Optimization of neural network parameters by Stochastic Fractal Search for dynamic state estimation under communication failure

Abstract

A powerful metaheuristic technique is proposed that uses a mathematical concept called Stochastic Fractal Search (SFS) to ensure fast convergence along with accuracy. This paper intends to determine the optimal set of multilayer perceptron neural network (MLP) parameters (weights and thresholds) to improve the performance of MLP by using the SFS technique. The SFS is used because of its effective search in finding the global minima and therefore, it avoids the MLP neural network trapped in local minima. The hybrid approach (MLP-SFS) is applied to solve the dynamic state estimation (DSE) problem at the filtering stage. DSE amalgamates forecasting procedure with measurement data to precisely assess the system state. The approach classifies the process into three stages. In the first stage, a short term hourly load forecasting is applied using support vector machine (STLF-SVM) for time series to forecast the unavailable load data due to communication failure from the previous hourly historical data load. The second stage constitutes an optimal power flow (OPF) that is used to determine the minimum cost generation dispatch to serve the given load and convert the obtained loads, and generations into measurement data. The third stage has a filtering process, which uses SFS technique to optimize the MLP neural network parameters (weights and thresholds) to estimate the system state. The hybrid MLP-SFS is used to find the optimal connection weights and thresholds for the MLP neural network. Following this, a simple backpropagation neural network (BPN) will adjust the final parameters. The approach is tested on IEEE 14-and 118-bus systems using realistic load patterns from the New York Independent System Operator (NYISO) under several scenarios of measurement error and communication failure. The mean absolute percentage error index of the system state (phase and magnitude voltage) is used to determine the accuracy of the approach (MLP-SFS). Results of the proposed approach (MLP-SFS) are compared with non-optimized MLP (random weights and thresholds) and other methods, such as, optimized MLP based on genetic algorithm (MLP-GA) and Particle Swarm Optimization (MLP-PSO) individually. The results indicate that the hybrid (MLP-SFS) increases the precision by about 20%–50% and reduces the computational time around by 30%–50%, which is good for real-time applications, such as, security assessment and contingency evaluation. Details of the models of generation and distribution level are not part of the state estimation in the high voltage transmission problems.