Prediction and Feature-Level Analysis of Power Quality Indicators in Industrial Power Systems with Reactive Power Compensation Active and Inactive Using Ensemble Learning

Reactive power compensation systems are widely used in industrial power grids for power factor improvement and reactive power optimization. However, the reliable estimation of power quality indicators under conditions where compensation is active or inactive is a critical engineering problem that has not been sufficiently addressed in the literature. This study proposes a prediction approach that provides high accuracy, interpretability, and noise robustness for both operating conditions. The input space consists of 19 physical parameters, including active (P), reactive (Q), and apparent (S) powers for each of the three phases, RMS current and voltage values per phase (I1rms-3rms and U1rms–3rms), interphase voltages (U12, U13, U23), and neutral current (Ineutral). The predicted targets are 11 fundamental power quality indicators: power factor (PF1–3, Pftot), displacement power factor (dPF1–3, dPFtot), and total harmonic distortion ratio per phase (ITHD1–3). The dataset was split into training and test subsets using a hold-out strategy, and independent bagging-based ensemble regression models were constructed for each target variable. Experimental results showed that the R² values ranged from 0.987 to 0.997 and the mean absolute percentage error (MAPE) values ranged from 0.68% to 3.57% for all targets. Normalized feature importance scores revealed the dominant role of RMS current components per phase. Model robustness was maintained with ΔR² < 0.004 under 5% Gaussian noise. The findings prove that the proposed method can predict power quality indicators with high reliability in both compensation cases and is suitable for predictive maintenance applications with real-time monitoring on an industrial scale.