A. Genovese, V. Bernardoni, V. Piuri, F. Scotti, F. Tessore, Photovoltaic energy prediction for new-generation cells with limited data: A transfer learning approach, Proc. of the 2022 IEEE Int. Instrumentation and Measurement Technology Conf. (I2MTC 2022), pp. 1-6, Ottawa, ON, Canada, May 2022, ISSN 978-1-6654-8360-5.
Environmental informatics
Photovoltaic Energy Prediction
The Photovoltaic Energy Prediction project investigates data-driven methods for forecasting the electrical output of photovoltaic systems and assessing variations in their operating performance. Accurate prediction is important for planning renewable-energy production, balancing distributed generation, reducing wasted energy, and improving the stability and management of smart grids.
Photovoltaic production depends on several interacting factors, including solar irradiance, air and panel temperature, geographical position, orientation, weather variability, and the condition of the panel surface. Dust, dirt, snow, and prolonged periods without rainfall may reduce conversion efficiency, while direct on-site measurement of all relevant quantities can require costly instrumentation and specialised maintenance.
A first research direction therefore estimates photovoltaic power using measurements collected by nearby public weather stations. The approach combines global irradiance, temperature, rainfall information, panel temperature, and the duration of dry periods to predict the maximum power point of the observed photovoltaic module.
Several computational-intelligence paradigms are compared, including k-nearest-neighbour regression, support vector regression, and feed-forward neural networks with one or two hidden layers. These models learn the nonlinear relationship between environmental conditions and power production directly from measured data.
The experiments used 373 hourly samples collected during two measurement periods and combined local panel measurements with weather information from an ARPA Lombardia station located approximately two kilometres from the plant. The input variables included panel temperature, global irradiance, and the number of days since significant rainfall.
Among the tested methods, support vector regression obtained the lowest average absolute error, approximately 0.247 W, followed closely by the neural predictors. The results also confirmed that production is strongly related to irradiance and can decrease as the drought period becomes longer, suggesting that deviations between expected and measured output can support maintenance scheduling.
A second major research direction addresses photovoltaic technologies for which only limited historical data are available. This is especially relevant for third-generation cells such as dye-sensitized solar cells, which offer simpler fabrication and less intrusive installation but do not yet provide the large operational datasets available for conventional silicon technologies.
The project introduces a transfer-learning approach that creates a digital twin using data from several photovoltaic technologies and multiple geographic locations. A fully connected neural network is first trained to learn the general relationship between meteorological variables and photovoltaic output across different cells and years.
The digital twin is then calibrated using only a small amount of data from the target photovoltaic technology. In the reported experiments, the calibration stage used data corresponding to a single month, reproducing a realistic scenario in which a new-generation cell has only recently entered operation.
The methodology was validated across different locations, including Edinburgh, Milan, Piacenza, Palermo, and Dubai. Calibration consistently reduced the mean absolute error, with reported values decreasing from approximately 22.74–37.30 before calibration to about 13.48–14.56 after transfer learning.
These results show that knowledge learned from conventional and heterogeneous photovoltaic systems can be transferred successfully to new cell technologies. The digital-twin approach therefore reduces the amount of target-specific training data required while preserving useful prediction accuracy across different climates and operating conditions.
The project also considers the real-world degradation of third-generation solar cells under environmental stress. Experimental activities expose dye-sensitized solar cells to combinations of meteorological conditions, gaseous pollutants, and atmospheric particulate matter, and evaluate their degradation through changes in radiation-to-current conversion efficiency.
The resulting measurements are intended to be combined with environmental information from regional monitoring agencies and analysed through digital-twin models. This creates a unified framework for studying both expected energy production and long-term performance loss under realistic environmental exposure.
Overall, the project integrates photovoltaic modelling, weather data, machine learning, transfer learning, digital twins, and experimental degradation analysis. Its main contribution is a flexible methodology for predicting photovoltaic output, detecting persistent losses, supporting maintenance decisions, and extending reliable forecasting to emerging photovoltaic technologies with limited available data.
Photovoltaic production depends on several interacting factors, including solar irradiance, air and panel temperature, geographical position, orientation, weather variability, and the condition of the panel surface. Dust, dirt, snow, and prolonged periods without rainfall may reduce conversion efficiency, while direct on-site measurement of all relevant quantities can require costly instrumentation and specialised maintenance.
