The intermittent nature of solar energy disrupts grid stability, leading to voltage fluctuations and power outages. Accurate intra-hour prediction of photovoltaic power output would allow grid operators to better balance supply and demand, while simultaneously monitoring the system’s performance.  

Most methods for photovoltaic power prediction rely on black-box models, which are hard to interpret. Purely physics-based models, on the other hand, struggle to adapt to dynamic weather conditions (e.g., cloud movement). Therefore, we explore novel hybrids of physics-based models and neural networks to obtain less-biased, noise-sensitive models.

1. Transform Global Horizontal Irradiance into Plane of Array Irradiance using established physical models

2. Adapt the physical model for future prediction 

3. Build a sequential Deep Learning model that captures temporal dependencies in meteorological data 

4. Merge the model predictions taking into account sky conditions, to create a more accurate hybrid

• Explore the strengths and weaknesses of physical and DL models to create a hybrid model that outperforms them 

• Manage model size to account for small datasets, mimicking early-stage deployment 

• Develop the hybrid to a proof-of-concept (TRL 3). 

The physical model adapts well-known mathematical frameworks to compute future PV output using the Global Horizontal Irradiance (GHI). Most of the other variables are computed mathematically. To account for panel tilt, the GHI is decomposed and then transposed using specific models (Rendl 3, Hay-Davies). After the transformations, the GHI is now the Plane of Array Irradiance (POAI), which is specific to the tilted surface of the panels. Considering relevant PV system characteristics and losses, the PV power output is estimated mathematically. 

To support future prediction, the ESRA model is used to compute the GHI components under clear-sky conditions. To this end, the physical model makes use of 2 novel mathematical formulas:

A Temporal Convolutional Network (TCN) is trained on meteorological input (irradiance, humidity, wind speed, etc.) to predict the PV output. This network type was chosen for its ability to capture both short-term and long-term dependencies. As the available data spans only three months, keeping the model lightweight and efficient is a priority. 

The model uses a sliding-window approach, in which it is trained only on data with seasonality similar to that of the validation set. This method ensures that model performance remains consistent as model size increases. 

After extensive parameter tuning, the best configuration for the TCN is:

• sequence length = 24 (4 hours in the past)

• output horizon = 12 (2 hours in the future)

• Adam optimizer 

• ReduceLRonPlateau scheduler 

• 3 TCN layers with a [8, 16, 8] channel configuration 

• kernel size = 3 

• dropout rate = 0.2 

• learning rate = 1e^(-3) 

The predictions of the TCN and the physical model are merged to create a hybrid-aware model that considers sky conditions. The sky conditions are given by the values of the diffuse fraction Kd, which is a direct indicator of cloudiness. 

Given that the error characteristics are horizon-dependent, two versions of the regime-aware hybrid are created to consider the 10-minute interval and the 2-hour one separately. The coefficients, obtained through linear regression, highlight that the TCN excels at short-term prediction, while the physical model is more stable in the medium-term.  

As of April 2026, a conference paper detailing the full methodology and evaluation process is currently under peer review for KES 2026. The paper is named "A Regime-Aware Hybrid for Photovoltaic Power Forecasting Using Physical Modeling and Deep Learning" and any further updates regarding its publication status will be made on the website.  

The PROMETHEUS team aims to communicate its findings to the scientific community and other interested stakeholders. Results were shared through an academic publication (submitted to KES 2026) and an in-depth presentation for both specialized and non-specialized audiences. 

The results of the PROMETHEUS project were presented at the Scientific Seminar hosted by the Department of Computer Science from the West University of Timișoara, scheduled for April 29th, 2026, at 6 pm EET. The presentation could be attended in person in room 048 of the university or online. More details about the Scientific Seminar are available here.