Abstract

Accurate forecasting of household water consumption is essential for sustainable water-resources planning; however, guiding model selection remains challenging due to fragmented and heterogeneous evidence. Following the PRISMA methodology, we screened 4264 records from major electronic databases, including ScienceDirect and Web of Science, and identified 80 primary studies published between 2009 and 2024. The evidence is systematically synthesized to characterize dominant machine-learning model families, predictor patterns, and evaluation practices, and to translate these findings into actionable guidance for model selection. The analysis shows that multilayer perceptrons and recurrent architectures (e.g., LSTM) are predominantly used for sequence-based forecasting tasks relying on historical consumption and climatic predictors, whereas tree-based ensemble models are more frequently adopted in heterogeneous tabular settings incorporating property-related descriptors. RMSE and MAE emerge as the most commonly reported evaluation metrics across studies. Finally, the review highlights key research gaps and future directions, including the limited use of socio-economic predictors, the need for standardized hybrid modeling pipelines, and the importance of improving model interpretability for operational deployment.