Loading...

Urban Water Demand Forecasting with Dynamic Artificial Neural Networks

Fa'al, Fatemeh | 2010

622 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 41127 (09)
  4. University: Sharif University of Technology
  5. Department: Civil Engineering
  6. Advisor(s): Abrishamchi, Ahmad
  7. Abstract:
  8. The water demand forecasting is an important activity for successful planning, utilization and operation of urban water supply and distribution systems. The population growth result in water consumption growth, also the restriction of water resources lead to pay more attention to the water demand management. The unexpected droughts, financial crises, over-use of water resources, or inessential infrastructure development are the outcome of poor water demand prediction and inflexible water resource management. This research is addressed the daily short-term (two week ahead), weekly medium-term (six months ahead), and monthly long-term (two years ahead) water demand. The dynamic artificial neural network (DAN2), focused time-delay neural network (FTDNN), and K- nearest neighbor models are used to forecast these demands. The number of layers is defined in DAN2 architecture so that with increasing each layer, the accumulated knowledge is monotonically increased, total error is decreased, and the network training is improved. The FTDNN model is a type of dynamic neural networks in which the dynamics was appeared only at the input layer of a static multilayer feedforward network. The K- nearest neighbor is based on finding the historical nearest neighbor of current feature vector and resampling from their successors using a discrete resampling kernel. These models have been executed using the daily production and monthly water consumption data of Tehran metropolis. The results of daily model have shown that partitioning into weekdays and weekends did not provide better outcomes. In the weekly model, the improved results achieved by straight forecasting using daily data then summing up them to produce corresponding weekly values. For the monthly models, dividing the year into highseason and lowseason enhanced the accuracy and square of correlation coefficient of forecasting. The more accurate performance has been represented in the DAN2 model. Furthermore the results have shown that, the three models have desirable performance generally
  9. Keywords:
  10. Forecasting ; Neural Networks ; Urban Water Demand ; K-Nearest Neighbor Method

 Digital Object List

 Bookmark

No TOC