Digitalisation of PV systems
Rationale for support
The mainstream PV industry produces cells and modules that are standard in size and shape, but are not optimised for the many emerging applications (e.g. agri-PV, floating PV, vehicle-integrated PV, ...) and do not deliver maximal potential electricity when shaded, a common situation in urban areas and indoor spaces. To fully harvest all available solar energy and to extend the variety of PV applications (towards “PV everywhere”), so-called digital PV-modules need to be developed. The combination of PV and various digital technologies can provide virtual inertia, support the grid and smooth the output of systems (e.g. by very accu- rate solar energy forecast and energy storage at the level of the PV module or system).
At present, most PV modules fulfil only one function, namely electricity generation, and therefore they usually contain only simple electrical actuators such as power op- timisers.
Field uptake of techniques for production forecasting (foremost among them operation and maintenance opti- misation, and sizing) is still limited, with TRL in the range 7-8 for forecasting applications and 3-4 for energy trading agents.
Application of AI to PV modeling and sizing relies on black- box models while recent findings suggest that integration of physical (“whitebox”) modeling (and AI will significantly increase accuracy. Work targets c-Si technology.
Forecasting algorithms are mostly developed for short term horizons (up to 48hrs), often relying on well-estab- lished machine learning and system control methodologies with limited uptake of deep learning methodology. Most AI-based models have shown to perform well for sunny days, while for cloudy days the forecasting accuracy de- creases significantly. (33) and are usually limited to local predictions. (34) No accurate general regional model has been proposed to date. (35) EU research groups are among the most active in the forecasting field, foremost among them Italian and Spanish groups who produce 35 % of the world’s publications on machine learning for PV forecast- ing (China: 23 %; US: 13 %).
AI in production optimisation is currently mostly based on MPPT or solar tracking.
PV plants are complex systems and fault sources are mul- tiple, including optical degradation or fault, electrical mismatches (including shading), potential induced deg- radation (PID), defective/short-circuited bypass diodes, short-circuited modules or strings, and junction box fail- ure. Most research work on AI advanced maintenance, however, operates on single fault detection with fault clas- sification accuracy rapidly decreasing with the number of considered classes and detected faults are not precisely localized. Human investigation is needed. Aerial inspection with IR Cameras and artificial vision components equipped UAVs can improve the fault localization but has been little tested in the field.
Targets, Type of Activity and TRL
Novel digital PV-systems will be developed combing PV technology with photonics, micro- and power-electronics, sensors technology, energy storage, wireless communica- tion, and computer science.
AI and Big Data for PV techniques are essentially in their development phase having been tested on a limited scale in the field and mostly as an off-line data processing tool. The main step to be taken is to favor their actual imple- mentation as a real time field deployed asset. The ongoing setting up of large-scale PV plant data collection, moni- toring and performance analysis will contribute, through semantic extraction capability of Big Data techniques, to enlarging the knowledge about real time and long-term behavior of PV installations. As an enabler, IoT technology is expected to play a major role in increasing the availabil- ity of real time data streams for monitoring and diagnosis of PV plants, particularly in remote locations.
Digitalisation of PV modules themselves will improve data collection, widening the range of available parameters and helping the development of advanced AI4PV technologies which in turn represent a significant opportunity for:
- Modeling of Energy Yield and Sizing (Wu, 2020) (Ghannam R., 2019) (Hesan Ziar, 2021) (A. Mellit, 2009)
- Development and optimisation of PV modules pro- duction (Ghannam R., 2019) (Project SelFab Web- site, 2021)
- Operational Optimisation (e.g. MPPT) and Yield Fore- casting also in the context of energy communities
- Extending Life and Long-term Yield maximization: Fault detection & Predictive Maintenance (Muham- mad Hussain)
- Fast Energy trading
The expected increase in cell and module technologies field readiness will increase the needs and hopefully the development of custom energy yield modeling and pre- diction of site suitability in urban scenarios. To accelerate the integration of PV in urban environments at large scale, we need to improve methods to predict surfaces (mostly roofs) potential yield for different cells and modules PV technologies including bifacial modules.
Following the increase in the digitalisation of PV modules and sensors integration it is expected that AI can play a significant role in automating the PV module operational optimisation. An embedded AI can improve efficiency by adapting MPP to shading and temperature.
Fault diagnosis algorithms will be better at multi-fault identification, and at isolating the contribution of each to underperformance, e.g. Particulate and Ozone (Fattoruso et al., 2020) or shading (Johnson, 2021).
By 2030, AI will be capable for accurate forecasts in cloudy weather and for time horizons beyond 72hrs. Deep learning algorithms fed with air temperature, solar irradiance, rela- tive humidity, wind speed, and/or remote sensing data (e.g. cloud cover) will be capable of regional scale forecasting.
Investigate whether LiFi is a technology for passing info around a PV plant.
Modeling performances under realistic conditions to identify materials and bandgaps adapted to spe- cific realistic conditions.
Integrate multi-aspect sensing (optical, thermal, electrical) into a PV module to suppress degrada- tion, detect unwanted operating conditions and avoid failures.
Digitalisation of PV modules: integrated sensors (optical, electronic), self-diagnosis, reconfigurable modules, self-cleaning, self-cooling with emphasis on achieving high MWh/Wp (shade tolerant; more advanced electronic design with in-module compo- nents).
Develop wireless power transmission of electricity directly from a PV module to the energy system for additional saving on energy losses, potentially in combination with maximum power point tracking.
Develop and apply edge AI and Big Data to:
- improve the energy yield (advanced module con- trol, self-reconfigurable topologies, etc.)
- improve module and plant models
- improve yield forecasting (deep black box model, data driven white box models)
- implement predictive maintenance and early detection of failures in PV technologies (digital twin)
- enable AI agents-based energy trading at plant level, taking care of specific climates /applica- tions / conditions (snow, dust, environmental pollution, water…)
Improved and more accurate ways of creating a dig- ital twin of a PV system or energy system to pre- dict the output and utilization of real distributed PV technology
Build large (time and scale dimension), wide (in- cluding not only yield but multisensorial operation- al, thermal, mechanical and environmental data) and possibly publicly available datasets to enable, foster and empower research in AI for Digital PV at EU scale.
Demonstrate PV modules with integrated storage (e.g. solid state batteries) and new energy manage- ment systems for coupled PV-battery systems.
Demonstrate automated and predictive PV asset management software based on sensor-data-image fusion to reduce human effort and increase trust- worthiness of current PV asset management soft- ware.
Improved energy yield prediction and forecasting software based on physical models (“whitebox models”) that can provide more accurate and faster predictions on very short timescale.