Predicting Oil Temperature in Electrical Transformers Using Neural Hierarchical Interpolation

Abstract

Effective electricity consumption planning is critical for power distribution. Ensuring the distribution network aligns with expected demand fluctuations is a challenging task influenced by various time-related and seasonal variables. This study focuses on improving transformer oil temperature forecasting, an indicator of transformer health, using the neural hierarchical interpolation for time series (NHITS) model. The NHITS model’s architecture is designed to handle long-term forecasting efficiently, making it ideal for capturing extended trends in transformer oil temperature. It incorporates multirate signal sampling via MaxPool layers and hierarchical interpolation to merge predictions across different time scales. The proposed methodology involves two key phases: data preparation and model development. In the data preparation phase, the electricity transformer temperature (ETT) datasets are used, normalized with a standard scaler, and essential features such as oil temperature and external power load are selected. During the model development phase, the proposed NHITS model is trained and its hyperparameters are optimized for optimal performance. The study evaluates the model’s performance under various conditions, including the comparison of multivariate and univariate time series, the effects of short and long-term forecasting horizons, and the impact of temporal resolution. The model was validated using the ETT dataset, and our results were benchmarked against a previous study that employed the same dataset and used the Informer model. The results indicate that the NHITS model outperforms the Informer model, showing an average decrease of 51.37% in mean squared error (MSE) and 37.83% in mean absolute error (MAE). These findings highlight the model’s ability to capture both long-term and short-term characteristics of time series data, making it a promising solution for forecasting transformer oil temperatures.