What question did this study set out to answer?

The aim is to enhance the performance of LSTM models in time series predictions by optimizing hyperparameters using a novel algorithm.

April 12, 2026International Journal of Advanced Engineering Research and Science0 citationsOpen Access

Research on Construction and Performance Optimization of the LEA-LSTM Model

WWWen-Min WenGuangdong University of Petrochemical Technology YCYan-Zuo Chang JCJin-Ping ChenGuangdong University of Petrochemical Technology

Key Points

The aim is to enhance the performance of LSTM models in time series predictions by optimizing hyperparameters using a novel algorithm.
Developed an LEA-LSTM model using the Love Evolution Algorithm for hyperparameter tuning.
Analyzed the influence of LSTM's core hyperparameters on model performance.
Compared LEA-LSTM performance with traditional models across prediction accuracy, convergence speed, and robustness.
LEA-LSTM reduced the MAE by 68.3% compared to LSTM on the Jena Climate Dataset.
MAE reductions of 42.1%, 37.5%, and 29.8% were observed against PSO-LSTM, WOA-LSTM, and BWO-LSTM, respectively.
In power load forecasting, LEA-LSTM demonstrated an 18.6% MAE reduction compared to IGWA-ADConv1D-LSTM.
Convergence speed improved by over 35% compared to traditional optimization methods.

Abstract

Aiming at the problems of the Long Short-Term Memory network (LSTM) in time series modeling, such as hyperparameter adjustment relying on experience, being prone to falling into local optimum, and slow convergence speed, an LSTM model optimized by the Love Evolution Algorithm (LEA), namely LEA-LSTM, is proposed. First, the gating mechanism and time series processing principle of the LSTM network are elaborated, and the influence of its core hyperparameters on model performance is analyzed. Second, the LEA algorithm is introduced, and the adaptive optimization of the key hyperparameters of LSTM is realized through the five-stage evolution mechanism of encounter, stimulation, reflection, value and role, which solves the defect of insufficient global search capability of traditional optimization algorithms. Finally, the Jena Climate Dataset, a general time series dataset, and scenario-specific dataset such as power load are used for performance verification. The proposed model is compared with LSTM, PSO-LSTM, WOA-LSTM, BWO-LSTM and IGWA-ADConv1D-LSTM models in three aspects: prediction accuracy, convergence speed and robustness. The results show that the Mean Absolute Error (MAE) of the LEA-LSTM model on the Jena Climate Dataset is reduced by 68.3% compared with LSTM, and by 42.1%, 37.5% and 29.8% compared with PSO-LSTM, WOA-LSTM and BWO-LSTM respectively; in the power load forecasting scenario, the MAE is reduced by 18.6% compared with IGWA-ADConv1D-LSTM; the convergence speed is increased by more than 35% compared with traditional optimized models, and the coefficient of determination (R²) remains 99.1% even in small sample scenarios.

Ask AI

Helpful

Bookmark

View Full Paper