Forecasting Aggregate GB Electricity Demand with Hybrid CNN and BiLSTM Methods in Stack and Parallel Structures and Temperature Dependency

• Published in: 2024 IEEE Power & Energy Society General Meeting (PESGM) pp. 1 - 5
• Main Authors: Zhu, S. Z., Jedidia, S. J., Djokic, S. Z.
• Format: Conference Proceeding
• Language: English
• Published: IEEE 21-07-2024
• Subjects: Accuracy; convolutional neural network; Convolutional neural networks; Deep learning; Demand forecasting; Electricity; Electricity demand forecasting; hybrid model; Load modeling; Long short term memory; long short-term memory neural network; Predictive models; Recurrent neural networks; Temperature dependence
• ISSN: 1944-9933
• DOI: 10.1109/PESGM51994.2024.10689059

As many other countries, the Great Britain (GB) is in the midst of an energy transition towards more sustainable energy supply from renewable energy resources and more flexible energy consumption through controllable loads. For transition to be successful, it is vital to accurately forecast day-ahead domestic aggregated electricity demand. This paper proposes two novel model architectures of hybrid convolutional neural network (CNN) and bidirectional long-short term memory (BiLSTM) layer models. The first structure connects the BiLSTM layer before the CNN layer, which is different from the commonly used hybrid CNN-BiLSTM approach. The second structure is a parallel mode, allowing CNN and BiLSTM networks to handle input variables simultaneously. For temperature dependency, two sampling approaches are used to derive average temperature datasets at the national GB level. The results show that the hybrid model with parallel structure has the best performance and that values of mean temperature obtained from sampling nine urban areas across the GB helps to improve the model performance during the winter season.