South Korean researchers have developed a groundbreaking guided-learning model for predicting PV power without the need for irradiance sensors. This innovative approach uses routine meteorological data, outperforming conventional methods and proving especially effective in noisy or inconsistent data conditions. The model's unique strength lies in its ability to learn an irradiance proxy from meteorological signals, enabling deployment without irradiance sensors while maintaining accuracy. The research team, led by Sangwook Park, has introduced a two-component system: a solar irradiance estimator and a power regressor. The system processes weather time series data, generating internal features that are then used to learn irradiance representations. After training and validation, the model is deployed without irradiance inputs, instead estimating irradiance internally and calculating PV power output. The model's performance was demonstrated using a dataset from Gangneung, South Korea, over a year, with three PV plants analyzed. The double-stacked LSTM architecture delivered the best results, with statistically comparable outcomes from the attention-augmented variant. The guided-learning method showed strong out-of-sample performance, with average improvements of 0.06 kW in hourly RMSE and 1.07 kW in daily RMSE over baseline approaches without irradiance data. When compared to reference methods using irradiance data, the improvements were even more significant at 1.03 kW and 15.33 kW, respectively. One of the most surprising findings was the model's better generalization at the test site compared to models using direct irradiance data during inference. The guided model remained stable and achieved lower error across both hourly and daily metrics, even when irradiance inputs were noisy or inconsistent. The research team is now expanding their study to diverse climates and installation types, exploring multi-station data fusion to enhance model robustness. They also plan to add missing-input robustness, uncertainty quantification, and out-of-distribution detection for extreme weather and sensor faults. Finally, they are scoping pilot deployments with grid operators to assess the operational value of the new model.
