South Korean researchers have developed a supervised learning framework that accurately predicts PV power without the need for radiation sensors during operation, but instead using routine meteorological data. The model reportedly showed strong out-of-sample performance, while outperforming conventional radiation-based approaches, especially under noisy or inconsistent data conditions.
A research team from South Korea has developed a new supervised learning framework that jointly estimates irradiance and regresses PV power.
“The model first learns a proxy for the irradiance of routine meteorological signals and then uses that proxy for PV power regression,” said corresponding author Sangwook Park. pv magazine. “This enables deployment in locations without radiation sensors, while maintaining the accuracy benefits that these signals typically provide.”
The proposed framework only uses solar radiation measurements during training and does not require them during operation. According to the researchers, it consistently delivers the same level of accuracy even when applied to scenarios outside the training dataset.
The method consists of two main components: a solar irradiance estimator, which predicts the irradiance from meteorological inputs, and a power regressor, which supplements its inputs with the estimated irradiance and provides PV power normalized by the installed capacity. The system initially collects data such as temperature, humidity and wind speed and also records irradiation data during training.
A deep sequence model processes the weather time series to generate internal features. These features are passed to an estimation block and a region block, allowing the model to learn internal radiation representations. After training and validation, the model is deployed without input from irradiation. Instead, the irradiance is estimated internally and used to calculate the PV power.
The framework was demonstrated using a dataset collected in Gangneung, South Korea, over one year, from January 1, 2022 to December 31, 2022. Three PV installations were analyzed: C9 for training, N19 for validation, and C3 for testing. Several deep sequence models have been evaluated within the framework, including double-stacked models long short term memory (LSTM), attention-based LSTM, and convolutional neural network, long short-term memory CNN-LSTM architectures. The double-stacked LSTM delivered the best overall performance, with the attention-augmented variant showing statistically similar results.
“The proposed supervised learning method showed strong out-of-sample performance on the test set,” the researchers said. Statistical comparisons using t-tests and bootstrap methods showed average improvements over the baseline approaches without irradiation data of 0.06 kW in hourly root mean square error (RMSE) and 1.07 kW in daily RMSE. Compared to reference approaches using radiation data in both training and testing, improvements reached 1.03 kW and 15.33 kW, respectively.
Park noted that one of the most unexpected findings was that the guided model generalized better to the test site than models that used radiation data directly during inference. “When the irradiation input was noisy or inconsistent, conventional models degraded, while the guided model remained stable and achieved lower error on both hourly and daily metrics,” he said.
The research team is now preparing a multi-regional study covering different climates and installation types and investigating data fusion between multiple stations to further improve the robustness of the model. “We also plan to add missing input robustness, uncertainty quantification with calibrated forecast intervals, and out-of-distribution detection for extreme weather and sensor errors,” Park added. “Finally, we are investigating pilot implementations with grid operators to assess the operational value.”
The new model was introduced in “Supervised learning for photovoltaic energy regression in the absence of key information”, published in Measurement. Scientists from South Korea’s LG Electronics and Gangneung-Wonju National University participated 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.
