A HYBRID CNN & BiLSTM WITH AM MODEL TO PREDICT AND ESTIMATE LOAD HARMONICS AT NESTLE EAST LONDON SOUTH AFRICA.

Abstract

Prediction of electrical power harmonics promotes development and provision of clean power supply. However, the elevated frequency combined with noise of harmonics make precise forecasting challenging and demanding. Due to the stochastic nature of harmonics occurrence and the challenge of attaining a dependable, and efficient working model, to this date research efforts have not been sufficient. Although various statistical and machine learning algorithms have shown very interesting and promising results, but work is still being done on algorithms that would produce the least possible error. This research uses hybrid convolutional neural network (CNN) and bidirectional long short term memory (BiLSTM) with Attention Mechanism (AM) model, to predict load harmonics at Nestle, a confectionary manufacturing plant in South Africa. Historical load harmonics data is used as the dataset. The results show that the hybrid deep machine learning method CNN – BiLSTM – AM has better performance compared to other five prediction models when detecting and forecasting load harmonics. The hybrid algorithm has a prediction accuracy of 92.3569 % and the lowest RMSE of 0.0000002215.