Nuclear energy plays an increasingly crucial role in global decarbonization and sustainable economic development. Among the essential systems in a nuclear power plant, the nuclear reactor coolant system (RCS) is critical for maintaining operational safety by removing the heat generated by the reactor and serving as a secondary barrier to prevent radioactive material leakage. Accurate fault diagnosis of the RCS is crucial for enhancing the safety of nuclear power plants. In order to address the issues of low training efficiency and insufficient accuracy in traditional RCS fault diagnosis models. A novel RCS fault diagnosis model that integrated marine predator algorithm (MPA), temporal convolutional network (TCN) and gated recurrent unit (GRU) was used to investigate. First, the TCN model was applied to extract critical feature information from RCS sample data. Subsequently, the features extracted by TCN were input into the GRU model to capture temporal dependencies within the sample data, thereby constructed the TCN-GRU model. Finally, to further enhance the model’s performance, the MPA was applied to optimize the hyperparameters of the TCN-GRU model. The results show that the proposed model converges more rapidly, requires fewer training epochs and exhibits significantly reduced fluctuations compared to traditional TCN, GRU, and unoptimized TCN-GRU models. Specifically, diagnostic accuracy improvements of 2.64%, 2.93%, and 2.01% are observed, respectively. It is concluded that the integration of MPA, TCN, and GRU significantly enhances fault diagnosis performance. By addressing questions in training efficiency and accuracy, the proposed model demonstrates high efficiency and reliability, making it a promising solution for fault diagnosis in complex systems.