Automatic prediction modeling for Time-Series degradation data via Genetic algorithm with applications in nuclear energy

Abstract

Time-series data prediction, a predominant problem in fault prediction, enables effective and efficient predictive maintenance. For time-series data, hybrid methods combining data decomposition, component-wise prediction, and aggregation are frequently reported. However, most hybrid models merely use a single model for all components and adopt the equal-weighted aggregating strategy. This may easily overfit or underfit the decomposed data and reduce the generalization capability, especially in long-term predictions. This study proposed a novel decomposition-based framework for time-series data fault predictions. Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) disintegrates the raw data into several components, and different prediction models were considered as the modeling candidates. Joint modeling optimization was performed automatically based on the Genetic Algorithm to optimize the prediction model, modeling parameters, and aggregation weight of each component. To verify the effectiveness of the proposed method, a case study concerning leakage of reactor coolant pumps in nuclear power plants was carried out. The experimental results indicated the superior performance of the proposed method for various prediction horizons.