Comparative Time Series Machine Learning Model for Nuclear Reactor Water Plant Coolant System Data at RTP Malaysia

The secondary cooling system in a nuclear reactor plays a critical role in ensuring operational stability, safety, and performance. Accurate forecasting of key process parameters such as temperature, pressure, and flow rate within this system is essential for early anomaly detection and predictive maintenance. This study presents a comprehensive comparative evaluation of ten time series forecasting models, spanning statistical models such as autoregressive integrated moving average (ARIMA), seasonal autoregressive integrated moving average, exponential smoothing state space model (ETS), and Prophet; machine learning like support vector machine for time series, random forest, and gradient boosted trees; as well as deep learning architectures such as gated recurrent unit (GRU), long short-term memory, and WaveNet, using real-world supervisory control and data acquisition (SCADA) sensor data from the PUSPATI TRIGA Reactor (RTP) research reactor in Malaysia. The models were trained and evaluated across five essential signals, which are TT003, TT004, PT002, PT003, and FT002, and their performances were benchmarked using multiple metrics, i.e. root mean squared error (RMSE), mean absolute error, R-squared (R2), mean absolute percentage error, residual analysis, uncertainty quantification, anomaly response, and computational feasibility, which is model size, inference, and training time. Results demonstrate that GRU outperforms all other models in accuracy (e.g. RMSE < 0.1), resilience to transient anomalies, and computational efficiency; that is, model size = 0.125 MB, inference time = 77.5 ms/sample, and training time = 70.3 s. WaveNet, while competitive in accuracy, exhibited higher latency and resource demands. Statistical models such as ARIMA and ETS failed to capture nonlinear and abrupt behavior under reactor transients, resulting in poor generalization and negative R2 scores. Residual and uncertainty analysis further confirmed that deep learning models maintain robust prediction quality under dynamic operating conditions. The ability of GRU to adapt to anomalies and provide confidence-aware forecasts highlights its suitability for integration into nuclear SCADA systems. This study is the first to provide an end-to-end comparison of forecasting models, including deep neural architectures, with full-spectrum evaluation for nuclear reactor cooling system data. These findings support the deployment of GRU-based models in safety-critical environments and pave the way for intelligent, real-time, and energy-efficient monitoring frameworks in next-generation nuclear reactor operations.