Overview

IT companies handle huge amounts of data that have to be processed constantly. Data Processing Centers do exactly that, and these buildings have to be kept at a constant supply y electricity while maintained at an optimal temperature.

All this translates into a very high energy and power demand (which means a significantly high electricity bill), and that leads these companies into setting up devices which monitor the energy and power consumption of certain machines in the buildings. This is how they keep track of the real consumption of these stations for alater comparison with the electrical bill, which most of the times is ponderated negatively against them, making them pay sustancial amounts of extra money.

The Challenge

The main problem with this systems is that measuring errors happen often, leading to missleading conclussions...

That's where AI comes in handy. Time series algorithms can be very helpful in this sort of situations, where daily lost data can be easily corrected using certain period of past data for present prediction.

This algorithm had to process the data of all the technical buildings in the country. That ment that it had to be able to handle different scenarios, as not all Data Centers are reporting all the wanted variables. Also, new buildings are being built while others are taken down constantly. The ability to identify new centers with enough useful data for processing, was a big requierment in order to full automate this algorithm.

My Solution

Many different strategies where though of. The main goal was to find the best fitting algorithm to fit these specific time series, where not all Data Centers behave the same way (note that the locations weather can be very significat to the buildings electricity demand). The best outcomes where found in the following models:

• Recurrent Neural Networks (RNN)
• XGBoost Regressor
• Facebook Prophet

The best performance came from FB Prophet, which when trained with only a month of data, was able to accurately adjust to the distribution and the patterns. The main problem with RNN was that the ammount of data providede seemed to not be enough for the algorithm to behave moderately against energy peaks and significant variantions in consumptions... Also, the training time made it arduous to itterate and find the best models. As I said before, not all Centers behave the same, so generalizing wasn't much of an option. That ment 10h trainings.

On the other hand, Prophet's training time is around one minute, which means the algorithm can be trained daily without necessarily saving it for later use. That simplifies the strategy and assures the usage of the latest data without weight handeling.

‍

Later on, the resuts where contrasted with the bills from the supplier, realizing that tens of thousands of € could be lowered, saving the company great resources.

‍