Machine Learning and ArtificialIntelligence techniques for data-informed energy efficiency management in thetelecommunication sector

The ecological and economical concerns for consumes reduction, the huge amount of data collected and the widespread of Machine Learning techniques have enhanced the employment of Artificial Intelligence methods as tools for optimal energetic management of facilities, for fault detection and consumes prediction. Many algorithms have been developed in the attempt of successfully deal with Big Data, with the perspective of supporting analysis, classifying elements and providing predictions and reference values. Time-serie analysis is an important field of application of Machine Learning algorithms, with specific features and constrains. Clustering of time-series, sub-sequences analysis, regression techniques, correlation analysis, decomposition and pattern discovery, development of appropriate prediction tools are some typical tasks regarding analysis of consumes. In this thesis, several Machine Learning techniques are employed with the perspective of supporting efficient energy management of Telecommunication facilities. Telecommunications represent a fast-growing energy-intensive sector, constantly dealing with the opportunity of improving efficiency of plants, with a particular regard to indirect consumes, mainly due to the consumption of the cooling system. The thesis deals with a consistent dataset, comprehending 1819 Telecommunication facilities and electricity consumption hourly measurements for one year. The investigated structures include Central Offices, Data Centers, Radio Based Stations and offices buildings located all around Italy. Three specific goals were identified, namely re-classification of plants according to usage, fault detection and consumes prediction. A methodology is proposed in order to cover all the preparatory steps, from raw consumption hourly data to the final goals. The employed Machine Learning tools include clustering algorithms, namely K-Means, K-Shape and DBSCAN, linear and multiple regression, auto-regressive model and Artificial Neural Networks. A robust methodology for re-classification of plants, employing K-Means algorithm and periodical components of time-series, was developed and the resulting outputs were compared to real usage of facilities identified as anomalous by the clustering algorithm. Long-short Term Memory neural network was employed for prediction of consumes and results were compared, by means of proper metrics, with a traditional Machine Learning tool, namely Auto-regressive model. Long-Short Term Memory algorithm was proved to be more effective, both regarding short-term prediction, which is employed for fault-detection, and long-term prediction.