Reliable Estimation of Neutral Current in Industrial Power Systems Using Genetic Algorithm–Based Ensemble Learning and Multimethod Explainability Analysis

İngilizce (OpenAlex)

Accurate estimation of neutral current ( I n ) in industrial three‐phase power systems is critical for harmonic suppression, equipment protection, and operational safety. This study proposes an ensemble regression framework optimized by a multiobjective genetic algorithm (GA) using 12,328 real‐field measurements based on 29 electrical characteristics ( P , Q , S ; I rms ; U rms ; PF, dPF; I THD , etc.). The GA simultaneously determines the selection and weights of the base learners (SVR, ANN, GPR, RF, GBR, XGB, DT, and GPR‐RQ), improving eight performance metrics together: RMSE, MAE, SMAPE, MdAPE, R 2 , EVS, maximum error, and PBIAS. Comparative analyses show that GA achieves high accuracy in 10‐fold cross‐validation compared to PSO, SA, random search, and average voting strategies (e.g., R 2 = 0.9972, RMSE = 1.83, and SMAPE = 10.31%); unseen test data maintained competitive overall performance (e.g., R 2 = 0.9820; SMAPE = 56.17%). In noise robustness, R 2 = 0.9933 was achieved in target‐injected disturbance scenarios. Optimization reached Pareto convergence in approximately 50 generations. In the explainability analysis, SHAP and LIME outputs showed significant differences ( p < 0.05) in 28 out of 29 variables; despite low inter‐method correlation (Pearson ≈ −0.022), they provided complementary insights. The results demonstrate that the GA‐XAI–supported ensemble provides high accuracy, interpretability, and applicability for I n prediction. To the best of our knowledge, this study presents the first I n prediction framework that statistically compares SHAP and LIME when used together with a GA‐optimized ensemble and reports the process in a reproducible MATLAB script. We translate these distinctions into a practical protocol: SHAP for global monitoring and policy and LIME for case‐level triage, thus enabling practitioners to confidently leverage complementary XAI signals during operations.