Long Time-Series Prediction and Anomaly Detection of Electricity Transformer Oil Temperature

A precise prediction and anomaly detection of electricity transformer oil temperature is the effective method of the fire preventing. Long short-term memory (LSTM) was built for handling long time-series sequence. However, the performance of LSTM drops sharply as length of sequence gets longer. A method based on Transformer model is introduced here achieving long sequence time-series forecasting (LSTF). 2-year data consists of 6 load features as input and oil temperature as output is collected, which is a perfect indicator of electricity transformer during long-term deployment. The results of experiments show that Informer is superior than LSTMa [1] which is based on Recurrent Neural Network (RNN). To be specific, the evaluation metrics Informer accquired on the datasets are much more better LSTMa. The results have shown that Informer is qualified on LSTF task and for electricity transformer oil temperature prediction.