Water resource management in urban areas relies on precise demand prediction, in-time anomaly monitoring, and intelligent decision-making. This study develops a Fusion-based hybrid framework based on deep learning and probabilistic reasoning for improved water demand prediction and wastewater system management. Fusion-based model integrates Long Short-Term Memory (LSTM) for temporal water demand forecasting, Convolutional Neural Networks (CNN) to analyze spatial patterns, and Autoencoders (AEs) for unsupervised anomaly detection, with the emphasis on the detection of potentially contaminated water, and also the framework includes a Bayesian Network-based decision support system. Through experiments with synthetic and real datasets, the Fusion-based methodology achieves significantly better prediction accuracy and anomaly detection of water contamination precision, and is more reliable for decision-making than single model initiatives. A hybrid model with smart sensors provides an auditable, scalable, and smart approach for sustainable urban water management.