Deep-learning-based short-term electricity load forecasting: A real case application

Abstract

The rising popularity of deep learning can largely be attributed to the big data phenomenon, the surge in the development of new and novel deep neural network architectures, and the advent of powerful computational innovations. However, the application of deep neural networks is rare for time series problems when compared to other application areas. Short-term load forecasting, a typical and difficult time series problem, is considered as the application domain in this study. One-dimensional Convolutional Neural Networks (CNNs) use is rare in time series forecasting problems when compared to Long Short Term Memory (LSTM) and Gated Recurrent Unit (GRU), and the efficiency of CNN has been rather remarkable for pattern extraction. Hence, a new method that uses one-dimensional CNNs based on Video Pixel Networks (VPNs) in this study, in which the gating mechanism of Multiplicative Units of the VPNs is modified in some sense, for short term load forecasting. Specifically, the proposed one-dimensional CNNs, LSTM and GRU variants are applied to real-world electricity load data for 1-hour-ahead and 24-hour-ahead prediction tasks which they are the main concerns for the electricity provider firms for short term load forecasting. Statistical tests were conducted to spot the significance of the performance differences in analyses for which ten ensemble predictions of each method were experimented. According to the results of the comparative analyses, the proposed one-dimensional CNN model yielded the best result in total with 2.21% mean absolute percentage error for 24-h ahead predicitions. On the other hand, not a noteworthy difference between the methods was spotted even the proposed one-dimensional CNN method yielded the best results with approximately 1% mean absolute percentage error for 1-h ahead predictions.