AI Use Cases for Renewables Developers

With the increasing deployment of wind and solar, as well as the growing interest in emerging generation technologies like geothermal, small nuclear, and hydrogen, it’s an exciting time for renewable energy developers. But that doesn’t necessarily mean that developers can put their feet up on the desk and eat ice cream while watching a stream of revenue cash in – throughout the project lifecycle, developers typically interact with a complex matrix of stakeholders that are demanding on multiple fronts, such as regulators, grid operators, utilities, builders, and technology providers. They also face unique challenges across each phase of the project lifecycle, typically categorized into 4 phases: 1) pre-development, 2) development, 3) construction, and 4) operation. 

So, it’s not unreasonable to see the developers drooling over the ways AI can unlock new value. In fact, the Department of Energy recently published a comprehensive and practical report called AI and Energy: Opportunities for a Modern Grid and Clean Energy Economy, in which some of the opportunities outlined are directly applicable to developers. For this blog, we’ll cover various AI technologies that can help developers across their project lifecycle, and then explore more in-depth use cases. 

AI use cases for renewables developers

We considered five different types of AI tools: classification, forecasting, digital twins, reinforcement learning, and generative AI. 

• Classification refers to the process where algorithms are trained to categorize data into distinct classes, such as categorizing renewable project sites based on various risk factors. 

• Forecasting involves the use of statistical models and machine learning to predict future variables and trends, such as predicting energy demands and market prices. 

• Digital twins are advanced simulations that create a real-time digital counterpart of a physical object or system, which can provide a monitoring tool during the construction and operations phase of renewable projects. 

• Reinforcement learning is a type of machine learning where algorithms learn to make decisions through defined reward-penalty criteria. In energy systems, this might be used for optimizing bidding strategies in markets or managing the distribution of energy resources. 

• Generative AI can process and generate texts and images, which is helpful when humans in the loop interpret compliance documents, develop permit applications, or engage stakeholders through interactive Q&A. 

Each of these technologies plays a strategic role in enhancing decision-making, operational efficiency, and innovation in renewable energy. The AI use case map below is our take on specific tasks that AI could be useful for renewable energy developers. 

We found that classification and forecasting could be more appropriate for the earlier two phases of development. This is mostly because classification is best used at filtering and funneling opportunities, and forecasting is best utilized to quantify profitability and risk, leading up to financial closure. In contrast, digital twins and reinforcement learning could be more applicable during the later phases. This is mostly because digital twins require physical counterparts first (quite obvious!) and reinforcement learning typically handles operational tactics that aren’t applicable during the earlier phases of development. Lastly, we found that generative AI (further broken into document understanding, copilots, and interative/Q&A) could be applied across the project lifecycle.

We then surveyed our developer friends on how AI could improve the speed, cost, and quality of the development process, and learned that three areas were particularly fit for AI applications: 1) site selection & control, 2) permitting & interconnection, and 3) revenue modeling. We’ll lay out deep-dives of each below.

1. Site selection & control

The site selection process generally follows a series of mapping and filtering, with 1) preliminary desktop analysis, 2) detailed technical assessment, 3) onsite surveys/inspections, and 4) land control agreements. All of these require manual analysis that demands extensive expertise and a thorough understanding of local nuances – which is possibly the reason why all of the developer friends we talked to have pointed site selection as a top pain point. In fact, according to Transect’s 2024 Renewable Energy Developer Report, the top 5 concerns for surveyed respondents are all related to the site selection: access to transmission lines, local sentiments, protected waterways, endangered species, and protected cultural sites. Combined with the race for grabbing the prime real estate in the solar- and wind-abundant regions, challenges in the site selection pose one of the most critical issues for the renewables developers. 

