Application and Improvement of LSTM in Power System Prediction

In previous power system predictions, the accuracy of predicting extreme electricity consumption peaks was often insufficient. This study aims to explore the differences in power data prediction among three models: BiLSTM, BiLSTM with self-attention mechanism, and self-attention BiLSTM optimized by WOA. By analyzing and predicting three years of continuous power data, this paper compares the prediction accuracy and stability of these three models and provides a detailed comparison and evaluation of their performance. The BiLSTM model with the self-attention mechanism performs better in terms of prediction accuracy and stability in power data prediction, according to experimental findings. This model can better capture the long-term dependency relationship of time series data and automatically adjust attention, thus predicting future power loads more accurately. In addition, this study found that the self-attention BiLSTM model optimized by WOA can further improve prediction accuracy and stability, and its performance is better than that of traditional BiLSTM models. The experimental findings of this study demonstrate that the WOA optimization and self-attention mechanism play a significant role in enhancing the performance of BiLSTM models in power data prediction, and can increase the accuracy to over 80%. These results can provide reference for future deep learning-based power load prediction and valuable guidance for other time series prediction tasks.