Over the past decade, solar generation in Europe has grown rapidly. To maintain security of supply and balance the grid, it has become essential to forecast solar power production accurately.
Given the rapid build-out of solar in recent years, improving forecast accuracy remains a priority for the security of supply and a well-functioning power system. Meeting Europe’s net-zero ambitions will also depend on building workforce capability to apply AI and data methods in practice. To support this goal, Europe is implementing initiatives to strengthen digital and energy skills. This includes the AI4GreenDeal, which will equip students and professionals through industry-aligned education.
The limits of traditional forecasting models:
Historically, PV forecasts have often been produced using white-box models. They rely on well-understood physical processes to map irradiance and weather inputs to solar power generation. These models are interpretable, but they do not learn from operational observations. Essentially, this means that they can miss site-specific, real-world effects. Environmental factors such as soiling and shading can reduce PV yield and introduce systematic bias if not captured. In addition, operational behaviour can shape power delivery. This includes curtailment, where generation is intentionally reduced due to grid constraints or market conditions such as low or negative prices.
Data-driven methods and the cold-start problem:
Machine learning and other data-driven methods are extremely popular alternatives to traditional forecasting models because they learn plant-specific patterns directly from observational data. They do face several limitations, though. Notable among these is data availability (i.e., a large amount of data is needed before reliable, accurate forecasts can be made). This is especially pronounced in newly commissioned systems, where observational data has not yet been recorded, leading to the well-known cold-start problem. Performance usually improves with history, so in cases of data scarcity, early forecasts are less reliable
Bridging the gap with machine learning and simulation:
A third option involves combining training models from nearby systems with extensive data and adapting them to new situations. This can work well, but it presupposes the existence and accessibility of similar systems in the neighbourhood.
Another option, developed at KU Leuven and published recently in the International Journal of Forecasting, addresses this challenge by pre-training deep neural networks on copious amounts of simulated PV generation data. The simulated data is generated using PVGIS and the models are subsequently fine-tuned with available (sparse) observational data. If a model has already learned the core relationships governing PV generation from a wide range of realistic simulations, it can be useful much earlier when only limited real data is available.
Sample-efficient machine learning for solar forecasting:
Pre-training provides a strong starting point, while fine-tuning enables rapid adaptation with limited site data. Simulation, however, requires assumptions about PV system characteristics. In SolNet, synthetic data is generated across variations in panel specifications and the approach is reported to be robust to misspecification in system parameters. It aligns with the broader principle of domain randomisation (training across diverse simulated conditions to improve model generalisation under uncertainty). This often results in faster learning curves and smoother integration of new solar capacity.
Many modern machine-learning approaches rely on abundant data. However, this is often not feasible in the energy sector. Sample-efficient approaches that combine simulation-based pre-training with careful fine-tuning offer a practical path to improve PV forecasting in data-scarce conditions, supporting faster and more reliable renewable integration as solar continues to scale.
This is exactly the type of real-world challenge AI4GreenDeal is designed to tackle through the Advanced Energy Systems and AI Master’s programme that is being developed within the project. By translating advanced research into challenge-based learning for students and professionals, these methods can move from the pages of papers into practice. In turn, it will strengthen Europe’s power systems as the energy transition accelerates.
Hussain Syed Kazmi is the incoming Programme Director of the Master’s in Advanced Energy Systems and AI. He is also an Assistant Professor at KU Leuven, working at the intersection of energy systems and data science. His research focuses on large-scale data-driven forecasting and optimisation, and on how human behaviour, climate change, and market design affect the global energy system.
References
[1] Kazmi, Hussain, and Zhenmin Tao. “How good are TSO load and renewable generation forecasts: Learning curves, challenges, and the road ahead.” Applied Energy 323 (2022): 119565.
[2] Depoortere, Joris, et al. “SolNet: Open-source deep learning models for photovoltaic power forecasting across the globe.” International Journal of Forecasting (2025).
AI4GreenDeal is co-funded by the European Union under the DIGITAL programme, supporting Europe’s strategic priorities in advanced digital education, innovation and the green transition.