Short-Term Power Load Forecasting Based on DE-IHHO Optimized BiLSTM

Abstract

Accurate short-term power load forecasting is the key to determining the grid company’s dispatch plan and system operation mode. Aiming at the problem of low prediction accuracy due to the difficulty in selecting hyperparameters of BiLSTM, a hybrid parallel Harris hawk optimization algorithm (DE-IHHO) is proposed to choose the optimal hyperparameters of BiLSTM to improve the prediction accuracy of the model. In this paper, several load forecasting models are tested and BiLSTM with better performance is chosen as the baseline model. Aiming to solve the problem of the complex selection of hyperparameters for BiLSTM, the Harris Hawk Optimization (HHO) algorithm is used to obtain better hyperparameter combinations and improve prediction accuracy. To further explore the optimal hyperparameter combinations of BiLSTM, running in parallel with differential evolutionary algorithm (DE) to enhance the search diversity, adopting chaotic dyadic learning strategy and mutation operation strategy to improve the global search capability of HHO, and finally smoothing the optimal solution to reduce the influence of abnormal solutions. The results show that the convergence speed and optimization ability of DE-IHHO are significantly improved, and the BiLSTM optimized in this way improves considerably in all three metrics of MAE, MAPE, and RMSE, proving this prediction model’s effectiveness.

Keywords

• Short-term load forecasting
• bidirectional long and short-term memory neural network (BiLSTM)
• hybrid parallel Harris Hawk optimization algorithm (DE-IHHO)
• chaotic dyadic learning strategy
• variational manipulation strategy