Detection of Harmonic and Interharmonics Contents in Water Desalination Plants' Distribution System Based on Deep Learning Algorithms

ABSTRACT Water desalination plants are significant consumers of electric power, making them some of the largest energy users in power grids. Their electricity consumption presents an urgent challenge for efficient and sustainable operation, and they are among the most power‐quality‐threatening customers for utilities. This study presents three distinct strategies to enhance prediction accuracy and extend forecasting horizons, aiming to reduce algorithmic and hardware delays. Additionally, it suggests effective methods to compensate for voltage fluctuations, voltage flicker, and dips arising from desalination plants. The paper also discusses forecasting harmonics and interharmonics in current signals. Furthermore, it integrates the above techniques into a comprehensive computing system, along with an active power filter (APF) scheme, within the Simulink framework. A comparison is drawn between the performance of predictive techniques in an APF and a conventional, non‐predictive APF. The proposed data augmentation method successfully increases prediction accuracy. By effectively forecasting upcoming waveforms, it reduces algorithmic and hardware delays. These techniques are designed to address multiple power quality issues simultaneously, including harmonics, interharmonics, flicker, and voltage dips, which often coexist in the spectrum as interharmonics. The suggested approach employs Long Short‐Term Memory (LSTM) networks combined with the Jetson TX2 embedded artificial intelligence computer to accelerate machine learning applications. This method has proven effective in predicting and classifying time series data, including harmonics, interharmonics, and raw current signals, achieving 100% accuracy. This eliminates the need for designing specific low‐pass filters. The evaluation results for this time‐domain deep learning‐based technique will be reported in the subsections below. The implementation is conducted in Python using the KERAS deep learning framework and TensorFlow backend, and it is evaluated on a workstation equipped with an Intel i7 processor running at 4.0 GHz and 48 GB of RAM.