Data-driven forecasting of transformer oil temperature using deep learning and ensemble models on multivariate time-series data

Abstract Transformers must be maintained within their operating temperature range within their optimal temperatures. Monitoring oil temperature is one way for utilities to know the condition of the insulation and reliability of transformers. Long-term thermal stress on the insulation creates accelerated aging and increases the risk of transformer failure. Existing methods of oil temperature monitoring (i.e., thermal threshold alarm systems, physics-based models of oil temperature, simple linear regression) do not typically account for any of the nonlinear behaviours or long-associative temporal dependencies usually found in the operational data of transformers. Therefore, the authors of this paper propose the use of a comprehensive data-driven approach to forecasting transformer oil temperature based on applying deep-learning and ensemble machine-learning methodologies to multivariate time-series data. The Electricity Transformer Temperature (ETT) Dataset (i.e., ETTh1, ETTh2, ETTm1, and ETTm2) provides both hourly- and minute-level electrical load and thermal parameter measurements; therefore, it is used to evaluate the proposed approach. The authors systematically assess recurrent neural network (RNN) models based on Long Short-Term Memory (LSTM) technology, transformer-based Informer architectures, and tree-based ensemble models (Random Forest and XGBoost) in accordance with an overall experimental protocol. Model training and validation employ only chronologically based train–validation–test splits with early stopping, regularization, and residual diagnostics to avoid overfitting and promote generalization. The experimental results demonstrate that when ensemble models were employed, there were very low root mean squared errors (RMSE) associated with predictions of transformer oil temperature.