A Comparative Study of AutoML Approaches for Short-Term Electric Load Forecasting

Abstract

Deep learning is increasingly used in short-term load forecasting. However, deep learning models are difficult to train, and adjusting training hyper-parameters takes time and effort. Automated machine learning (AutoML) can reduce human participation in machine learning process and improve the efficiency of modelling while ensuring the accuracy of prediction. In this paper, we compare the usage of three AutoML approaches in short-term load forecasting. The experiments on a real-world dataset show that the predictive performance of AutoGluon outperforms that of AutoPytorch and Auto-Keras, according to three performance metrics: MAE, RMSE and MAPE. AutoPytorch and Auto-Keras have similar performance and are not easy to compare.