Artificial neural networks versus conventional methods for boiling water reactor stability monitoring
- Delft Univ. of Technology (Netherlands). Interfaculty Reactor Inst.
The application of an artificial neural network (ANN) for boiling water reactor (BWR) stability monitoring was studied. A three-layer perceptron was trained on synthetic autocorrelation functions to estimate the decay ratio and the resonance frequency from measured neutron noise. Training of the ANN was improved by adding noise to the training patterns and by applying nonconventional error definitions in the generalized delta rule. The performance of the developed ANN was compared with those of conventional stability monitoring techniques. Explicit care was taken for generating unbiased test data. It is found that the trained ANN is capable of monitoring the stability of the Dodewaard BWR for four specific cases. By comparing properties such as the false alarm ratio, the alarm failure ratio, and the average time to alarm, it is shown that it performs worse than model-based methods in stability monitoring of exact second-order systems but that it is more robust (better resistant to corruptions of the input data and to deviations of the system at issue from an exact second-order system) than other methods. The latter explains its good performance on the Dodewaard BWR and is promising for the application of an ANN for stability monitoring of other reactors and for other operating conditions.
- OSTI ID:
- 109976
- Journal Information:
- Nuclear Technology, Vol. 109, Issue 2; Other Information: PBD: Feb 1995
- Country of Publication:
- United States
- Language:
- English
Similar Records
Comparisons of conventional techniques and neural networks in power system load forecasting and stability
Adaptive conventional power system stabilizer based on artificial neural network