Day-ahead building-level load forecasts using deep learning vs. traditional time-series techniques

Abstract

Load forecasting problems have traditionally been addressed using various statistical methods, among which autoregressive integrated moving average with exogenous inputs (ARIMAX) has gained the most attention as a classical time-series modeling method. Recently, the booming development of deep learning techniques make them promising alternatives to conventional data-driven approaches. While deep learning offers exceptional capability in handling complex non-linear relationships, model complexity and computation efficiency are of concern. A few papers have explored the possibility of applying deep neural networks to forecast time-series load data but only limited to system-level or single-step building-level forecasting. This study, however, aims at filling in the knowledge gap of deep learning-based techniques for day-ahead multi-step load forecasting in commercial buildings. Two classical deep neural network models, namely recurrent neural network (RNN) and convolutional neural network (CNN), have been proposed and formulated under both recursive and direct multi-step manners. Their performances are compared with the Seasonal ARIMAX model with regard to accuracy, computational efficiency, generalizability and robustness. Among all of the investigated deep learning techniques, the gated 24-h CNN model, performed in a direct multi-step manner, proves itself to have the best performance, improving the forecasting accuracy by 22.6% compared to that of the seasonal ARIMAX.