3ETS+RD-LSTM: A New Hybrid Model for Electrical Energy Consumption Forecasting

This work presents an extended hybrid and hierarchical deep learning model for electrical energy consumption forecasting. The model combines initial time series (TS) decomposition, exponential smoothing (ETS) for forecasting trend and dispersion components, ETS for deseasonalization, advanced long short-term memory (LSTM), and ensembling. Multi-layer LSTM is equipped with dilated recurrent skip connections and a spatial shortcut path from lower layers to allow the model to better capture long-term seasonal relationships and ensure more efficient training. Deseasonalization and LSTM are combined in a simultaneous learning process using stochastic gradient descent (SGD) which leads to learning TS representations and mapping at the same time. To deal with a forecast bias, an asymmetric pinball loss function was applied. Three-level ensembling provides a powerful regularization reducing the model variance. A simulation study performed on the monthly electricity demand TS for 35 European countries demonstrates a high performance of the proposed model. It generates more accurate forecasts than its predecessor (ETS+RD-LSTM [1]), statistical models such as ARIMA and ETS as well as state-of-the-art models based on machine learning (ML).