Distributed rooftop solar installations present a growing challenge for electric utilities: these systems typically lack real-time monitoring, leaving grid operators unable to accurately quantify how much solar generation exists within their distribution networks. This visibility gap complicates forecasting, voltage management, and grid stability. A new research approach addresses this problem using machine learning to infer distributed PV generation from aggregate measurements already available at distribution transformers.
The method employs a multiscale convolutional neural network trained to recognize the distinctive temporal patterns of solar generation. By analyzing aggregated power flows entering a transformer alongside solar irradiance data from nearby weather stations, the algorithm learns to isolate the PV component from overall load demand. The multiscale architecture captures both rapid fluctuations from cloud cover and longer-term solar cycles, mimicking how solar output naturally varies throughout the day.
Key to the approach is its reliance on existing data sources. Rather than requiring installation of dedicated meters at thousands of residential systems—an economically unrealistic solution—the method works with measurements already routinely collected. The algorithm establishes a mathematical relationship between reference solar station data and transformer-level measurements, then separates distributed PV generation from customer load without individual system visibility.
Testing confirms the method outperforms traditional statistical approaches in identifying PV output. Sensitivity analysis shows the framework performs reliably even with incomplete or uncertain sensor inputs, addressing real-world conditions where weather data or measurements may contain gaps or errors.
For utilities managing rapid solar adoption, this technique offers practical benefits: better demand forecasting, improved voltage regulation, and more accurate assessment of distributed energy resources. As solar penetration continues rising globally, non-intrusive identification methods become increasingly valuable for grid operators seeking to maintain reliability while accommodating renewable variability without massive monitoring infrastructure investments.
