Forecasting Data Center Load Using Hidden Markov Model

The number of data centers is increasing at an alarming rate, causing an ever increase in the operating costs. Several works in literature propose the use of data centers as a virtual power plant (VPP) and their participation in the power market. A day-ahead load forecasting is an integral part of the energy management system (EMS) of data centers providing a baseline to schedule energy resources and thus reducing operating costs. Traditional methods of load forecasting are not suitable for a data center load due to its high variability and a difference in service from the utility load. Hidden Markov model (HMM) is a very flexible tool for modeling heterogeneous data, which is especially useful when the response is highly variable. Here, we propose an HMM to forecast the day-ahead load of a data center to assist in scheduling of available resources. A case study on the data center at National Renewable Energy Laboratory (NREL) Research Support Facility (RSF) resulted in an annual average mean percentage absolute errors (MAPE) of 2.93% and 3.52% for two models proposed.