TL;DR
This study compares four machine learning (ML) models (ANN, RF, SVR, DNN) and a physical model for estimating solar radiation across South Korea using satellite data. The research highlights that more complex, deeper neural network architectures like Random Forest (RF) and Deep Neural Networks (DNN) are better at capturing the intricate spatial patterns of solar radiation, especially under challenging cloud conditions, compared to simpler models or traditional physical approaches.
What Problem Does This Paper Address?
Accurately estimating solar radiation across a region is crucial for various applications, including agriculture, hydrology, and renewable energy. While ground-based pyranometers provide precise measurements, their limited spatial coverage makes them insufficient for comprehensive regional assessments. Satellite data offers broad spatial and temporal coverage, but translating raw satellite signals into accurate surface solar radiation estimates is challenging. Existing methods often struggle with atmospheric complexities, particularly cloud cover, which significantly impacts radiation reaching the surface. This paper aims to evaluate and compare the spatial assessment capabilities of different data-driven (ML) and physical models for solar radiation estimation using geostationary satellite data, focusing on their ability to capture spatial variations and handle cloud-related uncertainties.
Methods
The study employed four machine learning approaches:
- Artificial Neural Network (ANN): A shallow network using the Levenberg-Marquardt backpropagation algorithm.
- Support Vector Regression (SVR): A regression version of SVM, utilizing kernel functions for high-dimensional mapping.
- Random Forest (RF): An ensemble method combining multiple decision trees, known for its robustness and ability to handle complex relationships.
- Deep Neural Network (DNN): A more complex, multi-layered neural network architecture, leveraging deep learning principles.
These ML models were trained and validated using data from the COMS MI geostationary satellite. A physical model (Kawamura model) was also implemented as a benchmark. The models were evaluated using two validation strategies:
- Hold-out validation: Splitting data into training, validation, and testing sets based on location.
- k-fold cross-validation: Randomly partitioning the data into k subsets, using k-1 for training and one for testing, repeated k times.
The study also analyzed variable importance from the RF model to understand which input features (e.g., solar zenith angle, satellite spectral bands) were most influential in estimating solar radiation.
Datasets
- COMS MI Satellite Data: L1b data from the Communication, Ocean, and Meteorological Satellite (COMS) Meteorological Imager (MI) was used. This included visible and infrared channels, acquired between April 2011 and December 2017, with a spatial resolution of 1 km for visible and 4 km for infrared bands.
- Ground Pyranometer Data: 35 ground pyranometers (CM21 type) from South Korea, maintained by the Korea Meteorological Administration (KMA), provided reference data for training and validation. The observation period ranged from April 2011 to December 2017.
Key Insights and Ideas
- Deep Architectures Excel: The study strongly suggests that more complex, multi-layered neural network architectures (RF and DNN) are superior in capturing the intricate spatial patterns of solar radiation, especially under cloudy conditions, compared to simpler ML models or even the physical model.
- Cloud Masking is Critical: The physical model's performance was significantly hampered by cloud masking errors, leading to overestimation of solar radiation in areas with thin clouds. This highlights the challenge of accurately identifying and accounting for cloud effects using physical models alone.
- Data-Driven Models Adapt: ML models, particularly RF and DNN, demonstrated a greater ability to simulate spatial variations, even when faced with limited ground data and without explicit understanding of the physical processes.
- Variable Importance Matters: Solar zenith angle and visible spectral bands were identified as the most crucial predictors for solar radiation estimation, underscoring the importance of geometric and atmospheric factors.
- Temporal Sensitivity: While all models showed temporal variations in accuracy, RF and DNN exhibited less sensitivity to the number of data samples compared to ANN and SVR, suggesting better generalization capabilities.
Key Results
- Accuracy: In the five-fold cross-validation, the RF model achieved the best performance among data-driven models (RMSE = 87.441 W·m⁻², R² = 0.888). The DNN model was a close second (RMSE = 88.219 W·m⁻², R² = 0.886). The physical model performed well but was slightly less accurate in cross-validation (RMSE = 91.787 W·m⁻², R² = 0.882). ANN and SVR showed significantly lower accuracies.
- Spatial Simulation: Visual comparison of solar radiation maps revealed that RF and DNN accurately simulated complex cloud spatial distributions, including thin clouds, where the physical model struggled due to cloud masking errors.
- Variable Importance: Solar zenith angle (SZA) and the visible spectral band (VIS) were the most important variables for RF-based estimation, followed by infrared bands (IR1, IR3). Geographically fixed variables like latitude and longitude were less important.
Limitations and Future Work
- Limited Ground Data Domain: The study acknowledges that the generalization of results to larger regions (like Northeast Asia) might be limited due to the training data being restricted to South Korea. Obtaining well-calibrated ground-measurement data from other countries is a challenge.
- Physical Model's Cloud Masking: The physical model's reliance on cloud masking proved to be a significant limitation, especially for thin clouds.
- Under/Overestimation: While RF and DNN showed good statistical performance, their predicted spatial distributions were sometimes underestimated (RF) or overestimated (DNN) compared to the physical model.
- Future Work: The authors suggest future studies should focus on expanding the spatial coverage of ground reference data to improve the generalization capabilities of data-driven models. They also highlight the potential of combining ML methods with satellite spectral imagery for producing more useful spatial radiation maps, particularly for regions with complex cloud patterns. The study concludes that RF and DNN, with their deeper network structures, are promising for future solar radiation field estimation.
Summary updated