Virtual Booth

Efficient monitoring of nuclear reactor cooling systems is critical for operational safety. The primary cooling system of the TRIGA PUSPATI Reactor is maintained on fixed schedules, limiting the ability to detect anomalies in advance. This study develops a multi-input, multi-output (MIMO) machine learning framework to jointly predict six interdependent cooling parameters including three temperatures, two pressures, and one flow rate, recorded at 0.5 second intervals, resulting in a total of 1,270,707 samples. Nine ML models were evaluated, including LSTM, RNN, RF, GRU, MLP, TCN, SVR, GBM, and XGB. Model performance was assessed using mean squared error (MSE), root mean squared error (RMSE), mean absolute error (MAE), and the coefficient of determination (R²), along with percentage-based error metrics and uncertainty estimation through Monte Carlo (MC) Dropout. Results show that RNN achieved the lowest error on the temperature data across all performance metrics, with the highest R². Specifically, TT005 produced an MSE of 0.2713 ºC², RMSE of 0.5209 ºC, and MAE of 0.3523 ºC when predicting rapid interval temperature data, with R² exceeding 0.9824, while RF achieved the lowest errors for PT004 across all metrics, with an MSE of 0.1469 kPa², RMSE of 0.3832 kPa, MAE of 0.2555 kPa, and R² of 0.9987. The RF model also produced the lowest error values for FT001 data. Other models achieved reasonable accuracy, but their prediction errors increased when processing longer time sequences, indicating reduced stability for extended temporal dependencies. MC Dropout analysis with 50 stochastic runs produced average 95% confidence interval widths of 1.93 °C for TT005 using RNN and 0.44 kPa for PT004 and 0.36 m³/s for FT001 using RF, indicating low uncertainty and stable predictive performance. These findings highlight the complementary strengths of RNN and RF, where RNN excels at learning dynamic time-based patterns, while RF demonstrates reliable performance in continuous real-time monitoring tasks. Overall, this study demonstrates the potential of the proposed MIMO framework integrated with Monte Carlo Dropout for early anomaly detection, predictive maintenance, and improved operational safety in nuclear reactor cooling systems.