Power Flow Prediction for a Steel Plant with a Dynamic Regression Model
AbstractWe present a power flow forecast for an industrial complex consisting of more than 20 different companies. The predominant consumer of electricity in the industrial complex is a steelwork company with an electric arc furnace. For the mentioned production system, we have developed a robust power flows forecast for a month in the future, with a high level of confidence in the forecast. A steelwork with an electric arc furnace is a very specific example of an energy consumer. Other companies in the industrial complex are not connected to the steel plant technologically, but they are on the same energy connection. They have a weekly power flow profile significantly different from the steel plant. To calculate the forecast model and perform the forecast of power flows on the common power connection, we need only two inputs of data: Historical measurements of power flows and the number of loads of the electric arc furnace in the following days. These are input data in a forecast model that we get without any problems. We first showed a prediction with classical linear regression. The next model to predict was the seasonal ARIMA model with a regressor, also called a dynamic regression model. The dynamic regression model improved the prediction by 15% compared to static linear regression, according to the RMSE measure. This was followed by an improvement in the dynamic regression forecasting model by considering the seasonality 7/5 in the time series. We did this with a model with superimposed noise. This model incorporates the ARMA methodology and considers superimposed noise. With this model, we improved the forecasting by 30% to linear regression. Logically, the filter model of the prediction also improved, gaining more Lag coefficients, and losing a constant. Qualitatively, the result is a forecast of power flow for one month with prediction error MAPE 8% and measure R2 is 0.9.
Original Research Papers
This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.