Progress reports

Analysis using descriptive statistics on the data used for the ITL networks and on the data used in the Topology problem (Power system)

The training of neural networks must consider three datasets: train, test and validation. For each experiment, the three datasets must belong to the same population. The assessment of populations’ similarity for these three datasets is the main goal of this report. In this work, the Smirnov and Cramér-von Mises statistical tests were employed to three case studies. The analysis here presented is conducted on a neuron basis, meaning that the train, test and validation datasets are analyzed considering each neuron separately.

Theoretical Concepts of ITL Neural Networks

This report summarizes the theory applied to the ITL neural networks developed. The pseudo-code is presented, alongside with the main functions developed.

On the equivalence of maximizing entropy and mutual information

This study is conducted under the context of unsupervised training of neural networks with two layers, using the concepts of information theory to perform the training. The two criteria here addressed are: i) maximizing the entropy of the outputs (MaxEnt) and ii) maximization the mutual information between the inputs and outputs (MaxMI). The research question pursued is “are these two approaches equivalent?”. With base on the existing literature, it is possible to conclude that the two approaches are theoretically equivalent provided the system is noiseless.

Theoretical Concepts of BackPropagation Neural Networks

This report summarizes the theory applied to the PROP neural network developed. The pseudo-code is presented, alongside with the main functions developed.

Comparative Analysis between ITL and BackPropagation neural networks to power Systems Applications

This report includes the results obtained by applying the ITL and PROP networks to different case studies.

Training Neural Networks Theory of Practical Issues

Through the work conducted concerning the implementation of neural networks, some theoretical issues emerged. This work arises from the need to explore these issues, covering the main topics grounding the learning process and data processing. A review is provided on adequate methods of data normalization, synaptic weights normalization and initialization, window width estimation, activation functions selection, saturation and overfitting prevention and calibration of other parameters.

Breakers’ state estimation using autoassociative neural networks

This work addresses the breakers’ state estimation in power systems. A classification methodology is applied, which considers the competition of two autoencoders, each one trained to learn a specific manifold concerning a breaker’s status: “open” and “closed”. The classification decision is made with the adoption of the state referring to the autoencoder with the lowest error. The methodology was applied for 11 breakers, under two modes: the first includes, for the training of the autoencoders, all information that is immediately adjacent to the breaker and its buses, the second excludes information concerning the flows directly connected to the respective breaker. The first technique returned performances that can be considered equivalent to the empirical approach. The second approach tested (without direct flows), represents a significant gain: without information on direct flows, the empirical approach is unfeasible.

Wind-Hydro Coordination Problem

This working report includes a preliminary approach to the wind-hydro coordination problem. It addresses the theory of the problem, and some results obtained by solving this problem with EPSO under different assumptions.

Wind – Hydro Coordination Problem solved with linear programming

This report includes the theoretical formulation of an attempt to solve the wind-hydro coordination problem using linear programming. However, this is not a linear problem, with relaxation of some constraints, feasible solutions could be found. The results obtained are presented and discussed.

CPLEX tutorial

A report with a simple example on how to use IBM CPLEX with the excel interface.

Application of autoassociative neural networks to solve large scale problems in power systems

This paper synthetizes the work conducted concerning the application of autoassociative neural networks (autoencoders) to solve large scale problems in power systems. Autoencoders have been shown effective in providing a useful tool to compress and expand information in various domains. This work explores the autoencoders' ability to reduce dimensionality by applying an evolutionary optimization metaheuristic into a space of reduced dimension. This idea was firstly pursued with the goal of reducing computational effort, which is known to be significantly high for large scale problems. The bew methodology developed, “Hybrid”, allows an evolutionary metaheuristic to evolve in a reduced dimension space S’, controlling its evolution in the original space S. The results obtained are detailed.