Optimizing long short-term memory network with genetic and Bayesian optimization algorithms for accurate forecasting

Accurate load forecasting is crucial for effective grid management and strategic decision-making in the energy sector, particularly due to the inherent volatility and nonlinearity in load demand. This paper introduces a hybrid forecasting framework that combines advanced feature selection and Bayesian optimization (BO) to tune the long short-term memory (LSTM) model. The feature selection employs a genetic algorithm-based wrapper to systematically eliminate irrelevant and redundant features, enhancing computational efficiency and addressing dimensionality challenges. Unlike conventional approaches, the proposed framework uses BO for LSTM hyperparameter tuning, overcoming manual tuning limitations and reducing the risk of suboptimal performance. Integrating the search capabilities of the genetic algorithm with LSTM’s nonlinear modeling strengths and the optimization precision of BO, the framework achieves superior accuracy, enhanced stability, and accelerated convergence. The proposed model achieves a mean absolute percentage error of 0.5% by iteration 12, converging 20–40% faster than counterpart algorithms. Whereas, the other models exhibit slower convergences with an error of 1.4–1.6%. Statistical analysis validates the performance of the proposed algorithm marking it as a robust solution for dynamic forecasting, with precision and stability for real-world applications.