Application of Long Short-Term Memory Neural Networks for Electric Arc Furnace Modelling

Abstract

Worldwide steelmaking industry strongly relies on the use of electric arc furnaces (EAFs). EAFs make use of electric arc phenomenon for melting scrap steel and consequently they can be sources of power quality issues, such as harmonics or voltage flickering. In order to design and implement effective systems for power quality improvement, it is necessary to dispose of an adequate model. Due to the complicated nature of the electric arc phenomenon, it is difficult to develop such an accurate model. Researchers around the world use different approaches, mostly relying on deterministic modelling with the addition of a stochastic ingredient. In this paper, we propose an approach which similarly is based on a deterministic equation enhanced with stochastic ingredients describing its coefficients. The identification of the time series of the equation is carried out by means of genetic algorithms. Next, we developed two models using long short-term memory artificial neural network (LSTM) for recreating the time series of the coefficients while remaining their stochastic properties. The second model also applies another LSTM for the reduction of stochastic-like residuals emerging from comparison of the first model with measurement data.