In a panel discussion on “AI for O&M” at Renewable Watch’s 2nd edition of the AI in Renewables conference, the discussion focused on the use of artificial intelligence (AI) in asset monitoring, diagnostics and maintenance planning across solar, wind, transmission and battery energy storage systems (BESSs). The panellists were Sameer Ahmad, Head, Data Analytics, Juniper Green Energy; Prosenjit Chakraborty, Business Head, Technology, Jakson Green; Mathew Oommen, Director, Operations, Apraava Energy; Anil Pareek, Zonal Head, Solar O&M, Tata Power Renewable Energy; and Rashmi Shringi, Head, Asset Management, Gentari India. Key takeaways from the session…
Asset monitoring and predictive maintenance
Key use cases of AI discussed in the panel were in asset monitoring, fault detection and predictive alerts. AI is not replacing operations and maintenance (O&M) teams, but giving them faster and more structured information. AI tools are helpful for daily asset monitoring, generating sensor-based alerts, predicting underperformance, and identifying when a site has drifted away from expected performance.
Some developers have integrated AI into a central monitoring system that combines conventional monitoring functions with predictive tools designed to flag failure, shutdown or underperformance at granular asset levels. Additionally, companies are developing custom in-house tools on top of existing data, rather than relying only on off-the-shelf platforms. The aim is simple: use the data already coming from the site and extract more value from it.
Furthermore, AI is increasingly being used in day-to-day O&M decisions. Instead of relying on periodic and manual scheduling, systems are now being trained to identify the right time for module cleaning, especially in soiling-prone sites, by comparing generation loss with the cost of cleaning. The aim is to clean only when the benefit exceeds the cost, rather than following a fixed calendar. This kind of data-driven scheduling has already helped reduce unscheduled downtime by about 20 per cent and improve equipment life by roughly 15 per cent.
Forecasting and scheduling
Forecasting and scheduling emerged as the use case with the strongest consensus among the panellists on the measurable gains derived from AI. In this space, AI shows credible results as the output can be compared directly with actual generation, and the difference is easy to quantify.
Moreover, forecasting also has a direct revenue impact. Forecasting penalties can range from 0.5 per cent to 10 per cent, depending on the project and the regulatory framework. Tighter forecasting rules increase the pressure on developers to improve accuracy, because errors lead to high penalties.
AI is helping forecasting models improve, with companies now using in-house data sets rather than relying fully on external forecasting sources. Such models combine current conditions with historical patterns from the same time in previous years, including the same day and season, to improve accuracy. This is especially useful in regions with more complex weather patterns, like Haryana and Uttar Pradesh, because of fog, monsoon variability and other local conditions.
Furthermore, AI is being used to develop site-specific power curves and understand where energy is being lost at the turbine level for wind projects, rather than looking only at plant averages.
That is particularly relevant in the wind sector, where the long-term performance of each turbine may drift over time, even when plant availability appears strong on paper. Wind plants may still achieve 99.5 per cent availability, but that does not mean each machine is performing at its full potential. Over a 25-year life, it can lose efficiency gradually, and the challenge is to identify where the loss is occurring.
Key challenges
The most pressing issue faced in the use of AI for O&M is data quality. Several panellists highlighted that AI models are only as useful as the data fed into them, and that data quality is still far from consistent across sites. The backbone of any AI system is the data stack, and if the data is weak, the model cannot deliver the intended result. The use of AI in O&M is simply not just about the algorithm but also about the quality of the site data that enters the algorithm in the first place.
Open Platform Communications Unified Architecture protocols and store-and-forward architecture can help mitigate the data quality challenge. If a site loses internet connectivity, the system should still store the data locally and transmit it later when the connection returns. This matters especially in remote or weather-affected locations. It is also useful for many older plants that were not designed for the level of data exchange now required by AI tools. However, there are still some cases where older hardware cannot communicate with current digital systems without replacement or major retrofitting.
Limited data access and a lack of data sharing are additional challenges, with O&M teams not always receiving the full set of tags or raw information they need from systems. In some cases, data is shared only at a high level or on a delayed basis. For meaningful predictive maintenance, micro-level data is required, which is not always available.
Hence, AI adoption is not only about software readiness, but also depends on how much operating history has been accumulated, how granular the records are, and whether the site architecture was built for analytics in the first place. If historical data is not available for the right season or time period, the model cannot accurately predict the next failure.
This is especially important for equipment like gearboxes, inverters and turbines, where failure patterns are not random. They are often tied to specific operating conditions, temperature ranges or vibration signatures. Without access to the right data, the AI tool can only describe the problem after it has occurred.
Additionally, the results seen in a pilot often do not translate fully when the same system is rolled out across a larger portfolio. The reason is usually data quality, loss of connectivity, incomplete tags or weak site discipline. In pilots, the data is often cleaner and the project is closely managed. At scale, the system has to work across different sites, different conditions and different operating teams.
The way forward
AI in O&M has not replaced human judgment, and will probably not do so in the near future. However, it is already changing how teams monitor assets, schedule cleaning, assess performance and forecast output.
Future progress will depend on three factors. First, better data architecture, including local storage, stronger sensors and more reliable connectivity. Second, greater access to original manufacturer equipment data, especially for older and more complex assets. Third, more in-house capability, so that operators can build tools suited to their own plants rather than depending fully on external platforms.
The emerging pattern is clear: AI is most valuable when it reduces the time to detect, diagnose and act. The technology is useful only when the data is good enough to make it work. For now, AI in O&M is delivering the most value in well-defined use cases. The next step will be to turn those gains into something broader: better plant life, lower downtime, fewer blind spots and more consistent output across the asset base.
