The Transformer Oil Temperature Prediction Method Based on the CEEMDAN-PatchTST-Transformer Hybrid Model

Transformers are a key component of power systems due to increasing electricity demand. Transformers are the nerves of the power system, and the operation of the transformers directly affects the stability of the power supply, the safety and economy of the electrical equipment and their life. High-precision transformer oil temperature prediction helps the transformer to make proactive protective measures, avoid accidents. Many existing methods for oil temperature prediction are unable to extract significant information or learn complex trends from past oil temperature information. In this paper, we propose a hybrid prediction model based on CEEMDAN (Complete Ensemble Empirical Mode Decomposition with Adaptive Noise) and an improved model of PatchTST (Time Series Patch Transformer) to solve this problem. CEEMDAN decomposes the original temperature series into a few of subsequences with different frequencies, providing the opportunity for adaptive analysis of the periodic variation trends in the oil temperature data. Following that, an improved model of PatchTST is used to handle the data, by altering the naive linear prediction head of the initial PatchTST model. The model enhances the correlations among the multiple variable input data for a single-variable prediction task via a Transformer-based prediction head. Finally, the ETTh1 and ETTh2 datasets are used to give a comparison of the prediction results of the proposed CEEMDAN-PatchTST-Transformer model and other models, which verifies that the proposed model has a significant enhancement on the accuracy of prediction, proving the effectiveness of the proposed method in transformer oil temperature forecasting.