Comparative Analysis of Hybrid Deep Learning Frameworks for Energy Forecasting

Abstract

As energy forecasting is paramount to efficient grid planning, this work presents a comparative analysis of different hybrid deep learning frameworks for energy forecasting in applications such as energy consumption and trading. Specifically, we developed hybrid architectures comprising of Convolutional Neural Network (CNN), an Autoencoder (AE), Long Short-Term Memory (LSTM) and Bi-directional LSTM (BLSTM). We use the individual household electric power consumption dataset by University of California, Irvine to evaluate the proposed frameworks. We evaluated and compared the result of these frameworks using several error metrics. The results show an average MSE of ∼ 0.01 across all developed frameworks. In addition, the CNN-LSTM framework performed the least with a 20% and 10% higher RMSE and MAE to other frameworks respectively, while CNN-BiLSTM achieved the least computation time.