Experimental research on short-term forecasting based on DOA-BiLSTM model

Abstract

Precise short-term electric load prediction plays a vital role in power system analysis, scheduling, generation planning and secure grid operation. Nevertheless, load sequences typically show nonlinear fluctuations, time-dependent correlations and noise interference, which bring challenges to direct forecasting tasks. To raise prediction precision, this research establishes a combined prediction framework using bidirectional long short-term memory neural network and dream-inspired optimization algorithm. BiLSTM network is capable of capturing temporal features from both forward and backward directions, which helps the model depict the dynamic change rules of load sequences more effectively. Meanwhile, DOA algorithm is utilized to optimize the core hyper-parameters of BiLSTM, covering hidden unit count, learning rate and training rounds. Moreover, singular spectrum decomposition is adopted to separate the raw load sequence, so as to extract primary trend components and eliminate local disturbance signals. Tests on Belgian load records of January 2022 indicate that the presented model yields MAE, RMSE and MAPE of 124.46, 145.51 and 1.17% in turn. In contrast with LSTM, BiLSTM and RIME-BiLSTM, the proposed approach shows better prediction performance, which proves that the integration of SSA decomposition, DOA optimization and BiLSTM modeling distinctly boosts the accuracy and stability of short-term load forecasting.