Anomaly-Regularized Foundation Models for Robust Electricity Market Forecasting
Abstract
In recent cases of multivariate time series forecasting, models are facing challenges that are due to complex interactions between multiple variables, rapid regime shifts and extreme events that cause problems, especially in the domain of electricity markets like the Australian National Electricity Market (NEM). In our research, we tried to improve the existing singular value decomposition autoencoder (SVD-AE) by introducing a spectral aspect to the anomaly scoring and coming up with a regularisation framework aimed at increasing the accuracy of the time series foundation model (TSFM). We used 5-minute resolution NEM data to evaluate our framework; specifically, the anomaly detection part was done on multiple variables influencing the Regional Reference Price (RRP), such as demand, generation, interconnector, temporal and weather, but the zero-shot forecasting and fine-tuning of the forecasting part were done solely on raw and regularised RRP against the same original ground truth. The forecasting outputs from our Hybrid-Scored SVD-AE (HS-SVD-AE) are then compared to the existing baseline SVD-AE. The experiments showed that in the fine-tuned scenario of TSFM, our proposed framework outperforms the existing SVD-AE regularisation by a good margin, but in the zero-shot scenario, SVD-AE performs better by a slight difference. Our findings in this research show that explicitly scored spectral information structure can complement multivariate anomaly detection, which produces more robust foundation model forecasting.