In a recent article published in the journal Sustainability, researchers investigated a new deep learning (DL) approach for forecasting solar power generation (SPG) across multiple sites. They aimed to develop a scalable and accurate SPG forecasting model that could be applied to various locations using a single common model, addressing the limitations of traditional site-specific models.
Image Credit: leolintang/Shutterstock.com
Background
In recent years, renewable power generation, particularly solar power, has witnessed remarkable growth due to its potential to address environmental concerns and its increasing economic viability. However, integrating solar power into the energy grid presents unique challenges because its variability depends on factors like sunlight intensity and duration. Therefore, accurate SPG forecasting is crucial for ensuring grid stability and maximizing solar energy utilization efficiency.
Previous work in SPG forecasting has primarily focused on developing site-specific models. These models require collecting and employing location-specific input data, such as training data and weather conditions, to produce forecasts for an individual site. This approach limits scalability and efficiency when extending forecasting capabilities across multiple sites.
About the Research
In this paper, the authors developed a scalable and accurate SPG forecasting model that can be applied across multiple sites using a single model. They introduced a novel DL-based model that leverages common meteorological elements, such as humidity, temperature, and cloud cover, to extract site-specific features and improve forecasting accuracy.
The proposed model consists of two subsystems: a feature encoder and a regressor. The feature encoder uses convolutional neural networks (CNNs) and long short-term memory (LSTM) networks to extract features from 24 hours of weather data, including solar elevation and azimuth angle. The regressor, a multilayer perceptron (MLP), interprets these encoded features to predict solar power output.
To address the variability of different sites, the study integrated a classifier module within the Encoder system. This classifier helps the encoder to implicitly understand the types of weather data, embedding local site characteristics into the model. This approach enhances the model's robustness and adaptability, enabling reliable forecasts for previously unknown sites by leveraging feature similarities to infer local environmental conditions.
Furthermore, the researchers evaluated the performance of their proposed system using SPG data from seven sites across the Republic of Korea. They compared the performance of a site-specific model, which was trained and tested separately for each site, with a common model, which was trained on data from multiple sites.
Research Findings
The outcomes showed that the site-specific model achieved a mean absolute error (MAE) of 3.43. This significantly surpassed the regulatory requirement of an 8% MAE threshold for participation in the Republic of Korea's renewable energy generation forecasting system. However, the common model without a classifier experienced a drop in prediction accuracy for unknown sites.
Including the classifier module in the common model led to a 3-6% improvement in performance. This demonstrated its effectiveness in using site-specific information to enhance forecasting accuracy across new and diverse locations. The classifier module also reduced the mean squared error (MSE) and root mean squared error (RMSE). This helped minimize larger errors and maintain robust forecasting performance across various sites.
Furthermore, the authors investigated the effectiveness of transfer learning (TL) by retraining the common model with a small subset of site-specific data. The TL scenario improved prediction accuracy across all sites, especially for those with unique meteorological conditions. Including the classifier module in the TL scenario further boosted performance. This emphasized its crucial role in using site-specific information to enhance the model's adaptability and generalization capabilities.
Applications
The presented model has significant implications for the renewable energy sector, especially in improving the operational efficiency of solar power systems. By providing accurate SPG forecasts across multiple sites, the model aids in better grid management and planning, making it easier to integrate solar energy into power grids. This leads to improved stability and efficiency in energy systems, supporting the global shift towards sustainable energy sources. Additionally, the model's scalability and adaptability make it valuable for expanding solar power infrastructure in various geographical regions.
Conclusion
In summary, the DL-based novel approach was effective for forecasting solar power across multiple sites. The paper highlighted the model's robustness and its potential to support the integration of renewable energy into power grids.
Future work should focus on determining the best combination of sites for configuring a common model, exploring hybrid models that combine the strengths of common and site-specific models, and incorporating seasonal variations to enhance accuracy and reliability across different climates. These efforts could enhance the effectiveness and scalability of solar power forecasting models, thereby advancing the integration of renewable energy and ensuring grid stability.
Disclaimer: The views expressed here are those of the author expressed in their private capacity and do not necessarily represent the views of AZoM.com Limited T/A AZoNetwork the owner and operator of this website. This disclaimer forms part of the Terms and conditions of use of this website.
Source:
-
Jang, S.Y.; Oh, B.T.; Oh, E. A Deep Learning-Based Solar Power Generation Forecasting Method Applicable to Multiple Sites. Sustainability 2024, 16, 5240. DOI: 10.3390/su16125240, https://www.mdpi.com/2071-1050/16/12/5240