Some say site selection is more art/intuition than science, e.g. we heard “we knew that the land for our solar farm wasn’t right because the color of the grass was off”. However, we believe that the decision variables are a closed set when it comes to the probabilistic outcome of decisions at a population level (e.g. developers could apply the 80/20 rule on whether to move forward with a particular site given enough historical data points). In addition, many of the challenges in this space revolve around data processing. Because the necessary data for site identification criteria (e.g. grid, wetland designation, flood plains, topography/topology, transportation, title & ownership, etc.) are in various formats, understanding the layered data in navigable and executable views currently requires lots of manual work. Besides larger players developing their software in-house, some software companies tackling these issues include Transect, LandGate, Glint Solar, Paces, and Esri.

In more detail, AI can bring some differentiated solutions, such as below.

Opportunities for AI-enabled time and cost savings

• AI-assisted pre-screening & site identification: Classification algorithms can analyze vast amounts of satellite imagery and other geospatial data to help developers create a shortlist of high-quality potential sites. This down-selection process requires not only analyzing a wide range of different data layers (e.g. solar/wind resource, land ownership, environmental restrictions, and more) but also increasingly querying optimization tools to understand the detailed production potential or interconnection risk of a site. AI tools can help developers quickly sift through these data.

• Retrieval-assisted generation for deed & title search: Although GIS layers exist to show land parcels and ownership, these data often do not tell the full story. Before securing an option to lease, developers want to know that there aren’t any skeletons in the closet that could introduce headaches down the road. A common example is when land is jointly owned by relatives who aren’t involved day-to-day (e.g. if a farm was left jointly to all descendants, but one sibling runs the farm and another lives in a distant city). AI can help developers conduct deed and title searches faster and at lower cost, identifying potential issues earlier in the process and providing the information developers need to navigate these interactions.

Opportunities for AI value creation

• Robot-assisted site assessment: With the decreasing cost of both data processing and portable mobility hardware like drones, detailed onsite assessment could leverage the power of AI for various applications. One common example is using a convolutional neural network to process the drone-captured images and videos, which then are used to flag site-related risk factors. Another example is using a hyperspectral camera or a magnetometer on a drone to capture more precise environmental data, such as water, soil, minerals, and land cover. This could also help the developer get a head-start on the environmental study process, albeit limited to preliminary assessment.

• New data sources: Beyond simply providing a better interface to existing data, AI pattern recognition and computer vision tools have the potential to provide new data layers to support site selection. For example, computer vision tools can extract information on existing grid infrastructure (particularly at the distribution level, where data is scarce), and sentiment analysis tools can help assess the likelihood of local opposition to new projects.

2. Permitting & interconnection

If you’ve been following the energy transition, it should come as no surprise that permitting and interconnection represent two of the largest binary risks for new renewables projects. Before a new project can be financed and built, it needs to receive permission from both the local government (“permitting”) and from the relevant utility or independent system operator (“interconnection”). Although the direct cost of permitting and interconnection is relatively small, these processes play an outsized role in determining a project's likelihood of success. As a result, although we see several opportunities for AI to reduce the time and expense of preparing permit and interconnection applications, the more impactful use cases will be those where AI can help reduce risk and improve a developer's odds of success.

Startups working in this space, with varying degrees of AI integration, include Spark, Learnewables, Transect, Paces, Pearl Street Technologies, Nira Energy.

Opportunities for AI-enabled time and cost savings

• Summaries and Q&A from legal documents (ordinances and tariffs): details about process, requirements, and timelines for both permitting and interconnection are often buried in long, dense documents like zoning ordinances and utility tariffs, which can easily stretch to hundreds of pages long. Although developers can specialize (to some extent) in particular interconnection jurisdictions like MISO or ERCOT, permitting regulations vary from county to county (except for the few states with legislation requiring state-level permitting for renewable energy projects). Given this context, AI tools like large language models (LLMs) can help developers quickly get up to speed in new jurisdictions and save time when looking for information about project requirements. The challenge for this application will be scaling LLM-assisted search to long, complex legal documents and building trust that AI assistants can accurately summarize and answer questions without hallucinating.

