To enhance the accuracy of machine learning algorithms in urban water demand forecast, this study proposes a systematic forecasting method based on rolling cross-validation. The method integrates the construction of an influencing factor indicator system, the development of forecast models, hyperparameter optimization guided by rolling cross-validation, and comprehensive model performance evaluation and model selection. The method is validated using Hengyang City as a case study. The results indicate that the forecast urban water demand of Hengyang City for 2025 closely aligns with the planned values, demonstrating the applicability and practical value of the method. Moreover, the method exhibits strong generation ability, allowing flexible adaptation of indicator systems and model structures to accommodate varying regional socio-economic trends and water use characteristics. The incorporation of rolling cross-validation into the hyperparameter optimization process significantly enhances the generalization capability and forecast accuracy of the models, making the method well-suited for real applications in urban water demand management.