Colloquia

Summary

For many data centers (DC)s, the cooling system is one of the most energy-intensive components in the facility. Thus, in order to increase the energy efficiency of a DC, one requires optimization of a cooling system. This can be achieved by setting the optimal parameters of a DC such that the total cooling energy is minimized while the servers are protected from overheating. However, such an optimization problem is hard to mathematically describe as one does not have a proper model. In this thesis, the data-driven approaches are used for the prediction of the cooling capacity of DCs based on a set of its controllable parameters. In particular, the focus is on feedforward neural networks (FFNN), long-short term memory (LSTM) networks and Bayesian neural networks (BNN).

Since the design of DCs is not standardized and server capacity varies between DCs, each DC needs an individual neural network (NN) to ensure accurate predictions of the cooling energy and the reduced power consumption under different settings. Due to the increasing number of DCs as well as a large amount of data, it is computationally expensive to train a new model for each new DC. Therefore, there is a need for an online algorithm that can use existing (pre-trained) NN models, and only modify part of its architecture when needed, i.e. when the new DC is considerably different than the observed one. Therefore, this work presents the potential of different transfer learning techniques such as fine-tuning, pretraining, partial freezing and architecture modification as a possible substitute for training multiple NNs for each new DC. The transfer learning is further compared with the Bayesian type of learning since this method allows the incorporation of prior knowledge into the training of the model architecture. As both architecture and the training algorithm depend on a set of hyperparameters, the corresponding numerical analysis is undergone in order to find an optimal baseline model. The results show that FFNNs outperform LSTM neural networks in accurately predicting the cooling energy and total energy consumption of a DC. More importantly for this research, it can be observed that architecture modification is the best-performing transfer learning technique for FFNNs while fine-tuning is the best-performing technique for BNNs.