Power Transformer Forecasting in Smart Grids Using NARX Neural Networks

Abstract

Inthenextyears, withagrowingpresenceof electric vehicles and a massive penetration of renewable sources and low levels of voltage for self-consumption, it will be essential that medium- and low-voltage distribution networks be planned, operated, and supervised as transportation networks have been managed for decades, from the distributor to be a simple agent of distribution assets to be the operator of the network. This paper shows a non-linear autoregressive neural network with exogenous inputs (NARX) for time-series forecasting and power transformers monitoring. The NARX network model provides a description of the system by means of a non-linear function of lagged inputs, outputs, and prediction errors that can be interpreted as a recurrent dynamic network, with feedback connections enclosing several layers of the network. The prediction model consists of a multilayer perceptron (MLP) in the hidden layer that takes as input a window of past independent (exogenous) inputs and past outputs followed by an output layer that finally forecast the target time series. A previous study was carried out in order to select the most important electrical measurements enabling the prediction of the safe operation of the power transformer. The selection of the electrical measurements that have more influence on the transformer temperature was based on the computation of the pairwise Pearson’s correlation coefficient, the Kendall’s rank correlation coefficient as well as the cumulative conditional Granger causalities. The proposed NARX network was trained and evaluated in open-loop and closed-loop modes showing a high accuracy when predicting and monitoring the operation of power transformers.