.

This content is only available via PDF.
You can access this article if you purchase or spend a download.