Differential Learning for Outliers: A Case Study of Water Demand Prediction