AI for climate: a real world test

Debugging

Debugging AI-generated code presents unique challenges, as Duncan Campbell discovered while working with ChatGPT to optimize battery dispatch software. He found that AI models require precise guidance, often leading to unexpected paths if not directed correctly 1. Seyed Madaeni highlighted that even experienced engineers can overlook obvious constraints, such as a battery's inability to charge and discharge simultaneously 2. This underscores the necessity of human oversight in AI-driven projects, as Seyed noted, "There still needs to be humans involved" 3.

   

Collaboration

The collaborative process between Duncan and ChatGPT resembled a partnership, akin to working with a junior coder. He described the experience as exchanging messages with an analyst, emphasizing the need for clear communication and iterative refinement 4. Despite initial errors, the AI's ability to generate a structured script was evident, though it required human intervention to address issues like simultaneous charging and discharging 5. Seyed likened the process to flying a plane with an autopilot, where human pilots are still essential for oversight and guidance 6.

   

Optimization

Using AI for optimization tasks revealed both strengths and limitations. Seyed praised the textbook approach of the optimization model, noting its effective use of parameters and constraints, though he pointed out minor inconsistencies in variable definitions 7. The task of optimizing battery scheduling involved creating a system to forecast market prices and generate a schedule, a complex process traditionally handled by teams of engineers 8. Despite challenges, Duncan successfully developed a working optimizer, though it focused solely on energy markets, excluding ancillary services 9.