Advances in evolutionary neural networks with applications in energy systems and the environment

Abstract

Evolutionary neural networks combine two of the most powerful areas of computing, evolutionary algorithms and neural networks. There are a number of benefits of using evolutionary algorithms to train a network over traditional methods. These include: no need for target outputs and error gradients, applicable to both supervised learning and reinforcement learning problems, and robustness to noise and local optima.

This thesis presents both a number of novel applications of evolutionary neural networks to energy systems and also the development of algorithms for evolving neural networks for these problems. The applications of evolutionary neural networks to energy systems and the environment include: Watershed Management, Dynamic Economic Emission Dispatch, forecasting Ireland's power demand, wind power generation and carbon dioxide levels, and also forecasting CPU utilization in data centers. Each of these problems are of critical importance. The work described in this thesis demonstrates how evolutionary neural networks can have a positive impact on these problems.

Aside from applications to real world problems, this thesis also makes fundamental contributions to evolutionary neural networks. A meta optimisation analysis of the largest collection of Particle Swarm Optimisation velocity update equations is conducted, revealing that many uncommon parameter settings can enhance performance. A Neuro Differential Evolution algorithm is proposed that combines Genetic Algorithms with Differential Evolution to evolve the size, topology and weights of a neural network. This thesis also presents a Multi-Objective Neural Network trained with Differential Evolution that can successfully produce Pareto fronts for dynamic multi-objective optimisation problems.