• Retrieval-assisted generation for preparing permit applications and environmental reviews: once developers begin using LLMs to summarize and answer questions about permit requirements, the natural next step is to use LLMs to help prepare permit applications. Today, large developers will commonly use past permit applications as a starting point for new ones, and it is not difficult to imagine AI being used to generate a first draft based on a developer’s previous projects and mountains of data from engineering and environmental reviews. This use case will likely see LLMs deployed as copilots, with humans reviewing and approving all AI-generated content, but technical challenges around recall and reliability will still need to be solved to enable this use case.

Opportunities for AI as a lever to reduce risk and increase upside

• Optimizing interconnection strategy: currently, although developers consider interconnection considerations while siting new projects, the time and expense of running injection studies means that they must rely on proxies (e.g. hosting capacity maps, substation age, etc.) rather than ground-truth injection capacity. AI tools can help in two ways. First, by learning from the success and failure of past projects, ML models can help pre-screen potential sites based on interconnection risk. Second, and likely more impactfully, the new software will enable developers to run injection studies in near-real time, adding high-fidelity hosting capacity information on top of existing site selection data layers. Using this information, developers will be able to better manage interconnection risk across a portfolio of early stage projects and ultimately optimize the design of individual projects for minimal interconnection costs and delays.

• Support for community engagement: one of the core functions of project developers, particularly at the early stage, is engaging with local stakeholders, including landowners and local officials, to build support for the project. As more and more projects are delayed and canceled due to community opposition, community engagement will only become more critical. AI tools can help developers supercharge their community engagement efforts in a couple of ways, including summarizing and extracting key themes from dozens of hours of community meetings, providing sentiment analysis for local community members, and helping developers reach out and develop relationships with key stakeholders in local communities. The primary challenge is that community engagement is a fundamentally human process; it is neither feasible nor helpful to fully automate outreach efforts, but there is an opportunity for targeted automation to allow human developers to “cover more ground” and be more effective in managing a larger number of relationships with diverse stakeholders. 

3. Revenue modeling

AI can help improve financial workflows for energy projects both by incremental efficiency improvements and step-change improvements in unlocking new revenue streams.

On the efficiency side: as in other domains, AI copilots can help developers build and maintain financial models for their projects, helping to adapt models from previous projects and keeping massive spreadsheets up-to-date with the latest material costs and policy incentives. Given the extensive discussion of copilot-style applications so far, we won’t spend much additional time on this topic.

Beyond copilot use cases, AI has the potential to unlock additional revenue, particularly for energy storage projects. Currently, battery storage projects earn revenue by providing frequency regulation services and capacity reserves and through energy price arbitrage across time (i.e. charge low, discharge high). Because frequency regulation and capacity payments provide more predictable revenue streams, they tend to be preferred from a project-financing perspective. However, as the energy storage market matures and new battery projects come online, these revenue opportunities will shrink, potentially forcing storage to rely more heavily on price arbitrage. AI prediction, optimization, and risk-management models will be key enabling technologies not only for improving the performance of arbitrage-based battery management strategies, but also for de-risking these strategies by providing better uncertainty quantification and scenario analysis. Moreover, AI optimization and control tools like reinforcement learning have the potential to unlock arbitrage opportunities beyond intraday peak shifting (e.g. augmenting a peak-shifting strategy with an AI policy that takes advantage of sub-hour volatility in energy prices).

Particularly when it comes to these advanced charging management strategies, there is a significant trust gap for new AI technologies. It is not enough for an individual software vendor or developer to trust their AI model’s projected returns; they also need to convince their project finance partners to accept these projections. There is a lack of “bankable” AI models for making these trustworthy projections, but more importantly, there is a need for an accepted set of best practices for AI model development and risk management to pave the path to the adoption of these technologies.

Cross-cutting challenges for AI adoption

Although these use cases show the potential for AI to accelerate renewables deployment, there are several challenges facing any application of AI, particularly in the energy sector. First, it should come as no surprise that not only are AI models extremely data-hungry, but they are also highly sensitive to data quality. For use cases like forecasting and reinforcement learning, where AI models must be trained on historical and simulated data, respectively, there is a risk of distribution shift during practical deployments (where the data seen in reality deviates from the data used during training). As a result, best practices for data management, data and model versioning, careful validation, and online monitoring will be critical enabling components for AI applications.

