Application of Artificial Intelligence in Short-Term Load Forecasting at Low-Voltage Substations

Abstract

This study presents a Short-Term Load Forecasting (STLF) process in Vietnam. A total of 13 features, including 9 electricity-related features and 4 time-related features, are extracted to predict the Total Active Power (TAP) of a three-phase low-voltage transformer station. The Mutual Information (MI) analysis results show that the Active Power (AP) features in the three phases have a significantly higher importance level than the other features in predicting TAP. Two algorithms, including Deep Neural Network (DNN) and Artificial Neural Network (ANN), are used to predict TAP. In addition to the two models using all 13 features, two corresponding models using only 3 AP features are also trained for comparison. Results show that, with 17 047 data points, models using 3 AP features have a suitable level of complexity for better results than similar models using all 13 features. The DNN model with 3 features yields the best results, with MAE, RMSE, and R2 values of 1.9089, 4.2880, and 0.9971, respectively. In future research, the number of data points will be improved to explore better features, and other features could also be considered.