Machine learning as a tool to forecast the power quality of a wind energy power plant

Abstract

The large-scale integration of renewable energy sources, particularly wind power, into national grids introduces non-linear and intermittent behaviour that challenges the maintenance of acceptable power quality (PQ). In South Africa, these challenges must be managed within the limits and measurement requirements of NRS 048-2 and related IEC standards at the point of common coupling (PCC). This dissertation investigates how effectively machine learning (ML) models can forecast key PQ parameters at a utility-scale South African wind power plant, and how these forecasts perform when judged not only by conventional regression accuracy but also by compliance-oriented metrics aligned with NRS 048-2 (Missed Violation Rate, MVR, and False Alarm Rate, FAR). Using 10-minute operational SCADA data from the wind-farm PCC, a six-phase, leakage-aware analytical workflow is implemented in JMP Pro 18. The pipeline covers standards-aligned preprocessing, domain-informed feature engineering, chronological train/validation/test splits, and multi-model benchmarking. Four model families are compared: Boosted Trees (BT), Support Vector Machines (SVM), a feed-forward Neural Network (NN), and a SARIMAX baseline. Forecasts are evaluated using RMSE, MAE, MAPE, R² and compliance metrics (FAR, MVR), including a ±5 % sensitivity analysis around key NRS 048-2 limits. Results show that non-linear models are essential for PQ forecasting in this wind-integrated grid. BT consistently delivers the best or near-best performance for Voltage Unbalance, Frequency, long-term flicker (P_lt), and V_RMS (p.u.), achieving high R² (often > 0.95) and low MVR (≈ 5 % for Unbalance, ≈ 0–5 % for P_lt) with FAR typically below 1 %. SVM is particularly strong for V_THD and competitive for P_lt, while the NN performs best on short-term flicker (P_st) in terms of regression accuracy but exhibits higher MVR. SARIMAX underperforms on non-linear and volatile parameters, confirming the limits of purely linear approaches. The study provides the first detailed, NRS 048-2-aligned PQ forecasting case study on a South African wind plant, demonstrates a reproducible, standards-aware modelling pipeline, and shows how FAR/MVR-based evaluation can translate ML forecasts into actionable early-warning tools for grid operators, supporting proactive PQ management in renewable-rich networks.