Mohammadreza Ahmadipour, Esmat Rashedi, Maryam Amoozegar,
Volume 13, Issue 2 (7-2024)
Abstract
In recent years, electricity consumption forecasting has gained significant importance by relying on new technologies and utilizing big data processing methods. Various methods including classical statistical time series analysis techniques, support vector machines, and recurrent neural networks with long short-term memory have been studied in this field. Statistical methods may not be suitable for forecasting and modeling some complex phenomena due to their inability to account for sudden changes. On the other hand, the support vector machine method operates based on increasing the data dimensions. Therefore, in cases where the data has high dimensions, this leads to increasing the complexity of the problem space. Various types of neural networks also face limitations such as vanishing gradients and the inability to account for temporal relationships. To achieve more accurate electricity consumption forecasting, this paper proposes a hybrid approach using the Transformer model and long short-term memory neural network. By addressing the gradient problem and learning complex patterns, this approach offers higher accuracy compared to other methods. Additionally, the Transformer model, with its attention mechanism, has the ability to focus on important components of the data, creating a more interpretable model with high resistance to noise. The proposed method has been evaluated on a standard dataset and compared with existing methods. The results show that this method achieves higher accuracy and lower error in metrics such as mean squared error and mean absolute percentage error.