A novel outlier calendrical heterogeneity reconstruction deep learning model for electricity demand forecasting

Abstract

The development of an accurate electricity demand forecasting model is of paramount importance for promoting global energy efficiency and sustainability. Nonetheless, the presence of outliers and inappropriate model training can result in suboptimal performance. To tackle these challenges, this study explores the potential of Convolutional Neural Network (CNN) and active learning theory as forecasting solutions, offering high efficiency and advantages for long time series. In this study, a hybrid model that combines Isolation Forest (IF), Outlier Reconstruction (OR), CNN and Random Forest (RF) is conducted to mitigate computational complexity and enhance the accuracy of electricity demand forecasting in the presence of outliers. IF is employed to detect outliers in electricity demand time series, while OR is used to reconstruct subsequences based on calendrical heterogeneity for training. CNN is applied for both training and forecasting, and the final output is combined using RF. The effectiveness of the proposed IF-OR-CNN-RF model is validated using electricity data collected from recent sources in Australia at different sampling frequency. The experimental results demonstrate that, in comparison with other popular CNN-based electricity demand forecasting models, IF-OR-CNN-RF model outperforms with significantly improved performance metrics. Specifically, Mean Absolute Error (MAE), Root Mean Squared Error (RMSE) and R-squared values are 77.92, 179.18 and 0.9769 in 5-minute frequency; 162.67, 353.96 and 0.9775 in 10-minute frequency; 841.27, 1374.79 and 0.9622 in 30-minute frequency; 2746.01, 3824.00 and 0.9262 in 60-minute frequency; 9106.08, 12269.04 and 0.8044 in 120-minute frequency. IF-OR-CNN-RF model represents a valuable framework for future electricity demand forecasting, particularly in scenarios involving outliers.