Prognozowanie zapotrzebowania na energię elektryczną: Ewolucja hybrydowego modelu łączącego wygładzanie wykładnicze i rekurencyjne sieci neuronowe [Electricity demand forecasting: Evolution of a hybrid model combining exponential smoothing and recurrent neural networks] Grzegorz Dudek (Department of Electrical Engineering, PCz, Częstochowa) Accurate electricity demand is critical for efficient power system operations, planning, and maintenance. The intricate nature of electricity demand time series data, characterized by nonlinear trends, triple seasonality, evolving variance, and random fluctuations, poses substantial challenges for forecasting models. This presentation introduces the evolution of a hybrid and hierarchical model designed for electricity demand forecasting. The model combines the strengths of exponential smoothing and gated recurrent neural networks, incorporating specialized mechanisms to tackle short- and long-term dynamics, as well as complex seasonality. It involves exponential smoothing dynamically extracting primary components of individual series, facilitating the learning of series representations. A multi-layer recurrent neural network, featuring dilated recurrent cells, efficiently captures short-term, long-term, and seasonal dependencies inherent in time series data. Our most advanced model introduces an attentive dilated recurrent cell, incorporating an attention mechanism for dynamic input vector weighting. Furthermore, the model employs a double-track architecture, featuring a context track that generates supplementary inputs for the main track. These supplementary inputs are derived from representative or exogenous series, enhancing the model's capability to capture and adapt to the complexities of electricity demand forecasting.