Introduction: The "How" Problem in Climate Solutions

• Climate instructions often tell "what" to do (transition to renewables, electrify everything) but not "how"
• Similar to furniture assembly instructions that lack practical guidance
• Need to understand feasibility within current infrastructure, regulatory environment, dependencies, constraints
• Climate change is scientific challenge, socio-political issue, economic problem - solution depends on framing
• Speaker: scientist, techno-optimist, AI product manager approaching from technological perspective

Three Ways AI Can Help Climate Action

• Understanding climate change: Help comprehend effects on Earth's ecosystems
• Optimizing current systems: Can't start from scratch, must work with existing infrastructure
• Accelerating breakthrough science: Enable fusion as carbon-free energy source
• Focus: optimizing current systems, specifically wind energy through AI

The Renewable Energy Predictability Challenge

• Renewables key to sustainable future but unpredictable (sun/wind intermittency)
• Electricity systems operators need supply to meet demand 24/7 in real time
• Renewables can't be 100% reliably scheduled
• Fossil fuel plants opposite: burn specific coal amount at set time for exact electricity delivery in predictable window
• Power systems managers more confident depending on predictable fossil fuels
• AI solution: powerful forecasting tool to predict wind availability more accurately

Project Implementation Process

Research Phase

• Team composition: research scientists, engineers, product manager, program manager, impact analyst
• Solution: neural net trained on historical weather data and turbine power production
• Identified two core needs: training data and deployment partner

Data Acquisition Challenges

• Massive gaps in climate-critical data across electricity, agriculture, transportation, industry
• Some data purchasable/free (weather forecasts)
• Proprietary data needed: turbine power production, operational wind farm data
• Required proprietary data to train models on weather-power production relationships
• Used few years of hourly resolution data, not massive climactic time scales

Partner Identification

• Needed domain expertise partner willing to test new systems at scale
• Major hurdle: wind farm managers reluctant to test AI on multi-million/billion dollar systems
• Real-world deployment opportunities essential to prove AI effectiveness
• Google became ready, willing partner despite DeepMind being part of Google

Google Partnership and Testing

• Google provided 700 megawatts wind power capacity (equivalent to large US wind farm)
• Excellent proxy for external wind farm operators
• Expert team provided: metrics/benchmarks advice, data sharing
• Critical AI deployment components: domain expert collaboration for needs assessment, constraints, safety requirements, performance metrics, cost-benefit analysis

Results and Performance

• AI system performed 20% better than Google's existing systems
• Improved forecast accuracy crucial: overestimation leads to insufficient supply, driving carbon-intensive fossil fuel backup purchases
• Google decided to scale technology - scaling essential for climate countdown urgency
• Solution developed into software product piloted by French company OMG

Broader Applications and Examples

• Small UK non-profit Open Climate Fix: AI for demand-side forecasting
• Partnership with UK National Grid
• Deploying forecasts 2x more accurate than UK Grid's previous systems
• Demonstrates scalability beyond major research organizations

Call to Action for Different Stakeholders

Domain Experts

• Share problems and challenges faced
• Ensure AI pursuits have real-world impact, not purely academic
• Build competitions to incentivize ML researchers
• Provide clear metrics

Data Holders

• Share climate-critical data where safe and responsible
• Check Climate Change AI website for wish list of critical datasets
• Data access would unblock crucial research and innovation

Deployment Partners

• Identify willingness to test innovative systems
• Essential for proving AI effectiveness in real-world scenarios

General Participants

• Technical background not required for AI climate work
• Diverse skill sets needed: ethicists, policy experts, communication teams, product managers, program managers
• Variety of backgrounds essential beyond research scientists and engineers

Warning Labels and Limitations

• AI not silver bullet for all climate problems
• Not right tool for many challenges faced
• Technology has tensions, needs safe and responsible deployment
• Until grids run on clean energy, AI carries carbon footprint
• Any energy-intensive technology has environmental impact

Conclusion

• AI can be transformational tool in climate fight
• Requires effective wielding by all stakeholders
• "Why" is harrowing, "what" is exciting, but "how" illuminates feasibility and drives impact
• Advance climate conversations from exciting "what" to impactful "how"