Water resource management is critical for sustainable development and is essential for the smooth functioning of ecosystems and human activities. Yet, water scarcity and drought can severely impact water supplies and agricultural productivity. Consequently, there is an urgent need for accurate predictions of water scarcity and drought conditions. Despite this, few studies have used machine learning to predict water scarcity and drought accurately. Therefore, this study collected daily hydrological data from January 1, 2016 to March 24, 2023, totaling 21,120 observations. In light of this, we developed a deep learning integrated model (CNN-BILSTM-AM) that combined an attention mechanism (AM), a convolutional neural network (CNN), and a bi-directional long-term and short-term memory network (BILSTM) to predict water scarcity and drought. The outcomes demonstrate that the integrated CNN-BILSTM-AM model effectively captures the nonlinear and time-varying characteristics of water scarcity and drought. With excellent adaptability to random sample selection, data frequency, and sample structure breaks, its prediction accuracy (with a value of 95.09%) significantly outperforms that of the traditional and single models. This study expands the knowledge of machine learning in water resource management and extends the research into the prediction of water scarcity and drought. It provides decision support and risk management tools for water resource managers, agricultural stakeholders, and government decision-makers.