A first research direction therefore estimates photovoltaic power using measurements collected by nearby public weather stations. The approach combines global irradiance, temperature, rainfall information, panel temperature, and the duration of dry periods to predict the maximum power point of the observed photovoltaic module.
Several computational-intelligence paradigms are compared, including k-nearest-neighbour regression, support vector regression, and feed-forward neural networks with one or two hidden layers. These models learn the nonlinear relationship between environmental conditions and power production directly from measured data.
The experiments used 373 hourly samples collected during two measurement periods and combined local panel measurements with weather information from an ARPA Lombardia station located approximately two kilometres from the plant. The input variables included panel temperature, global irradiance, and the number of days since significant rainfall.
Among the tested methods, support vector regression obtained the lowest average absolute error, approximately 0.247 W, followed closely by the neural predictors. The results also confirmed that production is strongly related to irradiance and can decrease as the drought period becomes longer, suggesting that deviations between expected and measured output can support maintenance scheduling.
A second major research direction addresses photovoltaic technologies for which only limited historical data are available. This is especially relevant for third-generation cells such as dye-sensitized solar cells, which offer simpler fabrication and less intrusive installation but do not yet provide the large operational datasets available for conventional silicon technologies.
The project introduces a transfer-learning approach that creates a digital twin using data from several photovoltaic technologies and multiple geographic locations. A fully connected neural network is first trained to learn the general relationship between meteorological variables and photovoltaic output across different cells and years.
The digital twin is then calibrated using only a small amount of data from the target photovoltaic technology. In the reported experiments, the calibration stage used data corresponding to a single month, reproducing a realistic scenario in which a new-generation cell has only recently entered operation.
The methodology was validated across different locations, including Edinburgh, Milan, Piacenza, Palermo, and Dubai. Calibration consistently reduced the mean absolute error, with reported values decreasing from approximately 22.74–37.30 before calibration to about 13.48–14.56 after transfer learning.
These results show that knowledge learned from conventional and heterogeneous photovoltaic systems can be transferred successfully to new cell technologies. The digital-twin approach therefore reduces the amount of target-specific training data required while preserving useful prediction accuracy across different climates and operating conditions.
The project also considers the real-world degradation of third-generation solar cells under environmental stress. Experimental activities expose dye-sensitized solar cells to combinations of meteorological conditions, gaseous pollutants, and atmospheric particulate matter, and evaluate their degradation through changes in radiation-to-current conversion efficiency.
The resulting measurements are intended to be combined with environmental information from regional monitoring agencies and analysed through digital-twin models. This creates a unified framework for studying both expected energy production and long-term performance loss under realistic environmental exposure.
Overall, the project integrates photovoltaic modelling, weather data, machine learning, transfer learning, digital twins, and experimental degradation analysis. Its main contribution is a flexible methodology for predicting photovoltaic output, detecting persistent losses, supporting maintenance decisions, and extending reliable forecasting to emerging photovoltaic technologies with limited available data.
Relevant publications
S. Ferrari, M. Lazzaroni, V. Piuri, A. Salman, L. Cristaldi, M. Faifer, A data approximation based approach to photovoltaic systems maintenance, Proc. of the 2013 IEEE Workshop on Environmental Energy and Structural Monitoring Systems (EESMS 2013), pp. 1-6, Trento, Italy, September 2013, ISSN 978-1-4799-0628-4.
V. Bernardoni, S. Valentini, G. Valli, F. Crova, Alice C. Forello, A. Genovese, R. Vecchi, F. Tessore, Sviluppo di un set-up sperimentale per l'esposizione di celle solari a diversi fattori di stress ambientale, Proc. of the X Convegno sul Particolato Atmosferico (PM 2022), Bologna, Italy, May 2022.
R. Donida Labati, A. Genovese, V. Piuri, F. Scotti, Towards the prediction of renewable energy unbalance in smart grids, Proc. of the 2018 IEEE 4th Int. Forum on Research and Technology for Society and Industry – Innovation to shape the future (RTSI 2018), pp. 1-5, Palermo, Italy, September 2018, ISSN 978-1-5386-6282-3.