ES-dRNN: A Hybrid Exponential Smoothing and Dilated Recurrent Neural Network Model for Short-Term Load Forecasting.

Area of Science:

• * Electrical Engineering
• * Data Science
• * Time Series Analysis

Background:

• * Short-term load forecasting (STLF) is crucial for power grid management but is complicated by complex time series (TS) with multiple seasonalities and nonlinear trends.
• * Existing methods often struggle to accurately model these intricate TS patterns, especially with limited data.

Purpose of the Study:

• * To propose a novel hybrid hierarchical deep-learning (DL) model for STLF.
• * To address challenges posed by multiple seasonality and nonlinear trends in TS.
• * To generate both point forecasts and predictive intervals (PIs).

Main Methods:

• * A hybrid model combining Exponential Smoothing (ES) for dynamic component extraction and on-the-fly deseasonalization.
• * A multilayer Recurrent Neural Network (RNN) with a novel dilated recurrent cell for modeling short and long-term dependencies.
• * Simultaneous learning of ES parameters and the main forecasting function by the RNN to enhance TS representation.

Main Results:

• * The proposed DL model demonstrated high expressive power for nonlinear stochastic forecasting problems with multiple seasonality and random fluctuations.
• * Empirical studies on STLF for 35 European countries showed superior performance compared to classical statistical and state-of-the-art machine learning (ML) models.
• * The model achieved higher accuracy in forecasting complex time series data.

Conclusions:

• * The novel hybrid hierarchical DL model effectively handles multiple seasonality and nonlinear trends in STLF.
• * The approach offers a significant advancement in forecasting accuracy for complex time series.
• * This method provides a robust solution for accurate short-term load forecasting in electrical grids.