Time Series Foundation Model for Improved Transformer Load Forecasting and Overload Detection

Tasks such as transformer load forecasting, and heavy overload prediction have always been handled in previous work by simple models such as LSTM and XGBoost, which are not able to handle the increasing amount of data in power systems. Recent Foundation models have changed the task-specific modeling approach in time series analysis, they can scaling up to large time series variables and datasets across domains, however the simple pre-training setting makes it unsuitable for complex downstream tasks. To address this problem, this paper proposes the FreqMixer framework for adapting recently proposed time series foundation models to complex tasks in power scenarios, such as the task of predicting transformer load fluctuations caused by the Spring Festival returnees. Experiments demonstrate that the FreqMixer large model is able to adapt to a variety of downstream tasks, obtaining optimal performance in tasks such as overload prediction and load rate prediction (23.65% MAPE reduction, as well as 87% Recall improvement and 72% Precision improvement), while being able to memorize all the large amount of transformer data in a parametrically efficient manner (0.4%) for better performance.