Elite GA-based feature selection of LSTM for earthquake prediction
Hong, Liang,Yu, Xinguo,Gao, Yunxuan
DOI: https://doi.org/10.1007/s11227-024-06218-2
IF: 3.3
2024-06-09
The Journal of Supercomputing
Abstract:Earthquake magnitude prediction is an exceptionally challenging task that has been explored through various machine learning approaches. However, the development of prediction models is hindered by redundant features and time series properties. The Elite Genetic Algorithm (EGA) excels in searching for optimal feature subsets; while, Long Short-Term Memory (LSTM) is specialized in processing time series and complex data. Hence, we introduce an EGA feature selection and LSTM model (EGA-LSTM) for time series earthquake prediction. Initially, acoustic and electromagnetic data from our developed AETA system are fused and preprocessed using a roulette-based EGA to search strong correlation indicators. Subsequently, LSTM is employed to perform magnitude prediction with the selected features. Specially, the Root Mean Square Error (RMSE) of LSTM and the ratio of selected features are chosen as fitness components of EGA. Finally, we evaluate the proposed EGA-LSTM on AETA data from four regions from China, considering the influence of data in different time periods ( timePeriod ) and fitness function weights ( and ) on prediction results. Furthermore, our proposed approach outperforms state-of-the-art methods in evaluation metrics such as EV, MAE, MSE, RMSE, and R2. Nonparametric tests reveal that EGA-LSTM is significantly different from others and outperforms the standard LSTM.
computer science, theory & methods,engineering, electrical & electronic, hardware & architecture