Forecasting System Marginal Price Using Multilayer Perceptron and Nonlinear Autoregressive exogenous model

With the introduction of competition in the power generation sector in Korea, determining how much each electricity company should develop and maintain a stable system marginal price (SMP) has become an important issue. In this study, SMP was predicted by comparing the values for each company using a multilayer perceptron (MLP) and nonlinear autoregressive exogenous (NARX) model among artificial neural networks (ANN), and setting electricity demand as an external variable. The Autoregressive Integrated Moving Average (ARIMA) was also compared. Each method predicted 30, 40, and 50 days. In the case of MLP, optimal results were obtained when the number of hidden layer units were 90, 85, and 80, whereas in the case of NARX, optimal results were obtained when the number of hidden layer units were 95, 80, and 70. Overall, the forecast errors were small in the order of ARIMA, MLP, and NARX for all predictions.