Accurate prediction of urban water consumption supports sustainable resource management and aligns with the United Nations Sustainable Development Goals. This study applies the inclusive multiple modeling using deep learning models, implemented in MATLAB, to improve prediction robustness in the Mahabad water distribution network in West Azerbaijan Province, Iran. At the first level, Long Short-Term Memory and Group Method of Data Handling models predicted water consumption using precipitation, temperature, and population data. The second level employed a deep neural network, taking both inputs and outputs from the first-level models. Model performance was evaluated using the Nash–Sutcliffe efficiency, root mean square error, residual homoscedasticity, Taylor diagram, and peak demand analysis. The results indicate that the inclusive multiple modeling framework enhances prediction robustness by leveraging model strengths and reducing individual errors. The deep neural network at the second level outperformed other models, achieving an overall Nash–Sutcliffe efficiency of 0.956 and root mean square error of 0.046, compared to Long Short-Term Memory (Nash–Sutcliffe efficiency = 0.887, root mean square error = 0.074) and Group Method of Data Handling (Nash–Sutcliffe efficiency = 0.875, root mean square error = 0.078). Residual analysis confirmed stable error distribution for the deep neural network and Group Method of Data Handling, while Long Short-Term Memory showed heteroscedasticity. The Taylor diagram confirmed higher correlation and better standard deviation match for the deep neural network. The approach also accurately predicted extreme demand peaks, demonstrating its potential as a robust and transferable tool for urban water management under changing climatic and demographic conditions.
