Comparison of Traditional ML Algorithms for Energy Consumption Prediction Models

Abstract

Data centers consume a large amount of energy to meet the increasing demand for IT infrastructure and software due to the IT equipment and cooling infrastructure involved. Therefore, it is necessary to have energy consumption control strategies for DC computer equipment that will allow infrastructure upgrades to reduce energy consumption and to meet the requirement of Green IT. In this way, energy consumption is reduced and the use of technological resources can be optimized. In this paper, we propose to evaluate several traditional Machine Learning algorithms as prediction models using three different temporal windows (i.e. minute, hour and day) taking into account several features such as voltage, energy, frequency, current, power, power factor, and temperature. A comparison of the root square mean error (RMSE) during the validation stage is carried out in order to select the most appropriate algorithm for each time window. In addition, running times are calculated to determine the feasibility of the selected algorithms. Moreover, the suitable predictive model can be the key to the ensure a fair distribution of the workload among the different servers in a Datacenter.