머신러닝과 시계열 기법 기반의 초단기 시간단위 수요예측방법론 개발 연구

초록

Demand forecasting is an important field and it is safe to say that forecasting is a key component of economic activity. An accurate forecasting is the key to determining the competitiveness of all economic players. Forecasting an uncertain future is a difficult task and radical change in the external environment are adding to the difficulty of forecasting. Amid the increasing demand for accurate demand forecasting, the emergence of Big data, AI, ML, and DL following the development of computing power is becoming a major turning point in the demand forecasting field as well. In addition to the traditional forecasting methodologies, the use of dataming techniques is also rapidly increasing. And various efforts have been continued to improve the forecasting accuracy. In this paper, a hybrid forecasting methodology which is combined time series model and data mining technique and a multistage methodology are presented for short-term forecasting. Specifically, we developed a hybrid forecasting model that combines SARIMA(Seasonal Autoregressive Integrated Moving Average) and Random Forest, and a multistage methodology that utilizing the forecasting result of the upper-category as a variable in the forecasting process of the sub-category. In order to verify the methodologies presented in this paper, we use the rental data of ‘Seoul bike’(shared bicycle in Seoul) as verification data. As a result of the forecasting ‘Seoul bike’ demand for the next 7 days(every 3 hours) of rental point clusters, the average forecasting accuracy was 81.5%. It is high accuracy level considering that the forecasting unit was 3hours, forecasting horizon was next 56 steps, and the average accuracy by Random forest was 65%. In addition, it was confirmed that high accuracy was maintained steadily regardless of the time difference from the forecasting point unlike the characteristics of general demand forecasting, And the high accuracy level was confirmed as a forecasting model not only a 3 hours forecasting, but also daily(90.1%) and weekly(91.7%) forecasting. The research shows the forecasting methodologies of this paper is worth to use as a short-term forecasting model. And we confirmed that the methodologies are very useful to forecasting daily and weekly demand as well. It is expected that the methodologies proposed in this paper will be widely used as an accurate forecasting model in more diverse fields.

키워드