AI applications involving LLMs face a distinct set of challenges. Rather than being trained from scratch on domain-specific data (as a forecasting model might be), LLMs are trained as foundation models on large amounts of data from the internet. Although LLMs may be fine-tuned to domain-specific applications, in practice most LLMs are not fine-tuned and instead gain domain-specific knowledge by a combination of direct prompting and retrieval-augmented generation. This technical difference implies different risk and reliability challenges for applications of LLMs to the energy sector. In particular, there is a risk of hallucinations, which means that additional care must be taken for alignment and quality assurance for LLMs intended to be used in legal applications like summarizing regulations and preparing permit applications.

In all applications of AI to the energy sector, particularly renewables development, we will likely see a three-part approach to adoption, with increasing attention paid to these risk management challenges as users progress through the steps.

1. Experimentation: Users experiment with publicly available models (e.g. ChatGPT, Claude, Bing Copilot) and software offered by third-party providers without relying on the results for their operations. This phase can be considered pre-pilot; users are interested in seeing what value AI tools can provide, but the technology is not sufficiently de-risked to be relied upon. During this stage, users should consider developing policies for data security and sharing, especially if they are working with a third-party provider.

2. Infrastructure Development: Once users decide that AI provides value for them in a particular use case, they will begin formalizing some of their earlier experimental applications. They will work either internally or with software providers to develop the infrastructure to support further development of AI tooling, with a particular focus on data management, and they will continue to refine AI use cases and workflows.

3. Deployment & Standardization: After gaining a clear sense of the value of AI and investing in the infrastructure needed to support AI development, users will move towards broader deployment of AI applications and begin to rely more upon AI for their regular operations. For safety- or enterprise-critical applications such as legal document summarization, demand forecasting, and energy trading, there will be an additional validation phase, often including a pilot or “sidecar” deployment, where AI tooling is run in parallel to existing workflows, before expanding to wide-scale adoption.

AI is one of the most promising innovations in our time, bringing step-change improvements on the biggest pain points like site selection and permitting. But as with any technology, the Spider Man quote applies: with great power comes great responsibilities. To avoid falling into the hype and false promise of AI, developers must carefully assess their use cases, and precisely deploy appropriate AI technologies with thoughtful planning and execution. And before reaping the benefits of better management of stakeholders, cost savings, and overall risk, developers should also remember the prerequisite for technology implementation, such as robust data collection & management, partnering with the right software/service vendors, and building trust in AI across their stakeholders. And to end on a hopeful note – as these tools mature and gain acceptance, we hope that they will significantly accelerate the transition to sustainable energy, ensuring a resilient and greener future. 

For any inquiries, suggestions, or collaboration requests – feel free to contact us at energytransitionnotes@gmail.com. 

---

Title image generated with DALL-E on ChatGPT using the prompt: “Draw me a futuristic image with a background of a solar farm and a wind farm, and in the front a diverse group of renewable energy developers using various functions of artificial intelligence such as large language model, forecasting, and digital twin on large computer screens and holograms, to improve their productivity and revenue through better project site selection, project permitting, and revenue modeling.”

About the author(s)

Dennis Cha is an MS/MBA candidate at Harvard Business School focused on energy and climate. Prior to Harvard, he worked at Google on hardware supply chain, with PG&E on electric infrastructure operations and wildfire analytics, with SoCalGas on asset decarbonization, and at Taslimi Construction on sustainable construction management. He is trained in civil engineering and data science, and began his career in structural engineering. Outside of work, he enjoys scuba diving, youth mentoring, and traveling.

Charles Dawson is a PhD candidate at MIT working on risk assessment and reliability for AI/ML systems. While at MIT, he has worked on reliable AI applications for electrical power systems, self-driving cars, unmanned aerial vehicles, industrial robots, and large-scale transportation networks. Prior to MIT, Charles received a BS in engineering from Harvey Mudd College, with a focus on mechanical and industrial engineering. Outside of research, he enjoys rock climbing and cooking.