System of automatic detection of PV mounting configuration

Read full document

Overview

Accurate knowledge of photovoltaic (PV) system mounting parameters is essential for optimal system operation, maintenance, performance evaluation, and reliable power forecasting. However, in practice, configuration data is often incomplete, inaccurate, or outdated. This study presents an automatic system developed by Solargis for detection of fundamental mounting parameters of PV installations using time series of measured PV production and location data.

The proposed system uses a combination of data quality control, machine learning classifiers, and a simulation-driven optimization approach. Its goal is to determine key mounting configuration parameters—including array type (fixed or tracking), azimuth, tilt, installed capacity, relative row spacing, and tracking rotation limits—based solely on energy output records and geographic coordinates. This automatic detection improves the reliability of forecasting, monitoring, and asset management in cases where ground truth metadata is unavailable or unreliable.

Key elements of the detection system

Input data

• PV yield time series (production measurements)

• Geographic coordinates of the site

Configuration parameters detected

• Mounting geometry type (fixed-tilt, 1-axis, or 2-axis tracking)

• Azimuth angle

• Tilt angle

• Installed power capacity

• Relative row/column spacing (inverse ground coverage ratio)

• Tracker rotation limits

Approach and methodology

Data quality control pipeline

The detection process starts with automatic quality assessment and filtering of the input time series. This includes:

• Removal of missing, duplicate, and out-of-bounds values

• Consistency checks for time references and data resolution

• Exclusion of periods impacted by hardware, communication, or environmental anomalies

• Filtering for clear-sky and high-elevation conditions to ensure valid inference (thresholds: Kt > 0.6, solar elevation > 5°)

Mounting type identification

Classification of the mounting system (fixed, 1-axis tracker, 2-axis tracker) is handled by an ensemble of machine learning models (Random Forest, Gradient Boosting), trained on production data from over 1,400 PV sites and synthetic samples. Model features include production statistics and relationships with modelled solar parameters. High classification accuracy (F1-score near 1.0) is achieved, provided data quality is sufficient.

Parameter inference and optimization

With the mounting type identified, solar resource models and PV simulation engines generate simulated yield series for a range of possible configurations. The best-fit configuration is found by minimizing a custom error metric (ESIGN, a weighted RMSE variant focused on penalizing overperformance) between simulated and measured series across the parameter search space. The method systematically finds the most plausible effective configuration for the site.

Evaluation and results

The system was validated using 86 operational PV sites with known metadata (tilt, azimuth, capacity). Typical deviations between predicted and true parameters were within acceptable limits, varying by geographic region and data resolution. Additional benefits were demonstrated in a portfolio forecast case study: for 24 central European sites, using estimated configuration parameters improved yield forecasting accuracy in 23 of 24 cases. The magnitude of improvement varied, but was significant in some instances, especially where metadata was inaccurate.

Limitations

• Best results require at least one year of sub-hourly granularity production data (15-minute preferred)

• Inference for relative spacing and tracker limits is less robust and dependent on data quality

• Cannot handle interval-based (time-varying) configurations; recommended to segment data if system characteristics change mid-series

• The detected configuration represents an “effective” best fit in the context of the simulation model, which may differ from the true on-site hardware configuration, especially in complex or mixed sites

Relevance and application

Having access to accurate PV configuration data underpins improved monitoring and forecasting for grid operators, asset managers, and service providers. The proposed system facilitates the reconciliation and ongoing validation of PV asset metadata, reducing forecast errors associated with misconfigured input parameters and enabling a feedback loop for continuous records improvement.

Solargis incorporates this approach in the Forecast service during the forecast enhancement process.

Further reading

• “System of Automatic Detection of PV Mounting Configuration” by Oliver Osvald, Tomáš Hruška, Artur Skoczek, Tomáš Cebecauer, Markéta Hulík Jansová