AI Outperforms Traditional Methods in Solar Forecasting

As solar energy rapidly expands, predicting its short-term variability has become crucial for grid stability and efficient power management. Solar photovoltaics have more than doubled their share of renewable energy over the past five years, increasing the demand for accurate forecasts to handle production fluctuations. These forecasts are essential for applications like power plant control, energy trading, and integrating storage systems, where unexpected changes can lead to penalties or equipment degradation. Traditional s, such as numerical weather prediction, are useful for longer horizons, but for the critical 0 to 30-minute range known as nowcasting, deep learning models using ground-based sky images have emerged as a promising tool. However, these models often struggle with the chaotic nature of the atmosphere, highlighting the need for more reliable approaches that can better leverage visual data from the sky.

Researchers from the Institute for Energy Technology and the University of Oslo have discovered that a specific of processing sky images significantly enhances the accuracy of solar irradiance forecasts. In a comparative study, they evaluated three different techniques for incorporating all-sky imager (ASI) data into deep learning models designed to predict global horizontal irradiance (GHI) up to 15 minutes ahead. The key finding was that C, which uses engineered features aggregated into time series, outperformed the others. This achieved an average root mean squared error (RMSE) of 87.0 W/m² across seven test days, compared to 87.8 W/m² for A (using raw RGB images) and 90.3 W/m² for B (using engineered feature maps). The skill score, which measures improvement over a baseline persistence model, also favored C, with an average of 5.7% for horizons over one minute, versus 5.3% for A and 2% for B. This demonstrates that simplifying image input through feature aggregation can lead to more effective predictions, especially in variable weather conditions.

Ology involved collecting data from a high-frequency, 29-day dataset in Kjeller, Norway, using a pyranometer to measure GHI and two ASI systems to capture sky images every ten seconds. The researchers compared three approaches: A fed raw RGB images into a convolutional neural network (CNN) combined with long short-term memory (LSTM) layers; B used engineered 2D feature maps, such as cloud segmentation, cloud motion vectors, solar position, and cloud base height, processed through a CNN-LSTM hybrid; and C aggregated these engineered features into time-series data, which were then input directly into LSTM layers without spatial processing. All models were trained on historical GHI data and evaluated on seven test days with diverse atmospheric conditions, using metrics like RMSE and skill score to assess performance across 90 forecast horizons from 10 seconds to 15 minutes. The training involved optimizing hyperparameters through trial-and-error and grid searches, with models kept relatively small to ensure computational efficiency.

Analysis of , detailed in Figures 5 and 6 of the paper, reveals that C consistently delivered superior forecasting performance, particularly on days with high variability in solar irradiance. For instance, on June 18, 2024, a day with mixed cloud conditions, C achieved a maximum skill score of 13% at a horizon of 13 minutes and 20 seconds, outperforming the other s. In contrast, on overcast days like April 5, 2024, characterized by thin altostratus clouds, A showed an advantage, suggesting that raw image data can capture nuances like cloud optical thickness that engineered features might miss. Feature importance analysis, shown in Figure 7, indicated that cloud segmentation, solar azimuth, and clear-sky irradiance were the most critical inputs for C, with their importance increasing at longer forecast horizons. This underscores the value of physically informed features, while also revealing that some engineered elements, such as cloud base height and motion vectors, had minimal impact on model performance, potentially allowing for computational savings in future implementations.

Of this research are significant for the renewable energy sector, as more accurate short-term solar forecasts can enhance grid reliability and reduce costs associated with energy imbalances. By showing that aggregated engineered features outperform complex deep learning architectures, the study offers a simpler, more explainable approach that could be easier to deploy in real-world settings, such as solar farms or energy trading platforms. This reduces the need for computationally intensive image processing and large datasets, making it accessible for regions with limited resources. However, the study acknowledges limitations, including of evaluating forecasts under diverse atmospheric conditions and the reliance on a specific dataset from a high-latitude location, which may not generalize to all climates. Future work could focus on improving feature extraction, such as estimating cloud optical thickness, and benchmarking against public datasets to ensure broader applicability and robustness in solar energy management.