Document Type : Research Paper

Authors

• Zahra Ghafari Moghadam ¹
• Ali Sardar Shahraki ²
• Abdulhaseeb Azizi ³
• Somayyeh Mirshekari ⁴

¹ Assistant Professor, Agricultural Research Institute, Zabol University

² professor of agriculture economics, University of Sistan and Baluchestan

³ Center for Development Research (ZEF), University of Bonn, Germany

⁴ Agriculture Institute, Research Institute of zabol, Zabol, Iran.

Abstract

Accurate forecasting of urban water demand is essential for effective water resources management, especially in arid and water-stressed regions. This study evaluates the performance of Recurrent Neural Networks (RNN) and Long Short-Term Memory (LSTM) models for forecasting monthly urban potable water demand in the Sistan region of Iran using data from April 2006 to March 2021.
Two forecasting approaches are examined: a univariate approach based only on past water demand and a multivariate approach incorporating selected climatic and socioeconomic variables. A one-step-ahead monthly forecasting strategy is applied, with the dataset divided chronologically into training and testing subsets. The deep learning models are compared with two benchmark methods, Naïve and Seasonal Naïve, using MAE, RMSE, MAPE, and Pearson’s correlation coefficient. The results show that LSTM provides more stable predictions and higher correlation with observed demand than RNN. However, the simple benchmark models produce lower forecast errors overall, with the Naïve model achieving the best performance. Adding climatic and socioeconomic variables slightly improves correlation in some cases but does not consistently reduce errors. Overall, the findings suggest that urban water demand in the study area is strongly persistent and seasonal, indicating that simple forecasting methods can outperform more complex deep learning models under data-limited conditions.

Keywords

• Water demand forecasting
• Deep learning
• LSTM neural network
• RNN neural network
• Climate variability