Živko Sokolović

Forecasting electricity consumption is a key aspect of planning, operation, maintenance and management of the power system, as it allows for the optimal use of generation resources, increasing system reliability and reducing operating costs. In recent years, the development of artificial intelligence and machine learning methods has found wide application in this field, providing advanced tools for analyzing and modeling complex time series. Deep learning models have outperformed traditional statistical methods, especially recurrent neural networks (RNN) and convolutional neural networks (CNN). These models have dominated time series forecasting tasks due to their ability to model nonlinear dependencies in data. However, these models have certain limitations, especially in capturing long-term dependencies and efficiently processing long sequences. In natural language processing (NLP), a new deep learning architecture has emerged – Transformer, which was originally introduced to overcome the limitations of RNN models arising from their sequential nature. Thanks to the self-attention mechanism, Transformer allows for direct modeling of the dependencies between all elements in a sequence, regardless of their mutual distance. This feature makes Transformer particularly suitable for application in time series forecasting problems, where the observation of long-term and complex patterns is of key importance. In this work, the Transformer model is applied to forecast electricity consumption. The implementation and experimental evaluation of the model was realized using the Darts Python library for time series, while the development and training of the model was performed in the Google Colab environment, which allowed for easy use of GPU resources and accelerated the training process. The experiments were conducted on a dataset including electricity consumption for 370 customers in Portugal, with hourly time resolution, taking into account aggregated electricity consumption. The model evaluation was performed using standard accuracy metrics, including MAE, RMSE and MAPE. The model was tested for different forecast horizons, namely 24, 48, 96 and 168 hours. The obtained results show that the model achieves stable performance in all considered scenarios, with a MAPE of 7.2%, which confirms high forecasting accuracy. In addition to the accuracy analysis, the computational efficiency of the model was also examined by measuring the training time for different input window lengths. The analysis showed that the training time grows superlinearly with the input window length, which is consistent with the quadratic complexity of the self-attention mechanism.

Keywords: Load forecasting, Time series, Transformer, Deep learning