A Comparison of Various Deep Learning Methods for Household Load Forecasting

Forecasting energy consumption is vital for smart grid operations to manage demand, plan loads, and optimize grid operations. This work aims at reviewing and experimentally evaluating six univariate deep learning architectures to forecast load for a single household using a real-world dataset. Multi-layer perceptron (MLP), Convolutional neural network (CNN) and recurrent neural networks (Simple RNN, Long Short Term Memory (LSTM)) were the neural network methods that were analysed along with robust LSTM architectures like Bi- directional LSTM and CNN-LSTM Hybrid. All the models were tuned using Bayesian optimization and evaluated using root mean squared error (RMSE) as the metric. In addition to neural network models, Seasonal ARIMA (SARIMA) a statistical model is also presented to observe the performance. As a result, Bi-directional LSTM was observed to have achieved the best performance with the smallest value of RMSE; however, it was also observed that differences in performances between other neural network models were quite low, especially between the RNN architectures. Additionally, although machine learning methods performed better than SARIMA the former model was more complex and computationally intensive.