부산대학교 도서관

Load forecasts are becoming increasingly important in an increasingly digitalized world, even for smaller companies, for energy procurement or operational optimization. At the same time, it is unclear how much historical data is required to calculate a sufficiently good forecast. To answer this question, this work investigates the impact of training set size (historical load, weather, and calendar information) on the predictive performance of a day-ahead load forecast in the industrial context. For this purpose, a use case study on the data of seven companies from the manufacturing sector using six model classes was conducted. The results suggest that a forecast can produce meaningful results on 18 months of data whereas a period of less than six months yields results of high variance. For six out of seven companies, the best model trained on historical data less than or equal to one year was at most 1.5% (MAPE) worse than the overall best model.