Training a Multi-Layered Perceptron using Moth Swarm Algorithm for Predicting Energy Demand of a Data Centre and Weights-Based Analysis of Input Parameters

Abstract

Multi-Layered Perceptron is a type of artificial neural networks and obtaining the optimal weights and biases of the model is critical to achieving good performance. In this study, Moth Swarm Algorithm has been proposed to train a Multi-Layered Perceptron neural network by finding the best combination of weights and biases that produce outputs with the least possible Mean Squared Error. The model has been applied for predicting the energy demand of a data centre. The simulations have been conducted using real life data obtained from an anonymous data centre operator in South Africa. The input parameters considered in the model are the ambient temperature, ambient relative humidity, chiller output temperature and CRAC air supply temperature. The performance of the proposed method has been evaluated based on the Mean Squared Error, Root Mean Squared Error, Mean Absolute Error, Mean Absolute Percentage Error and accuracy values obtained for the training and testing set. By comparing the results obtained with other models like Moth Flame Optimization, Ant Lion Optimization and Whale Optimization Algorithm, it was found that the Moth Swarm Algorithm-trained Multi-Layered Perceptron outperformed the other models. Further, a Percentage Relative Contribution analysis has been conducted to highlight the level of influence each of the input parameters considered has on the energy demand pattern of the data centre. Analyses show that the ambient temperature has the highest influence of 31.7% on the energy demand of the building.