Short-term Load Forecasting With Clustered Hybrid Models Based On Hour Granularity

Although the recent technological achievements have noticeable impact on several aspects of daily life, more and more challenges are raised in practice. As it concerns the Energy field, the need for accurate predictions over time-dependent use cases of large scale remains high. Deep learning approaches have already found great acceptance in energy time-series signals, but there is still much space for improvement. Contributing to the task of short-term load forecasting we compose a hybrid method; first it exploits the statistical profiling of input raw-signals validating them through various complexity metrics; then a series of feature-engineering processes are applied, before fitting a specified recurrent neural network (RNN) architecture. During the first stage, we use time series clustering to separate time periods in order to capture better temporal patterns. We evaluate our approach using a public dataset that regards the total load consumption of Spain, thus supporting our assumptions about the benefits of leveraging hybrid models for short-term load forecasting. The proposed method outperforms other competitors, including a different RNN architecture and some representative Machine Learning regressors.