PUE is one of the most important metric for a data center. PUE is the ratio of total power consumed by the data center by power consumed by IT devices. Most DCIM software will report on PUE by tapping the points where these measurements can be done. DCIM software is able to show the historical trend of PUE. Data center managers track PUE carefully to keep an eye on optimal cooling that is necessary as well as spike in IT load, either of which can cause an undesirable change in PUE.
Now, some data centers may want to see a projected future value of PUE. We believe DCIM software can forecast PUE by using the historical trends effectively with ML techniques. With this in mind a project was initiated, where the historical data set was pre-processed with outlier detection and removal, which plays an important role in any machine learning project. Then several regression models and Ensemble learning on Random forest regressor and Gradient boosting regressor were applied.
To demonstrate the efficacy of projection, 1 year’s data set of historical PUE data set was used from a data center containing the RECORD_TIME and PUE values. The model also takes as input the date for which PUE projection is desired. After applying the regression models, mentioned above, PUE value for the projected date was predicted by the model, with trained data set. This method has been tried with data sets from several data centers and projected value came within reasonable approximation of the actual value observed.
Based on the projected value of PUE from historical data sets for several data centers, we believe in most cases it is possible to project reasonably accurate PUE. However, as this is based on historical data there is an underlying assumption that the historical trend will continue in future. For e.g. in summer months, temperature will be similar and hence use the same amount of cooling as was done in the past. Also, IT load will follow a trend that has been observed in the past. If these underlying variables change drastically, the projected value will be off from the actual.