Aplicação de um modelo computacional para predição de distorções harmônicas monitoradas em sistema geração eólica: um estudo de caso da planta de Casa Nova CHESF

Abstract

In an ideal scenario, power systems should operate with constant, balanced, and uninterrupted currents, both in amplitude and frequency. However, reality often presents significant variations due to multiple factors, impacting power quality. This study aimed to develop an effective model for predicting harmonic distortions using a Long Short-Term Memory (LSTM) neural network approach, combined with Dropout techniques to mitigate overfitting. In the first phase, focused on the 04S3 230 kV transmission line of the Sobradinho Substation (EletrobrasChesf), data were collected between July and October 2023 every 10 minutes using the ION7650 power quality meter. After corrections, harmonic distortions of 4.00% (THDa), 4.56% (THDb), and 1.04% (THDc) were recorded, with phases A and B exceeding the 1.5% limit set by ONS. The use of Active Power Filters (APFs) was proposed, following studies by Akagi (2005). In the second phase, a predictive model was developed in the laboratory using Python and Keras on a system with an Intel® Core™ i7 processor and an NVIDIA GeForce RTX 3060 graphics card. The data underwent preprocessing and were split into 80% for training and 20% for evaluation. With two hidden layers and 95 neurons each, the model was trained for 500 epochs, presenting a total error of less than 5%. The results show high accuracy in predicting harmonic distortions, with correlation coefficients close to 0.99 and mean absolute errors of 0.0067 (THDa), 0.0085 (THDb), and 0.0062 (THDc). The metrics indicate good performance in relative accuracy. Therefore, the LSTM model with Dropout proved effective in predicting harmonic distortions, with an average error of 0.0067, suggesting potential for practical applications in power systems. However, further research is needed to improve these techniques and adapt them to various operational scenarios. The primary results demonstrate that the LSTM-based model provides reliable predictions of harmonic distortions, establishing its effectiveness in wind power generation systems.