Demand Forecasting of a Fused Magnesia Smelting Process Based on LSTM and FRA

Abstract

In a Fused Magnesia Smelting Process(FMSP), its electricity demand is defined as the average electric power consumption over a fixed period of time and often used to calculate the electricity cost. The power supply has to be switched off once the demand value exceeds one specific threshold for safety and economic reasons. However, it has been shown that through appropriate current control of the FMSP, the demand can be reduced hence avoiding the shut-down of the process. A key issue to adopt the control strategy to avoid switch-off of electricity is to forecast the power demand and its trend However, this is technically challenging given the complexity and unknown dynamics of the process. In this paper, a hybrid approach combining a linear model with an unknown high order function is proposed. The linear model is used to capture the priori information from the domain knowledge and historic data, while the unknown dynamics in FMSP embedded in the error of the linear model are approximated with a high order nonlinear function. The Recursive Least Square algorithm (RLS) is used for identifying the unknown parameters in the linear model. A Long-Short Term Memory (LSTM) trained by the Fast Recursive Algorithm (FRA) is proposed to fit the unknown high-order function. Finally, the output weights of LSTM is updated by the RLS again. Experimental studies reveal that compared with other hybrid models such as a linear model combined with Radial Basis Function Neural Network (RBF), the proposed model offers the better performance.