Researchers in Slovakia have demonstrated a machine learning framework that predicts PV inverter output and detects anomalies using only electrical and temporal data, achieving 100% accuracy in classifying inverter output states under static operating conditions at a solar installation.
A research team led by scientists from Slovakia’s Constantine the Philosopher University in Nitra has developed a new predictive and anomaly detection model for PV inverters in commercial installations. The new machine learning-based framework uses only temporal and electrical data, without relying on environmental sensors.
“The chosen algorithms, Random Forests for prediction and Z-score analysis for anomaly detection, were selected for their robustness, interpretability and suitability for small but high-frequency datasets, making them well suited to practical PV monitoring implementations,” the academics said. “Additionally, the absence of irradiation or temperature data is explicitly addressed by constructing time-based proxies (hourly, daily and weekday patterns) to capture the cyclical behavior of solar energy.”
The model uses real-world operational data from a grid-connected PV plant in western Slovakia, including two inverters with nominal power of 30 kW and 40 kW. Inverter, grid current and grid voltage data were collected at five-minute resolution from January 1 to February 1, 2025 using inverter and grid monitoring sensors.
Image: Constantine the Philosopher University in Nitra, Results in Engineering, CC BY 4.0
To enable machine learning analysis, preprocessing was required. A Random Forest Regressor was then trained to predict the inverter’s actual power (kW) at each five-minute step. A Random Forest Classifier was then used to map continuous power into operational states, namely low, medium and high. It could classify both the current state and a future state, one hour ahead. Finally, a Z-score analysis was used to quantify the extent to which actual power differed from predicted power. Values exceeding a statistical threshold were marked as anomalies.
“A Random Forest Regressor achieved high reliability in power prediction (R² = 0.995, mean absolute error = 0.12 kW), while classification models categorized the output levels under static conditions with 100% accuracy,” the results showed. “Anomaly detection using Z-score analysis identified significant outliers, especially during high production intervals. However, one-hour ahead classification revealed significant drops in predictive performance (accuracy = 36.4%), highlighting the inherent difficulty of forecasting under variable environmental conditions.”
In conclusion, the research team added that “unlike other recent work, which integrates meteorological and contextual data for multi-level diagnosis, the proposed model works exclusively on electrical measurements on the inverter and grid side. This distinction highlights the practical value of the presented approach in scenarios without environmental sensors, providing a transparent and computationally efficient alternative for interpretable anomaly detection.”
The framework was presented in “Predictive modeling and detection of anomalies in PV inverters using machine learning”, which was recently published in Results in technology. Scientists from the Slovak Constantine the Philosopher University in Nitra, the Hungarian Obuda University and the Czech University of South Bohemia in České Budějovice took part in the study.
This content is copyrighted and may not be reused. If you would like to collaborate with us and reuse some of our content, please contact: editors@pv-magazine.com.
Popular content
