Mine water poses serious threats to mine safety and the sustainable exploitation of resources. Efficient identification of the water source of roof water inflow in coal seams is crucial. Hydrochemical analysis is key for mine water source identification, yet it is challenged by large, complex datasets. Deep learning offers an effective solution with its powerful data representation capabilities. This study develops a water source identification model based on a deep neural network (DNN), using hydrochemical and organic composition indicators including Na⁺, K⁺, Ca²⁺, Mg²⁺, SO₄²⁻, HCO₃⁻, Cl⁻, TDS, TOC, UV254, and dissolved organic matter. Bayesian optimization is applied to tune key hyperparameters of the DNN, such as learning rate, number of neurons in each layer, and training epochs, to achieve an optimal network architecture. The model is validated using 197 water samples collected from three representative coal mines located on the border between Inner Mongolia and Shaanxi Province, China. The proposed model achieves an identification accuracy of 96.31%, outperforming traditional classifiers such as support vector machines and random forests. The results indicate that this method has high accuracy and reliability, and can provide new ideas for quickly and accurately identifying the water source of coal seam roof water inflow.