Accurate prediction of steam generator (SG) water levels under various operating conditions enhances the safety and economic efficiency of nuclear power plants. Based on existing machine learning and intelligent algorithm theories, this paper proposes an ECRBM-GRU-SSA ensemble model for SG water level prediction. The model employs an enhanced continuous restricted Boltzmann machine (ECRBM) for data augmentation and combines the prediction results of multiple gated recurrent unit (GRU) sub-models, optimized by the sparrow search algorithm (SSA), to generate the final output. To validate the model’s predictive performance, this study uses simulated SG water level data. Experimental results demonstrate that the proposed model exhibits a rational and effective structure, with predicted water levels closely aligning with actual variations. Furthermore, its evaluation metrics outperform those of other mainstream models. This model provides an effective tool for fault diagnosis and fault-tolerant control in SG-related systems, offering significant engineering value.