As solar energy becomes an increasingly important part of the global energy mix, accurately predicting photovoltaic (PV) power generation is essential for maintaining grid stability and optimizing energy operations. However, rapidly changing weather conditions, cloud movement, and variations in solar irradiance can cause significant fluctuations in power output, making reliable forecasting a complex challenge.
Ubiik addresses this challenge with an AI-powered forecasting platform that combines real-time satellite imagery, meteorological data, physical radiation models, and deep learning algorithms to predict solar power generation with high accuracy. By integrating multiple data sources into a unified forecasting system, the platform delivers actionable insights to support energy planning and solar power plant operations.
A Multistage Approach to Solar Forecasting
Ubiik’s forecasting platform combines several complementary technologies to improve prediction accuracy across different time horizons. The system processes multispectral satellite imagery from the Himawari-8 geostationary satellite, which provides updates every 10 minutes across 16 spectral bands. Convolutional neural networks (CNNs) analyze these images to identify cloud patterns, track atmospheric changes, and extract localized weather features that influence solar power generation.
For longer-term predictions, the platform incorporates data from the Global Forecast System (GFS), enabling forecasts up to 72 hours ahead. These predictions are complemented by physical clear-sky models that calculate expected solar radiation based on the sun’s position, atmospheric conditions, and seasonal factors.
By dynamically combining satellite-based short-term forecasts with numerical weather predictions, Ubiik’s system maintains continuity across forecasting horizons ranging from minutes to several days.
Turning Meteorological Data into Accurate Predictions
Weather conditions affect solar generation in different ways. Ubiik’s approach integrates local solar power measurements with meteorological parameters to identify relationships between environmental conditions and actual photovoltaic output.
The research identifies solar irradiance as the strongest correlated variable, with a correlation coefficient of 0.9840. Other relevant parameters include air temperature, dew point, cloud type, relative humidity, and precipitable water. Incorporating these variables allows the platform to capture a broader range of atmospheric influences on solar generation.
The system also addresses one of the key challenges in deploying AI at new solar installations: limited historical data. By combining satellite imagery from multiple geographic locations within a shared training framework, Ubiik applies transfer learning techniques to improve forecasting at new sites without relying exclusively on years of local operational records.
Continuous Learning and Forecast Optimization
To maintain reliable performance under changing operating conditions, the platform uses ensemble learning and K-fold cross-validation to reduce overfitting. Its active learning process incorporates new operational data into monthly model updates, allowing forecasts to adapt as additional observations become available.
A dedicated PV evaluation mechanism also helps identify potential discrepancies between expected and observed plant capacity, including possible panel degradation or incorrect capacity information.
Before forecasts are delivered to users, a post-processing stage converts model outputs into time-series estimates of solar power generation and continuously monitors prediction errors. This enables operators to assess forecast performance and better understand changes in expected generation.
Forecasting Performance
The research evaluates forecasting accuracy using Root Mean Squared Error (RMSE) and Normalized Root Mean Squared Error (NRMSE), metrics commonly used to assess prediction performance in the energy sector.
For one-hour-ahead forecasting, the satellite-driven model achieves an intrinsic error rate of 6.05%. Under real-world operating conditions, the reported NRMSE increases to 6.85% when using commercial satellite feeds with approximately 20 minutes of latency, and to 7.50% when using public data feeds with 30–40 minutes of latency.
These results demonstrate the importance of both model performance and data availability when deploying forecasting solutions in operational environments.
Designed for Enterprise Integration
Beyond its forecasting engine, Ubiik provides a cloud-based infrastructure designed to support reliable service delivery, ongoing research, and controlled software updates. The platform includes separate production and development environments, automated data validation, and preprocessing mechanisms to address missing timestamps, mislabeled observations, and inconsistencies across data sources.
Forecasts are accessible through a web-based dashboard featuring time-series visualizations, as well as RESTful JSON APIs that enable integration with existing energy management systems. Security measures include HTTPS encryption, token-based authentication, rate limiting for failed login attempts, and secure password hashing.
Enabling Smarter Solar Energy Operations
By combining satellite-based computer vision, numerical weather forecasting, physical modeling, and continuously updated AI algorithms, Ubiik’s platform offers a comprehensive approach to solar power generation forecasting. Its multistage architecture is designed to support both short-term operational decisions and longer-term energy planning, helping utilities and solar plant operators anticipate generation fluctuations and manage uncertainty more effectively.
The full technical paper, “Forecasting of Solar Power Generation using Real-time Meteorological Information,” by Siraj Sabihuddin, Alexander Grishin, and Ho Ko of Ubiik Inc., provides further details on the system architecture, methodology, and evaluation results.
