Electrical Load Prediction by an Improved Long Short-Term Memory Based on Variable Dimension Reduction

Electrical power load plays an important role in keeping safety and security of power grid and systems. Tremendous attentions have been given in line with how to predict electrical load in an accurate way. However, there exists various uncertainties and influential factors leading to complexity of prediction or inferior accuracy. This paper has proposed an improved long short-term memory (LSTM) neural network with variable dimension reduction by employing gray correlation principle. The computational test results verify the effectiveness of the proposed scheme in short-term load prediction. The accuracy of the proposed method increases from 4.97% to 1.60% in MAPE while R2 promotes from 0.64 to 0.97 by comparing with BP neural network under the same case study. Moreover, the proposed algorithm has saved computational cost due to variable dimension reduction where the redundant dimensions are removed from the original historical data set by employing correlation analysis. The finding of the study will encourage real-time applications in future power grid since the proposed algorithm has relatively high prediction accuracy with less computational costs. The both benefits are essential to real-time applications in practical electrical power grid.