Global Adaptive Input Normalization for Short-Term Electric Load Forecasting
Nikolaos Passalis, Anastasios Tefas
Abstract
Recent advances in Deep Learning (DL) provided immensely powerful tools for various time series forecasting tasks. However, DL models are often quite sensitive to the used input normalization method, requiring extensive experimentation and fine-tuning to identify and apply the most appropriate method for different tasks. Even though trainable adaptive normalization methods have been recently proposed to overcome this to a certain extend, these methods also tend to remove useful information from the data, ignoring the global statistics of the input time series. To overcome this limitation, in this paper we propose a trainable adaptive normalization approach that is capable of both maintaining important mode information, since global statistics are employed for the normalization, as well as taking into account the current behavior of the time series and adjusting the normalization scheme to this. The effectiveness of the proposed method over baseline and state-of the-art normalization methods is demonstrated using extensive experiments on two different electric load forecasting datasets.