Forecasting of Winding and Oil Temperatures in Power Transformers Using a Hybrid Deep Learning Model With Attention Feedback

Industrial time series forecasting in near real-time involving multiple covariates remains a challenge. Although machine learning models have been used to solve the problem, there are several challenges that require attention. Most of the machine learning models do not provide any interpretability and these models can lose their relevance over time as the underlying machine characteristics change due to wear and tear, upgrades, preventative maintenance etc. This paper presents a hybrid model based on feedforward deep neural network with attention feedback, 1D convolutional layer and non-evolutionary deterministic symbolic regression, which transforms a multi-covariate forecasting into a univariate forecasting problem. In addition, the deterministic symbolic regression allows for the discovery of interpretable relationship between the predicted variable and the co-variates. The forecasting model is carefully managed over its entire life cycle by a self-exploratory orchestration mechanism which ensures that the models are always kept in synch with the changes happening to the underlying machine characteristics through continuous, autonomous monitoring, decision making and learning. This paper introduces the model and the orchestration framework and presents a case study using the model described here to predict winding temperature and oil temperature in power transformers by monitoring the basic transformer parameters of load current, active and reactive power, and ambient conditions. This paper demonstrates the effective use of the Deep Feedforward Learning model and how it outperforms the traditional sequence-oriented models e.g. LSTM.