Short term energy consumption forecasting using neural basis expansion analysis for interpretable time series

Area of Science:

• Energy Systems Engineering
• Artificial Intelligence
• Data Science

Background:

• Smart grids and smart homes face energy efficiency challenges due to increasing demand and data uncertainty.
• Existing machine learning models, particularly neural networks, show limitations in predicting energy consumption with large, diverse customer datasets.
• Current research often relies on small datasets, hindering the scalability of predictive models for real-world smart grid applications.

Purpose of the Study:

• To enhance the Neural Basis Expansion Analysis for interpretable Time Series (N-BEATS) model for accurate energy consumption prediction in smart grids.
• To address the limitations of existing models in handling large datasets with diverse customer profiles.
• To validate the performance of the enhanced N-BEATS model against other leading time series forecasting methods.

Main Methods:

• The study proposes an enhanced N-BEATS model, integrating a large dataset of energy consumption from 169 customers.
• Performance comparison was conducted against Long Short Term Memory (LSTM), Blocked LSTM, Gated Recurrent Units (GRU), Blocked GRU, and Temporal Convolutional Network (TCN).
• The model incorporated covariates, such as 'day', to improve the learning of past and future energy consumption patterns.

Main Results:

• The enhanced N-BEATS model demonstrated improved prediction accuracy on a large dataset encompassing multiple customer energy consumption profiles.
• Incorporating covariates significantly boosted accuracy by enabling the model to learn historical and future energy usage trends.
• The proposed model outperformed comparative methods in daily, weekly, and monthly energy consumption forecasting, with superior 1-day-ahead prediction accuracy when using 'day as covariates'.

Conclusions:

• The enhanced N-BEATS model offers a robust and interpretable solution for energy consumption prediction in smart grid environments.
• The model's ability to handle large, multi-customer datasets overcomes previous limitations, paving the way for more reliable energy management.
• Future research can leverage this interpretable model for enhanced energy efficiency and demand-side management strategies in smart cities.