Despite recent efforts to apply machine learning (ML) for water demand modeling, overcoming the black-box nature of these techniques to extract practical information remains a challenge, especially in developing countries. This study integrated random forest (RF), self-organizing map (SOM), and artificial neural network (ANN) techniques to assess water demand patterns and to develop a predictive model for the city of Fortaleza, Brazil. We performed the analysis at two spatial scales, with different level of information: census tracts (CTs) at the fine scale, and census blocks (CBs) at the coarse scale. At the CB scale, demand was modeled with socioeconomic, demographic, and household characteristics. The RF technique was applied to rank these variables, and the most relevant were used to cluster census blocks with SOMs. RFs and ANNs were used in an iterative approach to define the input variables for the predictive model with minimum redundancy. At the CT scale, demand was modeled using HDI and per capita income. Variables which assess the education level and economic aspects of households demonstrated a direct relationship with water demand. The analysis at the coarse scale provided more insight into the relationship between the variables; however, the predictive model performed better at the fine scale. This study demonstrates how data-driven models can be helpful for water management, especially in environments with strong socioeconomic inequalities, where urban planning decisions should be integrated and inclusive.