A hybrid power load forecasting model based on evolutionary strategy and long short term memory

Abstract

Accurate short-term load forecasting (STLF) is crucial for maintaining supply-demand balance and ensuring stable power grid operations. This paper proposes a hybrid model, the Evolutionary Strategy Long Short-Term Memory (ES-LSTM), to enhance the accuracy of STLF. The model combines genetic algorithm (GA) optimization with LSTM neural networks, tackling significant challenges related to data quality and hyperparameter tuning. Missing load and meteorological data are restored using Newton interpolation, and principal component analysis (PCA) is employed to reduce feature redundancy. GA optimizes key LSTM hyperparameters, such as the number of hidden layer units, time steps, and learning rate, to maximize prediction performance. Validated on a data set of 475 low-voltage users, ES-LSTM achieves a mean absolute percentage error (MAPE) of 0.02, significantly outperforming benchmark models like LSSVM (MAPE: 0.036) and BPNN (MAPE: 0.115). Experimental results confirm the model’s robustness, generalization ability, and suitability for real-world applications. This research provides a reliable solution for power utilities to improve operational efficiency and grid stability.