Daily electricity consumption forecast for a steel corporation based on NNLS with feature selection

Electricity consumption forecast is very important for both suppliers and large consumers. However, the electricity consumption of a large enterprise is quite different with regional consumption, and has not been studied sufficiently, especially for an energy intensive corporation. In this paper, we investigate the daily electricity consumption forecast of a large steel corporation. By our observation, the electricity consumption is inversely proportional to maintenance duration and directly proportional to production quantity. Therefore, the production and maintenance schedules are considered as input data of the forecast model. The nonnegative least squares (NNLS) method is applied to build a linear regression forecast model with nonnegative coefficients. In addition to NNLS, random approximated greedy search (RAGS) based feature selection method is applied to select the relevant input features on the available items of maintenance and production schedules. Then the ensemble forecast models are built based on the selected feature subsets by bagging approach. Numerical testing results on the real data from a steel corporation show that results obtained by the NNLS are stable, and the forecast accuracy is greatly improved by our ensemble forecast model.