A Modified Energy Demand Forecasting Model using Hybrid CNN-LSTM with Transformer for Univariate Time Series

Precise energy demand forecasting is important in managing electrical power systems, particularly if univariate time series analysis can be applied. To overcome the shortcomings of traditional hybrid models, this paper proposes an improved deep learning architecture that combines Transformer layers, Long Short-Term Memory (LSTM), and Convolutional Neural Networks (CNN). The proposed architecture was trained and validated on historical hourly energy demand data from 2015 to 2018. Performance evaluation revealed that the CNN-LSTM-Transformer model significantly improved forecasting accuracy compared to the baseline CNN-LSTM model. Specifically, the hybrid model achieved a Mean Absolute Error (MAE) of 234.25, Root Mean Squared Error (RMSE) of 386.15, and Mean Absolute Percentage Error (MAPE) of 0.84%, alongside an R² score of 99.28%. These results confirm the model’s robustness in capturing both local temporal dynamics and long-range dependencies, making it a promising solution for real-time energy forecasting applications.